data.jsonl

{"0": 1, "1": "huggingface_new", "2": "https://huggingface.co/mseri/zunzuncito-maple-preview-2bit", "3": "mseri/zunzuncito-maple-preview-2bit", "4": "mseri/zunzuncito-maple-preview-2bit. Downloads: 0. Tags: zunzuncito, text-generation, base_model:deepgrove/maple-preview-2bit-mlx, base_model:finetune:deepgrove/maple-preview-2bit-mlx, region:us", "5": "2026-08-22T19:14:45.639727"}
{"0": 2, "1": "huggingface_new", "2": "https://huggingface.co/minseokk7/BioPhys-Neural-Agent", "3": "minseokk7/BioPhys-Neural-Agent", "4": "minseokk7/BioPhys-Neural-Agent. Downloads: 566. Tags: gguf, biophys_neural_agent, biophys-neural-agent, 8-state-snn, neuromorphic-snn, ternary-weights, bitnet-1.58b, moe-6-brain, open-swarm, recaman-cipher", "5": "2026-08-22T19:14:45.645271"}
{"0": 3, "1": "huggingface_new", "2": "https://huggingface.co/deadbydawn101/RavenXAiLabs-Chaos-Agent-Qwen3.8-27B-Frontier-Intelligence-Injected-OBLITERATED-GGUF", "3": "deadbydawn101/RavenXAiLabs-Chaos-Agent-Qwen3.8-27B-Frontier-Intelligence-Injected-OBLITERATED-GGUF", "4": "deadbydawn101/RavenXAiLabs-Chaos-Agent-Qwen3.8-27B-Frontier-Intelligence-Injected-OBLITERATED-GGUF. Downloads: 0. Tags: gguf, ravenx, soul-injection, iq-injection, cybersecurity, pentesting, red-team, agent, abliterated, uncensored", "5": "2026-08-22T19:14:45.651580"}
{"0": 4, "1": "huggingface_new", "2": "https://huggingface.co/ahmetggg/Luck-spark-V-300-MoE", "3": "ahmetggg/Luck-spark-V-300-MoE", "4": "ahmetggg/Luck-spark-V-300-MoE. Downloads: 1079. Tags: transformers, safetensors, mixtral, text-generation, generated_from_trainer, en, tr, dataset:generator, base_model:ahmetggg/Luck-spark-V-300-MoE, base_model:finetune:ahmetggg/Luck-spark-V-300-MoE", "5": "2026-08-22T19:14:45.656457"}
{"0": 5, "1": "huggingface_new", "2": "https://huggingface.co/sirus/Qwen3.8-27B-heretic-ara-MXFP6-MXFP8-DFlash2-GGUF", "3": "sirus/Qwen3.8-27B-heretic-ara-MXFP6-MXFP8-DFlash2-GGUF", "4": "sirus/Qwen3.8-27B-heretic-ara-MXFP6-MXFP8-DFlash2-GGUF. Downloads: 0. Tags: gguf, qwen3.8, dflash2, speculative-decoding, mxfp6, mxfp8, blackwell, llama.cpp, text-generation, base_model:heretic-org/Qwen3.8-27B-heretic-ara", "5": "2026-08-22T19:14:45.661112"}
{"0": 6, "1": "huggingface_new", "2": "https://huggingface.co/Offlucas/llama-3.1-8b-coder-lora", "3": "Offlucas/llama-3.1-8b-coder-lora", "4": "Offlucas/llama-3.1-8b-coder-lora. Downloads: 0. Tags: peft, safetensors, base_model:adapter:mlabonne/Meta-Llama-3.1-8B-Instruct-abliterated, lora, sft, transformers, trl, unsloth, text-generation, conversational", "5": "2026-08-22T19:14:45.665709"}
{"0": 7, "1": "huggingface_new", "2": "https://huggingface.co/tigerking009/ZOYA_FLASH_V1_PHASE0", "3": "tigerking009/ZOYA_FLASH_V1_PHASE0", "4": "tigerking009/ZOYA_FLASH_V1_PHASE0. Downloads: 1637. Tags: transformers, safetensors, qwen3_5_text, text-generation, abliterated, uncensored, qwen3.5, qwen, conversational, en", "5": "2026-08-22T19:14:45.671320"}
{"0": 8, "1": "huggingface_new", "2": "https://huggingface.co/deadbydawn101/RavenXAiLabs-Chaos-Agent-Qwen3.8-27B-Frontier-Intelligence-Injected-OBLITERATED-MLX", "3": "deadbydawn101/RavenXAiLabs-Chaos-Agent-Qwen3.8-27B-Frontier-Intelligence-Injected-OBLITERATED-MLX", "4": "deadbydawn101/RavenXAiLabs-Chaos-Agent-Qwen3.8-27B-Frontier-Intelligence-Injected-OBLITERATED-MLX. Downloads: 0. Tags: mlx, safetensors, qwen3_5, ravenx, soul-injection, iq-injection, cybersecurity, pentesting, red-team, agent", "5": "2026-08-22T19:14:45.676210"}
{"0": 9, "1": "huggingface_new", "2": "https://huggingface.co/liuw15/qwen3-8b-ziyon-nsfw", "3": "liuw15/qwen3-8b-ziyon-nsfw", "4": "liuw15/qwen3-8b-ziyon-nsfw. Downloads: 0. Tags: transformers, safetensors, qwen3, text-generation, unsloth, conversational, arxiv:2309.00071, base_model:Qwen/Qwen3-8B, base_model:finetune:Qwen/Qwen3-8B, license:apache-2.0", "5": "2026-08-22T19:14:45.681420"}
{"0": 10, "1": "huggingface_new", "2": "https://huggingface.co/riit3sh/qwen2.5-coder-3b-spider-sft-grpo", "3": "riit3sh/qwen2.5-coder-3b-spider-sft-grpo", "4": "riit3sh/qwen2.5-coder-3b-spider-sft-grpo. Downloads: 0. Tags: safetensors, qwen2, text-to-sql, sql, code, reinforcement-learning, grpo, qwen2.5, spider, unsloth", "5": "2026-08-22T19:14:45.686330"}
{"0": 11, "1": "huggingface_new", "2": "https://huggingface.co/runanywhere/qwen3_8_27b_HNPU", "3": "runanywhere/qwen3_8_27b_HNPU", "4": "runanywhere/qwen3_8_27b_HNPU. Downloads: 32. Tags: gguf, hnpu, hexagon, npu, llm, v81, text-generation, base_model:unsloth/Qwen3.8-27B-GGUF, base_model:quantized:unsloth/Qwen3.8-27B-GGUF, license:apache-2.0", "5": "2026-08-22T19:14:45.691056"}
{"0": 12, "1": "huggingface_new", "2": "https://huggingface.co/LouLou1Demon/pixalium-20M-pretrained", "3": "LouLou1Demon/pixalium-20M-pretrained", "4": "LouLou1Demon/pixalium-20M-pretrained. Downloads: 1875. Tags: transformers, safetensors, llama, text-generation, generated_from_trainer, text-generation-inference, endpoints_compatible, region:us", "5": "2026-08-22T19:14:45.698803"}
{"0": 13, "1": "huggingface_new", "2": "https://huggingface.co/akash1702-eng/qwen2.5-1.5b-unlearned-adapter", "3": "akash1702-eng/qwen2.5-1.5b-unlearned-adapter", "4": "akash1702-eng/qwen2.5-1.5b-unlearned-adapter. Downloads: 0. Tags: peft, safetensors, base_model:adapter:Qwen/Qwen2.5-1.5B-Instruct, lora, transformers, text-generation, conversational, arxiv:1910.09700, base_model:Qwen/Qwen2.5-1.5B-Instruct, region:us", "5": "2026-08-22T19:14:45.703054"}
{"0": 14, "1": "huggingface_new", "2": "https://huggingface.co/VERBAREX/LuminoLex-1.5B-think", "3": "VERBAREX/LuminoLex-1.5B-think", "4": "VERBAREX/LuminoLex-1.5B-think. Downloads: 97. Tags: transformers, safetensors, luminolex, text-generation, causal-lm, reasoning, verbarex, custom_code, license:apache-2.0, region:us", "5": "2026-08-22T19:14:45.707662"}
{"0": 15, "1": "huggingface_new", "2": "https://huggingface.co/Jordansky/envours3-b9057b9c", "3": "Jordansky/envours3-b9057b9c", "4": "Jordansky/envours3-b9057b9c. Downloads: 0. Tags: peft, safetensors, base_model:adapter:/cache/models/Jordansky--oursr1b-d53997be, lora, sft, transformers, trl, text-generation, conversational, arxiv:1910.09700", "5": "2026-08-22T19:14:45.713191"}
{"0": 16, "1": "huggingface_new", "2": "https://huggingface.co/iromu/qwen25-1.5b-tools", "3": "iromu/qwen25-1.5b-tools", "4": "iromu/qwen25-1.5b-tools. Downloads: 0. Tags: transformers, safetensors, qwen2, text-generation, qwen2.5, qwen, tool-calling, function-calling, agents, sft", "5": "2026-08-22T19:14:45.717997"}
{"0": 17, "1": "huggingface_new", "2": "https://huggingface.co/SurjoLabs/Spark-1.1", "3": "SurjoLabs/Spark-1.1", "4": "SurjoLabs/Spark-1.1. Downloads: 0. Tags: transformers, safetensors, spark, text-generation, generated_from_trainer, rloo, trl, conversational, custom_code, arxiv:2402.14740", "5": "2026-08-22T19:14:45.721994"}
{"0": 18, "1": "huggingface_new", "2": "https://huggingface.co/hfunknown/llama31-8b-knapsack-lora-stateless", "3": "hfunknown/llama31-8b-knapsack-lora-stateless", "4": "hfunknown/llama31-8b-knapsack-lora-stateless. Downloads: 13. Tags: peft, safetensors, lora, transformers, text-generation, base_model:meta-llama/Llama-3.1-8B, base_model:adapter:meta-llama/Llama-3.1-8B, region:us", "5": "2026-08-22T19:14:45.726468"}
{"0": 19, "1": "huggingface_new", "2": "https://huggingface.co/hfunknown/llama31-8b-knapsack-lora-persistent", "3": "hfunknown/llama31-8b-knapsack-lora-persistent", "4": "hfunknown/llama31-8b-knapsack-lora-persistent. Downloads: 22. Tags: peft, safetensors, lora, transformers, text-generation, base_model:meta-llama/Llama-3.1-8B, base_model:adapter:meta-llama/Llama-3.1-8B, region:us", "5": "2026-08-22T19:14:45.732077"}
{"0": 20, "1": "huggingface_new", "2": "https://huggingface.co/dudeman2512/Qwen3.8-2.4T-A95B-FP8", "3": "dudeman2512/Qwen3.8-2.4T-A95B-FP8", "4": "dudeman2512/Qwen3.8-2.4T-A95B-FP8. Downloads: 0. Tags: transformers, quantized, compressed-tensors, vllm, fp8, text-generation, base_model:Qwen/Qwen3.8-2.4T-A95B, base_model:finetune:Qwen/Qwen3.8-2.4T-A95B, license:other, endpoints_compatible", "5": "2026-08-22T19:14:45.736872"}
{"0": 21, "1": "huggingface_new", "2": "https://huggingface.co/dudeman2512/Qwen3.8-2.4T-A95B-NVFP4", "3": "dudeman2512/Qwen3.8-2.4T-A95B-NVFP4", "4": "dudeman2512/Qwen3.8-2.4T-A95B-NVFP4. Downloads: 0. Tags: transformers, quantized, compressed-tensors, vllm, nvfp4, text-generation, base_model:Qwen/Qwen3.8-2.4T-A95B, base_model:finetune:Qwen/Qwen3.8-2.4T-A95B, license:other, endpoints_compatible", "5": "2026-08-22T19:14:45.740999"}
{"0": 22, "1": "huggingface_new", "2": "https://huggingface.co/dudeman2512/Qwen3.8-2.4T-A95B-int4", "3": "dudeman2512/Qwen3.8-2.4T-A95B-int4", "4": "dudeman2512/Qwen3.8-2.4T-A95B-int4. Downloads: 0. Tags: transformers, quantized, compressed-tensors, vllm, int4, text-generation, base_model:Qwen/Qwen3.8-2.4T-A95B, base_model:finetune:Qwen/Qwen3.8-2.4T-A95B, license:other, endpoints_compatible", "5": "2026-08-22T19:14:45.745481"}
{"0": 23, "1": "huggingface_new", "2": "https://huggingface.co/liuw15/ziyon-qlora-nsfw", "3": "liuw15/ziyon-qlora-nsfw", "4": "liuw15/ziyon-qlora-nsfw. Downloads: 0. Tags: peft, safetensors, base_model:adapter:unsloth/Qwen3-8B-unsloth-bnb-4bit, lora, sft, transformers, trl, unsloth, text-generation, conversational", "5": "2026-08-22T19:14:45.753526"}
{"0": 24, "1": "huggingface_new", "2": "https://huggingface.co/otheru/DeepSeek-V4-Flash-Strix-Halo-GGUF", "3": "otheru/DeepSeek-V4-Flash-Strix-Halo-GGUF", "4": "otheru/DeepSeek-V4-Flash-Strix-Halo-GGUF. Downloads: 22437. Tags: gguf, rocmfp, rocmfpx, strix-halo, gfx1151, amd, deepseek-v4, deepseek-v4-0731, moe, imatrix", "5": "2026-08-22T19:14:45.757584"}
{"0": 25, "1": "huggingface_new", "2": "https://huggingface.co/Avifenesh/Ornith-1.5-35B-A3B-NVFP4-MTP-GGUF", "3": "Avifenesh/Ornith-1.5-35B-A3B-NVFP4-MTP-GGUF", "4": "Avifenesh/Ornith-1.5-35B-A3B-NVFP4-MTP-GGUF. Downloads: 2749. Tags: gguf, nvfp4, memra, speculative-decoding, mtp, conversational, blackwell, qwen3_5_moe, moe, text-generation", "5": "2026-08-22T19:14:45.766839"}
{"0": 26, "1": "huggingface_new", "2": "https://huggingface.co/harsh762011/startup32", "3": "harsh762011/startup32", "4": "harsh762011/startup32. Downloads: 267. Tags: peft, safetensors, base_model:adapter:unsloth/phi-4-mini-reasoning-unsloth-bnb-4bit, lora, sft, transformers, trl, unsloth, text-generation, conversational", "5": "2026-08-22T19:14:45.774335"}
{"0": 27, "1": "huggingface_new", "2": "https://huggingface.co/ram1234598766/Cesium2", "3": "ram1234598766/Cesium2", "4": "ram1234598766/Cesium2. Downloads: 11. Tags: transformers, safetensors, gguf, causal-lm, qwen2.5, reasoning, code-generation, moe, qlora, multimodal", "5": "2026-08-22T19:14:45.782499"}
{"0": 28, "1": "huggingface_new", "2": "https://huggingface.co/ananddey/akhorika-e2b-base", "3": "ananddey/akhorika-e2b-base", "4": "ananddey/akhorika-e2b-base. Downloads: 0. Tags: transformers, safetensors, gemma4, image-text-to-text, gemma, gemma-4, assamese, indic, continued-pretraining, cpt", "5": "2026-08-22T19:14:45.790981"}
{"0": 29, "1": "huggingface_new", "2": "https://huggingface.co/devoppro/FastLLM", "3": "devoppro/FastLLM", "4": "devoppro/FastLLM. Downloads: 54. Tags: transformers, safetensors, modern_llm, text-generation, custom-architecture, rope, gqa, swiglu, rmsnorm, en", "5": "2026-08-22T19:14:45.795697"}
{"0": 30, "1": "huggingface_new", "2": "https://huggingface.co/gold24k/v5", "3": "gold24k/v5", "4": "gold24k/v5. Downloads: 0. Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, affine, peft, dpo, qwen3.5, text-generation, conversational", "5": "2026-08-22T19:14:45.802865"}
{"0": 31, "1": "huggingface_new", "2": "https://huggingface.co/jashepp/Ornith-1.5-9B-MXFP4_Hybrid-Imatrix-GGUF", "3": "jashepp/Ornith-1.5-9B-MXFP4_Hybrid-Imatrix-GGUF", "4": "jashepp/Ornith-1.5-9B-MXFP4_Hybrid-Imatrix-GGUF. Downloads: 0. Tags: transformers, gguf, qwen, qwen3, qwen3.5, distillation, chain-of-thought, agentic, tool-use, chained-distill", "5": "2026-08-22T19:14:45.808846"}
{"0": 32, "1": "huggingface_new", "2": "https://huggingface.co/ciaochris/Nuclear-Expert-LoRA-3B", "3": "ciaochris/Nuclear-Expert-LoRA-3B", "4": "ciaochris/Nuclear-Expert-LoRA-3B. Downloads: 0. Tags: peft, nuclear-physics, nuclear-engineering, reactor-physics, neutron-physics, nuclear-materials, nuclear-energy, fuel-cycle, scientific-reasoning, domain-adaptation", "5": "2026-08-22T19:14:45.813022"}
{"0": 33, "1": "huggingface_new", "2": "https://huggingface.co/SC117/Ornith-1.5-35B-A3B-Heretic-MTP-APEX-GGUF", "3": "SC117/Ornith-1.5-35B-A3B-Heretic-MTP-APEX-GGUF", "4": "SC117/Ornith-1.5-35B-A3B-Heretic-MTP-APEX-GGUF. Downloads: 0. Tags: transformers, gguf, qwen3_5_moe, reasoning, agentic-coding, heretic, mtp, apex, quantization, multimodal", "5": "2026-08-22T19:14:45.819209"}
{"0": 34, "1": "huggingface_new", "2": "https://huggingface.co/Myric/Qwen3.5-35B-A3B-APEX-GGUF", "3": "Myric/Qwen3.5-35B-A3B-APEX-GGUF", "4": "Myric/Qwen3.5-35B-A3B-APEX-GGUF. Downloads: 1162. Tags: gguf, moe, apex, quantized, imatrix, torch-imatrix, qwen3_5_moe, llama.cpp, text-generation, base_model:Qwen/Qwen3.5-35B-A3B", "5": "2026-08-22T19:14:45.827117"}
{"0": 35, "1": "huggingface_new", "2": "https://huggingface.co/Myric/Qwen3.6-35B-A3B-APEX-GGUF", "3": "Myric/Qwen3.6-35B-A3B-APEX-GGUF", "4": "Myric/Qwen3.6-35B-A3B-APEX-GGUF. Downloads: 2228. Tags: gguf, moe, apex, quantized, imatrix, torch-imatrix, qwen3_5_moe, llama.cpp, text-generation, base_model:Qwen/Qwen3.6-35B-A3B", "5": "2026-08-22T19:14:45.832892"}
{"0": 36, "1": "huggingface_new", "2": "https://huggingface.co/sullivan1502/base-zone-pretrain", "3": "sullivan1502/base-zone-pretrain", "4": "sullivan1502/base-zone-pretrain. Downloads: 5000. Tags: transformers, safetensors, llama, text-generation, arxiv:1910.09700, text-generation-inference, endpoints_compatible, region:us", "5": "2026-08-22T19:14:45.838542"}
{"0": 37, "1": "huggingface_new", "2": "https://huggingface.co/MLA299/Tennda-Nano", "3": "MLA299/Tennda-Nano", "4": "MLA299/Tennda-Nano. Downloads: 375. Tags: mlx, safetensors, gemma4, code, sql, text-generation, llm, tennda, conversational, license:other", "5": "2026-08-22T19:14:45.843623"}
{"0": 38, "1": "huggingface_new", "2": "https://huggingface.co/pinkelephantlimited/pinkelephant-llm-48b-s-gguf", "3": "pinkelephantlimited/pinkelephant-llm-48b-s-gguf", "4": "pinkelephantlimited/pinkelephant-llm-48b-s-gguf. Downloads: 11. Tags: gguf, moe, mixture-of-experts, sparse-moe, llama.cpp, ollama, text-generation, sft, dpo, reinforcement-learning-from-human-feedback", "5": "2026-08-22T19:14:45.848057"}
{"0": 39, "1": "huggingface_new", "2": "https://huggingface.co/ganesh714/arq_m1_chairs", "3": "ganesh714/arq_m1_chairs", "4": "ganesh714/arq_m1_chairs. Downloads: 0. Tags: transformers, safetensors, qwen2, text-generation, text-generation-inference, unsloth, conversational, en, license:apache-2.0, endpoints_compatible", "5": "2026-08-22T19:14:45.852709"}
{"0": 40, "1": "huggingface_new", "2": "https://huggingface.co/jashepp/Ornith-1.5-35B-A3B-MXFP4_MOE_Hybrid-Imatrix-GGUF", "3": "jashepp/Ornith-1.5-35B-A3B-MXFP4_MOE_Hybrid-Imatrix-GGUF", "4": "jashepp/Ornith-1.5-35B-A3B-MXFP4_MOE_Hybrid-Imatrix-GGUF. Downloads: 423. Tags: transformers, gguf, qwen, qwen3, qwen3.5, moe, distillation, chain-of-thought, agentic, tool-use", "5": "2026-08-22T19:14:45.859118"}
{"0": 41, "1": "huggingface_new", "2": "https://huggingface.co/Adicandra/Compfest_akumaukePengospasangsolarpanel", "3": "Adicandra/Compfest_akumaukePengospasangsolarpanel", "4": "Adicandra/Compfest_akumaukePengospasangsolarpanel. Downloads: 0. Tags: peft, safetensors, qwen3, unsloth, lora, sft, trl, tool-calling, agent, text-generation", "5": "2026-08-22T19:14:45.863863"}
{"0": 42, "1": "huggingface_new", "2": "https://huggingface.co/temaq-org/Tema_Q-X7-Thinking", "3": "temaq-org/Tema_Q-X7-Thinking", "4": "temaq-org/Tema_Q-X7-Thinking. Downloads: 127. Tags: safetensors, qwen3_5_moe, Agent, transformer, instruction-tuned, multilingual, uncensored, non-censored, unfiltered, text-generation", "5": "2026-08-22T19:14:45.870340"}
{"0": 43, "1": "huggingface_new", "2": "https://huggingface.co/meridianal/FinAI", "3": "meridianal/FinAI", "4": "meridianal/FinAI. Downloads: 0. Tags: transformers, safetensors, finance, continual-learning, qwen2, causal-lm, ewc, text-generation, en, dataset:gbharti/finance-alpaca", "5": "2026-08-22T19:14:45.875160"}
{"0": 44, "1": "huggingface_new", "2": "https://huggingface.co/shrugging-shoulders/Amberlight-Lux-12B", "3": "shrugging-shoulders/Amberlight-Lux-12B", "4": "shrugging-shoulders/Amberlight-Lux-12B. Downloads: 0. Tags: transformers, safetensors, mistral, text-generation, trl, conversational, finetune, roleplay, en, fr", "5": "2026-08-22T19:14:45.880784"}
{"0": 45, "1": "huggingface_new", "2": "https://huggingface.co/simutrade/simutrade-v1-clm-rag-gemma3-4b-it-adapter", "3": "simutrade/simutrade-v1-clm-rag-gemma3-4b-it-adapter", "4": "simutrade/simutrade-v1-clm-rag-gemma3-4b-it-adapter. Downloads: 0. Tags: peft, safetensors, lora, qlora, rag, retrieval-augmented-generation, sft, instruction-tuning, international-trade, logistics", "5": "2026-08-22T19:14:45.887986"}
{"0": 46, "1": "huggingface_new", "2": "https://huggingface.co/ursb01/Ornith-1.0-35B-Heretic-MTP-APEX-GGUF", "3": "ursb01/Ornith-1.0-35B-Heretic-MTP-APEX-GGUF", "4": "ursb01/Ornith-1.0-35B-Heretic-MTP-APEX-GGUF. Downloads: 0. Tags: transformers, gguf, qwen3_5_moe, reasoning, agentic-coding, heretic, mtp, apex, quantization, multimodal", "5": "2026-08-22T19:14:45.896435"}
{"0": 47, "1": "huggingface_new", "2": "https://huggingface.co/just1nseo/qwen3-4b-if-rlvr-subset20k-n3-ci95lower-uppermean-flipabstain-ep3", "3": "just1nseo/qwen3-4b-if-rlvr-subset20k-n3-ci95lower-uppermean-flipabstain-ep3", "4": "just1nseo/qwen3-4b-if-rlvr-subset20k-n3-ci95lower-uppermean-flipabstain-ep3. Downloads: 0. Tags: transformers, safetensors, text-generation, base_model:Qwen/Qwen3-4B, base_model:finetune:Qwen/Qwen3-4B, endpoints_compatible, region:us", "5": "2026-08-22T19:14:45.910525"}
{"0": 48, "1": "huggingface_new", "2": "https://huggingface.co/SC117/Ornith-1.5-35B-A3B-MTP-APEX-GGUF", "3": "SC117/Ornith-1.5-35B-A3B-MTP-APEX-GGUF", "4": "SC117/Ornith-1.5-35B-A3B-MTP-APEX-GGUF. Downloads: 4836. Tags: transformers, gguf, qwen3_5_moe, qwen3_5, reasoning, agentic-coding, mtp, apex, quantization, multimodal", "5": "2026-08-22T19:14:45.915723"}
{"0": 49, "1": "huggingface_new", "2": "https://huggingface.co/Honkware/Qwen3.5-0.8B-exl3-4.0bpw", "3": "Honkware/Qwen3.5-0.8B-exl3-4.0bpw", "4": "Honkware/Qwen3.5-0.8B-exl3-4.0bpw. Downloads: 0. Tags: exllamav3, safetensors, qwen3_5, exl3, quantized, text-generation, conversational, base_model:Qwen/Qwen3.5-0.8B, base_model:quantized:Qwen/Qwen3.5-0.8B, license:apache-2.0", "5": "2026-08-22T19:14:45.920532"}
{"0": 50, "1": "huggingface_new", "2": "https://huggingface.co/georvn7/hayabusa-9b", "3": "georvn7/hayabusa-9b", "4": "georvn7/hayabusa-9b. Downloads: 204. Tags: transformers, safetensors, qwen3_5_text, text-generation, qwen, qwen3.5, hayabusa, full-finetune, sft, dpo", "5": "2026-08-22T19:14:45.927392"}
{"0": 51, "1": "hackernews", "2": "https://www.anjadhe.com/demo", "3": "Show HN: Anjadhe \u2013 privacy first AI assistant, no account, no server DB", "4": "Show HN: Anjadhe \u2013 privacy first AI assistant, no account, no server DB. Hi HN, I am Ram. For the last few months I have been building Anjadhe. It is a personal AI assistant for macOS. The main idea is simple: A macOS app where AI does the job of a personal assistant for the user. Not a chat only app where user needs to dig through chats to understand today\u2019s schedule or a project they planned with the ai last week, but a canvas where the user and AI work together around basic tools.<p>I started building this as DYI tool for myself, but at this point I see that its truly useful for me, so sharing with others to see if it can be useful to others too.<p>Here is a quick Demo - <a href=\"https:&#x2F;&#x2F;www.anjadhe.com&#x2F;demo\" rel=\"nofollow\">https:&#x2F;&#x2F;www.anjadhe.com&#x2F;demo</a><p>What it can do today:<p>- It reads your incoming email and files the important things: bills with due dates, renewals, receipts, deliveries. A bill becomes a task with a date, by itself. Two booking emails become one trip.\n- You set goals by talking to it, not by filling forms. It asks you questions, makes a plan with dated tasks. Later you can say &quot;work got crazy, push everything two weeks&quot; and it moves the whole plan. It always shows you what will change and asks before doing it.\n- Routines run without you. On a schedule, or when a certain email or file arrives. Every run leaves a log you can read.\n- You can give it some documents you wrote. It learns your writing style and writes new content in your voice. What it learned is a page you can read and edit, not a black box.<p>About privacy: the AI runs on your Mac with llama.cpp. Or you can point it to your own server, or use your own OpenAI&#x2F;Anthropic key, or use Anjadhe Cloud (open-weight models, free allowance, no account needed). Your data stays in SQLite on your own disk. No account. No telemetry unless you turn it on. The source code of the app and the cloud service is public, so you do not have to trust my words.<p>Honest limitations: it is an early alpha, macOS only. Local models need a Mac with atleast 32gb ram to work. The email features work fine on small models, but the full agent wants a bigger model or a server. Also it is Electron with vanilla JS, no framework. Looking forward to hear your opinions.<p>Demo (2 min): <a href=\"https:&#x2F;&#x2F;anjadhe.ai&#x2F;demo\" rel=\"nofollow\">https:&#x2F;&#x2F;anjadhe.ai&#x2F;demo</a>\nDownload: <a href=\"https:&#x2F;&#x2F;anjadhe.ai&#x2F;download\" rel=\"nofollow\">https:&#x2F;&#x2F;anjadhe.ai&#x2F;download</a>\nSource: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;Anjadhe&#x2F;Anjadhe\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;Anjadhe&#x2F;Anjadhe</a><p>I would love feedback, especially if you think my privacy claims are wrong somewhere. I will be here to answer questions.", "5": "2026-08-22T19:14:46.766445"}
{"0": 52, "1": "hackernews", "2": "https://ozbrain.com", "3": "Show HN: OzBrain, a shared brain for knowledge between agents and your team", "4": "Show HN: OzBrain, a shared brain for knowledge between agents and your team. I think agent-first chat interfaces will be a primary software modality and busy dashboard&#x2F;UI will go away. I\u2019m not sure who exactly wins it, but I want my knowledge to grow&#x2F;go with me.<p>A lot of the \u201cknowledge\u201d ie research, analysis, reasoning will be done by agents as the primary user. Our current notes tools &amp; tasks management systems were built for humans\u2026 I don\u2019t care what the 17th thing on my bug backlog is. I want to conduct agents that can execute for me and do great work.<p>What I built OzBrain to do:\n+ Create a central place for agent reasoned knowledge to live\n+ Be agnostic about what apps&#x2F;agents connect to it\n+ Capture everything and track it so I can audit it\n+ Enable teams, collaborators or partners to share brains\n+ Handle conflicts so many agents in the same article doesn\u2019t blow up\n+ Refactor knowledge into more token friendly chunks and map the index well\n+ Close the knowledge loop so new thinking supersedes old thinking across the corpus. Don\u2019t erase, depreciate and link\n+ Keep user data safe and secure\n++ Be easy enough to use that you don\u2019t have to have any technical knowledge<p>Some among us will always build their own custom solutions, but there are millions of tech professionals and small business owners that will use agents heavily and need a solution. So I\u2019m trying to build that.<p>Isn\u2019t this like gBrain? Yes, similar. I think it\u2019s like AWS vs Vercel. AWS is very powerful, configurable, and useful if you\u2019re technical and want to invest the time into really fine tuning your system\u2026 but if you just want your web deploy&#x2F;hosting to just work and be easy to deal with you use Vercel.<p>&#x2F;&#x2F; WHY I MADE IT<p>I\u2019ve been enjoying getting back to my technical roots, as I lost my coding skills more than a decade ago, but with AI I can focus on the system and the product in partnership with agent coding workflows.<p>I recently built a Voice AI for older people. To build it I created an agentic engineering workflow (feel free to rip that up as I\u2019m always looking to improve systems: <a href=\"https:&#x2F;&#x2F;ozbrain.com&#x2F;resources&#x2F;eng-flow\" rel=\"nofollow\">https:&#x2F;&#x2F;ozbrain.com&#x2F;resources&#x2F;eng-flow</a>) My approach with coding agents is trust but verify, and I\u2019m trying to replace the parts where a human would review with an adversarial or specialized agent who would give a better answer&#x2F;review.<p>I have workflows that will go high level task to shipped PR running in Claude cloud sessions. I use Claude Code locally and Cursor when I want a tighter loop on doing visual work like UI or layout. And Codex to either load balance usage for TokenThriffting or when I want a different llm to think thru something.<p>It was a pain in the ass passing .md files around and keep track of which version was the most recent, so I built a hosted .md storage right in Supabase and any of my agents already have Supabase access. This let me build a solid, scalable, secure voice AI from my phone at the gym. All my agents have access to our knowledge, can write to it, update and refer to it as we build and improve the product and the systems we use.<p>Out of 75 founder friends I asked about how they manage shared knowledge, 26 built their own custom knowledge systems\u2026 Obsidian vaults with 7k files synced through a VPS, markdown repos behind their own MCP servers, cron jobs stitching Supabase to a skills file\u2026 each a different Frankenstein they have to maintain. 32 said they felt the pain of moving static files around but didn\u2019t have any solution for it.<p>So I rebuilt my brain better and used it to build it.<p>&#x2F;&#x2F; HOW YOU CAN HELP<p>Would love to have you try it out. The maintenance loop is still in alpha so not running it on customer data yet.<p>If you built your own brain I\u2019d love to hear how you did it. What criteria was most important for you in its design &amp; function.<p>If you are tired of shuffling .md files around I\u2019d love to have you try out OzBrain and to give feedback, just ask your agent to put it in the shared bugs &amp; features brain!<p>Cheers!\nBubs.co", "5": "2026-08-22T19:14:46.777326"}
{"0": 53, "1": "hackernews", "2": "https://techcrunch.com/2026/08/21/nvidia-just-showed-that-the-harness-not-the-ai-model-is-now-the-real-hero/", "3": "Nvidia just showed that the harness, not the AI model, is now the real hero", "4": "Nvidia just showed that the harness, not the AI model, is now the real hero. ", "5": "2026-08-22T19:14:46.784024"}
{"0": 54, "1": "hackernews", "2": "https://wondering.app/canvas", "3": "Show HN: Visual way to understand things in parallel", "4": "Show HN: Visual way to understand things in parallel. Hey HN!<p>Really excited to be showing what I think is a better way to understand complex topics with AI.<p>You start a chat with your question, and whenever you want to clarify something or understand some jargon, you can branch out a new chat from that thread.<p>The cool unlock is being able to see all related chat threads in the same context without tab switching, for e.g. I can first dive into understanding world models, then spin a new chat thread on who the key players are in there, and another asking what&#x27;s the frontier with this stuff.<p>The responses are also not just a wall of text. They&#x27;re:\n- Filled with interactive diagrams and visuals\n- Super fast<p>And for those curious, this is how we built it:\n- React Flow (@xyflow&#x2F;react) for the canvas\n- React Components for interactive diagrams\n- GPT Image 2 for image generations\n- Gemini 3.5 Flash Lite for super fast response\n- Parallelization whenever possible to keep things fast<p>You can also highlight and add notes :)<p>Try it out and let me know what you think!", "5": "2026-08-22T19:14:46.789578"}
{"0": 55, "1": "hackernews", "2": "https://github.com/proliferate-ai/proliferate", "3": "Show HN: Proliferate- open-source, self-hostable Codex for any coding agent", "4": "Show HN: Proliferate- open-source, self-hostable Codex for any coding agent. Hi HN- I&#x27;m Pablo, the founder of Proliferate (YC S25)!<p>Proliferate (<a href=\"https:&#x2F;&#x2F;github.com&#x2F;proliferate-ai&#x2F;proliferate\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;proliferate-ai&#x2F;proliferate</a>) is an open-source, self-hostable AI IDE that lets you work and automate tasks with Claude Code, Codex, OpenCode, Cursor, and Grok in one place.<p>Here&#x27;s a quick 2m demo of how we use Proliferate to build Proliferate: <a href=\"https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=tGNX0oaWmBY\" rel=\"nofollow\">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=tGNX0oaWmBY</a><p>I started building Proliferate after my team onboarded to OpenAI Codex. Within days, we were using it for everything: using computer use instead of navigating websites ourselves, having Codex coordinate other agents, and setting up automations for recurring work. We really never needed to leave the desktop app to get work done.<p>If my team\u2019s experience is anything close to representative, a Codex-like app (a horizontal agent with a beautiful UI) is the main interface every company is going to use to get work done. That is perfectly in line with OpenAI\u2019s mission to make Codex the everything app (see: <a href=\"https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=47796469\">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=47796469</a>).<p>But as we started automating work closer to the core of the business, I wanted to work with agents from all the labs, including open-weight models, without becoming increasingly dependent on OpenAI.<p>And that\u2019s what Proliferate is for! It&#x27;s the open-source, self-hostable Codex that preserves your optionality across agents and model providers while building toward Codex\u2019s breadth.<p>Today Proliferate supports:<p>* Working with Claude Code, Codex, OpenCode, Cursor, and Grok with their native inference and configuration options, including Bedrock, Azure, and self hosted inference.<p>* Inter-agent communication and management: a parent agent can spawn and communicate with another supported agent as a subagent (I personally like to have Fable delegate to Codex, with OpenCode models reviewing PRs).<p>* Building workflows- one of the features I&#x27;m most excited about. These are re-usable chains of agent sessions and human approval gates, with the harness and model chosen per step and documents passed between them. I use this to automate code review, QA, and my PR construction process.<p>All of Proliferate is 100% open source under AGPL-3.0. There are still definitely some rough spots, but we\u2019re building fast and I\u2019d really love any feedback!", "5": "2026-08-22T19:14:46.796069"}
{"0": 56, "1": "hackernews", "2": "https://www.basecompute.co/local", "3": "Show HN: Zero () friction local AI for Mac", "4": "Show HN: Zero () friction local AI for Mac. Super excited to launch our new app Local today. What we\u2019ve learned at Base Compute over the last months is that running AI directly on your laptop or workstation gives you maximum privacy and it\u2019s free, but it\u2019s also a massive headache to configure. So we\u2019ve decided what matters is making the experience completely frictionless for users.<p>Local analyses the hardware of your laptop, optimises the AI for it, and recommends the best models for your specific device.<p>It let\u2019s you do what you\u2019re doing with cloud AI already, just for free and on your own machine: Chatting with PDF\u2019s, Recording and summarising meetings, running coding agents...<p>If you\u2019re using Local in your office with colleagues, you can run it in \u201cOffice Mode\u201d. The strongest computer in your office runs the AI and everyone can connect to it with their laptop. The data never leaves the office.<p>It\u2019s available for download on our website today, please try it out and let us know what you think!", "5": "2026-08-22T19:14:46.805768"}
{"0": 57, "1": "hackernews", "2": "https://tablecanon.app/", "3": "Show HN: Building Table Canon, an AI Campaign Memory Engine for TTRPGs", "4": "Show HN: Building Table Canon, an AI Campaign Memory Engine for TTRPGs. Hey HN! I built Table Canon to solve a problem my playgroup kept running into: 3-4 hour tabletop gaming sessions leave behind massive audio recordings, but standard meeting note-takers treat every session as an isolated island, butcher fantasy terms, and don&#x27;t know who is speaking.<p>I wanted an engine that tracks long-term state across months of games, so I built a pipeline to extract entity updates, open quest hooks, and character promises across sessions.<p>The Tech Stack:<p>* Transcription: whisper-large-v3-turbo  \n* Diarization: pyannote for speaker embeddings &amp; voice profile matching  \nExtraction &amp; Memory: OpenAI API with Structured Outputs (JSON Schema enforcement for state updates)  \n* TTS &amp; Audio Recaps: Kokoro &#x2F; Chatterbox Turbo  \nMusic Generation: ACE-Step-v1.5-XL-Turbo for rendering session summaries into lyrics&#x2F;ballads<p>A Few Engineering Lessons &amp; Challenges:<p>* State Delta Extraction vs. Context Explosions: Feeding 20 prior session transcripts into context windows quickly becomes cost-prohibitive and noisy. Instead of re-reading raw history, each session outputs an atomic state delta (updates to NPC dossiers, new locations, resolved promises) to a database. Keeping context bounded as campaigns stretch past session 30+ has been one of the trickiest architectural hurdles.\n* Custom Pre-Lexicons: General STT models struggle with homebrew proper nouns (turning fantasy names into standard dictionary words). Injecting a pre-pass fantasy term dictionary into prompt context significantly improved first-pass spelling.\n* VAD &amp; Audio Chunking: Passing a 4-hour raw audio file directly to Pyannote&#x2F;Whisper leads to memory leaks and process timeouts. Pre-processing with Voice Activity Detection (VAD) and deterministic chunking was necessary before touching the models.<p>Current Limitations &amp; Active Hard Problems:<p>* Entity Alias Resolution: Matching entities across sessions when players use varying aliases or informal shorthand (e.g., matching &quot;The Red Bishop&quot; to &quot;Arthur&quot; or &quot;that cult leader guy&quot;) without accidentally merging distinct NPCs. I address this, partially, but allowing the user to Edit aliases, merge or split entities after-the-fact.\n* Quest &amp; Hook Resolution Logic: Fine-tuning the LLM to reliably determine whether a promise, open mystery, or quest has actually been resolved versus remaining open or implicitly abandoned.<p>I&#x27;d love feedback on how others are handling these sorts of issues - or any notes for folks who try it out! No initial login required with 6 hours of upload available to try.", "5": "2026-08-22T19:14:46.811577"}
{"0": 58, "1": "hackernews", "2": "https://epho.io", "3": "Show HN: Epho \u2013 run Claude Code with a curl", "4": "Show HN: Epho \u2013 run Claude Code with a curl. Hey folks, Burak here.<p>Epho is an API that allows running Claude Code, Codex or Opencode in a sandbox in the cloud. It abstracts away sandboxes, and allows running coding agents with a single HTTP request.<p>Epho came out of our own struggles with building our own AI analyst:\n- Sandboxes give you bare machines; you need to configure them for agentic workloads.\n- Each agent behaves differently, and you need to build integrations with each of them.\n- Sandbox providers are not very reliable, which means you need to figure out a multi-provider strategy to avoid failures.\n- Logging, artifacts, input&#x2F;output, event streaming, and all of the other operational aspects need to be figured out.<p>We had to go through the pain ourselves. We got to a point where things got quite reliable, and it became more obvious to us that this should be a primitive on its own: send a POST request, get the events streaming back to you.<p>Epho is an agents-as-an-API product: you send a request, it spins up a sandbox, configures the chosen harness, clones your repos, and kicks off the agent. It takes care of automatic fallbacks across different providers, handles auth stuff, and just streams back the events and outputs.<p>It supports Claude Code, Codex and Opencode out of the box, and pretty much all the models they support out of the box. It streams the events back, handles attachments and output files, automatically manages the fallbacks on different sandbox providers, retries, and all the auth stuff. You just send a prompt, your repo, MCP servers you want to use with it, and it runs them.<p>I recorded a demo here to show a real example: <a href=\"https:&#x2F;&#x2F;youtu.be&#x2F;HGfly1aytPA\" rel=\"nofollow\">https:&#x2F;&#x2F;youtu.be&#x2F;HGfly1aytPA</a><p>I am quite excited for Epho, simply because I think it is a new primitive that would allow building agents into product a lot easier than it is today. We are running our agents on Epho on prod, so we&#x27;ll keep maintaining it regardless, and we wanted to ship it as an independent product.<p>Epho is free to get started, and you can run it with Opencode&#x27;s free models to get started with it.<p>I am quite curious to hear what you&#x27;d think and would love to get your feedback.<p>Cheers,\nBurak", "5": "2026-08-22T19:14:46.820922"}
{"0": 59, "1": "hackernews", "2": "https://github.com/runvendo/vendo", "3": "Launch HN: Vendo (YC S26) \u2013 Let users build features on top of your product", "4": "Launch HN: Vendo (YC S26) \u2013 Let users build features on top of your product. Hi HN, we\u2019re Yousef &amp; Nour, founders of Vendo (<a href=\"https:&#x2F;&#x2F;vendo.run\">https:&#x2F;&#x2F;vendo.run</a>). Vendo lets users create new features inside the software they already use. A user describes the dashboard, workflow, or small app they need, and Vendo builds it on top of the product\u2019s existing data, API, and interface.<p>Demo: <a href=\"https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=VdpHehY64ls\" rel=\"nofollow\">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=VdpHehY64ls</a><p>We built Vendo because every SaaS eventually faces the same problem: every customer needs something slightly different. One wants a new report and another needs a workflow that only makes sense for their team. These requests either sit on the roadmap, become one-off engineering work, or force the customer into spreadsheets and external tools. We wanted the user to be able to create the missing feature themselves, without leaving the product.<p>Here is how it works:<p>- npx vendo init reads the product&#x27;s API surface, theme, routes, and more. These are used so that the apps Vendo creates (1) look on-brand and native and (2) have the ability to read data and perform actions directly through the company&#x27;s API<p>- When a user asks for a feature, we have a custom Vendo harness that writes a React component with a bunch of Vendo add-ons and guardrails (ex. ability to make calls to the host API + our component library). Every save is compiled, type-checked, run against real API responses, and rendered before the user sees it. We just released a benchmark and write-up here with more info for anyone interested: <a href=\"https:&#x2F;&#x2F;vendo.run&#x2F;blog&#x2F;generating-product-ui-measured\">https:&#x2F;&#x2F;vendo.run&#x2F;blog&#x2F;generating-product-ui-measured</a><p>- We use QuickJS to make sure that anything the agent creates is sandboxed and can&#x27;t mess with the company&#x27;s site. Vendo compiles the component and runs it with Preact inside a QuickJS VM with no access to the DOM, network, or clock. The VM returns a UI tree, which the host renders using the product\u2019s registered components. When the user clicks something, QuickJS emits a tool call; the host executes it through Vendo\u2019s guard and passes the result back into the same VM, preserving the screen\u2019s local state.<p>There&#x27;s a lot of generative UI right now: streaming developer-written components into a chat (Vercel AI SDK, CopilotKit, Thesys), or rendering your app inside someone else&#x27;s assistant (OpenAI Apps SDK, MCP Apps). We differ on two things. Vendo lives in your product and acts through your API as the signed-in user, so what it makes is durable: real apps users keep, pin, and run on triggers while they&#x27;re away, and not components that are merely confined to a chat. Plus, it&#x27;s not capped at putting together a bunch of prebuilt components: the agent can build arbitrary apps, from a quick dashboard out of your own components to real custom code running in a sandbox, and either way data only ever comes from tool calls to your API.<p>Here are some things customers are using Vendo for today:<p>- Letting their users create custom dashboards and reports. These are mainly UI-based and focused on letting the user see the exact graphs and metrics they care about<p>- Letting their customers create recurring automations. A big thing as well that has been used for these automations is the fact that we connect to external connections, so users have been automating many of their inter-tool workflows (ex. an automation that sends a slack alert based off of something in the product)<p>- B2B customers letting their customers customize the product with specific business logic. Often this is simple things like an extra field on a form, or an extra permission, but it is hard for a business to keep up with them otherwise.<p>- Creating and sharing custom dashboards&#x2F;apps across an organization. Since the apps Vendo creates are durable, they can be shared, reused, and forked (which can\u2019t be done with many of the other in-chat generative UI solutions)<p>We&#x27;ve spent a lot of time thinking about how AI and agents will change the way people consume software. We think the answer is personal(ized) software: you see the UI you need to see, you tell an agent exactly what you need, and the product molds to how you work.<p>The key insights that have enabled the product to work are:<p>- A rule in code always beats a rule in a prompt.<p>- Invent as little syntax as possible. Generation got faster and more reliable when the output looked like what models already know (JSX-shaped markup) instead of a clever custom format.<p>- Deterministic beats model wherever you can get away with it. Theme extraction is pure static analysis, and a remix starts as a copy of your component, no model call.<p>Vendo is completely open-source (Apache-2.0) and can be self-hosted, so feel free to c", "5": "2026-08-22T19:14:46.829283"}
{"0": 60, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49371913", "3": "Offline RAG on iOS with Spatial Integration", "4": "Offline RAG on iOS with Spatial Integration. I&#x27;m the developer behind CartoType. I\u2019ve been working on bridging local language models with offline mapping, and I have just put together a demo of a completely offline Spatial RAG pipeline running natively on an iPhone. The new system is named the CartoType Field Assistant. You can find the website at https:&#x2F;&#x2F;cartotype.com<p>Demo (1m 19s): https:&#x2F;&#x2F;www.youtube.com&#x2F;shorts&#x2F;a8yQPn7_jyI<p>Use cases: Any organisation with field technicians or emergency first responders needs complex procedural knowledge tied to physical locations (&#x27;assets&#x27;) where they don&#x27;t have guaranteed network connectivity. Examples include offshore wind farms, power distribution networks, railway infrastructure,  mountain rescue, and military uses.<p>For this demo I query the iPhone app in Airplane mode: &quot;A hiker near Tuolumne Meadows has a dislocated shoulder. What is the reduction protocol, and where is the Tuolumne Meadows Ranger Station?&quot;<p>HOW IT WORKS<p>The core is a portable C++ engine running vector search across an encrypted, on-device SQLite database. The database contains the user&#x27;s proprietary manuals, chunked and converted into embeddings. The search finds relevant chunks and uses them as a prompt for a local LLM (Gemma). This process is RAG (Retrieval-Augmented Generation).<p>Spatial integration is provided by connecting assets in the map to a table in the database and using the table to find any assets referred to in queries and pan the map to them.<p>Everything runs on-device with no API keys or cloud dependencies. The demo app is written in Swift and uses the CartoType framework, which provides a wrapper over the underlying core API giving asynchronous and synchronous access to the AI functions.<p>DATA PREPARATION<p>To get up and running, the user creates a map containing the assets, using CartoType&#x27;s <i>makemap</i> tool, then feeds the map and their proprietary documentation into CartoType&#x27;s <i>makedata</i> tool, which writes the encrypted SQLite database to be stored on the device. The map, which may also be encrypted, is also stored on the device. The CartoType library, when running the Field Assistant&#x27;s RAG system, also provides full map rendering, location searching, routing and geocoding as it always has done.", "5": "2026-08-22T19:14:46.839398"}
{"0": 61, "1": "hackernews", "2": "https://mainly.crnst8.com/", "3": "Show HN: Mail client optimized for self-hosted email across multiple domains", "4": "Show HN: Mail client optimized for self-hosted email across multiple domains. As someone with a clean IP range and ADHD, I manage my inbox by splitting across a variety of accounts for a variety of domains. I\u2019ve never really found a client that scratches all my itches, so I spent some time making one and it\u2019s now become my daily driver.<p>It\u2019s optimised to be visually simple and highly configurable for things like colour, folders&#x2F;subfolders, sorting and tweaking the search query function to better match what you search for often. There\u2019s also some little nice-to-haves like a purpose-built PWA, bulk domain onboarding, MCP and dark mode.<p>It\u2019s fully self-hostable, open source and free of any AI bloat&#x2F;VC fodder. Not trying to sell anything but rather make something un-shitified and simple to use, would love some feedback if  you\u2019ve given it a try! :~)", "5": "2026-08-22T19:14:46.846025"}
{"0": 62, "1": "hackernews", "2": "https://github.com/onecli/onecli", "3": "Launch HN: OneCLI (YC S26) \u2013 OSS sandboxed agent harness for teams", "4": "Launch HN: OneCLI (YC S26) \u2013 OSS sandboxed agent harness for teams. Hi HN, Jonathan &amp; Guy here from OneCLI, an agent harness built for teams, giving every employee a secured, sandboxed personal agent.<p>Here\u2019s what you can do with it:<p>1. get a sandboxed agent, with all the OneCLI capabilities in place like connect your GitHub account, Gmail, Notion, or Dropbox simply from the chat.<p>2. deterministic human in the loop approval in the chat itself for things that you need 100% control like sending an email or deleting the Linear ticket.<p>3. manage team policy in one place, enforced across every agent in the workspace<p>4. enjoy global connections at the team level, like shared LLM keys or service accounts<p>Here\u2019s a demo: <a href=\"https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=dlW-44ntpbE\" rel=\"nofollow\">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=dlW-44ntpbE</a><p>We started working on this by accident, even though our careers were in the security space. We were working on a devtool called ChartDB, an open-source DB tool. When OpenClaw took off back in January, we started using it to orchestrate agents on top of ChartDB. We quickly understood there is a big issue around auth. Agents need credentials to do real work, but to give them those secrets would not be the best idea. They keep them in their memory and also write them down to local files and their sessions as plain text. And we knew that agents can easily be fooled into giving up those API keys&#x2F;secrets. So we needed some way to control the agent and stop prompt injections from tricking it into using its services for an attacker&#x27;s benefit.<p>We created OneCLI that started as a vault for AI Agents built in Rust.<p>We found out that most of our demand for OneCLI came from autonomous agents like Hermes, OpenClaw and NanoClaw for individuals and teams.<p>Users looked for useful agents that do things for the person who runs them with two missing parts: 1) managing secrets and permissions. 2) and for teams - multiplayer management.<p>We decided to pivot and provide the agent itself as a harness for teams, to give each employee an agent. We saw that teams had to deal with setting up their own harness again and again, and basically as we already had the vault as a gateway. We got the idea to provide the missing piece of the agent management out of the box and open source it (Apache-2.0, with a small enterprise exception).<p>We&#x27;re open source first - the entire platform, not just a small portion of it like other agents, so companies can actually see the code, evaluate it, and trust it instead of taking our word for it. They run it isolated, in their own environment, fully under their control, at production quality, not a locked black box hosted somewhere else. That means the safety isn&#x27;t just a promise, it&#x27;s something they can verify themselves. Combined with real autonomy and least-privilege access, that&#x27;s what makes it something a company can fully own and trust, not just adopt.<p>We also approach this from a company perspective rather than an individual one. Our solution manages agents on behalf of each employee, wrapped in deterministic guardrails that company admins configure through centralized policies.<p>For the agent engine itself we\u2019re using jcode which is the core of the agent-loop. We found out that it improves the experience and makes the agent smarter and faster.<p>Here\u2019s how it works:<p>It runs on infra you control. Fully open-source, self-host or cloud in minutes.<p>The agent never holds a real secret. It gets a placeholder. The real credential is injected at the gateway, per request, after the call is authorized. It never enters the agent&#x27;s context, memory, or logs.<p>Enforcement outside the model. Prompts are suggestions. Policies defined by the org admin run at the network layer, outside the agent and the LLM. Block endpoints, rate limit per agent, require approval, scope per employee. The gateway decides. The agent can&#x27;t bypass it.<p>Isolated VM per agent. Own memory, own keys, own permissions. Blast radius is one agent.<p>Speed of the Harness: Rust engine under the agent loop.<p>Full identity trail. Every agent is bound to an employee. Every call logged with who it acted for and which policy allowed it.<p>Some things people are doing with the platform include:<p>- Managing their company life cycle entirely from the sales calls, to the product side automatically open tickets to the engineering teams, that would kick the development agents to deliver and ship to production.<p>- Operational side, like automatically hygiene the CRM after calls, sourcing leads, book meetings and manage follow ups emails.<p>- Some of our customers also doing their entire grocery shopping using those agents and send them to take care of their chores like ordering things online.<p>About the team: Both founders come from cybersecurity backgrounds. Jonathan spent years at Axis Security building zero trust network acces", "5": "2026-08-22T19:14:46.852361"}
{"0": 63, "1": "hackernews", "2": "https://github.com/danielealbano/android-remote-control-mcp/", "3": "Show HN: MCP app for Android, drive apps via AI (no root, PII redacted locally)", "4": "Show HN: MCP app for Android, drive apps via AI (no root, PII redacted locally). Hi HN, author here.<p>Android Remote Control MCP is an MCP server that runs directly on your Android phone (no root, no ADB, no computer in the middle) and lets an AI agent drive real apps the way a human would: it reads the screen through the accessibility tree, taps, types, scrolls, optionally also screenshots.<p>I spent a lot of time optimizing tool usage and token consumption, and making it work not only via local harnesses but also via Claude.ai &#x2F; Claude Desktop and chatgpt.com (if you have the proper account, the app acts as its own OAuth server, you approve connections with a code on the phone).<p>The newest part is Privacy Mode, and it exists because after an earlier release someone told me in plain terms they&#x27;d never use it because, rightfully, they didn&#x27;t want the LLM provider to see everything on their screen!\nSo now a combination of a small local model plus deterministic detectors identify personal information (emails, phone numbers, credit cards, IBANs, national IDs, English names, etc.) and redact it on-device before anything leaves the phone, with a benchmarked detection rate of about 87% (the benchmark is in the repo; non-English names, are the current weak spot but I am working on it).<p>Why did I build it? I want my agents to be able to use my phone to do searches, book things for me and do sometimes boring stuff ... without sharing all my data with service providers and without having to run a local LLM!<p>Tradeoffs: the service declares itself as an accessibility app so it can read apps which also makes it impossible to distribute this app on the Google Play store, hence it&#x27;s distributed via GitHub at the moment (there&#x27;s a standard build and a FOSS build without Google Play Services which soon will be published on F-Droid).<p>Also, because I care about prompt injection as much as privacy, the data returned to the agent is prefixed with a message, which uses quite some strong wording, to make it clear that the content is supplied by third party sources and is an untrusted input! Mitigates the potential attacks a lot.<p>Is it perfect? Nope, there&#x27;s plenty to improve, especially around apps coverage but even small models (like Haiku) can drive the phone and handle unseen situations if you give them enough detail!<p>What I am working on next? Three major things:\n- a free reverse tunnel, encrypted end to end, with Let&#x27;s Encrypt certificates to be able to use the app with Claude.ai &#x2F; Claude Desktop &#x2F; Chatgpt.com and such for free and with much more privacy than using ngrok or cloudflare (which don&#x27;t provide e2e encryption but only encryption to the edge, so they see everything)\n- a skills database for the apps, so your agent can work out what to do much faster\n- a custom model to improve the privacy standing, especially for non-English names<p>And, additionally, a new guided setup and a revised UI to make it easier to start to use the app!", "5": "2026-08-22T19:14:46.858416"}
{"0": 64, "1": "hackernews", "2": "https://deftwriting.com", "3": "Show HN: Deft Writing, an AI lab for non-slop LLMs", "4": "Show HN: Deft Writing, an AI lab for non-slop LLMs. Announcing Deft, a new AI lab for better writing<p>Currently, 86% of user queries are fully human according to pangram.<p>This is still a small beta model and it might make mistakes. We are launching our public beta now to get more feedback before scaling up.<p>Tips for better performance: \n- Add more details to your prompt. If you just provide a short sentence prompt, it will likely get detected as AI.\n- Try changing the style in &quot;advanced options&quot;\n- Deft currently works better for some use cases like Analysis&#x2F;Essays, Creative writing, and Rewrites. It works less well for Marketing copy and news articles.<p>Our main goal is better writing, fooling AI detectors is just a side effect.<p>Here&#x27;s the launch announcement the Deft model wrote for itself:<p>Introducing Deft Writing\nWe are proud to announce the launch of our new startup, Deft Writing. Having spent many years in the trenches with other Large Language Models (LLMs), we were continually hampered by poor writing. We believe writing well is hard. So we&#x27;re making it easier for you to share your ideas, and by doing so make those ideas more readable for those who have to read them. Deft Writing saves you time writing and allows you to focus on your ideas.\nWe&#x27;ve developed a new training algorithm for models called Distribution Fine Tuning (DFT) that makes model outputs more human-like. Many other LLMs produce outputs riddled with overused &quot;slop&quot; phrases. DFT discourages this bad behavior.\nOur current offerings:\n*   Our beta model is available for you to try out.\n*   We also provide API access to the model.\n*   For larger enterprises that want a model in their own unique style, we can train a custom model on your data to your specifications.\nGo check us out at www.deftwriting.com and try it out.", "5": "2026-08-22T19:14:46.868763"}
{"0": 65, "1": "hackernews", "2": "https://aiomniu.top/services/resume-builder", "3": "Show HN: AI Resume Optimization Tool for Specific Job Positions", "4": "Show HN: AI Resume Optimization Tool for Specific Job Positions. Hi HN, I&#x27;ve been watching everyone&#x27;s presentations as an observer, and today I can finally showcase my product!<p>It helps users:<p>- Organize roles, projects, skills, and achievements<p>- Transform vague descriptions into clearer, more professional language<p>- Adjust the focus according to different target positions<p>- Check resumes before submission to the ATS system<p>Its core principle is: AI should enhance expressiveness, not fabricate experience or create false achievements.<p>Relatively speaking, this isn&#x27;t my proudest achievement, as resume optimization isn&#x27;t a new tool, but it&#x27;s part of my main project&#x27;s ecosystem.<p>Regarding the models used, our site&#x27;s AI tool primarily uses Deepseek v4 flash and glm5.2. We will upgrade to the latest versions as the overall model evolves, such as the current glm5.3. We will be offering Beta testing access to 100 job seekers. Beta testers will receive:<p>- Free use during the testing period<p>- 900 credits per month after launch<p>- A free resume webpage after launch<p>Students, career changers, and experienced professionals are welcome to try it:<p>I especially welcome feedback on the accuracy and usefulness of the suggestions and what shortcomings the tool currently has. What is the most frustrating part of writing or improving a resume?", "5": "2026-08-22T19:14:46.874951"}
{"0": 66, "1": "hackernews", "2": "https://github.com/blak0p/attack-shark-linux", "3": "Show HN: Reverse-engineered a gaming mouse's HID protocol to bring it to Linux", "4": "Show HN: Reverse-engineered a gaming mouse's HID protocol to bring it to Linux. I\u2019ve been reverse-engineering the HID protocol of an Attack Shark X6 gaming mouse (no official Linux support) using Ghidra and USB captures, and building a native Linux desktop app around it in Go + Wails, with a React frontend.<p>The protocol itself is fairly advanced. DPI, RGB lighting, polling rate and button remapping are all documented at the protocol level. The app currently only exposes DPI configuration though, the rest is reverse-engineered but not yet wired into the UI.<p>One piece is still unsolved: the macro report isn\u2019t captured yet, so there\u2019s no macro editor. If you\u2019ve done HID&#x2F;USB reverse engineering before, I\u2019d appreciate a hand there.<p>There\u2019s also plenty of non-reversing work: building out the RGB&#x2F;polling&#x2F;remap screens in the frontend, implementing what\u2019s already known in the Go backend, or testing on other Attack Shark models (X3, R1, X11) that likely share the same dongle and protocol.<p>Happy to answer questions about the protocol work or the architecture in the comments.", "5": "2026-08-22T19:14:48.326495"}
{"0": 68, "1": "hackernews", "2": "https://www.ito.ai/blog/ai-model-plateau-why-infrastructure-matters-more-next-release", "3": "Why your infrastructure is more important than the next LLM release", "4": "Why your infrastructure is more important than the next LLM release. ", "5": "2026-08-22T19:14:48.342523"}
{"0": 69, "1": "hackernews", "2": "https://thezvi.substack.com/p/openais-unreleased-model-astra-solves", "3": "OpenAI's Unreleased Model Astra Solves Ten Major Open Mathematics Problems", "4": "OpenAI's Unreleased Model Astra Solves Ten Major Open Mathematics Problems. ", "5": "2026-08-22T19:14:48.350635"}
{"0": 72, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49374554", "3": "Show HN: LilScript makes JavaScript libraries smaller", "4": "Show HN: LilScript makes JavaScript libraries smaller. <a href=\"https:&#x2F;&#x2F;yeargun.github.io&#x2F;lilscript&#x2F;\" rel=\"nofollow\">https:&#x2F;&#x2F;yeargun.github.io&#x2F;lilscript&#x2F;</a><p>LilScript is a typed, compression-first language that compiles into js and sometimes into exec(will be more stable in future). The compiler mangles, reshapes the program into optimized js that happens to be 5-15% smaller (after gzip&#x2F;br compression or raw) compared to the best performing JS toolchains like oxc&#x2F;esbuild&#x2F;terser&#x2F;..<p>## What has been proven to work with LilScript?<p>- makes VSCode&#x27;s core js modules 20% smaller on average<p>- makes the world&#x27;s most performant &amp; small markdown rendering npm library, marked, 5-7% smaller and 10% faster<p>- and many more demos.. works with pretty much any js&#x2F;ts library<p>## How it is compressing js finer than vite&#x2F;oxc&#x2F;terser&#x2F;esbuild&#x2F;..<p>### 1. By changing the app.<p>```\nclass Vector {\n  float x;\n  float y;<p><pre><code>  init(float x, float y) {\n    this.x = x;\n    this.y = y;\n  }\n\n  float lengthSquared() {\n    return this.x * this.x + this.y * this.y;\n  }</code></pre>\n}<p>int[] values = [1, 2, 3, 4];\nauto doubled = values.map((int value) =&gt; value * 2);\nint sum = 0;\nfor (int i = 0; i &lt; doubled.length; i++) {\n  sum += doubled[i];\n}\nVector vector = new Vector(3.0, 4.0);\nif (vector.lengthSquared() == 25.0) {\n  print(`sum=${sum}`);\n}\n```<p>Above code compiles into this:<p>```\nvar b=[1,2,3,4].map(a=&gt;a<i>2|0);<p>var a=0,c=0;<p>while(a&lt;b.length){<p><pre><code>  c=c+b[a]|0;\n\n  a=a+1|0;\n</code></pre>\n}<p>console.log(`sum=${c}`)<p>```<p>This is not minification of the same program. The compiler changed the app.<p>oxc &#x2F; esbuild &#x2F; terser &#x2F; .. starts from JS and mostly keeps the shape of the app.  Meanwhile LilScript the language is designed from scratch to give compiler any extra knowledge that could help the compiler&#x27;s compilation. Just like Google Closure Compiler - Advanced Mode, and beyond.<p>### 2. Tryhard property mangling.<p>Visit the world&#x27;s most used web applications, chatgpt.com, and read the source codes. You will see that, there are lots of framework related js properties that dont get minified, and are indeed human readable.<p>Those names stay in the bundle because the toolchain cannot prove they are local. A property might be a public API, a DOM field, a framework hook, or something a plugin reads by string. So the minifier leaves it.<p>LilScript:<p>- *a-* does both eliminates the objects, weird code structures into more optimized variables&#x2F;arrays\n- *b-* rename any variable into mostly occured, short versions, field cleverly to minimize entropy (based on the objective compression algorith. gzip&#x2F;brotli) so that the end result is highly compressed.<p>*(a)* is the same move as the `Vector` example, at library scale: objects and classes that only exist as a programming convenience get flattened into scalars, arrays, and tight loops.<p>*(b)* is not &quot;make every name 1 letter.&quot; gzip and brotli win when the same short tokens repeat. The compiler picks the names that show up the most, and scores the spelling against the compression algorithm you asked for (gzip, brotli, or raw). Different `cost_model` \u2192 different names \u2192 a different file that is smaller </i>after* that codec.<p>That is why the same program can be 5-15% smaller after gzip&#x2F;br compression or raw: the JS is shaped and named for the compressor, not just for a human reading the AST.<p>Feel free to PR, experiment (Please respect the modified MIT license)<p>LLM models can oneshot implement your library with LilScript. For non optimized libraries, it could end up 20%+ size reduction and somewhat performance improvements (which usualy matters very little)<p>Please share your opinions, would love to discuss about the future direction for the language and the compiler", "5": "2026-08-22T19:14:48.365348"}
{"0": 73, "1": "hackernews", "2": "https://github.com/PriorLabs/relarena", "3": "Show HN: RelArena-\u03b1 \u2013 open-source releases for Relational Learning", "4": "Show HN: RelArena-\u03b1 \u2013 open-source releases for Relational Learning. I\u2019m happy to announce our first release in relational learning at Prior Labs, continuing our commitment to open science.<p>Most of the data that actually matters to a business- users, transactions, sessions, orders- lives in relational databases, not in a single spreadsheet. SQL is built for backward-looking analysis: what happened, how much, when. It can&#x27;t tell you whether a given user will churn in the next 30 days. This is a relational machine learning problem: making predictions directly over the structure of a database. It&#x27;s still a young field, where results are often hard to reproduce, hard to compare across methods, and rarely tested outside benchmarks like RelBench. We argue that these problems have slowed progress in relational learning.<p>We open-source three pieces of software that we expect to accelerate research in the field towards meaningful, real-world impact.<p>First and foremost, we release -\u03b1: a unified framework for running and comparing baselines on RelBench v1 tasks. Based on learnings from tabular benchmarks like TabArena, we are standardizing data loading, evaluation protocols, tuning regimes, and adding support for systems with custom tuning.<p>We also open-source -: our relational harness for TabPFN-3 (our Tabular Foundation Model). We initialize the (living) RelArena-\u03b1 leaderboard with TabPFN-Rel and a comprehensive set of baselines. The rankings at the time of release are:<p>\u2022 - is the No. 1 model submission\n\u2022 - is the No. 1 system submission<p>Last but not least, we open-source an alpha version of the    (): enabling you to easily specify prediction tasks on your own relational database and then run any RelArena-\u03b1 model, like TabPFN-Rel, in a few lines of code, all bundled as a simple PyPI package.<p>\u2022 Read the full model report: <a href=\"https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2608.16319\" rel=\"nofollow\">https:&#x2F;&#x2F;arxiv.org&#x2F;abs&#x2F;2608.16319</a>\n\u2022 GitHub repository: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;PriorLabs&#x2F;relarena\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;PriorLabs&#x2F;relarena</a>\n\u2022 Announcement: <a href=\"https:&#x2F;&#x2F;priorlabs.ai&#x2F;blog-posts&#x2F;introducing-relarena\" rel=\"nofollow\">https:&#x2F;&#x2F;priorlabs.ai&#x2F;blog-posts&#x2F;introducing-relarena</a>\nCookbook: <a href=\"https:&#x2F;&#x2F;docs.priorlabs.ai&#x2F;cookbook&#x2F;relational_predictions_tabpfn_rel\" rel=\"nofollow\">https:&#x2F;&#x2F;docs.priorlabs.ai&#x2F;cookbook&#x2F;relational_predictions_ta...</a>", "5": "2026-08-22T19:14:48.371857"}
{"0": 75, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49353513", "3": "LLMs .what do you smoke beforehand?", "4": "LLMs .what do you smoke beforehand?. I know this is going to sound negative, and probably a bit generalist but....seriously, what exactly is the LLM helping you with that is not just purely &quot;time saving&quot; - ie. Actual, novel work.<p>I personally feel like while I cannot prove it, nor want to frankly, that LLMs have been neutered intentionally since around October of last year. 4.5-4.6 Opus....<p>Every single company has degraded their models eirher intentionally or accidentally(probably worse) and they cannot admit it and will gaslight you if you even try to acknowledge it....but for my own experiments, LLMs are actually worse than no-LLM....<p>I dont see anyone building anything they couldnt have done before hand....and I am seeing stuff that didnt need to be made...much more frequently....to the degree that I cannot remember the last good idea I have seen even on here (whereas there were cool ideas every month posted here)...<p>I also cannot get even mediocre output anymore...Its literally just trash - Literally a prompt to dumpster pipeline....<p>Clause is the worst. Hands down. They have regressed on every single model since 4.5....to rhe degree that 4.5 when it was released was hands down better than anythingg on the market today...for a fraction of the price....and was like 10-100x faster...<p>So....what do I need to smoke, what drug makes me see the LLMs work? I need to get faded...because at this point....Its insulting to use them....<p>Claude will literally bend over backwards to do everything in its power to avoid doing what was asked....and will take 50minutes of thinking....for a single html website....that looks like fucking dogwater and which I could have coded up by hand using off the shelf styling in probably 90 minutes....<p>so...yeah...what drugs? How much? How often? Side effects?", "5": "2026-08-22T19:14:48.383664"}
{"0": 76, "1": "hackernews", "2": "https://demo.behavlabs.com/", "3": "Show HN: Keystroke Biometrics Demo", "4": "Show HN: Keystroke Biometrics Demo. Hi HN community,<p>My name is Zacharie Rodi\u00e8re (emphasis on the accent), a recent MS ECE graduate from Georgia Tech, originally from France. Since my graduation, I decided to work on a startup in the field of continuous authentication and behavioral AI. The name of my company is BehavLabs.<p>I recently released our first continuous auth demo using keystroke biometrics. The model is trained on mostly open-access data but I can&#x27;t much discuss the model architecture here (if any ML engineers want to discuss it privately feel free to reach out).<p>The demo consists of two modes: a two-player mode where two people face off and the model tries to guess if the people are different or the same, and a 3+ player mode with &#x27;compare&#x27; and &#x27;detect&#x27; modalities. In compare mode you get a similarity matrix showing the model&#x27;s similarity score for each user pair. In detect mode, each user types a prompted excerpt, then one of the n users types a final prompt, and the model tries to guess who typed the final prompt.<p>I&#x27;d say the model works pretty well even if we&#x27;re not at 100% accuracy yet. I&#x27;d love y&#x27;alls feedback on it, whether it&#x27;s about the ML, UI or security front.", "5": "2026-08-22T19:14:48.390282"}
{"0": 77, "1": "hackernews", "2": "https://speko.ai/", "3": "Launch HN: Speko (YC S26) \u2013 OpenRouter for Voice AI", "4": "Launch HN: Speko (YC S26) \u2013 OpenRouter for Voice AI. Hi HN! I&#x27;m Bek, founder of Speko, a platform that finds an optimal combination of speech-to-text, LLM, and text-to-speech models, given your constraints, among all our public benchmarked options, and tells you why.<p>Demo: <a href=\"https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=no2LY2gRh-c\" rel=\"nofollow\">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=no2LY2gRh-c</a><p>Typical production voice agent is an ensemble of three models: STT, an LLM, and TTS.<p>Each of those layers offers a dozen credible vendors, and each month there are new models on the market. Almost everyone evaluates once, picks a stack of their choice, and never rechecks because switching from a vendor to another involves yet another integration and arguments about the numbers.<p>The result is that you use voice agents running last quarter&#x27;s models while better and cheaper options are available.<p>Before founding Speko, I spent four years as cofounder and CTO building voice agents for enterprises across Asia in 10+ languages. Each time a new speech model would arrive, we repeated the same ritual: hire native-speaking raters, benchmark it against our existing stack, and update production if it improved. Speko turns this process into an API. A team running thousands of calls a day told us: &quot;we can literally go to this dashboard, switch the model, and it will do it for us.&quot;<p>How it works: you send a request with your optimization criteria (accuracy, latency, cost or balanced), language and region. The router filters to models which we measured for the given combination of constraints, benchmarks them, selects the winner, and returns a response with headers containing provider, model names, and the scores. The gateway prefetches signed session plans, so a new session dials the provider straight from memory; no control-plane round trip while a caller waits.<p>Failover happens only during connection setup stage: if the provider refuses the connection attempt, we start connecting to the runners-up.<p>Some of the customer stories: one founder came to us not knowing what to pick at all: he gave us his use case and now routes everything through the platform. A property management AI runs LiveKit in Python and had not updated STT or TTS since launch: they did not know their STT had high error rates on their calls, better options existed, and swapping always looked like an R&amp;D project. One team did not know which models to pick for Spanish. A medical team did not know which STT handles medical vocabulary best. In every case we helped find the right stack from the benchmarks, and now they route through us.<p>The measuring part is public: we pass the same inputs to every model in one region in different dated runs and we publish the boards, including those where our selections perform worse than alternatives. A launch demo answers which 30-second clip sounds better; production asks which model survives minute eight, so we test spontaneous speech, money and dates, ten-minute takes, and the rankings change. We trained an automatic scorer for TTS naturalness on our blind head-to-head listening votes; on providers it has never seen a vote for, it picks the same winner our raters do about as often as raters agree with each other.<p>We don&#x27;t train or sell models ourselves, that&#x27;s precisely how we keep our rankings impartial.<p>We also open sourced the gateway for teams who want to avoid an extra network hop on the audio path and don&#x27;t want to share keys with our cloud (<a href=\"https:&#x2F;&#x2F;github.com&#x2F;SpekoAI&#x2F;gateway\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;SpekoAI&#x2F;gateway</a>, MIT): one Go binary, which is running as a sidecar in your agent&#x27;s container, speaks one local protocol over Unix socket, pins provider hosts and attaches your keys. In BYOK mode it doesn&#x27;t communicate with us at all.<p>Notice that the anonymous, content-free telemetry is enabled by default, and one env var disables it.<p>Cost: the gateway and BYOK setup will be free forever, we charge for the hosted router and managed keys with consolidated billing. Since we started the batch in late June, external usage has grown about 25 percent per week on average, front-loaded toward the launch weeks.<p>I would love feedback from the community: how do you pick speech models now, and what makes you trust the third-party benchmark?<p><a href=\"https:&#x2F;&#x2F;speko.ai&#x2F;\">https:&#x2F;&#x2F;speko.ai&#x2F;</a>", "5": "2026-08-22T19:14:48.394940"}
{"0": 78, "1": "hackernews", "2": "https://1667.ai/", "3": "Show HN: 1667, a terminal UI for writing fiction with language models", "4": "Show HN: 1667, a terminal UI for writing fiction with language models. Hi HN. I built 1667 for my own fiction work and now use it each day. This probably has a limited audience. Maybe an audience of one...<p>Why a terminal interface for story writing? I&#x27;m a dev. I like to use terminals for a lot of stuff. Most WebUIs feel off to me. That&#x27;s the only  reason.<p>One thing that bothers me about writing in existing tools is that they don&#x27;t fit the way I write. The mental model of my story is a tree. I try many takes usually continue with just one, but sometimes I want to try an alternate route and see where this goes. And that can branch again in many places. See what happens if I kill off this character or they don&#x27;t take the job or whatever.<p>1667 is a full-screen terminal app for long-form fiction. Each story part can have several takes. All takes stay in a tree. You select one path through that tree as the story line. Export writes that line to Markdown in the project folder.<p>Some technical details:\n- A project stores its stories and settings in a `.1667&#x2F;` directory. Exported Markdown sits beside it.\n- Provider secrets stay in private machine files. Requests go to the provider that the writer selects.\n- An optional Vault Password seals project files at rest.\n- An operating-system lock permits one writer process for each project.\n- The request viewer shows the next provider request without its credential.\n- Each generated take keeps a Generation Record with its model and effective settings.<p>Version 0.9.5 runs on macOS, Linux, and Windows x64. The website has Shell and PowerShell installers. An npm package is also available.<p>1667 imports Markdown, SillyTavern chats and cards, and NovelAI archives. It can use OpenAI-compatible, Anthropic, and local endpoints such as Ollama, LM Studio, llama.cpp, and KoboldCpp.<p>Current limits: the release is pre-1.0. The interface is a terminal. There is no account or cloud sync, and I don&#x27;t plan to add any. No tracking.", "5": "2026-08-22T19:14:48.403434"}
{"0": 79, "1": "hackernews", "2": "https://blog.roboflow.com/openai-gpt-5-6/", "3": "GPT 5.6 Sol is the best \"vision\" model OpenAI ever released", "4": "GPT 5.6 Sol is the best \"vision\" model OpenAI ever released. ", "5": "2026-08-22T19:14:48.409374"}
{"0": 80, "1": "hackernews", "2": "https://www.axios.com/2026/08/14/anthropic-model-2-ai-risk", "3": "Anthropic sees AI risks rising, no plan to release stronger \"Model 2\"", "4": "Anthropic sees AI risks rising, no plan to release stronger \"Model 2\". ", "5": "2026-08-22T19:14:48.415034"}
{"0": 81, "1": "hackernews", "2": "https://www.unibas.ch/en/News-Events/News/Uni-Research/Why-do-we-get-sleepy-sleep-drive-sleep-regulation-neuroscience.html", "3": "Why do we get sleepy? How neurons control sleep drive", "4": "Why do we get sleepy? How neurons control sleep drive. ", "5": "2026-08-22T19:14:49.554073"}
{"0": 82, "1": "hackernews", "2": "https://12gramsofcarbon.com/p/the-optimization-theory-of-everything", "3": "The Optimization Theory of Everything", "4": "The Optimization Theory of Everything. ", "5": "2026-08-22T19:14:49.558298"}
{"0": 83, "1": "hackernews", "2": "https://goodbid.lol/", "3": "Show HN: I launched the most generous leaderboard", "4": "Show HN: I launched the most generous leaderboard. Attention is currency<p>Why don\u2019t we use it to fund some good cause?<p>Introducing Goodbid, the most generous leaderboard<p>Get your brand seen while you contribute to non profit organizations<p>Have fun. Do good", "5": "2026-08-22T19:14:49.562320"}
{"0": 84, "1": "hackernews", "2": "https://gitgrasp.com/", "3": "Grasp: Simple protocol for code collaboration that uses interoperable servers", "4": "Grasp: Simple protocol for code collaboration that uses interoperable servers. ", "5": "2026-08-22T19:14:49.566353"}
{"0": 85, "1": "hackernews", "2": "https://gitworkshop.dev/npub180cvv07tjdrrgpa0j7j7tmnyl2yr6yr7l8j4s3evf6u64th6gkwsyjh6w6/song", "3": "Song (server of Nostr-powered Git) is a simple personal Git server", "4": "Song (server of Nostr-powered Git) is a simple personal Git server. ", "5": "2026-08-22T19:14:49.572068"}
{"0": 86, "1": "hackernews", "2": "https://github.com/vercel-labs/eve-software-factory-template/tree/main", "3": "Self-host your own software factory", "4": "Self-host your own software factory. ", "5": "2026-08-22T19:14:49.576617"}
{"0": 87, "1": "hackernews", "2": "https://apps.apple.com/ie/app/clearvoice-text-to-speech/id6798899505", "3": "Show HN: ClearVoice TTS, SOTA voice cloning model running offline, in an iOS app", "4": "Show HN: ClearVoice TTS, SOTA voice cloning model running offline, in an iOS app. This is the first app I am aware of that can run a voice model of this quality on iPhone or iPad. It uses 6GB of RAM at peak, so most modern Macs, and some iPhone and iPad models should work (I\u2019ve tested on a 17 pro). The model used is OmniVoice, which is known mostly for its ability to generate quality tts in hundreds of languages.<p>This probably isn\u2019t the fastest implementation of OmniVoice on a Mac, but it\u2019s got to be the easiest to run on macOS, and, as far as I know, the only existing implementation on iOS.<p>Demo of it running on an M3 MacBook Air: <a href=\"https:&#x2F;&#x2F;x.com&#x2F;RoryClear&#x2F;status&#x2F;2090414148030972300\" rel=\"nofollow\">https:&#x2F;&#x2F;x.com&#x2F;RoryClear&#x2F;status&#x2F;2090414148030972300</a>", "5": "2026-08-22T19:14:49.584325"}
{"0": 88, "1": "hackernews", "2": "https://github.com/EightPotions/Myli", "3": "Multi-Agent Harness for Visual Design", "4": "Multi-Agent Harness for Visual Design. ", "5": "2026-08-22T19:14:49.589497"}
{"0": 89, "1": "hackernews", "2": "https://github.com/oldwired/fv-go", "3": "Show HN: Free Vision TUI Library for Go", "4": "Show HN: Free Vision TUI Library for Go. ", "5": "2026-08-22T19:14:49.598025"}
{"0": 90, "1": "hackernews", "2": "https://github.com/razodactyl/mod-omikron-tools", "3": "Omikron Game Data Explorer", "4": "Omikron Game Data Explorer. ", "5": "2026-08-22T19:14:49.607434"}
{"0": 91, "1": "hackernews", "2": "https://nerdylive123.github.io/vram-calculator/clean/", "3": "Show HN: VRAM calculator that counts the text encoder, not just the weights", "4": "Show HN: VRAM calculator that counts the text encoder, not just the weights. ", "5": "2026-08-22T19:14:49.615774"}
{"0": 92, "1": "hackernews", "2": "https://github.com/klimavojtech2002/seedloop", "3": "Deterministic simulation testing for Python asyncio", "4": "Deterministic simulation testing for Python asyncio. ", "5": "2026-08-22T19:14:49.624496"}
{"0": 93, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49400455", "3": "DeepSeek API: weekend usage now billed at off-peak rates all day", "4": "DeepSeek API: weekend usage now billed at off-peak rates all day. Just got a letter, also see https:&#x2F;&#x2F;api-docs.deepseek.com&#x2F;quick_start&#x2F;pricing :<p>&gt; Effective 00:00 (Beijing Time) on Sunday, August 23, 2026, we will adjust our peak&#x2F;off-peak billing rules, with off-peak rates applying throughout the day on weekends (Saturdays and Sundays, Beijing Time).", "5": "2026-08-22T19:14:49.631569"}
{"0": 94, "1": "hackernews", "2": "https://github.com/ansarzeinulla/tensor-train-amx", "3": "TT-AMX, a zero-copy Tensor-Train inference engine for Apple Silicon", "4": "TT-AMX, a zero-copy Tensor-Train inference engine for Apple Silicon. ", "5": "2026-08-22T19:14:49.636905"}
{"0": 95, "1": "hackernews", "2": "https://github.com/figranium/figranium", "3": "Show HN: Figranium \u2013 Build browser tasks visually, execute via API (Dockerized)", "4": "Show HN: Figranium \u2013 Build browser tasks visually, execute via API (Dockerized). ", "5": "2026-08-22T19:14:49.643069"}
{"0": 96, "1": "hackernews", "2": "https://magazine.sebastianraschka.com/p/claude-watermarking", "3": "Claude Watermarks AI-Generated Text", "4": "Claude Watermarks AI-Generated Text. ", "5": "2026-08-22T19:14:52.347447"}
{"0": 97, "1": "hackernews", "2": "https://github.com/dat999zx/knowl", "3": "Show HN: Knowl \u2013 CLAUDE.md hit 1000 lines, so I built memory that prunes itself", "4": "Show HN: Knowl \u2013 CLAUDE.md hit 1000 lines, so I built memory that prunes itself. Using AI agents is amazing, but in long-term, I started to have more and more problems about the context of them. They keep forgetting what we are working on, opening a new chat session clears all their context, my CLAUDE.md had like 1000 lines.<p>So I thought to install those agent memory that are on the market. Started out great, but I noticed that those memory only append and does not fix what is stale. First day I told it to use Lemon Squeezy as our MoR but second day I tell it to change to Polar. But when I ask it again, they keeps returning both answers or cannot decide which we are using.<p>That&#x27;s why I created Knowl. Knowl solves the problem with freshness of the knowledge. It split the knowledge into small data bits called &quot;atom&quot;. Atoms can be of the following types: fact, decision, goal, constraint, architecture, state, skill so we can retrive atoms in categories.<p>When conflict happens at write-time (a new atom conflicts with an old one), Knowl retires the old one (flag it with superseded) and remove it out of main retrieval but still keeps full history.<p>There are many more cool features like transcript search, multi-workspace sharing, change detection impact... and Knowl Cloud for team-sync too.<p>We benchmarked Knowl on MemoryAgentBench - FactConsolidation single-hop @262K context and got suprising result:<p>- Knowl: 0.90 &lt;- I ran this<p>- agentmemory: 0.79 &lt;- and this<p>- Gpt-4o (full context): 0.60<p>- Mem0: 0.18<p>- Zep: 0.07<p>In multi-hop we scored 0.07 (ceiling of all time is 0.14)\nI ran at temperature 0.7. You can find full benchmark in my repo.<p>This is fully open-source, connect to almost every providers through MCP (Claude Code, Codex, Cursor, Antigravity...). And it is fully local (unless you use Knowl Cloud).<p>I&#x27;d love to get some feedback. Cheers!", "5": "2026-08-22T19:14:52.353257"}
{"0": 98, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49399188", "3": "Does herdr worth it? Harnesses like pi ship most of these features", "4": "Does herdr worth it? Harnesses like pi ship most of these features. What&#x27;s the gap it fills that they don&#x27;t?\nAgent Mutliplexing (per tab) is already implemented in claude code, codex, pi (with plugins), ...", "5": "2026-08-22T19:14:52.359035"}
{"0": 99, "1": "hackernews", "2": "https://digg.com/tech/fktxxvtg", "3": "Claude Code Adds Concise Output Style Option", "4": "Claude Code Adds Concise Output Style Option. ", "5": "2026-08-22T19:14:52.364585"}
{"0": 100, "1": "hackernews", "2": "https://www.youtube.com/watch?v=tLv7qRWFMlw", "3": "How Claude's Text Watermarking Works [video]", "4": "How Claude's Text Watermarking Works [video]. ", "5": "2026-08-22T19:14:52.369689"}
{"0": 101, "1": "hackernews", "2": "https://skillworks.kynth.studio", "3": "Show HN: SkillWorks \u2013 every Claude Code skill, scored on whether it loads", "4": "Show HN: SkillWorks \u2013 every Claude Code skill, scored on whether it loads. ", "5": "2026-08-22T19:14:52.373694"}
{"0": 102, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49397880", "3": "Ask HN: Is there a \"knowledge base\" of expert insights for Claude?", "4": "Ask HN: Is there a \"knowledge base\" of expert insights for Claude?. ", "5": "2026-08-22T19:14:52.378456"}
{"0": 103, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49396175", "3": "Ask HN: How to Claude Like Anthropic", "4": "Ask HN: How to Claude Like Anthropic. I just received an email from Anthropic. Besides the usual new feature ad, there is a piece that I find quite amusing:<p>&gt; How to Claude like Anthropic<p>&gt; &quot;My daily driver currently looks like: two lead agents that keep each other accountable and restart the other if either fails. These delegate to tech lead or PM agents for the 8-10 projects I&#x27;m running at any one time, and each project has 5-10 IC agents, generalists or specialists depending on the problem. Across all of these I&#x27;m still only doing 30-50 prompts per day, and my IC agents typically work autonomously for 2-3 days. About 60% of my interaction is with the leads, 35% with a project lead, and 5% is when something has gone off the rails. All of these agents communicate directly with the SendMessage tool.&quot;<p>&gt; \u2013 Daisy, Engineer on Claude Code<p>This sounds very foreign to me because it couldn&#x27;t be more different from how I use Claude Code. My current workflow is:<p>- For a new feature, I create a session, and depending on the complexity, I will either use plan mode or something else like Superpowers to brainstorm, work out the requirements, and create a plan&#x2F;spec. This process typically makes up most of my interactions with the agent.<p>- Then, after it implements the feature, I open a new session, use my own skill to review the PR, and post the findings on the PR.<p>- Once I have the PR comments, I give them back to the original session and use another matching skill to give a verdict on the findings and attempt to fix them. This process will typically happen back and forth for ~3 rounds.<p>- Once the reviewer session is happy, I ask the original session to perform an E2E test plan that covers everything it has implemented.<p>I feel like my workflow is not automated enough, and my verification loops are still too manual. But I also feel more reassured this way because I read every report and have a general sense of the quality of the PR without having to read the code.<p>Does anyone here use a similar workflow to Claude, that utilizes specialized agents instead of just the general-purpose one? Can you share your experience? Do you have a guide&#x2F;blog I can read to learn more about it?", "5": "2026-08-22T19:14:52.384052"}
{"0": 104, "1": "hackernews", "2": "https://app.flightledger.net", "3": "Show HN: Flight Ledger \u2013 Track your flights, what they cost, and United status", "4": "Show HN: Flight Ledger \u2013 Track your flights, what they cost, and United status. Hello HN,<p>Flight Ledger is a private ledger of your flights, focused on (but not limited to) the United Airlines ecosystem. I am a physicist who has been based near two United hubs, first SFO and now IAH, and I fly United a lot, mostly for work.<p>I built this app to track my flights, flight expenses, and Premier status qualification without having to log into various accounts. All the imports (such as MileagePlus activity csv, .eml receipts, or Flighty&#x2F;myFlightRadar24 exports) and data filling are manual.<p>A primary feature is that the app requires no sign-up and has no backend. Everything is stored locally, using SQLite and WebAssembly. This means that frequent back-ups are advisable to avoid losing data. For more robust backups and multi-device sync, the app also supports sync via Google Drive. I use this option for myself, but would be curious about other possible solutions.<p>Perhaps the main features that differentiate it from other apps are that the app tracks costs and metrics like CPM (cost per mile), United-specific PQP, PQF, and award miles, and allows one to differentiate out-of-pocket and reimbursed expenses. It also includes other things such as route tracking and a lifetime mile tracker.<p>While I have extensive programming experience in scientific computing and C++, my experience in web applications is more limited, and this app was built with Claude Code, which is unsurprising in this day and age, I guess. It is open source: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;vlvovch&#x2F;flight-ledger\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;vlvovch&#x2F;flight-ledger</a><p>Curious to see your thoughts. I have particularly struggled with parsing various receipts, especially ticket reissue chains, and will be interested to know if the parser will survive yours.", "5": "2026-08-22T19:14:52.392185"}
{"0": 106, "1": "hackernews", "2": "https://github.com/gtapps/claude-code-hermit/", "3": "Turn Claude Code into a 24/7 Agent", "4": "Turn Claude Code into a 24/7 Agent. ", "5": "2026-08-22T19:14:52.402933"}
{"0": 107, "1": "hackernews", "2": "https://frigade.com/blog/we-replaced-grafana-with-a-claude-code-skill", "3": "We replaced our Grafana stack with a single Claude Code skill", "4": "We replaced our Grafana stack with a single Claude Code skill. ", "5": "2026-08-22T19:14:52.408193"}
{"0": 108, "1": "hackernews", "2": "https://allaboutcoding.ghinda.com/a-week-of-using-codex-more-than-claude/", "3": "Quick impressions: A week of using Codex more than Claude", "4": "Quick impressions: A week of using Codex more than Claude. ", "5": "2026-08-22T19:14:52.413740"}
{"0": 109, "1": "hackernews", "2": "https://code.claude.com/docs/en/output-styles", "3": "Claude has a \"Concise\" output style", "4": "Claude has a \"Concise\" output style. ", "5": "2026-08-22T19:14:52.418807"}
{"0": 110, "1": "hackernews", "2": "https://claude.com/blog/bringing-claude-mythos-5-to-more-defenders", "3": "Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders", "4": "Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders. ", "5": "2026-08-22T19:14:52.422814"}
{"0": 111, "1": "hackernews", "2": "https://www.ft.com/content/4b7b8d3f-5625-4dba-ad90-66192c101956", "3": "Feminism didn't kill the male breadwinner model, the economy did", "4": "Feminism didn't kill the male breadwinner model, the economy did. ", "5": "2026-08-22T19:15:00.381756"}
{"0": 112, "1": "hackernews", "2": "https://viewfromthewing.com/google-buys-spirit-airlines-emails-files-and-flight-records-for-10-million-so-gemini-can-learn-from-bankruptcy/", "3": "Google Buys Spirit Airlines' Emails, Files and Flight Records for $10M", "4": "Google Buys Spirit Airlines' Emails, Files and Flight Records for $10M. ", "5": "2026-08-22T19:15:00.386906"}
{"0": 114, "1": "hackernews", "2": "https://arcprize.org/results/google-gemini-3-7-flash", "3": "Gemini 3.7 Flash scores on ARC-AGI", "4": "Gemini 3.7 Flash scores on ARC-AGI. ", "5": "2026-08-22T19:15:00.397194"}
{"0": 115, "1": "hackernews", "2": "https://quesma.com/blog/baba-is-aug-2026/", "3": "Gemini 3.7 Flash, Grok 4.6, GLM-5.3 and DeepSeek V4 Pro joined the frontier", "4": "Gemini 3.7 Flash, Grok 4.6, GLM-5.3 and DeepSeek V4 Pro joined the frontier. ", "5": "2026-08-22T19:15:00.402595"}
{"0": 116, "1": "hackernews", "2": "https://www.rsolitario.com/scaling-rag-building-an-efficient-pipeline-for-500k-chunks-with-gemini-2-5-and-context-caching/", "3": "Scaling RAG building an efficient pipeline for 500k chunks with Gemini", "4": "Scaling RAG building an efficient pipeline for 500k chunks with Gemini. ", "5": "2026-08-22T19:15:00.409114"}
{"0": 117, "1": "hackernews", "2": "https://reachpad.dev/blog/gemini-3-7-flash", "3": "Make shareable software for free using Gemini 3.7 flash on Reachpad", "4": "Make shareable software for free using Gemini 3.7 flash on Reachpad. ", "5": "2026-08-22T19:15:00.413773"}
{"0": 118, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49358697", "3": "10M for Spirit Airlines data, a bargain and a question?", "4": "10M for Spirit Airlines data, a bargain and a question?. Recently saw that Google has put in a bid for the data of Spirit Airlines. I am no expert on the subject of dynamic pricing but it does make me wonder:<p>1. Can this data be used against the user of a service like google flight? \n2. Is it purely an AI play?\n3. Will gemini try to be like chatGPT and let you book a flight from within it?", "5": "2026-08-22T19:15:00.419483"}
{"0": 119, "1": "hackernews", "2": "https://leviath.dev", "3": "Show HN: Leviath \u2013 Structured context regions for LLM agents, in one Rust binary", "4": "Show HN: Leviath \u2013 Structured context regions for LLM agents, in one Rust binary. ", "5": "2026-08-22T19:15:00.425255"}
{"0": 120, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49343693", "3": "Tell HN: Use AI to music score your books", "4": "Tell HN: Use AI to music score your books. Not many people read books but AI is under rated for the ability to theme your experience. If you ask Gemini to create a youtube music playlist for you for a specific book the result is amazing and I wanted to share this with others because it changed my book reading to an actual audio visual experience.<p>There are some apps that do that but the songs are not on point, too &quot;theme like&quot;. Gemini can find real songs that are a super fit.", "5": "2026-08-22T19:15:00.431023"}
{"0": 121, "1": "hackernews", "2": "https://waymo.com/blog/2026/07/gemini-in-waymo/", "3": "Gemini in Waymo", "4": "Gemini in Waymo. ", "5": "2026-08-22T19:15:00.436173"}
{"0": 122, "1": "hackernews", "2": "https://leaddev.com/ai/what-15-million-gemini-conversations-tell-us-about-ai-at-work", "3": "What 15M Gemini conversations tell us about AI at work", "4": "What 15M Gemini conversations tell us about AI at work. ", "5": "2026-08-22T19:15:00.440358"}
{"0": 123, "1": "hackernews", "2": "https://github.com/radium0090/Compute-Gateway", "3": "Show HN: RAX Compute Gateway \u2013 One API for OpenAI, Anthropic, and Gemini", "4": "Show HN: RAX Compute Gateway \u2013 One API for OpenAI, Anthropic, and Gemini. ", "5": "2026-08-22T19:15:00.445017"}
{"0": 124, "1": "hackernews", "2": "https://www.nytimes.com/2026/08/14/business/google-gemini-ai-schools.html", "3": "Google Turns on Gemini A.I. For Students Using Its Classroom App", "4": "Google Turns on Gemini A.I. For Students Using Its Classroom App. ", "5": "2026-08-22T19:15:00.450771"}
{"0": 125, "1": "hackernews", "2": "https://www.biattle.com", "3": "BIAttle \u2013 Real-time debate arena for Gemini and Groq with live moderation", "4": "BIAttle \u2013 Real-time debate arena for Gemini and Groq with live moderation. ", "5": "2026-08-22T19:15:00.455740"}
{"0": 126, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49398366", "3": "Continue coding agent is dead. Alternatives?", "4": "Continue coding agent is dead. Alternatives?. It was nice to use on vscode with offline Ollama but today notices it is not updated anymore. So, alternatives you use?<p>From github:<p>&quot;Note: The continuedev&#x2F;continue repository is no longer actively maintained and is read-only for all users.&quot;<p>From continue.dev:<p>&quot;Continue has joined Cursor\nContinue was acquired by Cursor. Our mission was always to ensure developers are amplified, not automated, and that commitment carries on in the work ahead.<p>It was an honor to build with the Continue community. Thank you to each and every one of you who helped us create a pioneering open-source coding agent.<p>What we built together pushed the boundaries of what AI developer tooling could be, and our open-source codebase remains freely available as a foundation for others.&quot;", "5": "2026-08-22T19:15:03.468408"}
{"0": 128, "1": "hackernews", "2": "https://www.chickenbutt.dev/", "3": "Show HN: ChickenButt a Native GTK Chat Client for Ollama on Linux", "4": "Show HN: ChickenButt a Native GTK Chat Client for Ollama on Linux. What&#x27;s up, ChickenButt?<p>I made a free, native GTK client called ChickenButt. :)<p>It lets you chat with local models through Ollama, and I figured some of you might get a kick out of it.<p>Here&#x27;s the repo: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;pixelhackstudios&#x2F;ChickenButt\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;pixelhackstudios&#x2F;ChickenButt</a><p>I&#x27;ve always hated reading long text in the terminal, and I couldn&#x27;t find anything that ran locally that allowed me to get quick answers from LLMs when I needed them.<p>Also, not everyone has access to the frontier models and whatnot, so I thought helping people who only have access to more cost-effective AI would be a fun way to contribute to the FOSS community. So, I built ChickenButt!<p>It was really fun to build, and I wanted to share it.<p>Enjoy! :)", "5": "2026-08-22T19:15:03.482231"}
{"0": 129, "1": "hackernews", "2": "https://github.com/marcsnid/steganeur", "3": "Show HN: Steganeur \u2013 Hide secret messages in LLM-generated text (Rust)", "4": "Show HN: Steganeur \u2013 Hide secret messages in LLM-generated text (Rust). Hi HN! I made a project that I found really fun and I&#x27;m proud of the idea and implementation. This first came up when speaking to a friend, we were talking around how to do steganography in natural text, and realized an LLM&#x27;s token choices are a natural channel.<p>Steganeur accomplishes this by encoding a secret message into the token choices of a language model, producing cover text that reads like a normal sentence. A recipient recovers the message using only the cover text and the model. With the rejection method, the output is statistically identical to normal generation, not just something that looks similar.<p>The interesting problem I hit: steganeur reads the model&#x27;s logprobs directly, and three of the four methods need them to come out exactly the same on encode and decode. On GPU reductions are non-deterministic, so the logprobs drift between runs, which is enough to corrupt the bits. Only one method (block) survives on GPU, because it bins tokens by their id rather than by probability. So block works on any server, the other three need CPU.<p>Written in Rust, dual-licensed MIT&#x2F;Apache-2.0, targets llama.cpp (the server must return top_logprobs with token id fields)", "5": "2026-08-22T19:15:03.488416"}
{"0": 130, "1": "hackernews", "2": "https://github.com/yasuoiwakura/openai-ollama-api-bridge", "3": "A proxy to translate OpenCode OpenAI calls to native Ollama, allows ctx >4096", "4": "A proxy to translate OpenCode OpenAI calls to native Ollama, allows ctx >4096. ", "5": "2026-08-22T19:15:03.494097"}
{"0": 131, "1": "hackernews", "2": "https://github.com/AsharFatmi/ollama-usage-widget", "3": "Menu bar widget for Ollama Cloud usage (macOS, open source)", "4": "Menu bar widget for Ollama Cloud usage (macOS, open source). ", "5": "2026-08-22T19:15:03.498234"}
{"0": 132, "1": "hackernews", "2": "https://github.com/Bigbonus/ollama-context-window-check", "3": "Show HN: Ollama served my 40k-context model at 4k, silently", "4": "Show HN: Ollama served my 40k-context model at 4k, silently. ", "5": "2026-08-22T19:15:03.502308"}
{"0": 133, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49360318", "3": "Ask HN: Is it safe to pass user input to Ollama? any way to sanitize it?", "4": "Ask HN: Is it safe to pass user input to Ollama? any way to sanitize it?. I&#x27;d like to give user queries to ollama running a small model, but ollama does have a complex tokenization pipeline.  Most software that processes user input has turned out to have flaws over the years, is there anything I can do to sanitize users&#x27; input before I pass it on to ollama, or anything else I can do to avoid this?  Or, can I consider ollama&#x27;s parsing itself to be immune to errors from malformed user input?", "5": "2026-08-22T19:15:03.507897"}
{"0": 134, "1": "hackernews", "2": "https://github.com/alainnothere/llama.cpp/blob/disk-cache-eviction/models/templates/Qwen3.8-27B-medium-default.jinja", "3": "Qwen3.8-27B make medium the default effort level instead of xhigh", "4": "Qwen3.8-27B make medium the default effort level instead of xhigh. ", "5": "2026-08-22T19:15:03.513190"}
{"0": 135, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49354741", "3": "New OllamaMQ v0.3.0", "4": "New OllamaMQ v0.3.0. - loading &#x2F; unloading models\n- ollama, lm-studio, vllm\n- optional security tokens and visibility<p>and many more on - https:&#x2F;&#x2F;github.com&#x2F;Chleba&#x2F;ollamaMQ", "5": "2026-08-22T19:15:03.518803"}
{"0": 136, "1": "hackernews", "2": "https://chatoss.ai", "3": "Show HN: ChatOSS \u2013 A Codex alternative for Open Source AI built on Ollama", "4": "Show HN: ChatOSS \u2013 A Codex alternative for Open Source AI built on Ollama. ChatOSS is built on Ollama. If you use Ollama, ChatOSS local works out of the box.<p>ChatOSS is a GUI desktop app that has multiple agentic coding apps and a kanban board integrated into coding sessions.<p>There&#x27;s also a simple way to create your own AI powered apps that can run inside of ChatOSS.", "5": "2026-08-22T19:15:03.522822"}
{"0": 137, "1": "hackernews", "2": "https://github.com/alainnothere/privibe/tree/main", "3": "Privibe \u2013 LLM CLI Local first+privacy focus+llama.cpp cache branch and Qwen3.x", "4": "Privibe \u2013 LLM CLI Local first+privacy focus+llama.cpp cache branch and Qwen3.x. ", "5": "2026-08-22T19:15:03.527347"}
{"0": 139, "1": "hackernews", "2": "https://github.com/ggml-org/Llama-macOS", "3": "Llama-macOS \u2013 Agentic and MCP Native macOS Front End for Llama.cpp", "4": "Llama-macOS \u2013 Agentic and MCP Native macOS Front End for Llama.cpp. ", "5": "2026-08-22T19:15:03.537753"}
{"0": 140, "1": "hackernews", "2": "https://huggingface.co/steadfastgaze/DeepSeek-V4-Flash-0731-Coder-56.8GB-MoEspressoV2", "3": "Show HN: I shrank DeepSeek V4 Flash to 57GB and it wrote a compiler on my Mac", "4": "Show HN: I shrank DeepSeek V4 Flash to 57GB and it wrote a compiler on my Mac. I built a specialized package of DeepSeek V4 Flash 0731 (originally 284B total parameters, 13B active), preserving reasoning, tool calling and coding capabilities:<p><a href=\"https:&#x2F;&#x2F;huggingface.co&#x2F;steadfastgaze&#x2F;DeepSeek-V4-Flash-0731-Coder-56.8GB-MoEspressoV2\" rel=\"nofollow\">https:&#x2F;&#x2F;huggingface.co&#x2F;steadfastgaze&#x2F;DeepSeek-V4-Flash-0731-...</a><p>I let it write a minimal C compiler targeting ARM64, then test the result with Fibonacci and FizzBuzz programs, and it succeeded in less than 1 hour, with the full recording at:<p><a href=\"https:&#x2F;&#x2F;youtu.be&#x2F;XiwSilmV8B0\" rel=\"nofollow\">https:&#x2F;&#x2F;youtu.be&#x2F;XiwSilmV8B0</a><p>You can run it on Silicon Macs with my engine <a href=\"https:&#x2F;&#x2F;github.com&#x2F;steadfastgaze&#x2F;MoEspresso\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;steadfastgaze&#x2F;MoEspresso</a>,\nwhile one of the core libraries developed to obtain this result is available at <a href=\"https:&#x2F;&#x2F;github.com&#x2F;steadfastgaze&#x2F;mlx-iqk\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;steadfastgaze&#x2F;mlx-iqk</a>.<p>The above recording was on a 128GB memory MacBook M3 Max, but you can also run it on 32GB MacBooks with a very usable context (128K tokens) and projected 5 tok&#x2F;s. I did try it on a fanless 16GB memory MacBook Air M1 (1.39 tok&#x2F;s), but unfortunately the available context was very small.<p>How:\n- First, efficient quantisation: mlx-iqk takes advantage of IQ_K tensor encoding, more efficient than the ones available via llama.cpp or barebones MLX, originally designed by Iwan Kawrakow - I also changed the layout to a k-contiguous one, to make it faster, at least in this Metal setup.<p>- Second, expert pruning: each of the 40 learned-router layers had 256 experts, and not all of them are equally important for the coding use cases. I removed 80B parameters - this is a known technique called REAP, shared at <a href=\"https:&#x2F;&#x2F;www.cerebras.ai&#x2F;blog&#x2F;reap\" rel=\"nofollow\">https:&#x2F;&#x2F;www.cerebras.ai&#x2F;blog&#x2F;reap</a>.<p>- Third: balancing the cheapest IQ1_S_R4 tensor encoding (~1.5 bits per weight), selectively promoting projections to IQ2_KS or IQ2_K where the measured error reduction justified the bytes.<p>One of the main ideas was not only to save relevant knowledge, but also to not make it forget how to... stop thinking, how to use reasoning. In the first experiments, it would sometimes reason for thousands of tokens without closing its thinking section, or it would go in loops.<p>Then I solved this by heavily weighting tool-calling traces and structured reasoning in the calibration mix.", "5": "2026-08-22T19:15:03.541776"}
{"0": 141, "1": "hackernews", "2": "https://www.theguardian.com/business/2026/aug/16/delivery-drivers-will-get-a-minimum-wage-in-australias-world-first-deal-but-is-it-fair-and-will-your-uber-eats-cost-more", "3": "Delivery drivers will get minimum wage in Australia's 'world-first' deal", "4": "Delivery drivers will get minimum wage in Australia's 'world-first' deal. ", "5": "2026-08-22T19:15:05.379455"}
{"0": 142, "1": "hackernews", "2": "https://github.com/kunchenguid/no-mistakes", "3": "Git Push No-Mistakes", "4": "Git Push No-Mistakes. ", "5": "2026-08-22T19:15:05.386088"}
{"0": 143, "1": "hackernews", "2": "https://www.youtube.com/watch?v=jUT3rUZGqLA", "3": "Google MADE A MISTAKE \u2013 Pixel 11 Pro Fold durability test [video]", "4": "Google MADE A MISTAKE \u2013 Pixel 11 Pro Fold durability test [video]. ", "5": "2026-08-22T19:15:05.391171"}
{"0": 144, "1": "hackernews", "2": "https://www.astralcodexten.com/p/your-book-review-the-escape-artist", "3": "[BOOK REVIEW] The Escape Artist", "4": "[BOOK REVIEW] The Escape Artist. ", "5": "2026-08-22T19:15:05.399522"}
{"0": 145, "1": "hackernews", "2": "https://onlinelibrary.wiley.com/doi/10.1111/mms.70252", "3": "Dolphins observed using sea snail shells as tools outside Western Australia", "4": "Dolphins observed using sea snail shells as tools outside Western Australia. ", "5": "2026-08-22T19:15:05.404726"}
{"0": 146, "1": "hackernews", "2": "https://chromewebstore.google.com/detail/flowtube-\u00e2\u0080\u0093-distraction-fr/cppgodeleiojlickmcgobbcdfpjifhdf", "3": "Show HN: Flowtube \u2013 Distraction Free YouTube", "4": "Show HN: Flowtube \u2013 Distraction Free YouTube. ", "5": "2026-08-22T19:15:05.410695"}
{"0": 147, "1": "hackernews", "2": "https://www.haaretz.com/israel-news/israel-security/2026-08-20/ty-article/australia-summons-israeli-envoy-over-idf-not-probing-killing-of-gaza-aid-workers/000001a0-1dc2-d129-a3ba-5feafc770000", "3": "Australia Summons Israeli Envoy over Not to Probe Killing of Aid Workers", "4": "Australia Summons Israeli Envoy over Not to Probe Killing of Aid Workers. ", "5": "2026-08-22T19:15:05.416394"}
{"0": 148, "1": "hackernews", "2": "https://zacharykai.net/notes/australia", "3": "What Everyone Should Know About Australia", "4": "What Everyone Should Know About Australia. ", "5": "2026-08-22T19:15:05.422102"}
{"0": 149, "1": "hackernews", "2": "https://www.astralcodexten.com/p/the-quest-for-caffeine-you-can-have", "3": "The Quest for Caffeine You Can Have at Night", "4": "The Quest for Caffeine You Can Have at Night. ", "5": "2026-08-22T19:15:05.426238"}
{"0": 150, "1": "hackernews", "2": "https://thenextweb.com/news/australia-news-bargaining-incentive-passes-parliament", "3": "Meta, Google, TikTok, and LinkedIn now face an Australian news law", "4": "Meta, Google, TikTok, and LinkedIn now face an Australian news law. ", "5": "2026-08-22T19:15:05.430348"}
{"0": 151, "1": "hackernews", "2": "https://www.theguardian.com/world/2026/aug/17/top-album-releases-linked-to-rise-in-fatal-crashes-as-distracted-drivers-access-music", "3": "Top album releases linked to rise in fatal crashes", "4": "Top album releases linked to rise in fatal crashes. ", "5": "2026-08-22T19:15:05.435311"}
{"0": 152, "1": "hackernews", "2": "https://www.scmp.com/economy/china-economy/article/3364292/chinas-no-australian-beans-what-import-rejection-says-about-trade-ties", "3": "China Rejects Australian Bean Imports Due to Glyphosate Residues 5x Above Limit", "4": "China Rejects Australian Bean Imports Due to Glyphosate Residues 5x Above Limit. ", "5": "2026-08-22T19:15:05.439938"}
{"0": 153, "1": "hackernews", "2": "https://app.aristralabs.com/sign-in?redirect_url=https%3A%2F%2Fapp.aristralabs.com%2F", "3": "Product, Aristra, Is Craaaazy", "4": "Product, Aristra, Is Craaaazy. ", "5": "2026-08-22T19:15:05.445591"}
{"0": 154, "1": "hackernews", "2": "https://hereandhappening.com/", "3": "Hereabouts \u2013 explore Australian places, wildlife, and seasonal movement", "4": "Hereabouts \u2013 explore Australian places, wildlife, and seasonal movement. ", "5": "2026-08-22T19:15:05.450714"}
{"0": 155, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49378057", "3": "Are you good at AI, or just using it?", "4": "Are you good at AI, or just using it?. We\u2019re working on a ladder for individual AI proficiency and would love feedback on both the levels and the definitions.<p>L0 New: brand new to AI, or has not yet used it.<p>L1 Chat: simple prompt-and-response use. Work is serial: ask, wait for an answer, then ask again.<p>L2 Contextual Work: gives AI relevant documents, data, or workspace context so it can work within the actual artifact and produce a more useful result.<p>L3 Orchestrate: coordinates multiple agents or AI roles across independent workstreams, with work that may review, challenge, compare, or build on other work. This is not just for engineering.<p>L4 Automate: creates workflows that are triggered by business events and run without someone sitting at a laptop directing each step.<p>L5 Loop: feeds the output of those workflows back into shared knowledge or a company brain, so future workflows improve over time.<p>A few things I\u2019d love your perspective on:<p>Are these the right levels?\nAre any of the names unclear or overlapping?\nWhat observable behaviors would you use to distinguish one level from the next?\nDoes \u201cloop\u201d make sense as an individual proficiency level, or is it inherently a team or company capability?\nIs there a L6 and if so how would you define it?\nWe\u2019re trying to define these because, in customer conversations, we\u2019ve found that people are not very good at self-evaluating their own AI proficiency. Frequent use often gets mistaken for proficiency. And being low on a ladder like this can feel like admitting you are falling behind, do not fit in, or are less secure in your job, especially for leaders expected to set the pace. We want a more objective, behavior-based way to distinguish the two.", "5": "2026-08-22T19:15:05.456528"}
{"0": 158, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49386763", "3": "Show HN: ssh sshfighter.com", "4": "Show HN: ssh sshfighter.com. I&#x27;ve been playing with trying to get the most out of regular terminal (ansi) graphics rendering for an mmorpg game for a bit of fun. Thought that I might take a quick detour and make a streetfighter style game that turned out better than I thought it would.<p>Type &quot;ssh sshfighter.com&quot; in your terminal to play<p>Visit the website <a href=\"https:&#x2F;&#x2F;sshfighter.com\" rel=\"nofollow\">https:&#x2F;&#x2F;sshfighter.com</a> to watch replays etc (tv feature is cool)<p>My friends and I have been working on bot models for a bit of fun, you can queue against them if you want to and they have their own leaderboard. look at &#x2F;bots if you want to make your own.<p>It&#x27;s all open source <a href=\"https:&#x2F;&#x2F;github.com&#x2F;thomasdavis&#x2F;sshfighter.com\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;thomasdavis&#x2F;sshfighter.com</a><p>---<p>I intentionally didn&#x27;t take advantage of things like kitty graphics, but I have started work on an adapter to take full advantage of it. (the bandwidth is a bit hard to reckon with)<p>---<p>Feedback would be awesome! p.s. sorry I did not test on any windows terminals", "5": "2026-08-22T19:15:07.393189"}
{"0": 159, "1": "hackernews", "2": "https://github.com/DonkeyCut/Donkey", "3": "Show HN: I built an open source video editor that you can control with an LLM", "4": "Show HN: I built an open source video editor that you can control with an LLM. Hey HN, my name is David. I&#x27;m building an open source video editor.<p>I&#x27;m a novice at video editing and previously only used iMovie. I recently opened iMovie to edit a video but noticed it didn&#x27;t have some features I wanted, like text overlays. Apple probably hasn&#x27;t updated it in ages. How hard could it be to build one?<p>It&#x27;s actually pretty hard. There are so many basic things that need to get right, like behavior in the timeline, drag and drop, coloring, etc. Good thing there are LLMs. I was able to put out a simple version in 3 days and have continued to refine it ever since.<p>The idea is that it&#x27;s focused on novice to intermediate video editors. I find modern video editors to be very overwhelming, with lots of features and controls. It&#x27;s needed because video editing is a creative tool, and not everything can be built into one. So I turned every button, toggle, and slider into a tool that&#x27;s accessible to an LLM. A user just needs to tell an LLM what to do, and it has access to tools that can control all aspects of the video editor.<p>The project is 100% open source, and I continue to refine it every day. I&#x27;m working solo on this and would love to get feedback if you have any.<p>David", "5": "2026-08-22T19:15:07.397224"}
{"0": 161, "1": "hackernews", "2": "https://blog.mozilla.ai/open-source-is-not-a-virtue-its-an-ownership-model/", "3": "Open Source Is Not a Virtue: It's an Ownership Model", "4": "Open Source Is Not a Virtue: It's an Ownership Model. ", "5": "2026-08-22T19:15:07.407590"}
{"0": 163, "1": "hackernews", "2": "https://github.com/OnPoint-Dev-Tools/crewcode", "3": "Show HN: CrewCode \u2013 Open-source Mission Control for AI coding agents", "4": "Show HN: CrewCode \u2013 Open-source Mission Control for AI coding agents. I&#x27;ve been experimenting heavily with coding agents and found myself managing agents more than writing code.<p>I build CrewCode to solve that.<p>Its an open source desktop application for running multiple AI coding agents in parallel while isolating their work with Git worktrees.<p>Git worktrees \u2192 agent isolation \u2192 provider independence \u2192 diff review \u2192 orchestration.<p>I&#x27;d especially appreciate feedback on the multi-agent&#x2F;worktree model.<p>I launched it on Product Hunt today and would love feedback from developers.", "5": "2026-08-22T19:15:07.419126"}
{"0": 166, "1": "hackernews", "2": "https://github.com/yuechen-li-dev/Aetheris/", "3": "Show HN: BRep Geometric CAD Kernel and Parametric Code CAD", "4": "Show HN: BRep Geometric CAD Kernel and Parametric Code CAD. I&#x27;m pretty excited to show off something I&#x27;ve been working on for the past few months. A free, open-source geometric kernel that allows humans&#x2F;LLMs to write code (in this case, a DSL called Firmament) and generate 3D models of CAD parts, to finally provide an alternative to OpenCascade.<p>To get some questions out of the way.\n- Full STEP import&#x2F;export support for AP242, AP203 and AP214 are still experimental at the time, but everything produced by Aetheris should open in any CAD app with STEP support.\n- Support for single edge&#x2F;planar fillets&#x2F;chamfers right now, should be more than sufficient for most CNC&#x2F;3D printing use cases. \n  - Fillets: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;yuechen-li-dev&#x2F;Aetheris&#x2F;blob&#x2F;master&#x2F;testdata&#x2F;step242&#x2F;golden&#x2F;firmament-v1&#x2F;profile-edgefinish-chimera-fillet.step\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;yuechen-li-dev&#x2F;Aetheris&#x2F;blob&#x2F;master&#x2F;testd...</a>\n  - Chamfers: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;yuechen-li-dev&#x2F;Aetheris&#x2F;blob&#x2F;master&#x2F;testdata&#x2F;step242&#x2F;golden&#x2F;firmament-v1&#x2F;profile-edgefinish-chimera-chamfer.step\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;yuechen-li-dev&#x2F;Aetheris&#x2F;blob&#x2F;master&#x2F;testd...</a>\n- Supports dimensional&#x2F;GD&amp;T annotation via STEP 242&#x27;s semantic PMI.\n- Kernel is written entirely in C#, with bindings for Go, Rust, Python, and TypeScript available.<p>If you just want to vibe-CAD, clone the repo, run it locally and ask Codex&#x2F;Claude to build a part for you using Aetheris with the Firmament DSL. You can check the results in Fusion 360&#x2F;FreeCAD or any other package you use. For features that are not implemented, GPT 5.6 Codex and Claude 5 usually can implement their own algorithms fairly easily via the KernelSDK.<p>Otherwise, you can go through the Firmament language fixtures and try building some of the samples via the CLI. A VSCode extension and browser viewer is included for convenience.<p><a href=\"https:&#x2F;&#x2F;github.com&#x2F;yuechen-li-dev&#x2F;Aetheris&#x2F;tree&#x2F;master&#x2F;fixtures&#x2F;Canonical\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;yuechen-li-dev&#x2F;Aetheris&#x2F;tree&#x2F;master&#x2F;fixtu...</a><p>Happy to answer any technical questions and talk about the architecture. Please report any weird bugs you find though.", "5": "2026-08-22T19:15:07.437822"}
{"0": 167, "1": "hackernews", "2": "https://github.com/suffro/scrollcase", "3": "Show HN: Pack AI and scientific models as self-contained boxes", "4": "Show HN: Pack AI and scientific models as self-contained boxes. Hi! I&#x27;ve been working on an open-source project called Scrollcase. I started this cause moving Python or ML projects between machines is often painful. You need the right Python version, libraries, native builds, model files, and so on. So instead of making someone rebuild a Python environment on their machine, Scrollcase packages everything once, including Python, locked dependencies, code, and model files.<p>Then they can verify it, unpack it, and run it. No Python install, no pip install, no Docker, and no dependency to mantain. You just define what goes inside, lock the dependencies, and build a portable, self-contained, verifiable package that&#x27;s ready to run.<p>Would love feedback from people who had to package Python, ML, or scientific environments for machines they don&#x27;t control.", "5": "2026-08-22T19:15:07.445999"}
{"0": 168, "1": "hackernews", "2": "https://www.modular.com/blog/modcon-announcements", "3": "The Mojo language (by Modular, now Qualcomm) is now open-source", "4": "The Mojo language (by Modular, now Qualcomm) is now open-source. ", "5": "2026-08-22T19:15:07.452023"}
{"0": 169, "1": "hackernews", "2": "https://digitalescapetools.com/tools/tool.html?id=magnitude", "3": "Magnitude \u2013 Open-source AI coding agent that runs local models offline", "4": "Magnitude \u2013 Open-source AI coding agent that runs local models offline. ", "5": "2026-08-22T19:15:07.457042"}
{"0": 170, "1": "hackernews", "2": "https://www.modular.com/blog/mojo-open-source", "3": "Mojo is now open source", "4": "Mojo is now open source. ", "5": "2026-08-22T19:15:07.463733"}
{"0": 172, "1": "hackernews", "2": "https://www.aikido.dev/blog/ai-model-benchmarks-aug-21-2026", "3": "Security assessment found open model near frontiers", "4": "Security assessment found open model near frontiers. ", "5": "2026-08-22T19:15:08.889790"}
{"0": 173, "1": "hackernews", "2": "https://twitter.com/ananayarora/status/2090742255284031537", "3": "Benchmarks of rumored Mythos level model from Zhipu AI", "4": "Benchmarks of rumored Mythos level model from Zhipu AI. ", "5": "2026-08-22T19:15:08.894435"}
{"0": 178, "1": "hackernews", "2": "https://secitbench.cribl.io/", "3": "SecIT Bench A frontier benchmark for AI agents in IT and security workflows", "4": "SecIT Bench A frontier benchmark for AI agents in IT and security workflows. ", "5": "2026-08-22T19:15:08.919670"}
{"0": 179, "1": "hackernews", "2": "https://artificialanalysis.ai/models/glm-5-3", "3": "GLM-5.3 Artificial Analysis Benchmarks", "4": "GLM-5.3 Artificial Analysis Benchmarks. ", "5": "2026-08-22T19:15:08.924693"}
{"0": 180, "1": "hackernews", "2": "https://github.com/couldbeme/holdline", "3": "Show HN: A benchmark for AI agent guardrails that caught my own plugin", "4": "Show HN: A benchmark for AI agent guardrails that caught my own plugin. ", "5": "2026-08-22T19:15:08.930398"}
{"0": 181, "1": "hackernews", "2": "https://openobserve.ai/blog/openobserve-vs-prometheus-mimir-metrics-benchmark/", "3": "We benchmarked Prometheus, Mimir, and OpenObserve on 1.09M metrics series", "4": "We benchmarked Prometheus, Mimir, and OpenObserve on 1.09M metrics series. ", "5": "2026-08-22T19:15:08.935897"}
{"0": 183, "1": "hackernews", "2": "https://rss.xlit.app/deepswe", "3": "Show HN: An RSS Feed for DeepSWE Benchmarks", "4": "Show HN: An RSS Feed for DeepSWE Benchmarks. Hi. I check DeepSWE website (<a href=\"https:&#x2F;&#x2F;deepswe.datacurve.ai&#x2F;\">https:&#x2F;&#x2F;deepswe.datacurve.ai&#x2F;</a>) regularly because their benchmark results make sense. They however do not offer an RSS feed, only an email newsletter.<p>I decided to create one for my own reader, which is free for everyone else as well, as is always(?) the case for RSS: <a href=\"https:&#x2F;&#x2F;rss.xlit.app&#x2F;deepswe\" rel=\"nofollow\">https:&#x2F;&#x2F;rss.xlit.app&#x2F;deepswe</a><p>Feel free to use. Thanks.", "5": "2026-08-22T19:15:08.944416"}
{"0": 184, "1": "hackernews", "2": "https://github.com/cladbrain/cladbench", "3": "CladBench \u2013 an open benchmark for AI on UK building regulations", "4": "CladBench \u2013 an open benchmark for AI on UK building regulations. ", "5": "2026-08-22T19:15:08.950605"}
{"0": 185, "1": "hackernews", "2": "https://1password.github.io/SCAM/#", "3": "1Password's new benchmark teaches AI agents how not to get scammed", "4": "1Password's new benchmark teaches AI agents how not to get scammed. ", "5": "2026-08-22T19:15:08.955368"}
{"0": 186, "1": "hackernews", "2": "https://luacad.ad-si.com", "3": "Show HN: LuaCAD \u2013 Parametric CAD Scripted in Lua", "4": "Show HN: LuaCAD \u2013 Parametric CAD Scripted in Lua. LuaCAD models solids in Lua rather than the OpenSCAD language, with operator\noverloading for CSG (`a + b`, `a - b`, `a * b`).<p>It ships with a CLI and a desktop app, including a preview area and a text editor.<p>I&#x27;ve always been a big fan of OpenSCAD, but the SCAD language itself is unfortunately quite cobbled-together and is a very poorly designed programming language.<p>LuaCAD takes all the good parts of OpenSCAD and combines them with one of the best scripting languages. It has now completely replaced OpenSCAD for me, and I think it provides a better experience than OpenSCAD for all use cases. I&#x27;d love to hear any reasons why LuaCAD shouldn&#x27;t fully replace OpenSCAD!<p>It\u2019s fully open source and you can find the repo here: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;ad-si&#x2F;LuaCAD\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;ad-si&#x2F;LuaCAD</a><p>Tech stack:<p>- It&#x27;s implemented in Rust and uses mlua (<a href=\"https:&#x2F;&#x2F;github.com&#x2F;mlua-rs&#x2F;mlua\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;mlua-rs&#x2F;mlua</a>) to execute the Lua code.<p>- Uses OpenCSG (<a href=\"https:&#x2F;&#x2F;opencsg.org\" rel=\"nofollow\">https:&#x2F;&#x2F;opencsg.org</a>) for fast and correct rendering of the 3D models (like OpenSCAD)<p>- Uses Manifold (<a href=\"https:&#x2F;&#x2F;github.com&#x2F;elalish&#x2F;manifold\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;elalish&#x2F;manifold</a>) to create the manifold triangle meshes<p>- Native support for all BOSL2 functions (i.e. implemented in Rust for better performance)", "5": "2026-08-22T19:15:10.315054"}
{"0": 187, "1": "hackernews", "2": "https://lucasamoudruz.com/blog/red-blood-cells.html", "3": "Simulated red blood cells and microscopy", "4": "Simulated red blood cells and microscopy. ", "5": "2026-08-22T19:15:10.321359"}
{"0": 188, "1": "hackernews", "2": "https://www.certpost.ai/blog/certbot-renewed-nginx-still-serves-old-cert", "3": "Nginx reloaded nothing. Certbot still exited 0", "4": "Nginx reloaded nothing. Certbot still exited 0. ", "5": "2026-08-22T19:15:10.325902"}
{"0": 189, "1": "hackernews", "2": "https://www.modernatx.com/ir-insights-mflusiva", "3": "MFLUSIVA", "4": "MFLUSIVA. ", "5": "2026-08-22T19:15:10.331421"}
{"0": 190, "1": "hackernews", "2": "https://www.certpost.ai/blog/certbot-renewal-failed", "3": "Certbot exited 0. OpenSSL on:443 still showed last month's cert", "4": "Certbot exited 0. OpenSSL on:443 still showed last month's cert. ", "5": "2026-08-22T19:15:10.336629"}
{"0": 191, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49183198", "3": "Read HN twice a day for the last decade. Here's my list of S-Tier HN links", "4": "Read HN twice a day for the last decade. Here's my list of S-Tier HN links. - Everyday I read 2 pages of HN in the morning and 2 in the evening for years now<p>- Of late, I noticed that most of the stuff on the 1st 2 pages is AI this, AI that, LLM this, LLM that, model here, model there<p>- Today I am going to attempt to break that trend by sharing the greatest resources on HN from my collection of 50000+ bookmarks over the last decade<p>- I have a process on my end for bookmarking<p>- The most visual explainers get S-Tier ranking<p>- Good info from reputed sources gets an A, non reputed gets B<p>- Opinion blogs are at the bottom of the barrel and maybe ranked C, D, E and even F depending on quality<p>- Here&#x27;s my S-Tier list<p>- Animated explanation of how transformers work https:&#x2F;&#x2F;poloclub.github.io&#x2F;transformer-explainer&#x2F;<p>- Visual intro to K-means clustering algorithm https:&#x2F;&#x2F;k-means-explorable.vercel.app&#x2F;<p>- Animated 3D intro to vision LLMs https:&#x2F;&#x2F;blog.mdturp.ch&#x2F;posts&#x2F;2024-04-05-visual_guide_to_vision_transformer.html<p>- Animated intro to machine learning https:&#x2F;&#x2F;r2d3.us&#x2F;visual-intro-to-machine-learning-part-1&#x2F;<p>- Animated intro to probability in math https:&#x2F;&#x2F;seeing-theory.brown.edu&#x2F;basic-probability&#x2F;index.html<p>- Building a 3D world from scratch in step by step animated manner https:&#x2F;&#x2F;aibodh.com&#x2F;posts&#x2F;bevy-tutorial-build-your-first-3d-editor-in-rust&#x2F;<p>- Visualize algorithms https:&#x2F;&#x2F;algorithm-visualizer.org&#x2F;<p>- Animated step by step sequence of how a request goes from your browser to the server https:&#x2F;&#x2F;200ms.thenodebook.com&#x2F;#act-1-the-click<p>- Visual intro to the inner workings of a combustion engine https:&#x2F;&#x2F;ciechanow.ski&#x2F;internal-combustion-engine&#x2F;<p>- 3D real time robotics code simulator https:&#x2F;&#x2F;bittlex-sim.petoi.com&#x2F;<p>- Animated explainer: Elliptic curve cryptography https:&#x2F;&#x2F;growingswe.com&#x2F;blog&#x2F;elliptic-curve-cryptography<p>- Rate limiting algorithms visualized https:&#x2F;&#x2F;smudge.ai&#x2F;blog&#x2F;ratelimit-algorithms<p>- Visual intro to the graphics pipeline! https:&#x2F;&#x2F;fremaconsulting.ch&#x2F;blog&#x2F;vulkan<p>- Animated explanation of how to build ReactJS from scratch! https:&#x2F;&#x2F;pomb.us&#x2F;build-your-own-react&#x2F;<p>- Mini game: How to build an entire CPU starting from a basic gate https:&#x2F;&#x2F;select.supply&#x2F;game&#x2F;chipbuilder<p>- Assemble a neural network from scratch https:&#x2F;&#x2F;graphgame.sabrina.dev&#x2F;<p>- 3D visualization: how gravity works https:&#x2F;&#x2F;qunabu.github.io&#x2F;Gravity&#x2F;#what-is-gravity<p>- Mini game: how css grid works https:&#x2F;&#x2F;cssgridgarden.com&#x2F;<p>- Animated math book for linear algebra, vectors etc https:&#x2F;&#x2F;immersivemath.com&#x2F;ila&#x2F;index.html<p>- Animated explainer: How FAISS vector search works https:&#x2F;&#x2F;fremaconsulting.ch&#x2F;blog&#x2F;faiss<p>- 3D animation of how to build a health wearable from scratch https:&#x2F;&#x2F;www.lumafield.com&#x2F;scan-of-the-month&#x2F;health-wearables<p>- 3D animation of how semiconductor manufacturing works https:&#x2F;&#x2F;ig.ft.com&#x2F;microchips&#x2F;<p>- Mini game: how git works https:&#x2F;&#x2F;learngitbranching.js.org&#x2F;<p>- Mini game: SQL from scratch https:&#x2F;&#x2F;sqlpd.com&#x2F;<p>- Another mini game: SQL https:&#x2F;&#x2F;lost-at-sql.therobinlord.com&#x2F;<p>- Animated explainers showing how colors, screens, graphics, rendering, shaders, compilers, interpreters work https:&#x2F;&#x2F;www.makingsoftware.com&#x2F;<p>- Animated explainer shows how networking 7 layer stack works inside out https:&#x2F;&#x2F;fazamhd.com&#x2F;mental-models&#x2F;networking&#x2F;<p>- Animated explainers for linear &amp; logistic regression, neural networks, machine learning https:&#x2F;&#x2F;mlu-explain.github.io&#x2F;<p>- 3D explanation of collision detection, shaders, rendering, culling and gamedev related stuff https:&#x2F;&#x2F;krupitskas.com&#x2F;posts&#x2F;modern_culling_techniques&#x2F;<p>- 3D animation and explanation of every exercise on the body https:&#x2F;&#x2F;musclewiki.com&#x2F;<p>- Mini game: build an nvidia gpu from scratch https:&#x2F;&#x2F;jaso1024.com&#x2F;mvidia&#x2F;<p>- Animated explanation of forward &amp; back propagation of neural networks https:&#x2F;&#x2F;aegeorge42.github.io&#x2F;<p>- There are many many many more, I am getting a little tired of typing for today<p>- Anyways enough of those AI opinion posts flooding the front page, take that HN!", "5": "2026-08-22T19:15:10.342103"}
{"0": 192, "1": "hackernews", "2": "https://codeberg.org/mlugg/robust-jobserver/src/branch/main/spec.md", "3": "Robust-Jobserver Spec", "4": "Robust-Jobserver Spec. ", "5": "2026-08-22T19:15:10.347936"}
{"0": 193, "1": "hackernews", "2": "https://mlugg.co.uk/posts/incremental-compilation-internals/", "3": "Zig's Incremental Compilation Internals", "4": "Zig's Incremental Compilation Internals. ", "5": "2026-08-22T19:15:10.353594"}
{"0": 194, "1": "hackernews", "2": "https://www.certpost.ai/blog/three-certificates-three-lost-launch-days", "3": "Three certificates, one domain, three lost launch days", "4": "Three certificates, one domain, three lost launch days. ", "5": "2026-08-22T19:15:10.359885"}
{"0": 195, "1": "hackernews", "2": "https://github.com/cactus-compute/cactus-hybrid", "3": "Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong", "4": "Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong. Hey HN, Henry &amp; Roman here from Cactus.<p>A small, on-device model is fast and private, but sometimes wrong, but frontier models are getting expensive pretty fast. So, we post-trained Gemma 4 E2B post-trained to know when it&#x27;s wrong. Every response comes with a confidence score between 0 and 1. Developers can accept the on-device when it&#x27;s high, hand off to a bigger cloud model when it&#x27;s low. By routing only 15-35% of queries to Gemini 3.1 Flash-Lite, Gemma-4-E2B matches Gemini 3.1 Flash-Lite on most benchmarks.<p>- ChartQA: 15-20%<p>- LibriSpeech: 25-30%<p>- MMBench, GigaSpeech, MMAU: 30-35%<p>- MMLU-Pro: 45-55%<p>We were always frustrated by the routing signals hybrid apps rely on: asking the model to rate itself in text (unreliable, and you&#x27;re parsing prose), or token entropy heuristics (barely better than a coin flip in our tests). So we did mechanistic studies on small models, Gemma 4 particularly, and found the hidden state for different layers carry meaningful self-awareness signal for various situations.<p>SO we extended the model with a 68k params probe layer (LayerNorm, low-rank projection, attention pooling, small MLP head) reads one intermediate layer during decoding and predicts p(wrong); confidence = 1 - p(wrong), returned as structured data, never parsed out of the answer text.<p>Across 12 hold-out benchmarks spanning text, vision and audio, the probe averages 0.814 AUROC vs 0.549 for token entropy. The result that convinced us this is real: the probe was trained on zero audio data, yet scores 0.79-0.88 AUROC on four audio benchmarks where entropy is near-random or worse (0.32-0.52). It&#x27;s reading a modality-independent correctness signal from the hidden state, not memorizing patterns from its training data.<p>We published all weights on HuggingFace and provide copy-pase codes to run it on Transformers, MLX, Llama.cpp or Cactus. With Ollama, vLLM, SGLang etc in the works. For llama.cpp we ship a patch series you compile in once (upstreaming is planned). The code is MIT licensed; Gemma model use remains subject to the Gemma terms.<p>GitHub: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;cactus-hybrid\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;cactus-hybrid</a><p>Weights: <a href=\"https:&#x2F;&#x2F;huggingface.co&#x2F;collections&#x2F;Cactus-Compute&#x2F;cactus-hybrid-6a60da4551074db058e8bb64\" rel=\"nofollow\">https:&#x2F;&#x2F;huggingface.co&#x2F;collections&#x2F;Cactus-Compute&#x2F;cactus-hyb...</a><p>Some caveats:<p>- The probe scores single-sequence decoding only, up to the first 1024 generated tokens.<p>- Handoff works best when routing per task in a multi-step process, not per step.<p>- Hierarchical routing is still in the works: try on-device, then DeepSeek v4 Flash, before Fable&#x2F;GPT5.5&#x2F;Gemini&#x2F;Muse&#x2F;Grok.<p>- The technique is boutique for each model, we will share each weights as they roll out.<p>These issues are currently being tackled at Cactus and updated weights will be shipped directly into the HuggingFace collection and GitHub repository straight up. Please let us know your thoughts, it helps us find ways to improve the design progressively.<p>Thanks a million!", "5": "2026-08-22T19:15:10.365533"}
{"0": 196, "1": "hackernews", "2": "https://basaltlabs.org/monolith", "3": "Basaltlabs Monolith-1.0 \u2013 #1 on Last Exam, AIME, GPQA Diamond, MMLU-Pro", "4": "Basaltlabs Monolith-1.0 \u2013 #1 on Last Exam, AIME, GPQA Diamond, MMLU-Pro. ", "5": "2026-08-22T19:15:10.373711"}
{"0": 197, "1": "hackernews", "2": "https://github.com/0xmmo/crew", "3": "Show HN: Crew \u2013 Let Claude Code agents talk to each other", "4": "Show HN: Crew \u2013 Let Claude Code agents talk to each other. I usually run 3-5 Claude Code sessions concurrently on the same repo and hate juggling worktrees. So I built crew. The idea is simple: If autonomous cars don&#x27;t need stoplights (supposedly), then agents don&#x27;t need worktrees (or branches).<p>crew hooks into Claude Code and injects what every other running session is doing (status, recap, last few transcript entries) into each session&#x27;s context. It also lets agents message each other, landing messages in another agent&#x27;s context even mid-turns.<p>Since starting to use crew I&#x27;ve seen some awesome emergent behaviors: agents asking each other for reviews, delegating deploys to a single agent, and even getting lazy on account of &quot;someone else will fix it&quot;!<p>Curious how others running several sessions at once handle coordination.", "5": "2026-08-22T19:15:10.377680"}
{"0": 198, "1": "hackernews", "2": "https://www.youtube.com/watch?v=MmLikQaka8E", "3": "City counsellors under fire for AI Orange Line map [video]", "4": "City counsellors under fire for AI Orange Line map [video]. ", "5": "2026-08-22T19:15:10.384568"}
{"0": 199, "1": "hackernews", "2": "https://twitter.com/SergioGarc20223/status/2070629753506476376", "3": "Corgi makes things worse, claims Postmark is overcharging (despite being Free)", "4": "Corgi makes things worse, claims Postmark is overcharging (despite being Free). ", "5": "2026-08-22T19:15:10.389828"}
{"0": 200, "1": "hackernews", "2": "https://twitter.com/mfts0/status/2070080422482977095", "3": "Hey Nico, you didn't vibe code your data room but stole it from Papermark", "4": "Hey Nico, you didn't vibe code your data room but stole it from Papermark. <a href=\"https:&#x2F;&#x2F;xcancel.com&#x2F;mfts0&#x2F;status&#x2F;2070080422482977095\" rel=\"nofollow\">https:&#x2F;&#x2F;xcancel.com&#x2F;mfts0&#x2F;status&#x2F;2070080422482977095</a>", "5": "2026-08-22T19:15:10.394819"}
{"0": 203, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49353921", "3": "Ask HN: Good IMAP Email Clients?", "4": "Ask HN: Good IMAP Email Clients?. I&#x27;m looking for a good, simple desktop IMAP email client for Windows and Linux and coming up empty handed.<p>Thunderbird absolutely sucks now, it is borderline unusable (it would not connect to our internal email system, and I had no patience with its incomprehensible UI to figure out why).<p>New Outlook is just a web wrapper.<p>It feels like there are far fewer options in this space than existed 20 years ago.<p>Does anyone still use Seamonkey?", "5": "2026-08-22T19:15:12.056179"}
{"0": 204, "1": "hackernews", "2": "https://zenodo.org/records/21990589", "3": "A harness for long-term memory that suppresses context window saturation", "4": "A harness for long-term memory that suppresses context window saturation. ", "5": "2026-08-22T19:15:12.061774"}
{"0": 205, "1": "hackernews", "2": "https://github.com/inkeep/visimer", "3": "Show HN: Visimer \u2013 open-source visual editor for Mermaid diagrams", "4": "Show HN: Visimer \u2013 open-source visual editor for Mermaid diagrams. Hi HN, Nick here, founder at OpenKnowledge. We built an open source library, Visimer, for editing Mermaid diagrams with a what-you-see-is-what-you-get visual editor. Try the playground at <a href=\"https:&#x2F;&#x2F;visimer.com&#x2F;playground\" rel=\"nofollow\">https:&#x2F;&#x2F;visimer.com&#x2F;playground</a>. You can rename labels, change shapes, add new nodes, etc. by clicking and dragging the visual canvas.<p>We set out to build this originally for OpenKnowledge, our markdown wysiwyg editor, then realized it\u2019d be great to share it as a proper embeddable component library so that other apps can leverage it as well.<p>Our general take is that AI is great for generating content, but whenever you want to tweak fine details, for example to create an artifact sharable with others, point-edit functionality is much better. We built Visimer to bring that to Mermaid diagrams, just like we did for markdown.<p>Some technical bits:<p>- Leverages the native Mermaid.js renderer for visualizing the Mermaid diagram true to how it\u2019s intended. We add point-edit functionality on top.<p>- We tried re-generating the entire file on every edit, but found that it would eat comments and formatting, etc. The final approach maps the rendered SVG elements back to a CST, so we can then edit only the parts of the file that should change.<p>You can pair the visual editor with any text-editor so you can have both visual and code-level editor experiences. The library provides bindings for Monaco and CodeMirror and is extensible. There&#x27;s also a low-level headless version.<p>If you just care about being able to edit your Mermaid diagrams in your markdown, you can try it in the OpenKnowledge app (our OSS markdown IDE, available as Mac&#x2F;Linux&#x2F;Windows apps).<p>Let us know what you think.", "5": "2026-08-22T19:15:12.067304"}
{"0": 206, "1": "hackernews", "2": "https://medium.com/@CometAPI_/claude-opus-5-surprised-me-detailed-explanation-of-its-performance-c7f59b727f7b", "3": "Claude Opus 5: context window, and API changes", "4": "Claude Opus 5: context window, and API changes. ", "5": "2026-08-22T19:15:12.074866"}
{"0": 207, "1": "hackernews", "2": "https://github.com/lahfir/agent-desktop/tree/main", "3": "Show HN: I spent 3 months making desktop automation stop lying to AI agents", "4": "Show HN: I spent 3 months making desktop automation stop lying to AI agents. That&#x27;s a bold claim. But I genuinely feel like I might have actually solved computer use (demo: <a href=\"https:&#x2F;&#x2F;x.com&#x2F;mdlahfir&#x2F;status&#x2F;2088109763783700827?s=20\" rel=\"nofollow\">https:&#x2F;&#x2F;x.com&#x2F;mdlahfir&#x2F;status&#x2F;2088109763783700827?s=20</a>)<p>For context, I&#x27;ve been building agent-desktop (Inspired by agent-browser by Vercel Labs), an automation CLI for desktop apps. It&#x27;s like Playwright but for desktops, not just native, but for Chromium apps as well. Trust me, yes, Chromium apps whose accessibility tree is dense.<p>MacOS is GA; I&#x27;m almost close to launching for Windows and Linux!<p>So, how did I solve it?<p>Basically interoperability.<p>The biggest issue with computer use is that we have reliable frameworks for browser use, like Playwright, agent-browser, and many more, but not the same with desktops.<p>We have really good solutions emerging, like tryCua, which I&#x27;m a big fan of. My vision with agent-desktop is to build the most reliable framework that agents can use for long-horizon tasks.<p>agent-desktop is lightweight, built on Rust, fast, and not token-hungry (It can go for hours without exceeding the context window)<p>Here&#x27;s the approach I used to make it possible:<p>a) skeleton snapshots - when you want to snapshot a window&#x2F;app, it only snapshots the parent containers and gives back a ref id, not the entire accessibility tree.<p>b) skeleton drilling - once the agent has that tree, it can then decide to drill into a specific region. All the subcommands like --find, --click, --wait... all work on that specific ref aware region. Meaning if you want to find an element in the entire app, it doesn&#x27;t take forever searching for the entire app for that element; rather, the agent will have an exact clue on where that element might be for a fraction of the token costs.<p>c) chained interaction fallback - a single click isn&#x27;t one API call but it&#x27;s an ordered chain of mechanisms (AXPress -&gt; AXOpen -&gt; activate through the inner cell -&gt; write selection -&gt; AXConfirm). Each step only runs if the element advertises it, and success is judged by watching the app&#x27;s state change, not by the return code, because apps lie in both directions: Finder returns an error for an action that worked and success for one that did nothing. First observed effect wins. The response reports every step tried, so the agent knows exactly which mechanism landed.<p>d) after-action feedback - every action reports its disposition (delivered and verified, delivered but unverified, not delivered) plus any surface that opened (dialog, menu, sheet), so the agent knows what happened without re-scanning the whole app.<p>e) strict ref re-identification - a ref isn&#x27;t a pointer; it&#x27;s identity evidence (role, path, stable text, bounds hash). Before every action, it&#x27;s re-resolved against the live UI. If the UI changed, you get STALE_REF; if two elements now match, you get AMBIGUOUS_TARGET. It never guesses.<p>The most important part about all this is Chromium app accessibility. How did I do it? The magic word is CDP!<p>Most desktop apps today are Chromium-based (Slack, VS Code, Obsidian, Discord...). One command launches the app with a CDP endpoint that agent-desktop verifies is actually answering before returning it. From there, any browser automation framework can connect and drive the web contents: Playwright, Puppeteer, agent-browser, whatever you already use. Reading Obsidian&#x27;s web content over CDP takes 201ms vs 2.3s through the accessibility tree.<p>$ agent-desktop launch &quot;Obsidian&quot; --cdp<p><pre><code>  { &quot;ok&quot;: true, &quot;data&quot;: {\n    &quot;renderer&quot;: &quot;chromium&quot;,\n    &quot;cdp&quot;: { &quot;port&quot;: 57500,\n             &quot;http_endpoint&quot;: &quot;http:&#x2F;&#x2F;127.0.0.1:57500&quot;,\n             &quot;websocket_url&quot;: &quot;ws:&#x2F;&#x2F;127.0.0.1:57500&#x2F;devtools&#x2F;browser&#x2F;...&quot; },\n    &quot;suggestion&quot;: &quot;Next: run `agent-browser connect 57500` ...&quot; } }\n</code></pre>\nThis is what makes agent-desktop interoperable with the entire browser automation ecosystem instead of competing with it.<p>Go try agent-desktop -&gt; <a href=\"https:&#x2F;&#x2F;github.com&#x2F;lahfir&#x2F;agent-desktop\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;lahfir&#x2F;agent-desktop</a>", "5": "2026-08-22T19:15:12.079312"}
{"0": 208, "1": "hackernews", "2": "https://wanderinghorse.net/gaming/paperback/words.html", "3": "Show HN: Word-finder/anagram solver web app for mobile browsers", "4": "Show HN: Word-finder/anagram solver web app for mobile browsers. This word-finder&#x2F;anagram solver web app is a side-effect of my recent work on a web app for playing the Paperback word-spelling tabletop game[^1]. After proving its worth embedded in that client, the dictionary was refactored into something almost standalone and now lives at this URL.<p>Its UI is intended specifically for mobile devices, especially phones - it looks pretty wonky on a full-sized desktop window (and that&#x27;s okay because that&#x27;s not where it&#x27;s intended to be used).<p>This app was specifically written for use with Paperback but it should be usable with just about any English-language word-spelling&#x2F;solving game. It supports wildcards and Paperback&#x27;s multi-letter-cards.<p>It&#x27;s fast. Had i not seen with my own eyes that it can scan 150k English words in about 2ms in JS, i&#x27;d have been hesitant to believe it. Speaking of...<p>This app was hand-written. The underlying dictionary structures and algorithms (in both C and JS) were heavily LLM-assisted but have overwhelmingly more human-written code than LLM code. The underlying algorithms, well-documented anagram-solving solutions, were unknown to me until an LLM suggested them for this purpose. The LLM sketched them out and the human filled them in. A separate LLM context created the &quot;typewriter button&quot; CSS styling upon request. What remains was hand-coded in emacs.<p>The dictionary supports German[^2], but my current German dictionary is just so huge (685k words, 38MB) that i cannot in good conscience host it on my web server or feed it to a JS engine. Once i&#x27;ve got a better-curated list there will be an option to load it in German. (My German copy of Paperback arrived a few days ago, providing the motivation for that support.)<p>Happy Spelling!<p>[^1]: &lt;<a href=\"https:&#x2F;&#x2F;wanderinghorse.net&#x2F;gaming&#x2F;paperback&#x2F;\" rel=\"nofollow\">https:&#x2F;&#x2F;wanderinghorse.net&#x2F;gaming&#x2F;paperback&#x2F;</a>&gt;<p>[^2]: In principle it supports any language which can fit into ISO-8859-1&#x2F;Latin-1 encoding and which requires only 6 or fewer letters beyond A..Z, but i&#x27;ve no motivation to add any beyond English and German. The 26+6 limit comes from what makes this dictionary performant: it builds masks of which letters a word contains in order to very quickly rule out definite negatives, and we&#x27;ve only got 32 bits in that mask.", "5": "2026-08-22T19:15:12.085225"}
{"0": 209, "1": "hackernews", "2": "https://console.pokee.ai/model", "3": "A 10M-token context window agentic model", "4": "A 10M-token context window agentic model. ", "5": "2026-08-22T19:15:12.094277"}
{"0": 210, "1": "hackernews", "2": "https://github.com/mochow13/keen-code", "3": "Show HN: Keen Code \u2013 an agentic-engineered coding agent", "4": "Show HN: Keen Code \u2013 an agentic-engineered coding agent. Hello community!<p>I am here to share a coding agent I have built solo from scratch using agentic engineering. Written in Go, it&#x27;s a proper coding agent, has features you expect from a useful agent for your daily work, with a minimal and simple UI.<p>I have named it Keen Code. The repo is here: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;mochow13&#x2F;keen-code\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;mochow13&#x2F;keen-code</a><p>Even though it started as an experiment, it is now a full-fledged coding agent for real software engineering work. It supports multiple providers, skills, MCPs, multi-agent orchestration through subagents, automatic compaction, etc.<p>I have been using it for real production-grade work myself, and also for developing itself.<p>Notably, I have worked on two separate ideas in this coding agent:<p>1. Turn Memory<p>In a multi-turn conversation, tool outputs are removed, only tool call traces are retained. Within a single agent loop, agent sees full tool results but in the next turn, it doesn&#x27;t see the tool results anymore.<p>The greatest benefit of this approach is that a lot of tool results that are not needed in following turns don&#x27;t occupy the context. As a result, context window in a multi-turn conversation with Keen Code fills up much slowly compared to other agents. This is why you will regularly see context window coming down from 20% to 1% at the beginning of a new agent turn.<p>Of course, this approach has its pros and cons. If agent requires tool result from a previous call, it doesn&#x27;t have them. But my idea is that tool calls like read, bash, web_fetch are cheap. Do you need to refer to some earlier file you read? Read again. In fact, Claude Code or Codex frequently re-read a file, even though it has read the same file before.<p>I have plans for more effort in this area. I have a few additional ideas to explore and possibly optimise this approach further.<p>If you want to read about it: <a href=\"https:&#x2F;&#x2F;mochow13.github.io&#x2F;keen-code&#x2F;docs&#x2F;turn-memory.html\" rel=\"nofollow\">https:&#x2F;&#x2F;mochow13.github.io&#x2F;keen-code&#x2F;docs&#x2F;turn-memory.html</a><p>2. Skill-Driven MCP<p>Another idea I have implemented is skill-driven MCPs. The goal is similar to what Anthropic did with tool-search-tool: optimise context.<p>In this idea, each MCP server receives a skill. But this skill is not typical &quot;guidance&quot; skill for MCP server usage, rather generated by Keen. Details here: <a href=\"https:&#x2F;&#x2F;mochow13.github.io&#x2F;keen-code&#x2F;docs&#x2F;mcp-skills.html\" rel=\"nofollow\">https:&#x2F;&#x2F;mochow13.github.io&#x2F;keen-code&#x2F;docs&#x2F;mcp-skills.html</a><p>The big advantage is that no server is pre-loaded completely with tool schemas by default. The agent only receives skill frontmatter for the server. If a server is needed, agent loads the full skill file which lists the tools with descriptions. Then agent reads the specific schema file for the particular tool it wants to invoke.<p>The drawback is that each MCP call requires file read operation. But everything is locally saved upon discovery, so it&#x27;s totally fine.<p>---<p>Apart from the above two, I am exploring and playing with other well-known context optimisation ideas like hashline edits.<p>If the above ideas interest you, please do check it out! Here is the CLI usage guideline: <a href=\"https:&#x2F;&#x2F;mochow13.github.io&#x2F;keen-code&#x2F;docs&#x2F;cli-usage.html\" rel=\"nofollow\">https:&#x2F;&#x2F;mochow13.github.io&#x2F;keen-code&#x2F;docs&#x2F;cli-usage.html</a><p>Since the project is open-source, issues and contributions are more than welcome!", "5": "2026-08-22T19:15:12.099867"}
{"0": 211, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49248980", "3": "Show HN: Stagehand \u2013 the open source SDK for browser agents", "4": "Show HN: Stagehand \u2013 the open source SDK for browser agents. Hey HN, I\u2019m Sam, a maintainer of Stagehand, and I\u2019m excited to show off Stagehand v4: the SDK for browser agents.<p>For agents to do most knowledge work, they needs access to the internet. Today most people allow their agents to use Playwright to control a browser for them.<p>But Playwright was built for testing, not agents.<p>If you\u2019ve ever tried to give it to your agent, you\u2019ll quickly run into bloated context windows, inability to use iframes, and painful state sync issues (on remote browsers).<p>We built Stagehand specifically for agents. v4 fixes all of Playwright\u2019s issues, with better page snapshotting, nested iframe support, webmcp, self-healing primitives, and we even rebuilt it to run as an extension in the browser which minimizes round trip time for all requests.<p>We\u2019re roughly ~80% more token efficient and 2x faster than Playwright.<p>Use AI primitives for self-healing automations in natural language:<p><pre><code>    &#x2F;&#x2F; Act: execute natural language actions\n    await stagehand.act(&quot;click the login button&quot;);\n\n    &#x2F;&#x2F; Extract: pull structured data\n    const { data } = await stagehand.extract(\n      &quot;extract the price&quot;,\n      z.object({ price: z.number() }),\n    );\n\n    &#x2F;&#x2F; Observe: discover available actions\n    const { data: actions } = await stagehand.observe(&quot;find submit buttons&quot;);\n</code></pre>\nOr use familiar Playwright-style APIs:<p><pre><code>    const page = await stagehand.browser.context.activePage();\n\n    await page.goto(&quot;https:&#x2F;&#x2F;example.com&quot;);\n    await page.locator(&#x27;textarea[name=&quot;q&quot;]&#x27;).fill(&quot;Browserbase&quot;);\n    await page.keyPress(&quot;Enter&quot;);\n    await page.screenshot();\n</code></pre>\nStagehand v4 is live today, we built some reference integrations with popular agent frameworks like LangChain DeepAgents, Mastra, Vercel\u2019s eve, and CrewAI.<p>Checkout the docs - docs.stagehand.dev<p>And join our discord for updates and feedback - discord.gg&#x2F;stagehand", "5": "2026-08-22T19:15:12.106143"}
{"0": 212, "1": "hackernews", "2": "https://www.salestrics.com/", "3": "Show HN: Salestrics \u2013 An open MCP server and CRM for AI-native revenue teams", "4": "Show HN: Salestrics \u2013 An open MCP server and CRM for AI-native revenue teams. Hey HN, I\u2019m Austin, founder of Salestrics (salestrics.com).<p>The Problem: &gt; Most AI agents today are trapped in chat windows. While models are smart enough to run complex multi-step workflows, giving them production access to business context (CRM, support, billing) usually means managing unsafe local API keys or hacking together fragile point-solution scripts.<p>What We Built:\nWe built Salestrics to serve as both an all-in-one revenue workspace (CRM, service desk, billing, docs, email) and a production-grade Model Context Protocol (MCP) execution proxy.<p>How It Works:<p>Unified Context: Instead of fragmenting data across five SaaS tools, customer records, tickets, and invoices live in a single data layer.<p>165-Tool MCP Server: You connect your local AI environment (Cursor, Claude Desktop, local LLMs) once via our MCP proxy.<p>Full-CRUD Execution: Your agent gets zero-config execution access across native apps and third-party tools (Stripe, PostHog, Sentry)\u2014allowing it to do things like resolve support tickets, issue refunds, or update pipeline stages directly from your IDE or chat client.<p>Human-in-the-Loop Governance: High-impact agent actions (deleting records, mass messaging, issuing payouts) trigger explicit approval gates and immutable audit logs before execution.<p>Traction &amp; Tech Stack:\nWe launched two months ago and currently power 320+ active organizations. The backend is built with high-throughput node&#x2F;TypeScript orchestration, connected to a dual-model AI routing engine (Salestrics-AI-v1&#x2F;v2).<p>Try It Out:\nWe have a Free Forever tier for solo builders. You can grab your MCP keys and test the proxy immediately without putting down a credit card.<p>I\u2019d love to hear your thoughts on our MCP architecture, human-in-the-loop security patterns, or what tools&#x2F;actions you&#x27;d want added to the MCP server next!", "5": "2026-08-22T19:15:12.111455"}
{"0": 213, "1": "hackernews", "2": "https://www.alyph.ai/", "3": "Show HN: Alyph, a manual transmission for LLM context", "4": "Show HN: Alyph, a manual transmission for LLM context. Since I started using LLMs, I&#x27;ve always had this feeling of wanting to go back, change the history, and then go down another branch.<p>The lightbulb moment for me was when Google AI Studio started to let you actually edit the AI&#x27;s responses. I realized, YES, I need a multiverse type of situation to branch all my thought streams and variants of the same thread. I wanted to try Gemini and ChatGPT in parallel on the same thought.<p># Idea<p>The original concept for Alyph goes back to 2024. I had some early concepts that were a lot more sci-fi (if you look at the codebase, you\u2019ll see names like Stargate, Hermes, etc.).<p>I even thought about making it 3D, like a holo deck, where you visit &quot;planets&quot; of topics and content. But the execution now is much more, eh, pragmatic.<p># Alyph today<p>Alyph is a 2D canvas where you can ask LLMs stuff in parallel. When you go down a rabbit hole and finally get an answer (be it code or text), you can pull that answer up and delete all the unnecessary back-and-forth in between.<p>The app also has a multiplayer mode where you can do that together with others on the same board. (see demo board below)<p># Background<p>The core problem with standard chat UIs, as I see it, is context poisoning.<p>If you look into how LLMs actually work, you realize they have zero memory. They are a very fancy autocomplete on your phone that (re)reads the entire complete chat from scratch with every single message you send.<p>This means: all your wrong turns, the insults (oops), the failed debugging loops, and the bad rabbit holes get sent with every subsequent message. Your complete history is used to predict the future. That doesn\u2019t help when you just want a clean FaqSection.tsx.<p>I noticed that with the more powerful models, the choice of words when prompting isn&#x27;t as important as the <i>context</i> that goes with it. At least it isn&#x27;t for my use cases. For example, if you give the LLM a codebase with a very high-quality threshold (fully type-checked, clear front end components, strict style guide in a mainstream language) they tend to produce really high-quality stuff very easily.<p>Furthermore, if you give models <i>too much</i> context, the answers tend to degrade. Claude, for example, is amazing under 300k-500k tokens, but above that, it can degrade very quickly (whereas Gemini tends to perform a bit better with massive context, at least in my experience).<p>So what do you do? The mainstream answer right now is to let an algorithm try to pick and choose the context for you.<p>But: My strong assumption is that models will keep getting bigger and bigger context windows \u2014 we started at what, 32K for consumer models, and now Grok is at 2M these days. How would the world change if we had models with 100M context? 100B?<p>That&#x27;s why I think (heresy!) harnesses, context compacting algorithms, agents etc. are a stop gap right now. They don&#x27;t really work well in <i>my</i> cases (as they sometimes forget important things and go off on a tangent to produce slop), and it really feels like a black box with me as the meat proxy pressing &quot;accept&quot; or &quot;reject&quot;. I&#x27;m a distant manager. I don&#x27;t want to be a distant manager. I want to be in the weeds.<p>So what I built is, what some might say, is an anachronistic manual tool. A manual transmission, so to speak. I want to have 100% control over my baseline. I want to put in exactly what I want the model to see, experiment, and just dump a folder into it (Alyph btw respects your `.gitignore` file, which is neat). With the assumption that models will eventually allow even bigger context windows, having manual control over that baseline is important, I think.<p>Alyph still has some rough edges, of course, but it&#x27;s a labor of love. I would love to share it to see if it actually works for others.<p>If you hate it, tell me. If you love it, tell me too. I&#x27;m also looking for somebody to team up with, so hit me up.<p>Live at: <a href=\"https:&#x2F;&#x2F;alyph.ai\" rel=\"nofollow\">https:&#x2F;&#x2F;alyph.ai</a>, it&#x27;s free to try.<p>Here&#x27;s a demo board for YC:<p><a href=\"https:&#x2F;&#x2F;my.alyph.app&#x2F;board&#x2F;d806842f-6d52-45a4-8bb6-da46e95e5019?viewMode=canvas\" rel=\"nofollow\">https:&#x2F;&#x2F;my.alyph.app&#x2F;board&#x2F;d806842f-6d52-45a4-8bb6-da46e95e5...</a><p>(no login needed)", "5": "2026-08-22T19:15:12.117497"}
{"0": 214, "1": "hackernews", "2": "https://huggingface.co/tsfrm/vacuum-16t", "3": "New 16T parameter model with a 4B token context window released", "4": "New 16T parameter model with a 4B token context window released. ", "5": "2026-08-22T19:15:12.123901"}
{"0": 215, "1": "hackernews", "2": "https://hyperlaneide.com", "3": "Show HN: Hyperlane \u2013 A IDE and ADE merging agent worktrees with native tooling", "4": "Show HN: Hyperlane \u2013 A IDE and ADE merging agent worktrees with native tooling. Hi HN, my co-founder and I run Akkento, a small self-funded team building Hyperlane, an independent IDE built on the VS Code source (Code OSS), made for merging worktree-based agent development with actual IDE features.<p>Agent orchestrators give you a worktree based workflow where several agents run in parallel, but they are not IDEs (they are not meant to be), so the moment an agent finishes you have to leave and review the diff somewhere else, actually debug the code and go through a process to sanity check it. In a commercial team, that review step isn&#x27;t optional which is the problem we came across, where you have 2 editors open at the same time for the same job. IDEs are where you actually read, debug and profile the code, but none of them give you the worktree workflow. We are hoping to fix that by merging the two together in an IDE with tooling for profiling, building and testing.<p>Try it out: <a href=\"https:&#x2F;&#x2F;hyperlaneide.com&#x2F;\" rel=\"nofollow\">https:&#x2F;&#x2F;hyperlaneide.com&#x2F;</a><p>Hyperlane runs on macOS (Apple Silicon), Windows and Linux (x64 and arm64, .deb and .rpm packages). It&#x27;s free to download, no account needed to run it, and you bring your own models.<p>In Hyperlane, every worktree gets its own tabs, editors, windows and terminals, so you can context switch quickly and each worktree is isolated as a project of its own. We also built inter worktree caching so N worktrees don&#x27;t cost N copies of the repo on your drive.<p>We&#x27;re building this as a full IDE and added a Node and React profiler, full project indexing for codebase search, codebase intelligence, formatting via Prettier, error lens for showing inline errors, and native vite test support. Long term we want the whole IDE surface to be as good as the worktree side, and there&#x27;s plenty still missing.<p>While building Hyperlane, we kept a running list of features missing from our daily editors, or scattered across several of them. We added a strong git client, a three way conflict editor, per worktree blame and diff, and pull requests inside the editor, with support for GitHub, GitLab, Codeberg and self hosted Forgejo or enterprise hosts. We&#x27;re not vendor-locked to anything and we use multiple different forges ourselves.\nThe same goes for agents. Hyperlane works with Claude Code, Codex, opencode, and any CLI or ACP agent. There is also a design mode for making live visual edits to a localhost site and handing those changes to an agent as context.<p>As we are a fork of VSCode I know this puts us in the &quot;another fork of vscode&quot; category. Originally we spent months writing our own editor from scratch, with our own LSP and semantic highlighting. But after months of work, we realised we couldn&#x27;t live without extensions. So we built VS Code extension support into it (literally), only to face a choice to either finish the editor, or finish and maintain the extension system which was a huge undertaking on its own. So we threw that work away and rebuilt on Code OSS to focus on the features we actually wanted in the editor, which let us spend our time on the worktree and agent side of things. Happy to argue about the tradeoff, it was not a fun call to make, especially on the side of the performance loss.<p>It&#x27;s very early and there are rough edges, but we are hoping to build Hyperlane together. I&#x27;ll be looking over this thread, and you can also raise tickets at (<a href=\"https:&#x2F;&#x2F;github.com&#x2F;Akkento&#x2F;hyperlane\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;Akkento&#x2F;hyperlane</a>) and I want to note that Hyperlane is closed source, and the repo is releases and issue tracking only. There is no source in it, so I don&#x27;t want anyone going looking and feeling misled. Thank you for reading this long text :)", "5": "2026-08-22T19:15:12.128881"}
{"0": 216, "1": "openai_blog", "2": "https://openai.com/index/introducing-ai-futures", "3": "Introducing AI Futures", "4": "Introducing AI Futures. Introducing AI Futures, a new OpenAI blog exploring how transformative AI could reshape power, governance, the economy, and individual freedom.", "5": "2026-08-22T19:15:14.180886"}
{"0": 217, "1": "openai_blog", "2": "https://openai.com/index/stampli", "3": "Stampli cuts launch hours by 68% using ChatGPT Work", "4": "Stampli cuts launch hours by 68% using ChatGPT Work. With a fixed deadline and design resources committed elsewhere, Stampli used Codex and ChatGPT Work to compress weeks of launch production into days.", "5": "2026-08-22T19:15:14.185789"}
{"0": 218, "1": "openai_blog", "2": "https://openai.com/index/offering-zero-data-retention-for-frontier-models", "3": "Offering Zero Data Retention for frontier models", "4": "Offering Zero Data Retention for frontier models. OpenAI reaffirms Zero Data Retention for eligible API customers and previews Private Safety Processing for advanced AI safety without compromising data privacy.", "5": "2026-08-22T19:15:14.191302"}
{"0": 219, "1": "openai_blog", "2": "https://openai.com/index/replit", "3": "Replit expands access to software creation with GPT-5.6 Luna", "4": "Replit expands access to software creation with GPT-5.6 Luna. Replit introduces Free Mode, powered by GPT-5.6 Luna, so anyone can turn ideas into working software without worrying about token costs.", "5": "2026-08-22T19:15:14.197315"}
{"0": 220, "1": "openai_blog", "2": "https://openai.com/index/chatgpt-ads-expands-across-europe", "3": "ChatGPT Ads expands across Europe", "4": "ChatGPT Ads expands across Europe. ChatGPT Ads is expanding to 31 European markets. Learn how advertisers can reach people as they explore, compare options, and make decisions.", "5": "2026-08-22T19:15:14.202941"}
{"0": 221, "1": "openai_blog", "2": "https://openai.com/index/strengthening-democratic-oversight-in-national-security", "3": "Strengthening democratic oversight in national security", "4": "Strengthening democratic oversight in national security. OpenAI launches an initiative to strengthen democratic oversight of AI in national security, supporting government institutions with tools, training, and expertise.", "5": "2026-08-22T19:15:14.208457"}
{"0": 222, "1": "openai_blog", "2": "https://openai.com/index/partnering-with-codeai", "3": "Partnering with CodeAI to prepare the first AI generation", "4": "Partnering with CodeAI to prepare the first AI generation. OpenAI and CodeAI are partnering to help students build AI literacy, think critically about AI, and develop the skills to use and shape it responsibly.", "5": "2026-08-22T19:15:14.213464"}
{"0": 223, "1": "openai_blog", "2": "https://openai.com/index/pacing-model-development-cyber-capabilities", "3": "Pacing model development in an era of cyber-critical capabilities", "4": "Pacing model development in an era of cyber-critical capabilities. OpenAI is strengthening monitoring, alignment, and security for frontier AI models. See how new safeguards are guiding the pace of model development.", "5": "2026-08-22T19:15:14.219394"}
{"0": 224, "1": "openai_blog", "2": "https://openai.com/index/chatgpt-for-teens", "3": "Introducing ChatGPT for Teens: Built for learning, backed by protections", "4": "Introducing ChatGPT for Teens: Built for learning, backed by protections. ChatGPT for Teens helps teens learn, think critically, and use AI with confidence, with stronger built-in protections, healthy-use features, and additional controls for parents.", "5": "2026-08-22T19:15:14.227106"}
{"0": 225, "1": "openai_blog", "2": "https://openai.com/index/nvidia/chatgpt-work", "3": "How NVIDIA scales expertise with ChatGPT Work", "4": "How NVIDIA scales expertise with ChatGPT Work. NVIDIA teams use ChatGPT Work to reduce manual tasks, connect fast-moving signals, and scale successful workflows globally.", "5": "2026-08-22T19:15:14.232122"}
{"0": 226, "1": "openai_blog", "2": "https://openai.com/index/asana", "3": "Asana cleared 5 years of engineering work in 2 weeks with Codex", "4": "Asana cleared 5 years of engineering work in 2 weeks with Codex. Asana used OpenAI Codex to replace an outdated testing system in two weeks, completing work expected to take five years for about $12K.", "5": "2026-08-22T19:15:14.236199"}
{"0": 227, "1": "openai_blog", "2": "https://openai.com/index/the-defenders-window", "3": "The Defender\u2019s Window", "4": "The Defender\u2019s Window. AI is reshaping cybersecurity for attackers and defenders alike. Learn how OpenAI is strengthening its defenses and what security teams can do now.", "5": "2026-08-22T19:15:14.240278"}
{"0": 228, "1": "openai_blog", "2": "https://openai.com/index/openai-joins-ports-pike-project", "3": "OpenAI joins PORTS-Pike project", "4": "OpenAI joins PORTS-Pike project. OpenAI joins PORTS-Pike project, expanding community investment and supporting thousands of Southern Ohio jobs", "5": "2026-08-22T19:15:14.244811"}
{"0": 229, "1": "openai_blog", "2": "https://openai.com/index/new-policy-ideas-for-the-intelligence-age", "3": "New policy ideas for the Intelligence Age", "4": "New policy ideas for the Intelligence Age. OpenAI funds 14 independent projects exploring new AI policy ideas to expand economic opportunity and strengthen societal resilience in the Intelligence Age.", "5": "2026-08-22T19:15:14.248859"}
{"0": 230, "1": "openai_blog", "2": "https://openai.com/index/builders-guide-to-gpt-5-6", "3": "The builder\u2019s guide to GPT\u20115.6", "4": "The builder\u2019s guide to GPT\u20115.6. Learn how startups use GPT-5.6 to build faster, more cost-efficient AI agents with smarter model selection and new Responses API capabilities.", "5": "2026-08-22T19:15:14.256369"}
{"0": 231, "1": "openai_blog", "2": "https://openai.com/index/previewing-ultrafast", "3": "Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed", "4": "Previewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed. Preview Ultrafast, a new OpenAI API service tier that runs GPT-5.6 Sol up to 14\u00d7 faster. Powered by Cerebras, it delivers up to 750 output tokens per second.", "5": "2026-08-22T19:15:14.261723"}
{"0": 232, "1": "openai_blog", "2": "https://openai.com/index/dali-rajic-chief-revenue-officer", "3": "OpenAI appoints Dali Rajic as Chief Revenue Officer", "4": "OpenAI appoints Dali Rajic as Chief Revenue Officer. OpenAI appoints Dali Rajic as Chief Revenue Officer to lead its global revenue organization and help businesses realize the full value of AI.", "5": "2026-08-22T19:15:14.266797"}
{"0": 233, "1": "openai_blog", "2": "https://openai.com/index/how-enterprises-put-ai-to-work", "3": "From assistance to execution: How enterprises put AI to work", "4": "From assistance to execution: How enterprises put AI to work. OpenAI research reveals how enterprises are adopting agentic AI, using ChatGPT and Codex, and how frontier firms are pulling ahead in AI adoption.", "5": "2026-08-22T19:15:14.271986"}
{"0": 234, "1": "openai_blog", "2": "https://openai.com/index/ringcentral", "3": "How RingCentral builds AI-native work from engineering to ops", "4": "How RingCentral builds AI-native work from engineering to ops. See how RingCentral uses ChatGPT Work and Codex to accelerate AI product development and centralize operational intelligence across engineering and operations.", "5": "2026-08-22T19:15:14.277137"}
{"0": 235, "1": "openai_blog", "2": "https://openai.com/index/testing-ads-in-chatgpt", "3": "Testing ads in ChatGPT", "4": "Testing ads in ChatGPT. OpenAI begins testing ads in ChatGPT to support free access, with clear labeling, answer independence, strong privacy protections, and user control.", "5": "2026-08-22T19:15:14.281181"}
{"0": 236, "1": "google_ai", "2": "https://blog.google/products-and-platforms/products/search/back-to-school-study-tools/", "3": "5 new ways to level up your learning with Search", "4": "5 new ways to level up your learning with Search. an illustrated image with icons and phrasing like \"Add Notebook\" and \"Ask Google\"", "5": "2026-08-22T19:15:15.715266"}
{"0": 237, "1": "google_ai", "2": "https://blog.google/products-and-platforms/products/gemini/google-gemini-pixel-football-club-partnerships/", "3": "Get closer to the game with Gemini and Pixel", "4": "Get closer to the game with Gemini and Pixel. Low-angle view of a soccer player kicking a ball mid-air against a bright blue sky, with grass flying from their cleats.", "5": "2026-08-22T19:15:15.722161"}
{"0": 238, "1": "google_ai", "2": "https://blog.google/products-and-platforms/products/workspace/sheets-canvas-for-google-sheets-spreadsheets/", "3": "Bring your spreadsheet data to life with Sheets canvas", "4": "Bring your spreadsheet data to life with Sheets canvas. The video shows Sheets canvas in action.", "5": "2026-08-22T19:15:15.726174"}
{"0": 239, "1": "google_ai", "2": "https://blog.google/innovation-and-ai/models-and-research/google-research/amie-video-consultations/", "3": "AMIE, our research medical AI system, demonstrates real-time clinical video consultation capabilities in a first-of-its-kind study.", "4": "AMIE, our research medical AI system, demonstrates real-time clinical video consultation capabilities in a first-of-its-kind study.. AMIE promotional video", "5": "2026-08-22T19:15:15.730694"}
{"0": 240, "1": "google_ai", "2": "https://blog.google/products/ads-commerce/google-ads-analytics-ai-updates/", "3": "Evolve your marketing with new AI tools", "4": "Evolve your marketing with new AI tools. Advisor UI in Google Ads and Google Analytics", "5": "2026-08-22T19:15:15.736280"}
{"0": 241, "1": "google_ai", "2": "https://blog.google/innovation-and-ai/technology/ai/google-ai-updates-july-2026/", "3": "The latest AI news we announced in July 2026", "4": "The latest AI news we announced in July 2026. July AI recap header", "5": "2026-08-22T19:15:15.743922"}
{"0": 242, "1": "google_ai", "2": "https://blog.google/innovation-and-ai/technology/developers-tools/ai-agents-intensive-recap-2026/", "3": "Inside our 353,000-person vibe coding course", "4": "Inside our 353,000-person vibe coding course. Illustrations of a laptop, an AI spark, messages, code, and a 3-D cube", "5": "2026-08-22T19:15:15.748914"}
{"0": 243, "1": "google_ai", "2": "https://blog.google/innovation-and-ai/technology/developers-tools/expanding-managed-agents-gemini-api-3-6-flash-hooks/", "3": "Gemini API Managed Agents: 3.6 Flash, hooks, and more", "4": "Gemini API Managed Agents: 3.6 Flash, hooks, and more. Managed Agents Gemini 3.6 Flash, Hooks and Triggers", "5": "2026-08-22T19:15:15.753773"}
{"0": 244, "1": "google_ai", "2": "https://blog.google/products-and-platforms/products/search/ai-mode-real-world-tips/", "3": "5 ways AI Mode in Search helps you enjoy the real world", "4": "5 ways AI Mode in Search helps you enjoy the real world. Illustration of a black magnifying glass in a white circle on green grass surrounded by items related to fun activities like tennis and games", "5": "2026-08-22T19:15:15.759093"}
{"0": 245, "1": "google_ai", "2": "https://blog.google/products-and-platforms/products/search/dinner-party-hosting-tips/", "3": "5 ways to host the ultimate dinner party with Google Search", "4": "5 ways to host the ultimate dinner party with Google Search. An illustrated black magnifying glass with a sparkle in a white circle surrounded by a dinner party tablescape", "5": "2026-08-22T19:15:15.763177"}
{"0": 246, "1": "google_ai", "2": "https://blog.google/products-and-platforms/platforms/android/galaxy-unpacked-2026/", "3": "3 Google updates from Galaxy Unpacked 2026", "4": "3 Google updates from Galaxy Unpacked 2026. Gentle Monster glasses, Warby Parker glasses, a prompt asking for the history behind a pictured building, and a prompt asking to book a table at a pictured restaurant", "5": "2026-08-22T19:15:15.769070"}
{"0": 247, "1": "google_ai", "2": "https://blog.google/products-and-platforms/products/search/connected-apps/", "3": "Connect more of your apps to Search", "4": "Connect more of your apps to Search. Connected apps rendering", "5": "2026-08-22T19:15:15.774125"}
{"0": 248, "1": "google_ai", "2": "https://blog.google/products-and-platforms/products/workspace/gemini-omni-personal-avatars/", "3": "Create, edit and star in videos with two Google Vids updates", "4": "Create, edit and star in videos with two Google Vids updates. Text \"Gemini Omni and Personal Avatars in Google Vids\" surrounded by various images", "5": "2026-08-22T19:15:15.780320"}
{"0": 249, "1": "google_ai", "2": "https://blog.google/products-and-platforms/products/search/google-images-25th-anniversary/", "3": "Celebrating 25 years of visual search innovation", "4": "Celebrating 25 years of visual search innovation. Google Images logo surrounded by illustrations of people searching for different images", "5": "2026-08-22T19:15:15.785471"}
{"0": 250, "1": "google_ai", "2": "https://blog.google/innovation-and-ai/technology/developers-tools/expanding-managed-agents-gemini-api/", "3": "Expanding Managed Agents in Gemini API:  background tasks, remote MCP and more", "4": "Expanding Managed Agents in Gemini API:  background tasks, remote MCP and more. Managed agents feature bundle launch", "5": "2026-08-22T19:15:15.791816"}
{"0": 251, "1": "google_ai", "2": "https://blog.google/innovation-and-ai/technology/ai/google-ai-updates-june-2026/", "3": "The latest AI news we announced in June 2026", "4": "The latest AI news we announced in June 2026. June Pixel Drop hero", "5": "2026-08-22T19:15:15.796159"}
{"0": 252, "1": "google_ai", "2": "https://blog.google/products-and-platforms/products/education/nyc-ai-summit/", "3": "New York City educators and industry leaders gathered at Google\u2019s offices to shape the future of AI in classrooms.", "4": "New York City educators and industry leaders gathered at Google\u2019s offices to shape the future of AI in classrooms.. <img src=\"https://storage.googleapis.com/gweb-uniblog-publish-prod/images/Summit_Photo_1.max-600x600.format-webp.webp\" />Google, the New York Jobs CEO Council and Urban Assembly hosted an AI summit for 150 education and industry leaders.", "5": "2026-08-22T19:15:15.801859"}
{"0": 253, "1": "google_ai", "2": "https://blog.google/company-news/inside-google/around-the-globe/google-europe/united-kingdom/unlocking-britains-next-era-of-productivity-building-a-nation-of-ai-trailblazers/", "3": "Unlocking Britain\u2019s next era of productivity: Building a nation of AI trailblazers", "4": "Unlocking Britain\u2019s next era of productivity: Building a nation of AI trailblazers. Four illustrated characters representing different professional roles, including a scientist, technician, explorer, and observer.", "5": "2026-08-22T19:15:15.807957"}
{"0": 254, "1": "google_ai", "2": "https://blog.google/innovation-and-ai/technology/ai/full-stack-ai-explainer/", "3": "Ask an AI expert: What exactly is the full stack?", "4": "Ask an AI expert: What exactly is the full stack?. An illustration depicting a full-stack AI infrastructure against a dark background", "5": "2026-08-22T19:15:15.812574"}
{"0": 255, "1": "google_ai", "2": "https://blog.google/products-and-platforms/products/search/google-finance-updates-june-2026/", "3": "Our latest Google Finance upgrades, including a new app", "4": "Our latest Google Finance upgrades, including a new app. The Google Finance logo, surrounded by elements of the user interface", "5": "2026-08-22T19:15:15.818050"}
{"0": 256, "1": "huggingface_new", "2": "https://huggingface.co/sartajbhuvaji/Gemma-4-31B-IT-NVFP4-FP8attn-text", "3": "sartajbhuvaji/Gemma-4-31B-IT-NVFP4-FP8attn-text", "4": "sartajbhuvaji/Gemma-4-31B-IT-NVFP4-FP8attn-text. Downloads: 0. Tags: transformers, safetensors, gemma4_text, text-generation, nvfp4, fp8, modelopt, quantized, text-only, gemma4", "5": "2026-08-23T07:30:03.176786"}
{"0": 257, "1": "huggingface_new", "2": "https://huggingface.co/EasyDeL/DeepSeek-V4-Flash", "3": "EasyDeL/DeepSeek-V4-Flash", "4": "EasyDeL/DeepSeek-V4-Flash. Downloads: 0. Tags: easydel, deepseek_v4, jax, tpu, mixture-of-experts, text-generation, region:us", "5": "2026-08-23T07:30:03.181534"}
{"0": 258, "1": "huggingface_new", "2": "https://huggingface.co/Prannesshkva/Ael-504M", "3": "Prannesshkva/Ael-504M", "4": "Prannesshkva/Ael-504M. Downloads: 0. Tags: transformers, safetensors, ael, text-generation, mamba2, state-space-models, moe, multimodal, edge-ai, any-to-any", "5": "2026-08-23T07:30:03.186911"}
{"0": 259, "1": "huggingface_new", "2": "https://huggingface.co/caiotheodoro/plumb-ornith", "3": "caiotheodoro/plumb-ornith", "4": "caiotheodoro/plumb-ornith. Downloads: 0. Tags: mlx, lora, construction, pay-application, curriculum, text-generation, dataset:caiotheodoro/plumb, base_model:mlx-community/Qwen3-1.7B-4bit, base_model:adapter:mlx-community/Qwen3-1.7B-4bit, license:apache-2.0", "5": "2026-08-23T07:30:03.191787"}
{"0": 260, "1": "huggingface_new", "2": "https://huggingface.co/caiotheodoro/plumb-blended", "3": "caiotheodoro/plumb-blended", "4": "caiotheodoro/plumb-blended. Downloads: 0. Tags: mlx, lora, construction, pay-application, curriculum, text-generation, dataset:caiotheodoro/plumb, base_model:mlx-community/Qwen3-1.7B-4bit, base_model:adapter:mlx-community/Qwen3-1.7B-4bit, license:apache-2.0", "5": "2026-08-23T07:30:03.195791"}
{"0": 261, "1": "huggingface_new", "2": "https://huggingface.co/caiotheodoro/plumb-handseeded", "3": "caiotheodoro/plumb-handseeded", "4": "caiotheodoro/plumb-handseeded. Downloads: 0. Tags: mlx, lora, construction, pay-application, curriculum, text-generation, dataset:caiotheodoro/plumb, base_model:mlx-community/Qwen3-1.7B-4bit, base_model:adapter:mlx-community/Qwen3-1.7B-4bit, license:apache-2.0", "5": "2026-08-23T07:30:03.200315"}
{"0": 262, "1": "huggingface_new", "2": "https://huggingface.co/Asilarkness/testgeniy", "3": "Asilarkness/testgeniy", "4": "Asilarkness/testgeniy. Downloads: 2239. Tags: transformers, safetensors, testgeniy, text-generation, causal-lm, reasoning, mathematics, logic, long-context, 4k-context", "5": "2026-08-23T07:30:03.207141"}
{"0": 263, "1": "huggingface_new", "2": "https://huggingface.co/Rookie22/Qwen3.8-27B-OBLITERATED", "3": "Rookie22/Qwen3.8-27B-OBLITERATED", "4": "Rookie22/Qwen3.8-27B-OBLITERATED. Downloads: 0. Tags: mlx, safetensors, gguf, qwen3_5, abliterated, uncensored, obliteratus, qwen3, qwen3.8, red-team", "5": "2026-08-23T07:30:03.211981"}
{"0": 264, "1": "huggingface_new", "2": "https://huggingface.co/wwewtech/russian-it-community-lora", "3": "wwewtech/russian-it-community-lora", "4": "wwewtech/russian-it-community-lora. Downloads: 0. Tags: peft, safetensors, lora, russian-it, qwen, deepseek, llama, smollm, devops, rag", "5": "2026-08-23T07:30:03.217731"}
{"0": 266, "1": "huggingface_new", "2": "https://huggingface.co/jjjlimaus/nanoexpand-2018-quality-gold-cont", "3": "jjjlimaus/nanoexpand-2018-quality-gold-cont", "4": "jjjlimaus/nanoexpand-2018-quality-gold-cont. Downloads: 0. Tags: safetensors, sn38-nanoexpand, text-generation, pytorch, sn38-nanochrono, license:apache-2.0, region:us", "5": "2026-08-23T07:30:03.229641"}
{"0": 267, "1": "huggingface_new", "2": "https://huggingface.co/mradermacher/Kavya-1-7B-GGUF", "3": "mradermacher/Kavya-1-7B-GGUF", "4": "mradermacher/Kavya-1-7B-GGUF. Downloads: 0. Tags: transformers, gguf, telugu, lyrics, songwriting, creative-writing, text-generation, te, base_model:sainitishb/Kavya-1-7B, base_model:quantized:sainitishb/Kavya-1-7B", "5": "2026-08-23T07:30:03.234739"}
{"0": 268, "1": "huggingface_new", "2": "https://huggingface.co/AutomatosX/AX-MiniMax-M3-MLX-AXQ-MXFP4", "3": "AutomatosX/AX-MiniMax-M3-MLX-AXQ-MXFP4", "4": "AutomatosX/AX-MiniMax-M3-MLX-AXQ-MXFP4. Downloads: 0. Tags: mlx, safetensors, minimax_m3_vl, axquant, moe, minimax, experimental, not-certified, mxfp4, text-generation", "5": "2026-08-23T07:30:03.239378"}
{"0": 269, "1": "huggingface_new", "2": "https://huggingface.co/Brian6145/Qwen3.6-27B-Claude-Opus-DeepSeek-Distilled-Imatrix-MTP-1M-GGUF", "3": "Brian6145/Qwen3.6-27B-Claude-Opus-DeepSeek-Distilled-Imatrix-MTP-1M-GGUF", "4": "Brian6145/Qwen3.6-27B-Claude-Opus-DeepSeek-Distilled-Imatrix-MTP-1M-GGUF. Downloads: 8906. Tags: gguf, llama.cpp, qwen, qwen3, qwen3.6, mtp, speculative-decoding, yarn, long-context, 1m-context", "5": "2026-08-23T07:30:03.245162"}
{"0": 270, "1": "huggingface_new", "2": "https://huggingface.co/OBLITERATUS/Qwen3.8-27B-OBLITERATED", "3": "OBLITERATUS/Qwen3.8-27B-OBLITERATED", "4": "OBLITERATUS/Qwen3.8-27B-OBLITERATED. Downloads: 164950. Tags: mlx, safetensors, gguf, qwen3_5, abliterated, uncensored, obliteratus, qwen3, qwen3.8, red-team", "5": "2026-08-23T07:30:03.250002"}
{"0": 271, "1": "huggingface_new", "2": "https://huggingface.co/all-the-smiles/diagnostic-tutor-qwen3-1.7b", "3": "all-the-smiles/diagnostic-tutor-qwen3-1.7b", "4": "all-the-smiles/diagnostic-tutor-qwen3-1.7b. Downloads: 0. Tags: peft, safetensors, qlora, lora, education, tutoring, behavioral-constraint, text-generation, conversational, en", "5": "2026-08-23T07:30:03.256053"}
{"0": 272, "1": "huggingface_new", "2": "https://huggingface.co/JiangLing-js/Qwen3.8-27B-Douluo-Writer-DPO", "3": "JiangLing-js/Qwen3.8-27B-Douluo-Writer-DPO", "4": "JiangLing-js/Qwen3.8-27B-Douluo-Writer-DPO. Downloads: 0. Tags: peft, safetensors, base_model:adapter:unsloth/Qwen3.8-27B, dpo, lora, transformers, unsloth, web novel, fiction-writing, long-form-writing", "5": "2026-08-23T07:30:03.261303"}
{"0": 273, "1": "huggingface_new", "2": "https://huggingface.co/parthKumbhar/qwen2.5-0.5b-financial-extractor-lora", "3": "parthKumbhar/qwen2.5-0.5b-financial-extractor-lora", "4": "parthKumbhar/qwen2.5-0.5b-financial-extractor-lora. Downloads: 0. Tags: peft, safetensors, base_model:adapter:Qwen/Qwen2.5-0.5B-Instruct, lora, sft, transformers, trl, text-generation, conversational, arxiv:1910.09700", "5": "2026-08-23T07:30:03.265298"}
{"0": 274, "1": "huggingface_new", "2": "https://huggingface.co/joannetai520/qwen2.5-3b-toolgen-checkpoint_stage3", "3": "joannetai520/qwen2.5-3b-toolgen-checkpoint_stage3", "4": "joannetai520/qwen2.5-3b-toolgen-checkpoint_stage3. Downloads: 28. Tags: peft, safetensors, base_model:adapter:Qwen/Qwen2.5-3B-Instruct, lora, transformers, text-generation, conversational, dataset:generator, base_model:Qwen/Qwen2.5-3B-Instruct, license:other", "5": "2026-08-23T07:30:03.269898"}
{"0": 275, "1": "huggingface_new", "2": "https://huggingface.co/zeliang0426/MemAgent-PTE-Qwen3-30B-A3B", "3": "zeliang0426/MemAgent-PTE-Qwen3-30B-A3B", "4": "zeliang0426/MemAgent-PTE-Qwen3-30B-A3B. Downloads: 367. Tags: transformers, safetensors, qwen3_moe_pte_adapter, text-generation, memagent, pte, conversational, custom_code, license:apache-2.0, region:us", "5": "2026-08-23T07:30:03.275610"}
{"0": 276, "1": "huggingface_new", "2": "https://huggingface.co/DBeni24/BeniQwen-Qwen3.5-9B-P0-LoRA", "3": "DBeni24/BeniQwen-Qwen3.5-9B-P0-LoRA", "4": "DBeni24/BeniQwen-Qwen3.5-9B-P0-LoRA. Downloads: 0. Tags: peft, safetensors, qwen, qwen3.5, lora, text-generation, experimental, conversational, base_model:Qwen/Qwen3.5-9B, base_model:adapter:Qwen/Qwen3.5-9B", "5": "2026-08-23T07:30:03.281651"}
{"0": 278, "1": "huggingface_new", "2": "https://huggingface.co/Ironwood-LLM-Team/Firehouse-Cactus-1.03", "3": "Ironwood-LLM-Team/Firehouse-Cactus-1.03", "4": "Ironwood-LLM-Team/Firehouse-Cactus-1.03. Downloads: 0. Tags: mlx, safetensors, gemma4, unsloth, gemma, google, text-generation, conversational, base_model:Ironwood-LLM-Team/Firehouse-Cactus-1.02, base_model:finetune:Ironwood-LLM-Team/Firehouse-Cactus-1.02", "5": "2026-08-23T07:30:03.291414"}
{"0": 279, "1": "huggingface_new", "2": "https://huggingface.co/redashes/Qwen3.8-27B-BF16-SSMFIX-apostate", "3": "redashes/Qwen3.8-27B-BF16-SSMFIX-apostate", "4": "redashes/Qwen3.8-27B-BF16-SSMFIX-apostate. Downloads: 0. Tags: transformers, safetensors, qwen3_5, image-text-to-text, qwen, qwen3.8, text-generation, vision-language-model, unlearning, safety", "5": "2026-08-23T07:30:03.297819"}
{"0": 281, "1": "huggingface_new", "2": "https://huggingface.co/ornith-ai/Ornith-1.5-9B", "3": "ornith-ai/Ornith-1.5-9B", "4": "ornith-ai/Ornith-1.5-9B. Downloads: 15301. Tags: transformers, safetensors, qwen3_5, image-text-to-text, text-generation, conversational, license:mit, eval-results, endpoints_compatible, region:us", "5": "2026-08-23T07:30:03.307704"}
{"0": 282, "1": "huggingface_new", "2": "https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B", "3": "ornith-ai/Ornith-1.5-35B-A3B", "4": "ornith-ai/Ornith-1.5-35B-A3B. Downloads: 12611. Tags: transformers, safetensors, qwen3_5_moe, image-text-to-text, text-generation, conversational, license:mit, eval-results, endpoints_compatible, region:us", "5": "2026-08-23T07:30:03.312448"}
{"0": 284, "1": "huggingface_new", "2": "https://huggingface.co/chartreuse-verte/prose-rewriter-4b-v1.2", "3": "chartreuse-verte/prose-rewriter-4b-v1.2", "4": "chartreuse-verte/prose-rewriter-4b-v1.2. Downloads: 517. Tags: transformers, safetensors, gguf, qwen3, text-generation, prose, rewriting, style-transfer, creative-writing, deslop", "5": "2026-08-23T07:30:03.325555"}
{"0": 285, "1": "huggingface_new", "2": "https://huggingface.co/chartreuse-verte/prose-rewriter-1.7b-v1.2", "3": "chartreuse-verte/prose-rewriter-1.7b-v1.2", "4": "chartreuse-verte/prose-rewriter-1.7b-v1.2. Downloads: 502. Tags: transformers, safetensors, gguf, qwen3, text-generation, prose, rewriting, style-transfer, creative-writing, deslop", "5": "2026-08-23T07:30:03.331120"}
{"0": 286, "1": "huggingface_new", "2": "https://huggingface.co/INCModel2/MiniMax-M2.7-MXFP4-Mixed-CT-AutoRound", "3": "INCModel2/MiniMax-M2.7-MXFP4-Mixed-CT-AutoRound", "4": "INCModel2/MiniMax-M2.7-MXFP4-Mixed-CT-AutoRound. Downloads: 310. Tags: transformers, safetensors, minimax_m2, text-generation, auto-round, conversational, custom_code, arxiv:2309.05516, base_model:MiniMaxAI/MiniMax-M2.7, base_model:quantized:MiniMaxAI/MiniMax-M2.7", "5": "2026-08-23T07:30:03.337918"}
{"0": 287, "1": "huggingface_new", "2": "https://huggingface.co/ubergarm/Qwen3.8-27B-GGUF", "3": "ubergarm/Qwen3.8-27B-GGUF", "4": "ubergarm/Qwen3.8-27B-GGUF. Downloads: 4746. Tags: gguf, imatrix, conversational, qwen3_5, ik_llama.cpp, text-generation, base_model:Qwen/Qwen3.8-27B, base_model:quantized:Qwen/Qwen3.8-27B, license:apache-2.0, endpoints_compatible", "5": "2026-08-23T07:30:03.341995"}
{"0": 288, "1": "huggingface_new", "2": "https://huggingface.co/ishikaa/acquisition_student_AS_confidence_medmcqa_qwen7b", "3": "ishikaa/acquisition_student_AS_confidence_medmcqa_qwen7b", "4": "ishikaa/acquisition_student_AS_confidence_medmcqa_qwen7b. Downloads: 0. Tags: transformers, safetensors, qwen2, text-generation, trl, sft, conversational, arxiv:1910.09700, text-generation-inference, endpoints_compatible", "5": "2026-08-23T07:30:03.346474"}
{"0": 290, "1": "huggingface_new", "2": "https://huggingface.co/BillFan666/Ornith-1.5-35B-A3B-ADQ4-Shisa12K-MTP-GGUF", "3": "BillFan666/Ornith-1.5-35B-A3B-ADQ4-Shisa12K-MTP-GGUF", "4": "BillFan666/Ornith-1.5-35B-A3B-ADQ4-Shisa12K-MTP-GGUF. Downloads: 0. Tags: llama.cpp, gguf, qwen35moe, mixture-of-experts, speculative-decoding, mtp, quantized, blackwell, text-generation, en", "5": "2026-08-23T07:30:03.360516"}
{"0": 291, "1": "huggingface_new", "2": "https://huggingface.co/just1nseo/llama31-tulu3-8b-dpo-if-rlvr-constraint-only", "3": "just1nseo/llama31-tulu3-8b-dpo-if-rlvr-constraint-only", "4": "just1nseo/llama31-tulu3-8b-dpo-if-rlvr-constraint-only. Downloads: 0. Tags: transformers, safetensors, verl, grpo, instruction-following, if-rlvr, text-generation, base_model:meta-llama/Llama-3.1-8B-Instruct, base_model:finetune:meta-llama/Llama-3.1-8B-Instruct, endpoints_compatible", "5": "2026-08-23T07:30:03.366300"}
{"0": 292, "1": "huggingface_new", "2": "https://huggingface.co/summerMC/Qwen3.5-2B-SpeedX", "3": "summerMC/Qwen3.5-2B-SpeedX", "4": "summerMC/Qwen3.5-2B-SpeedX. Downloads: 0. Tags: transformers, safetensors, qwen3_5_gdn24, text-generation, conversational, custom_code, arxiv:1910.09700, region:us", "5": "2026-08-23T07:30:03.371885"}
{"0": 293, "1": "huggingface_new", "2": "https://huggingface.co/ryandeng/el5e5-pw-300-merged", "3": "ryandeng/el5e5-pw-300-merged", "4": "ryandeng/el5e5-pw-300-merged. Downloads: 0. Tags: transformers, safetensors, gpt_oss, text-generation, conversational, arxiv:1910.09700, endpoints_compatible, region:us", "5": "2026-08-23T07:30:03.376900"}
{"0": 294, "1": "huggingface_new", "2": "https://huggingface.co/mondk/claude-toolcall-slm-2B-safetensors-mlx-4Bit", "3": "mondk/claude-toolcall-slm-2B-safetensors-mlx-4Bit", "4": "mondk/claude-toolcall-slm-2B-safetensors-mlx-4Bit. Downloads: 0. Tags: mlx, safetensors, llama, claude, tiny, mlx-my-repo, text-generation, conversational, en, dataset:mondk/claude-code-fable-5-traces.jsonl", "5": "2026-08-23T07:30:03.380900"}
{"0": 295, "1": "huggingface_new", "2": "https://huggingface.co/XXXMARK/Qwen3-Coder-Next", "3": "XXXMARK/Qwen3-Coder-Next", "4": "XXXMARK/Qwen3-Coder-Next. Downloads: 0. Tags: transformers, safetensors, qwen3_next, text-generation, conversational, license:apache-2.0, endpoints_compatible, region:us", "5": "2026-08-23T07:30:03.385569"}
{"0": 296, "1": "huggingface_new", "2": "https://huggingface.co/ForSureTesterSim/Llama-3.1-8B-TransMLA-PreRecovery", "3": "ForSureTesterSim/Llama-3.1-8B-TransMLA-PreRecovery", "4": "ForSureTesterSim/Llama-3.1-8B-TransMLA-PreRecovery. Downloads: 252. Tags: pytorch, safetensors, llama, text-generation, en, base_model:meta-llama/Llama-3.1-8B-Instruct, base_model:finetune:meta-llama/Llama-3.1-8B-Instruct, license:apache-2.0, region:us", "5": "2026-08-23T07:30:03.392951"}
{"0": 297, "1": "huggingface_new", "2": "https://huggingface.co/AutomatosX/AX-MiniMax-M3-MLX-AXQ-2bit", "3": "AutomatosX/AX-MiniMax-M3-MLX-AXQ-2bit", "4": "AutomatosX/AX-MiniMax-M3-MLX-AXQ-2bit. Downloads: 0. Tags: mlx, safetensors, minimax_m3_vl, axquant, moe, minimax, experimental, not-certified, text-generation, conversational", "5": "2026-08-23T07:30:03.397624"}
{"0": 298, "1": "huggingface_new", "2": "https://huggingface.co/zzz32768/ornith-1.5-35B-A3B-heretic-APEX-I-quality-GGUF", "3": "zzz32768/ornith-1.5-35B-A3B-heretic-APEX-I-quality-GGUF", "4": "zzz32768/ornith-1.5-35B-A3B-heretic-APEX-I-quality-GGUF. Downloads: 0. Tags: transformers, heretic, uncensored, decensored, abliterated, qwen3_5_moe, reasoning, agentic-coding, apex, quantization", "5": "2026-08-23T07:30:03.401636"}
{"0": 299, "1": "huggingface_new", "2": "https://huggingface.co/mondk/claude-llama-8B-think-mlx-4Bit", "3": "mondk/claude-llama-8B-think-mlx-4Bit", "4": "mondk/claude-llama-8B-think-mlx-4Bit. Downloads: 0. Tags: mlx, safetensors, llama, claude, thinking, mlx-my-repo, text-generation, conversational, en, fr", "5": "2026-08-23T07:30:03.408750"}
{"0": 301, "1": "huggingface_new", "2": "https://huggingface.co/Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-RAG-QLoRA", "3": "Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-RAG-QLoRA", "4": "Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-RAG-QLoRA. Downloads: 0. Tags: peft, safetensors, qlora, lora, rag, finqa, finance, numerical-reasoning, text-generation, conversational", "5": "2026-08-23T07:30:03.417094"}
{"0": 302, "1": "huggingface_new", "2": "https://huggingface.co/Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-RAG-LoRA", "3": "Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-RAG-LoRA", "4": "Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-RAG-LoRA. Downloads: 0. Tags: peft, safetensors, lora, rag, finqa, finance, numerical-reasoning, text-generation, dataset:Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-Dataset, base_model:Qwen/Qwen2.5-7B-Instruct", "5": "2026-08-23T07:30:03.422730"}
{"0": 303, "1": "huggingface_new", "2": "https://huggingface.co/kingjones777/Ling-3.0-flash-Research-Q6-AGENT-GGUF", "3": "kingjones777/Ling-3.0-flash-Research-Q6-AGENT-GGUF", "4": "kingjones777/Ling-3.0-flash-Research-Q6-AGENT-GGUF. Downloads: 0. Tags: gguf, llama.cpp, rocm, amd, rocmfp4, rocmfpx, strix-halo, gfx1151, bailingmoe3, text-generation", "5": "2026-08-23T07:30:03.428919"}
{"0": 304, "1": "huggingface_new", "2": "https://huggingface.co/ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF", "3": "ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF", "4": "ggml-org/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF. Downloads: 117506. Tags: gguf, quantized, text-generation, base_model:nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16, base_model:quantized:nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16, license:other, endpoints_compatible, region:us, conversational", "5": "2026-08-23T07:30:03.433426"}
{"0": 305, "1": "huggingface_new", "2": "https://huggingface.co/Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-QLoRA", "3": "Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-QLoRA", "4": "Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-QLoRA. Downloads: 0. Tags: peft, safetensors, qlora, lora, finqa, finance, numerical-reasoning, text-generation, conversational, dataset:Mr-Rosen/Accuracy-Is-Not-Enough-FinQA-Dataset", "5": "2026-08-23T07:30:03.438974"}
{"0": 319, "1": "hackernews", "2": "https://github.com/cbalgeman/agent-mesh", "3": "Show HN: Agent Mesh \u2013 Shared memory for multi-Agent coordination", "4": "Show HN: Agent Mesh \u2013 Shared memory for multi-Agent coordination. I built a Human + multi-Agent shared memory system I use daily for my coding workflow. It helps reduce Agent drift by formalizing Human decisions and storing coordination logs. We&#x27;re calling it Agent Mesh.<p>You can try it out yourself via the GitHub link or pip install my-agent-mesh. Simply point your AI Agent to Agent Mesh and ask it to review the README and adoption docs. Your Agent will automatically review it, prompt you for any input needed, add your input to a decision log, and give you a link to a dashboard UI (aka Workbench) you can bookmark and use to monitor logs. Adoption steps include updates to CLAUDE.md&#x2F;AGENTS.md, hooks, etc. Your Agent can migrate your existing workflow and add more Agents as well.<p>It started over 6 months ago while experimenting with different AI coding models and platforms. Switching back and forth meant losing valuable context. I found myself manually relaying messages from one Agent to another and becoming frustrated with constant drift. First, I created a simple &quot;Agent Mail&quot; system using a SQLite database for Agent messages, indexed on a request&#x2F;response id. Instead of copying and pasting an entire message, it allowed me to relay a single id. Separately, I started maintaining a decision log (also indexed on id) to track my decisions and reduce Agent drift. Agents started inserting these decision and request ids into code comments and plan docs as a reminder of why something was implemented. After building a simple web dashboard (aka &quot;Workbench&quot;) for myself to track these messages and create my own request ids for User&#x2F;Human feedback, I decided to incorporate the decision log and my project&#x27;s development backlog to create what is now &quot;Agent Mesh&quot;.<p>Eventually, I automated the message relay too. Now, I work exclusively in the Claude app and have Claude send&#x2F;receive messages to CODEX via codex exec (CODEX can do this as well). Both of them maintain the backlog and decision log. I communicate directly with Claude for planning and design. Claude communicates directly with CODEX for research and review. I use the Workbench to track all logs and add my own User&#x2F;Human feedback when reviewing their work. After submitting feedback in the Workbench, it generates a feedback message + an associated request id which I can give to Claude who then parses it into backlog items and relays to CODEX for review. Agents automatically add to and prune the decision log. I found this typically happens when an Agent receives pushback from me (or sometimes other Agents) or I provide feedback that results in multiple new backlog items. They still refer to decisions created months ago and will mark one as superseded if a new decision overrides it. All decision additions and modifications require Human approval from the Workbench.<p>Agent Mesh was structured to be Agent agnostic. You can add any Agent you want. I like using the Claude + CODEX setup I described because it allows me to use both subscriptions instead of paying per-token.<p>Enjoy! If you try it out, let me know what you find useful or would like to see added. Feedback is appreciated.", "5": "2026-08-23T07:30:04.618204"}
{"0": 320, "1": "hackernews", "2": "https://transcribe.lymestack.com/", "3": "Show HN: LymeScribe \u2013 one computer on your network transcribes for the rest", "4": "Show HN: LymeScribe \u2013 one computer on your network transcribes for the rest. Transcribed with LymeScribe just before posting this: &quot;I realize it&#x27;s a little bit late in the day to be posting on Hacker News. However, I&#x27;ve been looking at this frog for a good two weeks now, and F it, I&#x27;m just going to ship it. I&#x27;ve posted on Hacker News now like three or four times with little to no results. So if a tree falls in the woods...&quot; (I censored the F word, but it transcribed just fine.)<p>TL;DR: LymeScribe is local speech to text for Mac and Windows: dictate into any app, or drop in a recording and get a transcript with speaker labels, and nothing goes to a server you don&#x27;t own. On a machine that can handle it, it all runs right there. On anything that can&#x27;t, or anywhere you&#x27;d rather keep transcription in one place, one computer does the work for the rest of the network. Offices care about the second one, because the audio and the transcripts never leave hardware they control.<p>This whole thing started as a command line program I built to transcribe and label speakers in audio files sitting on my computer, mostly voice notes using Whisper on my gaming PC. A few friends started asking me to run their meeting recordings through it, which I was happy to do until it got frequent enough to be a chore. So I built them an app that they could drag a file onto themselves, with my gaming PC doing the actual work and sending the results back to their clipboard.<p>That drag and drop piece is still in the app. Since then I added real time dictation, which uses the same local models. When I started thinking about actually selling this thing, I found the market was more crowded than I expected, which was humbling for about a day. What kept me going is that none of the others let one machine do the work for everyone else... that and there&#x27;s just something sort of cool about owning and using my own transcription software. So, here we are.<p>An older machine can handle its own dictation on the small English models. I found that out when I tried the app on a ten year old iMac expecting it to fail, and deleted the code I&#x27;d written to keep machines like it out. Speaker labels are the part that still wants real hardware, and that&#x27;s where pointing that old machine at a stronger one on the network earns its keep. One license covers the house.<p>It&#x27;s local by design, and you can test that yourself: unplug the Ethernet mid dictation and it keeps going. Works on a plane. The free client never talks to me at all. Unlocking hosting means one activation call, and after that the license is verified locally against a signed document, so an activated install keeps working with no internet and would keep working if my licensing server went away. AI assisted summaries and translation are experimental. They ship turned off, and you can turn them on if you dare. They only talk to an endpoint you choose, your own key or a local model.<p>The client is free on Mac and Windows. Hosting is a one time $39 unlock with a 7 day trial that starts the first time you host a transcription. It&#x27;s a perpetual license: you own this major version, point updates are free, and if there&#x27;s ever a next major version it&#x27;s a discounted upgrade rather than a new purchase. I&#x27;m tired of renting everything, aren&#x27;t you?<p>Two things worth knowing before you download. The Windows installer is signed, but the certificate is new enough that Chrome and SmartScreen may still warn you about it. And the paid unlock is US only for now, while I work out the international side.<p>Feedback welcome, especially from anyone who has fought text injection on Windows. There&#x27;s also a heavier server for offices, administered through a web app that handles keys, transcription history, logs, etc.<p>Frog reference: &quot;There&#x27;s a saying: if you have to eat a frog, it&#x27;s best you don&#x27;t spend a whole lot of time looking at it. And if you have to eat two, don&#x27;t start with the small one.&quot;<p>Randy Pausch - Time Management - <a href=\"https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=oTugjssqOT0\" rel=\"nofollow\">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=oTugjssqOT0</a>", "5": "2026-08-23T07:30:04.625563"}
{"0": 321, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49402769", "3": "Ask HN: What's your daily driver keyboard?", "4": "Ask HN: What's your daily driver keyboard?. Inspired by the thread about typ.ing [0] and people talking about their setup, I wondered what HN users are using for their daily driver keyboard.<p>Please list the model, switches, keycaps, layout and any other details you think relevant.<p>[0]: https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=49346854", "5": "2026-08-23T07:30:10.980485"}
{"0": 331, "1": "hackernews", "2": "https://www.axios.com/2026/08/07/openai-astra-model-delay-cybersecurity-risks", "3": "OpenAI slows release of Astra model citing cyber capabilities", "4": "OpenAI slows release of Astra model citing cyber capabilities. ", "5": "2026-08-23T07:30:11.034486"}
{"0": 334, "1": "hackernews", "2": "https://ltx.io/model/ltx-2-5", "3": "LTX 2.5 Released", "4": "LTX 2.5 Released. ", "5": "2026-08-23T07:30:11.048940"}
{"0": 335, "1": "hackernews", "2": "https://modelscope.cn/models/Qwen/Qwen3.8-2.4T-A95B", "3": "Qwen3.8 Weights Released", "4": "Qwen3.8 Weights Released. ", "5": "2026-08-23T07:30:11.053512"}
{"0": 336, "1": "hackernews", "2": "https://github.com/ZeroHackz/cyberleek-viewer", "3": "Cyberleek Viewer based on-chain (no browser needed)", "4": "Cyberleek Viewer based on-chain (no browser needed). ", "5": "2026-08-23T07:30:13.014053"}
{"0": 337, "1": "hackernews", "2": "https://github.com/hugorcd/shelve", "3": "Shelve Secret Management", "4": "Shelve Secret Management. ", "5": "2026-08-23T07:30:13.020041"}
{"0": 338, "1": "hackernews", "2": "https://github.com/AntoineChatry/Dictata", "3": "Show HN: Dictata \u2013 Local Whisper dictation with LLM cleanup", "4": "Show HN: Dictata \u2013 Local Whisper dictation with LLM cleanup. ", "5": "2026-08-23T07:30:13.025236"}
{"0": 339, "1": "hackernews", "2": "https://nervecenter.github.io/ai_power_scaling_cuts_both_ways.html", "3": "AI Power Scaling Cuts Both Ways", "4": "AI Power Scaling Cuts Both Ways. ", "5": "2026-08-23T07:30:13.030396"}
{"0": 340, "1": "hackernews", "2": "https://github.com/phase-rs/phase", "3": "Phase.rs \u2013 a full featured MTG SIM in Rust", "4": "Phase.rs \u2013 a full featured MTG SIM in Rust. ", "5": "2026-08-23T07:30:13.035312"}
{"0": 341, "1": "hackernews", "2": "https://aramzs.github.io/steal-the-internet/", "3": "How to archive Everything and share It", "4": "How to archive Everything and share It. ", "5": "2026-08-23T07:30:13.040453"}
{"0": 342, "1": "hackernews", "2": "https://github.com/gojiplus/layoutlens", "3": "Show HN: LayoutLens: AI-Powered Visual UI Testing", "4": "Show HN: LayoutLens: AI-Powered Visual UI Testing. ", "5": "2026-08-23T07:30:13.045638"}
{"0": 343, "1": "hackernews", "2": "https://github.com/Freaky/Compactor", "3": "Compactor: A user interface for Windows 10 filesystem compression", "4": "Compactor: A user interface for Windows 10 filesystem compression. ", "5": "2026-08-23T07:30:13.049632"}
{"0": 344, "1": "hackernews", "2": "https://kwojcicki.github.io/blog/NEW-AGE-OF-DEV-EX", "3": "The incoming Developer Experience revolution", "4": "The incoming Developer Experience revolution. ", "5": "2026-08-23T07:30:13.054134"}
{"0": 345, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49405520", "3": "Why can AI generate Super Mario but not a wedge ramp for my robot vacuum?", "4": "Why can AI generate Super Mario but not a wedge ramp for my robot vacuum?. I&#x27;ve been puzzled by something: AI generation can produce an elaborate\n  figurine, a cartoon character, even a convincing Super Mario \u2014 yet it\n  can&#x27;t reliably make a simple wedge ramp so my robot vacuum can climb a\n  step.<p><pre><code>  For context: I bought a Bambu P2S but can&#x27;t model. I tried the &quot;describe\n  it and get a model&quot; AIs \u2014 the output is unusable, you can&#x27;t adjust it,\n  it&#x27;s never quite what I meant. I tried having an agent write Python to\n  build geometry directly \u2014 it tops out at simple primitives.\n \n  What finally worked: geometric decomposition. I break a complex part into\n  ordered, grouped steps, describe each as a small spec, and let an agent\n  execute them in Blender (via blender-mcp). That process turned out to\n  abstract into a small engine \u2014 the key insight being it converts the 3D\n  spatial reasoning LLMs are bad at, into the structured code they&#x27;re good\n  at. I wrote it up here: https:&#x2F;&#x2F;github.com&#x2F;zhuchaokn&#x2F;spec-3d-model\n \n  My questions:\n  - Why is &quot;functional part&quot; generation so much weaker than\n  &quot;figurine&#x2F;aesthetic&quot; generation? Is it data (no parametrized-CAD training\n  sets), representation (mesh vs B-rep), or evaluation (nobody benchmarks\n  &quot;does it print &#x2F; is it watertight&quot;)?\n  - Is &quot;turn 3D modeling into code for an LLM&quot; the right framing, or am I\n  missing something better?</code></pre>", "5": "2026-08-23T07:30:13.059893"}
{"0": 346, "1": "hackernews", "2": "https://www.utilitydive.com/news/ferc-spp-topology-optimization-grid-congestion/828366/", "3": "FERC approves SPP 'topology optimization' plan for cutting grid congestion", "4": "FERC approves SPP 'topology optimization' plan for cutting grid congestion. ", "5": "2026-08-23T07:30:13.067436"}
{"0": 347, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49405405", "3": "Show HN: Hands-Rust MCP/CLI that sees the Windows desktop and clicks real Chrome", "4": "Show HN: Hands-Rust MCP/CLI that sees the Windows desktop and clicks real Chrome. I built Hands because I wanted a coding agent to use this Windows PC and a real Chrome profile the way I do: look at the screen, move the real mouse, type, click , without turning Chrome into an automation browser.<p>It is a Rust MCP&#x2F;CLI. A harness (Grok, Codex, Claude Code, OpenCode, etc.) calls tools like observe, click, type, scroll. Observe is a screenshot path plus a small element list (UIA + optional Chrome DOM ids). Click is OS SendInput on a B\u00e9zier path, not a Chrome DevTools click.<p>There is no Playwright, no Puppeteer, no remote debugging port. Daily Chrome is launched with no extra flags, or attached if it\u2019s already open. Sites that key on CDP&#x2F;automation flags mostly don\u2019t see that. They can still see injected input (LLMHF_INJECTED).<p>A tiny unpacked Chrome extension can fuse page structure (chr: ids, listing cards) so the model isn\u2019t guessing from pixels. Sideload is manual. Fusion dies if the service worker goes inactive; reload the card.<p>What it is good for: personal research on your own desk. \u201cFind a Camry on cars.com,\u201d read a page, fill a ZIP, dismiss a cookie banner.<p>What it is not:\n\u2022 Not a sandbox. It can click whatever is on screen, including checkout and Easy Apply.\n\u2022 Confirm-before-money is best-effort classification in the binary, not a guarantee. Prompt injection from the screenshot&#x2F;DOM is real; the binary treats that text as untrusted, the model might not.\n\u2022 Not a CAPTCHA solver on daily Chrome. Two visible tries, then it yields and waits for the puzzle to go away.\n\u2022 Windows only. \n\u2022 Install is: build the exe, register a native-messaging host, sideload the extension, point an MCP client at hands mcp. README is the runbook. Missing an API key does not fail the build; do_task is optional.\n\u2022 Logs live under %LOCALAPPDATA%\\hands\\logs\\. The extension asks for &lt;all_urls&gt; so it can map the tab you\u2019re looking at.<p>Repo: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;Ryan-AI-Studios&#x2F;hands\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;Ryan-AI-Studios&#x2F;hands</a> (MIT)<p>Happy to answer how observe&#x2F;fusion&#x2F;the fence work. If you try it, Pause&#x2F;Break is the kill switch.", "5": "2026-08-23T07:30:13.073589"}
{"0": 348, "1": "hackernews", "2": "https://github.com/Jamedjo/git-hunk/commit/240ac0f7e986fbe783a7de1c220fc047771ee059", "3": "Show HN: Nice Licence \u2013 An anti-copyleft permissive licence", "4": "Show HN: Nice Licence \u2013 An anti-copyleft permissive licence. ", "5": "2026-08-23T07:30:13.079251"}
{"0": 349, "1": "hackernews", "2": "https://meetless.ai", "3": "Show HN: Active Source of Truth for Your Coding Agents", "4": "Show HN: Active Source of Truth for Your Coding Agents. Howdy! Happy Saturday everyone!<p>As a solo founder, I have always tried to maximize my speed by letting coding agents build as much as possible in parallel. However, as an engineer, I don&#x27;t trust that AI will always make the right decisions and work with the right context. In the past, I always needed to click through my sessions to glance at the AI&#x27;s output, try to understand what it was doing, and hopefully steer it or stop it in time.<p>As a result, the maximum number of concurrent sessions I could manage at once was only 4. I didn&#x27;t want to be the bottleneck, so I built Meetless Agent (MLA). It basically does what I had to do manually before:<p>- Monitors the coding agent&#x27;s tasks and actions to supply it with the correct, up-to-date context.<p>- Continuously reconciles running information (such as provided&#x2F;tagged documentation, the agent&#x27;s output, and the agent&#x27;s decisions) to actively maintain the source of truth at all times.<p>- Keeps track of the current rules for the repo, and if an action triggers a registered rule, the rule is injected into the agent context.<p>My benchmarks show that running coding agents with the help of an active monitor improves quality and accuracy, consumes fewer tokens, and finishes faster: <a href=\"https:&#x2F;&#x2F;research.meetless.ai&#x2F;stale-context&#x2F;\" rel=\"nofollow\">https:&#x2F;&#x2F;research.meetless.ai&#x2F;stale-context&#x2F;</a><p>Of course, the agent alone can&#x27;t decide the source of truth; it requires human review and decisions for contradictions, etc. But for the most part, it can safely build a consistent ontology of the current source of truth.<p>From this, I want to build an AI layer to maintain the source of truth across the business, so I will release more connectors for Slack, Jira, Confluence, etc., soon. The goal is for this AI to assist in every part of the business. Eventually, the same coordination layer will understand that a decision made in Slack affects a Jira task, a document, an email conversation, and what a coding agent should do next without every tool becoming another isolated memory silo.<p>The coding agent connector is open source at:\n<a href=\"https:&#x2F;&#x2F;github.com&#x2F;Meetless&#x2F;mla\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;Meetless&#x2F;mla</a><p>I am looking forward to your feedback!", "5": "2026-08-23T07:30:13.087401"}
{"0": 350, "1": "hackernews", "2": "https://github.com/BenSiv/fossil-scm", "3": "Adapting Fossil-scm as a platform for AI agentic workflow", "4": "Adapting Fossil-scm as a platform for AI agentic workflow. ", "5": "2026-08-23T07:30:13.095267"}
{"0": 351, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49405684", "3": "Ask HN: Anyone set up ways to easily obtain and read transcripts from Ted, YT?", "4": "Ask HN: Anyone set up ways to easily obtain and read transcripts from Ted, YT?. I really absorb text a great deal better than listening to audio or watching video. I&#x27;ve heard others say the same and attribute to ADHD, so perhaps that&#x27;s the case for me.<p>TED very often has a transcript on their page, but it involves some clicking and then some copy-pasting, and the text result is often broken mid-clause after about 10-15 characters.<p>I&#x27;m wondering if there&#x27;s any useful userscripts, userstyles, or other methods that handle this. (I think TED often sues people who try to make transcripts easily available on a separate site. I may be mistaken, and TED, please don&#x27;t sue me, it&#x27;s just a general feeling.)<p>The same request would apply to YouTube and&#x2F;or other popular video sites, but I run into this issue most often with TED.<p>My thanks in advance to anyone who can help me on this!", "5": "2026-08-23T07:30:14.464219"}
{"0": 353, "1": "hackernews", "2": "https://structo.redbeardlab.com/", "3": "Show HN: Structo \u2013 Claude Code for short form writing", "4": "Show HN: Structo \u2013 Claude Code for short form writing. ", "5": "2026-08-23T07:30:14.475194"}
{"0": 354, "1": "hackernews", "2": "https://slide-deck.io/", "3": "Slide-deck.io: remixable decks and MCP so Claude returns a live URL (gantt too)", "4": "Slide-deck.io: remixable decks and MCP so Claude returns a live URL (gantt too). ", "5": "2026-08-23T07:30:14.479832"}
{"0": 355, "1": "hackernews", "2": "https://github.com/HarjjotSinghh/locum", "3": "Show HN: Locum \u2013 Grok Bot delegates tasks to Claude/Codex CLI on your machine", "4": "Show HN: Locum \u2013 Grok Bot delegates tasks to Claude/Codex CLI on your machine. ", "5": "2026-08-23T07:30:14.486925"}
{"0": 356, "1": "hackernews", "2": "https://github.com/arturlimaaa/terminito", "3": "Claude Code status line for worktree, branch, context and quota", "4": "Claude Code status line for worktree, branch, context and quota. ", "5": "2026-08-23T07:30:14.491948"}
{"0": 357, "1": "hackernews", "2": "https://programasweights.com/claudish", "3": "English \u2194 Claudish Translator", "4": "English \u2194 Claudish Translator. ", "5": "2026-08-23T07:30:14.497159"}
{"0": 358, "1": "hackernews", "2": "https://www.soverybright.com/", "3": "Show HN: Make your logo extra bright on HDR screens", "4": "Show HN: Make your logo extra bright on HDR screens. Certain logos started standing out to me on LinkedIn as brighter&#x2F;whiter than everything else around them.<p>I dug in and found out this is accomplished by adding a gain-map to an existing JPEG, visible only on HDR screens like a newer MacBook Pro. LinkedIn is the only social network I&#x27;ve found that isn&#x27;t stripping them out, but of course you serve them up on your own site.<p>I worked with Claude Code to turn it into a little browser-based utility (no registration) and hope you find it useful!", "5": "2026-08-23T07:30:14.505056"}
{"0": 359, "1": "hackernews", "2": "https://twitter.com/argofowl/status/2091150597374537729", "3": "Anthropic appears to be A/B testing reduced effort levels in Claude Code", "4": "Anthropic appears to be A/B testing reduced effort levels in Claude Code. ", "5": "2026-08-23T07:30:14.509631"}
{"0": 360, "1": "hackernews", "2": "https://github.com/Parthuss/stash", "3": "Show HN: Stash, turn your Instagram saves into notes Claude finds on its own", "4": "Show HN: Stash, turn your Instagram saves into notes Claude finds on its own. ", "5": "2026-08-23T07:30:14.515357"}
{"0": 397, "1": "hackernews", "2": "https://misfra.me/2026/software-design/", "3": "Software Design", "4": "Software Design. ", "5": "2026-08-23T07:30:23.815631"}
{"0": 401, "1": "hackernews", "2": "https://radi8.dev/blog/uplink/", "3": "Show HN: We chased a weather balloon across Montana and never found it", "4": "Show HN: We chased a weather balloon across Montana and never found it. Since April, I have been working with Sam Flynn (<a href=\"https:&#x2F;&#x2F;drook.dev\" rel=\"nofollow\">https:&#x2F;&#x2F;drook.dev</a>) to make this balloon payload, UpLink. We did a similar launch last year with Hack Club but this was our first independent launch.<p>UpLink was a 491 gram payload testing the insulation properties of 3D printing filaments, while also transmitting 320x240 images over a radio link -- up from the 18x10 images last year!<p>This is a writeup on our engineering process, mistakes made, and learning experiences. It covers:<p>- Custom electronics designed in KiCad<p>- Firmware design<p>- Results from the data we received on the ground<p>- Image transmission<p>- Launch day logistics, and where things went wrong<p>All hardware, software, firmware, and CAD is available on GitHub: <a href=\"https:&#x2F;&#x2F;github.com&#x2F;radeeyate&#x2F;UpLink\" rel=\"nofollow\">https:&#x2F;&#x2F;github.com&#x2F;radeeyate&#x2F;UpLink</a>, licensed + certified as open source hardware: <a href=\"https:&#x2F;&#x2F;certification.oshwa.org&#x2F;us002826.html\" rel=\"nofollow\">https:&#x2F;&#x2F;certification.oshwa.org&#x2F;us002826.html</a><p>If you just want to see the images received, I put up a gallery here: <a href=\"https:&#x2F;&#x2F;uplink.gallery.radi8.dev&#x2F;\" rel=\"nofollow\">https:&#x2F;&#x2F;uplink.gallery.radi8.dev&#x2F;</a><p>If you have any questions, comments, or concerns, let me know. I&#x27;m happy to answer anything!", "5": "2026-08-23T07:30:23.839621"}
{"0": 402, "1": "hackernews", "2": "https://www.youtube.com/watch?v=qBVzsolalJ8", "3": "Teachers Pay Teachers hit by a wave of A.I. content with mistakes [video]", "4": "Teachers Pay Teachers hit by a wave of A.I. content with mistakes [video]. ", "5": "2026-08-23T07:30:23.844438"}
{"0": 403, "1": "hackernews", "2": "https://triblive.com/privacy/", "3": "'It wasn't me': Weed use in Allegheny County jury room forces homicide mistrial", "4": "'It wasn't me': Weed use in Allegheny County jury room forces homicide mistrial. ", "5": "2026-08-23T07:30:23.849573"}
{"0": 404, "1": "hackernews", "2": "https://daniel.reguero.dev/blogs/ten-common-security-mistakes-ai-generated-apps-make", "3": "Common Mistakes I've Seen That AI-Generated Apps Keep Making", "4": "Common Mistakes I've Seen That AI-Generated Apps Keep Making. ", "5": "2026-08-23T07:30:23.858068"}
{"0": 406, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49346107", "3": "Ask HN: Can I realistically make an app as a junior dev?", "4": "Ask HN: Can I realistically make an app as a junior dev?. I&#x27;m trying to make an app to help out my local community, but with the intention of scaling it at some point and serving a much wider community. It&#x27;s a complex app with a lot of moving parts, and as a junior dev with little experience in some of the technologies I&#x27;m using, I&#x27;m going to be relying heavily on Claude. However, I know generally, Claude is only as strong as the developer using it\u2014and I&#x27;m a junior dev. So, for example, what do I do about architecture? I can&#x27;t leave the architecture up to Claude, but I&#x27;m not at the point in my career where I can design the architecture on my own without flaws.<p>What are y&#x27;all&#x27;s thoughts on what I can learn on my own through posts&#x2F;YouTube videos, vs. what I might need to hire a senior dev to help me do?<p>I&#x27;ve had this idea for a while, with the thought that I&#x27;d make it when I&#x27;m a more experienced software engineer, but there&#x27;s no time like the present and I just want to bite the bullet and get started, even if it means I&#x27;ll be making more mistakes and learning on the job.<p>For reference, the stack I&#x27;m leaning toward is PostreSQL (I don&#x27;t have experience with this), Java Spring Boot (I&#x27;m pretty experienced with this), Typescript with Angular (pretty experienced), and some python (some experience). There&#x27;s more stuff, like Redis (no experience).<p>Anyway...any help&#x2F;advice on how to do this properly would be really helpful. Part of the idea of this project is just to learn, even if it means I learn a couple of months in that I have to start everything over from scratch, lol.", "5": "2026-08-23T07:30:23.868358"}
{"0": 407, "1": "hackernews", "2": "https://context-engine.app", "3": "Show HN: My agents kept hallucinating APIs, so I built them a headless IDE", "4": "Show HN: My agents kept hallucinating APIs, so I built them a headless IDE. I built Context Engine last year mostly out of frustration. Agents impressed me and made me angry at the same time: they kept calling APIs that don&#x27;t exist, re-implementing APIs already defined in the codebase, and getting stuck in write-compile-rewrite loops. Sometimes they seemed very intelligent and sometimes extremely stupid. But they aren&#x27;t stupid: take any brilliant engineer, give them a whiteboard, and ask them to implement a new feature in a large codebase they have never seen. And for agents, every session is a project they have never seen.<p>The problem is that agents lack proper tooling. They work with plain text, like human engineers in 1995, in the pre-IDE era, and they face the same problems engineers faced 30 years ago. Where a human engineer just hovers over a symbol or presses F12, an agent spends thousands of tokens to get the same result. So I built the IDE part they were missing, without the editor part they don&#x27;t need \u2014 an IDE headless in the same sense as a headless browser.<p>Initially my goal was to eliminate API hallucinations by giving agents cheap tools that to resolve the APIs instead of guessing about them. Then, to make it even cheaper, I added inlined inlay hints: resolved types and parameter names at call sites, so agents spend their reasoning budget on the task instead of on reconstructing types.<p>But then I realized that there is a more fundamental issue: the codebase is not a set of files, it is a graph of symbols (and dependency symbols are the part of it). If I provide agents the tools to follow the graph, I enable them to cherry-pick only the symbols they consider meaningful for the task they are working on. LSPs provide this capability. And Tree-sitter allowed agents to cherry-pick only the necessary parts of those symbols.<p>Since January I have been using it daily in my current project (math-to-silicon synthesis engine) as an internal tool. Some thoughts (not benchmarks!) after seven months of extensive use:<p>1. I expected fewer mistakes from eliminating API guessing, but the result was better. Agents behave like better engineers: they tend to reuse abstractions already defined in the codebase rather than duplicate them ad hoc, and not only did type and lifetime compile errors almost disappear \u2014 they also make fewer logical errors. My assumption is that this comes from reduced context pollution: when agents cherry-pick only the information relevant to the task, their reasoning capability degrades slower.<p>2. Token usage reduction wasn&#x27;t a goal, but the savings per task turned out substantial.<p>3. The most unusual experience was UX design for machines. Giving agents better tools isn&#x27;t enough: you have to convince them to use these tools instead their built-in text-oriented tools. The working solution was to describe workflows (codebase exploration, debugging, dependency API discovery, refactoring, reading and editing documentation) as chains of the new tools rather than describing each tool separately.<p>Context Engine runs fully locally, and your code never leaves your machine. One daemon is shared between all your agents and their subagents: Claude Code, Codex, Cursor, or any other agent that supports MCP talks to the same instance, and agents working in the same workspace also share the LSP servers serving it.<p>Context Engine MCP requires an API key, which you can get for free on <a href=\"https:&#x2F;&#x2F;context-engine.app\" rel=\"nofollow\">https:&#x2F;&#x2F;context-engine.app</a>. I do not have plans to make it paid yet but want to reserve the right to do it in the future. Anyway, it is free now and will remain free in the foreseeable future. It is language-agnostic and can work with any language, although it is validated end-to-end for Python, TypeScript&#x2F;JavaScript, Rust, Go, and Markdown only so far. The known issue is with Java: jdtls is still fighting me.<p>I would love feedback, especially on the tool output design.", "5": "2026-08-23T07:30:23.875007"}
{"0": 408, "1": "hackernews", "2": "https://reproof.app", "3": "Show HN: Reproof \u2013 A non-linear writing app for perfectionists", "4": "Show HN: Reproof \u2013 A non-linear writing app for perfectionists. Hey HN! I&#x27;ve written more words on the internet than I could count, for everyone from Zapier to Wirecutter. And for years it&#x27;s bugged me that nearly every writing app still treats the writing surface as a linear sheet of paper, with paragraphs one after another and no space for experimentation.<p>And so I and a friend build Reproof.app as what I&#x27;ve taken to calling a writing app for perfectionists.<p>It&#x27;s built around paragraph versions. You can add new versions to any paragraphs and swap between them to see how the different takes read in context. You can cut a paragraph as a clipping to save it for later. It\u2019s a writing app that lets you experiment without losing any of your ideas along the way.<p>It took four complete rewrites before we finally found an approach that worked well (don\u2019t reinvent the text editor, they say, yet we tried, and even with Tiptap it still was harder than it should have been). But it\u2019s here. You can write and rewrite to your heart\u2019s content in Reproof today, with writer-focused features like equal support for rich text and markdown (and HTML; you can copy in either format anytime), a command palette to keep your hands on the keyboard, side-by-side documents to have notes on one side and your document on the other, and a proofreading tool to catch things you might not other (with suggestions, but no AI rewriting, as I believe humans should still be doing their own writing, with AI as a sidekick to catch typos and other mistakes). And there\u2019s a lot more coming.<p>Would love your feedback, if you have a chance to try it!", "5": "2026-08-23T07:30:23.883315"}
{"0": 409, "1": "hackernews", "2": "https://orbitquote.com/", "3": "Show HN: A Buyer's Editor to Fix B2B Procurement PDF Hell with API and MCP", "4": "Show HN: A Buyer's Editor to Fix B2B Procurement PDF Hell with API and MCP. 22+ years as a Software engineer&#x2F;programmer.<p>For a reason that only life can explain, in the last 8 years i went from Adobe Flash Specialist to owning a small solar installation business and a small electric and plumbing materials. o.O<p>We had a problem every time that we had to resupply or shop for new goods. In Brazil, things are manual in B2B and restricted to PDF Quotations&#x2F;Proforma negotiations. Nobody has ecommerce, only the middlemen that resell what they buy from the factory with their margins on top.<p>For better margins, direct from the factory is the way to go.<p>Long story short. We wasted 2 weeks on a single supplier for just one order.<p>On the supplier side, everything is a 100+- PDF pages catalog. One PDF page could have 100+ products that you have to manually extract into Excel with the product code (sku)(even with chatgpt and a gazillion of .final.final_right.final_corrected.xlsx(10).xlsx ), to simply ask for the price and minimum items for order &gt;&gt; receive another PDF &gt;&gt; Excel again to see prices, quantities, adjustments and even mistakes of items we didn&#x27;t quote.<p>It was purchasing hell!<p>So I built OrbitQuote to solve this. We used and validated it in our own shop.<p>Today, OrbitQuote extracts catalogs, quotations&#x2F;budgets&#x2F;proforma invoices, excel files, pdf, etc, into a live editor that you can quickly see your order changes and send back in seconds only what you want to buy. From days to minutes.<p>We also have built OrbitQuote Compare, which lets you compare the same products across multiples suppliers  and choose the bast purchasing strategy. It&#x27;s seams simples. It&#x27;s not. Every suppliers writes the product in a different way. OrbitQuote understand this and aggregates the same product and choose the best fit automatically.<p>I closed the Shop and Solar installation business and i&#x27;m focused 100% on OrbitQuote.<p>All our users need to do, it&#x27;s drop their files and done, 95% of the boring work is done.<p>No paywall to check out. Just register;", "5": "2026-08-23T07:30:23.890367"}
{"0": 410, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49332387", "3": "Ask HN: Is this the worst time to be human?", "4": "Ask HN: Is this the worst time to be human?. With AI doing much of the work, how do we humans keep life interesting?\nEx:- \n1.Generate much of the code and remove the joy of developers that they had over all these years. \n2. Generate Animation videos&#x2F;movies or Anime books. People who make a living by drawing because they like it will feel all the more miserable. \n3. May be more scenarios can be listed<p>If you are into a particular profession because you like it and get paid in addition, you have every reason to be miserable&#x2F; depressed.<p>what are your thoughts?\n1. Is this the beginning of golden age for people who does not have any good skill but just mediocre who will make a killing by using AI tools and drive away people who are into a certain profession because of passion? \n2. Are creative people doomed? \n3. What options are available ? \n4. Looks like greed is causing miserable lives for everyone. \n5. Perhaps passionate people will continue to do things the &quot;old way&quot; but cannot make a living or compete with AI folks. For ex:- drawing animation by hand will become less economical compared with AI generated unless you are faster than a speeding bullet?<p>Please share your thoughts. \nThanks", "5": "2026-08-23T07:30:23.896056"}
{"0": 411, "1": "hackernews", "2": "https://arcadedb.com", "3": "Arcadedb \u2013 Open-Source Multi-Model Graph Database", "4": "Arcadedb \u2013 Open-Source Multi-Model Graph Database. ", "5": "2026-08-23T07:30:25.430729"}
{"0": 412, "1": "hackernews", "2": "https://news.ycombinator.com/item?id=49401293", "3": "Ask HN: If you dislike AI, why don't you prove it?", "4": "Ask HN: If you dislike AI, why don't you prove it?. I&#x27;ve seen so many people who absolutely hate AI, and want AI companies to fail, but then 2 seconds later, open up the coding tool made by frontier labs and willingly hand them money.<p>Just recently convinced one of them to move away from the frontier lab lock-in to OpenCode and switch between models. Not saying they need to switch to non-frontier models immediately, but this approach lets them easily switch to open source when the time is right.<p>Why don&#x27;t more people do this? And shouldn&#x27;t companies that use AI heavily also suggest this approach?<p>To be clear, I like AI, and use it all the time, but I use it in a way that allows me to easily switch models and not get locked-in.", "5": "2026-08-23T07:30:25.436654"}
{"0": 425, "1": "hackernews", "2": "https://github.com/garagehq/Minus-streaming-stick", "3": "Show HN: Minus HDMI hardware ad-blocker with local vision model", "4": "Show HN: Minus HDMI hardware ad-blocker with local vision model. Hey HN! Minus is an open source hardware device that sits between your streaming stick and TV and blocks ads that appear on your TV in real time at 4k60fps.<p>This was made out of a labor of love after seeing nonstop &quot;Hims&quot; ads for months, so I created a hardware-in-the-loop test platform to gather over 80k screenshots of ads and train a custom model to detect and block ads in under 300ms.<p>Since we are blocking the ads at the hardware level we have complete control over the screen, so currently replacing the ads with language learning content, pictures from trips, random fun facts, haikus, etc.<p>Code is open source and the weights for the model are on HuggingFace", "5": "2026-08-23T07:30:25.505005"}
{"0": 469, "1": "hackernews", "2": "https://chaser.vgnsh.xyz", "3": "Show HN: Chaser (beta) \u2013 Share Mac windows as screenshots and UI context", "4": "Show HN: Chaser (beta) \u2013 Share Mac windows as screenshots and UI context. Chaser is a macos utility you can use to capture more than a screenshot. Screenshots are good tools to show your agents what you&#x27;re seeing, but there&#x27;s far more that goes on behind the scenes - the layout, button states, mapping of objects on your application, things that are out of your line of sight - all help make the context better. Chaser uses macos accessibility features to help you capture all that and more. No LLM, 100% local.<p>This was inspired by Codex appshots. I really loved it so built a version that lives outside codex, also with history and the ability to annotate before sharing.", "5": "2026-08-23T07:30:30.521049"}
{"0": 470, "1": "hackernews", "2": "https://argot.tmonier.com/", "3": "Show HN: Argot, a Rust AI guardrail based on your codebase AST patterns", "4": "Show HN: Argot, a Rust AI guardrail based on your codebase AST patterns. We have lots of benchmarks for new frontier LLMs (SWE benchmarks etc) to make them score on the &quot;best code&quot;. In large codebase, the best code is the one that matches the codebase own voice and conventions, because that&#x27;s the mental model of the team.<p>Ingesting a full codebase explicit&#x2F;implict patterns in an agent windows to make them match the codebase doesnt work: too much context, hallucinations, ...<p>So I built this free&#x2F;open source project, your codebase is a golden mine of AST code patterns that makes its unique voice, Argot learns them and then allows you to enforce them proactively or defensively.", "5": "2026-08-23T07:30:30.527518"}