Post
81
The most important takeaway from the super agent paper by DeepSeek: finally someone admits a person can only attend to 3 to 5 agents.
AI & ML interests
None defined yet.
Post
131
To vibe code like Linus Torvalds, you must first become Linus Torvalds.
Post
98
I, and readers of this post should belong to the 20M people layer in the Karpathy funnel.
But we do so much more than described. We're the bridge to bring AI from the frontier layer to the mass layer.
But we do so much more than described. We're the bridge to bring AI from the frontier layer to the mass layer.
Post
116
I added ASD-STE100 to global claude.md
- 5 replies
Post
3,201
My takes on Jev
1. Very likely a small model. You can certainly pretrain, but I would grab an existing base model, say Qwen 3 class
2. The new RL method is a breakthrough, classification doesn't need to align with human preferences
3. The new output is an overstatement. It's just a new LM head. Of course autoregressive decoding can be used for classification: it takes just a few tokens to express the output. Think twice: are you sure classification doesn't need few-shot, CoT, or reasoning? All of these depend on auto-regressiveness
4. It carves out a market already existing, which is now served by oversized LLMs (hence overpaid), e.g. LLM as judge, labeling
5. Jevons effect will kick in, promoting more modeling efforts for small budget teams. It might even accelerate RSI
1. Very likely a small model. You can certainly pretrain, but I would grab an existing base model, say Qwen 3 class
2. The new RL method is a breakthrough, classification doesn't need to align with human preferences
3. The new output is an overstatement. It's just a new LM head. Of course autoregressive decoding can be used for classification: it takes just a few tokens to express the output. Think twice: are you sure classification doesn't need few-shot, CoT, or reasoning? All of these depend on auto-regressiveness
4. It carves out a market already existing, which is now served by oversized LLMs (hence overpaid), e.g. LLM as judge, labeling
5. Jevons effect will kick in, promoting more modeling efforts for small budget teams. It might even accelerate RSI
- 7 replies
Post
80
Not your weights, not your products.
This doesn't feel true at the beginning, but suddenly becomes true after PMF
This doesn't feel true at the beginning, but suddenly becomes true after PMF
- 1 reply
Post
4,497
I turned off subagents in Claude Code. Am I a minority?
- 5 replies
Post
1,009
My take on device-side inference: it's all about high bandwidth memory (thinking about it, this holds for the cloud too).
MacBooks enjoy incidental capacity of apple silicon, but per-device RAM is too low (16 to 24GB), only sufficient for a decent SLM. 512GB is the highest you can go (Kimi K2*). Counting MLX downloads of Kimi K2* on Huggingface, I estimate the user base to be <25K.
On the other hand, the newly debuted DGX station (Nvidia) has 748GB, which can fit in the latest Kimi, DS, and Qwen. Also the quantization options of CUDA is way better than MLX.
For high-end inferencing, I place my bet on workstations over Macs.
MacBooks enjoy incidental capacity of apple silicon, but per-device RAM is too low (16 to 24GB), only sufficient for a decent SLM. 512GB is the highest you can go (Kimi K2*). Counting MLX downloads of Kimi K2* on Huggingface, I estimate the user base to be <25K.
On the other hand, the newly debuted DGX station (Nvidia) has 748GB, which can fit in the latest Kimi, DS, and Qwen. Also the quantization options of CUDA is way better than MLX.
For high-end inferencing, I place my bet on workstations over Macs.
- 7 replies
Post
169
My biggest takeaway from Ben Thompson interview is the legacy of this round of "bubble".
For railroad, it was the track, and dot com, fiber. Compared to these two, GPUs depreciate too fast.
The answer is power plants.
For railroad, it was the track, and dot com, fiber. Compared to these two, GPUs depreciate too fast.
The answer is power plants.
Post
2,017
Back of the envelope calculation: Ox Alpha has been out for a week with 130K users and 7T tokens burned.
Method 1: Assuming Ox Alpha is the GLM 5.3 class, it has roughly the same active param count as DeepSeek V4 Pro (~40B vs 49B), derive the cost by the floor price DS has ever published.
Method 2: Assuming the users are concentrated within an 8-hour working window each day, derive number of H800 nodes needed (~700 nodes at $2/GPU-hour).
In both methods, I assume 90/10 IO split and 60% cache hit. Both methods come to $2M.
I think the ROI is awesome: (1) publicity and (2) data harvesting.
Method 1: Assuming Ox Alpha is the GLM 5.3 class, it has roughly the same active param count as DeepSeek V4 Pro (~40B vs 49B), derive the cost by the floor price DS has ever published.
Method 2: Assuming the users are concentrated within an 8-hour working window each day, derive number of H800 nodes needed (~700 nodes at $2/GPU-hour).
In both methods, I assume 90/10 IO split and 60% cache hit. Both methods come to $2M.
I think the ROI is awesome: (1) publicity and (2) data harvesting.
- 1 reply
Post
112
I for one believe DeepMind will ship one more frontier model.
In any case GCP is irrelevant here. GDM should still have more compute than DeepSeek, Kimi, or GLM, maybe all of them combined.
In any case GCP is irrelevant here. GDM should still have more compute than DeepSeek, Kimi, or GLM, maybe all of them combined.
Post
3,236
Lots of attentions are now on GPU residual value. I find car analogy to be useful.
* Both new and used cars can do the same job (you can run the latest model on 6-year-old A100)
* New cars are more efficient (higher performance per power draw)
* It's not the year, but mileage and maintenance
The last point is undeveloped. We need Carfax and KBB for GPUs.
And there will be lemon GPUs.
* Both new and used cars can do the same job (you can run the latest model on 6-year-old A100)
* New cars are more efficient (higher performance per power draw)
* It's not the year, but mileage and maintenance
The last point is undeveloped. We need Carfax and KBB for GPUs.
And there will be lemon GPUs.
- 1 reply
Post
2,775
This is an easy-to-remember pattern.
GLM and Kimi are in Beijing, DeepSeek and Qwen are in Hangzhou.
OpenAI and Anthropic are in SF, xAI and Meta are in the peninsula.
In both China and the Bay Area, token price of one area is ~1/3 of the other area.
GLM and Kimi are in Beijing, DeepSeek and Qwen are in Hangzhou.
OpenAI and Anthropic are in SF, xAI and Meta are in the peninsula.
In both China and the Bay Area, token price of one area is ~1/3 of the other area.
- 1 reply
Post
1,759
I think contributor pricing is a good idea. But things will get sophisticated very quickly.
* data sharing tiers from exclusive to syndication
* provenance and lineage
* scoping such as duration and geofencing
and their levels of enforcement (best practices, standards, laws)
* data sharing tiers from exclusive to syndication
* provenance and lineage
* scoping such as duration and geofencing
and their levels of enforcement (best practices, standards, laws)
Post
2,686
Looking forward to the Qwen 3.8 model drop, and congratulations on joining the trillion parameters club.
But my eyes are on the promised 27B model. Small (<50B) models decline on OpenRouter, because they are being run on local devices. If people see the family trees here on HF they will understand.
But my eyes are on the promised 27B model. Small (<50B) models decline on OpenRouter, because they are being run on local devices. If people see the family trees here on HF they will understand.
- 2 replies