Yi Cui

onekq

24 20 1

AI & ML interests

Benchmark, Code Generation Model

Recent Activity

posted an update about 2 hours ago

Many developers discovered that the native DeepSeek API has higher cache-hit rate than neocloud APIs hosting the same DS models. My speculation is that DS aggressively kills its old models. It has released 18 models thus far, and only 2 are being served now (v4 pro and flash). This is tough to customers who don't want to upgrade (migrate or leave), but effectively boost the serving capacity to the same model, i.e. more woods behind fewer arrows.

posted an update 4 days ago

There has been a leaked memo (now struck down) from the founder of DeepSeek. I'm not here to circulate it, but comment on the minimum-effort evolutionary path he proposed. LLM->CoT->Agent->Self-improvement->Singularity->Physical This makes sense to me: even at the agent stage I learn world models much faster than when I learned LLM at the LLM stage. But this means humans are still needed beyond the digital singularity, until robots can close their own loop: eval, manufacturing, self improvement, i.e. physical singularity.

posted an update 5 days ago

The Nvidia paper came down to this: remove synchronization barriers. DeepSeek has already done that with DeepEP (which this paper cited) one layer above. https://huggingface.co/papers/2607.16100 Nevertheless this is great. More users will benefit from this: Nvidia NCCL > SGLang/vLLM DeepEP

View all activity

Organizations

posted an update about 2 hours ago

Post

Many developers discovered that the native DeepSeek API has higher cache-hit rate than neocloud APIs hosting the same DS models.

My speculation is that DS aggressively kills its old models. It has released 18 models thus far, and only 2 are being served now (v4 pro and flash).

This is tough to customers who don't want to upgrade (migrate or leave), but effectively boost the serving capacity to the same model, i.e. more woods behind fewer arrows.

posted an update 4 days ago

Post

2193

There has been a leaked memo (now struck down) from the founder of DeepSeek. I'm not here to circulate it, but comment on the minimum-effort evolutionary path he proposed.

LLM->CoT->Agent->Self-improvement->Singularity->Physical

This makes sense to me: even at the agent stage I learn world models much faster than when I learned LLM at the LLM stage.

But this means humans are still needed beyond the digital singularity, until robots can close their own loop: eval, manufacturing, self improvement, i.e. physical singularity.

posted an update 5 days ago

Post

424

The Nvidia paper came down to this: remove synchronization barriers. DeepSeek has already done that with DeepEP (which this paper cited) one layer above.

Every Microsecond Matters: Achieving Near Speed-of-Light Latency in GPU Collectives (2607.16100)

Nevertheless this is great. More users will benefit from this:

Nvidia NCCL > SGLang/vLLM DeepEP

upvoted a paper 6 days ago

Every Microsecond Matters: Achieving Near Speed-of-Light Latency in GPU Collectives

Paper • 2607.16100 • Published 19 days ago • 3

posted an update 6 days ago

Post

101

This is definitely a minority opinion. World model MIGHT be the answer to enterprise AI.

https://arxiv.org/abs/2602.23643

* Tacit knowledge won't be captured in words, much less accessible dataset
* RL is not efficient at all, and labeling gets more and more expensive
* AGI won't cover your specialized tasks

PS: paper folks you missed one from Yann LeCun!!

posted an update 14 days ago

Post

144

You can stop goods but not ideas

https://substack.com/@onekq1/note/c-298811049

posted an update 16 days ago

Post

108

Hi, it's been a while 🤗

K3 stunned the world (again). But if you followed my post on K2, the architectural lineage is super clear

https://onekq1.substack.com/p/beyond-the-bonsai-how-kimi-k3s-quantization

I will next discuss the K3 optimizer (Muon again)

posted an update 8 months ago

Post

421

GPT 5.1 codex didn't make SOTA either. This should conclude 2025. No model has ever reached above 0.8.

onekq-ai/WebApp1K-models-leaderboard

Can this leaderboard be saturated in 2026?

updated a Space 8 months ago

WebApp1K Models Leaderboard

🥇

Display leaderboard for WebApp1K models

replied to their post 8 months ago

Hmm, I see.

posted an update 8 months ago

Post

282

I am starting a new series on matrix. The idea came to me when I wrote about the Muon optimizer.

Matrix itself has lots of fascinating properties and is applied in STEM fields for many decades. Its application in ML is just the beginning. There are lots of low hanging fruits. At the very least, I hope this math perspective will give you a new lens.

https://huggingface.co/blog/onekq/matrices-transformers-preface

2 replies

published an article 8 months ago

Article

🔀 Matrices in Transformers: Preface

onekq

•

Dec 5, 2025

posted an update 8 months ago

Post

236

DeepSeek v3.2 is worse than R1. This is quite puzzling. Why the regression with new GRPO and new attention?

onekq-ai/WebApp1K-models-leaderboard

I used reasoning mode against DeepSeek API

posted an update 8 months ago

Post

260

Hard-earned lessons to land your agent (some mine, most learned from others)

1. Clarify expectations. what do you mean by automating emails? auto drafting? replying via templates? extracting details into json?

2. Get access to your customer's corp/prod environment. Guest or sandbox won't cut it, much less your demo account.

3. Don't expect your agent to be turn-key. It will take at least a quarter to stabilize, if your customer actually uses it.

posted an update 8 months ago

Post

254

Claude Opus 4.5 didn't make SOTA either. So many models are stuck at 0.75 now

onekq-ai/WebApp1K-models-leaderboard

posted an update 8 months ago

Post

393

The second point re Ilya post is about RL pain point, i.e. sparse reward. I'm optimistic on this front.

Our actions are driven by unspeakable instincts, which left no traces in training set (pretraining or synthetic). These process rewards (motion sensing, vision etc.) help you master new skills quickly, like biking. Outcome reward only (falling off the bike) is indeed too sparse.

But lots of tasks can benefit from outcome rewards alone. Many latest RL works to upgrade SQL skills use success-failure reward only, with executable as optional reward.

Additionally, scale is the secret sauce for models to surpass humans. A human agent can learn a task quickly, but is capacity limited. But a model agent can process tasks in the scale of many human lifetimes. This made up for the inadequacy of process rewards.

Many such tasks happen to be economically viable, i.e. salary-making jobs.

posted an update 8 months ago

Post

279

Ilya's interview has been widely cited. I won't address meta points but share 2 cents on two mundane issues.

I will start with the leaderboard phenomena. This is a feature, not bug. Model training is a project under founder mode. But still like all projects, it needs north stars. And you guess right, (famous) leaderboards are the north stars.

For those startups which found PMFs, many maintain their own proprietary leaderboards/benchmarks condensed from user traffic. The path is blocked on both directions: startups won's share their moats, model makers won't prioritize either.

So instead of complaining, we should celebrate that our prompts work (most of the time)

posted an update 8 months ago

Post

330

Grok 4.1 didn't make SOTA, but improves a great deal over 3.
onekq-ai/WebApp1K-models-leaderboard

Members of the 70% club are the 4 big players (GPT, Claude, Gemini, Grok) and Kimi.

posted an update 8 months ago

Post

260

No SOTA from Gemini 3 either 😖

onekq-ai/WebApp1K-models-leaderboard

posted an update 8 months ago

Post

291

If RAG (by that I meant vectors and embeddings) transitions from QA to agents, is scalability (from wikipedia to personal memory) still an issue? What will be the new challenges?

Anyone care to share experience?

Yi Cui

AI & ML interests

Recent Activity

Organizations

onekq's activity

WebApp1K Models Leaderboard

🔀 Matrices in Transformers: Preface