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