Shrishailya Chavan

Machine Learning Engineer & Founder
Models, data, constraints
Systems thinking over hype

[Shrishailya Chavan @Shrishailya_5]

@simplifyinAI We’re not replacing researchers, we’re turning research into a search problem.

[Shrishailya Chavan @Shrishailya_5]

@kiaran_ritchie Models commoditize. Ecosystems don’t.

[Shrishailya Chavan @Shrishailya_5]

@fchollet Right now, AI is great at compressing human knowledge, but it still depends on humans to define the search space. Autonomy starts when it can generate its own problems

[Mustafa. @oprydai]

pre-gpt era hackathons were fun

[Shrishailya Chavan @Shrishailya_5]

@classiclarryd Speedrun culture is quietly doing real architecture research. Tiny changes, big gains.

[Andrej Karpathy @karpathy]

I packaged up the "autoresearch" project into a new self-contained minimal repo if people would like to play over the weekend. It's basically nanochat LLM training core stripped down to a single-GPU, one file version of ~630 lines of code, then:

- the human iterates on the prompt (.md)
- the AI agent iterates on the training code (.py)

The goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement. In the image, every dot is a complete LLM training run that lasts exactly 5 minutes. The agent works in an autonomous loop on a git feature branch and accumulates git commits to the training script as it finds better settings (of lower validation loss by the end) of the neural network architecture, the optimizer, all the hyperparameters, etc. You can imagine comparing the research progress of different prompts, different agents, etc.

https://t.co/YCvOwwjOzF
Part code, part sci-fi, and a pinch of psychosis :)

[Shrishailya Chavan @Shrishailya_5]

@bravo_abad From blind screening to geometry-guided discovery. That’s a real paradigm shift.

[Shrishailya Chavan @Shrishailya_5]

@tunguz What’s missing isn’t implementation skill, it’s algorithmic taste. Choosing the right model is the hard part, and LLMs have no prior for that yet.

[Shrishailya Chavan @Shrishailya_5]

@random_walker My guess is inertia more than theory. Fixed priors are easy, but learned ones seem like the obvious win when you’re trying to cut evaluations.

[Shrishailya Chavan @Shrishailya_5]

@Algomancer Hard-coding N(0,I) feels like assuming flat terrain before planning a route. Why not learn the landscape first?

[Shrishailya Chavan @Shrishailya_5]

@bravo_abad If this is right, then ‘understanding’ is less about symbols and more about where you sit in representation space.

[Shrishailya Chavan @Shrishailya_5]

@denny_zhou This feels right. Intelligence scales, but institutions, interfaces, and incentives decide who actually benefits.

[Shrishailya Chavan @Shrishailya_5]

@fchollet People confuse ‘AI is great at code’ with ‘AI is great at everything.’ Verifiable domains are a cheat code, not a template.

[Shrishailya Chavan @Shrishailya_5]

@kalomaze nanoGPT taught half the internet how AR transformers think. A nanoViT-AE would finally teach us how they see.

[Shrishailya Chavan @Shrishailya_5]

Yeah, on the timeline everyone’s building AGI. In production, they’re still patching their RAG pipeline and wondering why embeddings drift.

[Peter Yang @petergyang]

Do people even still do RAG and fine-tuning anymore

[Shrishailya Chavan @Shrishailya_5]

@jerryjliu0 Right now it looks like RAG on logs with a cooler name. The real test is whether it enables reasoning you can’t get from plain trace retrieval.

[Shrishailya Chavan @Shrishailya_5]

@yoavgo @lateinteraction I had the same reaction. Great idea, but the missing REPL details feel like where all the real engineering lives.

[Shrishailya Chavan @Shrishailya_5]

@0xDevShah The moment agents get real protocols, we’ll see whether they remove bureaucracy or become a new kind of bureaucracy themselves.

[Shrishailya Chavan @Shrishailya_5]

@0xlelouch_ This is the part nobody teaches. Once you know how to reason from intent -> invariants -> failure modes, new systems stop being scary

[Shrishailya Chavan @Shrishailya_5]

@rohanpaul_ai Honestly this explains half the weird agent behaviors. If the model can’t store a private belief, of course it keeps changing the secret.