ZeroNoise Logo zeronoise
Post
Recommended Reading from Tech Founders
Martin Casado recommends an MIT talk for an intuitive grasp of LLMs
•
1 min read
• 77 docs
Vishal Misra’s MIT talk explains how SFT, RLHF and RL reshape a distribution while next-token prediction remains underneath; Martin Casado recommends it for understanding in-context learning and LLMs.

MIT talk on LLMs (video; descriptive title)

Resource/creator: Vishal Misra’s MIT talk on LLMs. Recommended by: Martin Casado, who calls it the best talk on in-context learning and says it builds an intuitive grasp of LLMs.

Key takeaway: Misra describes a first-principles account that skips attention and transformers: SFT/RLHF/R reshape the distribution, while the underlying LLM remains a next-token predictor. Why it matters: It connects post-training methods to the model’s underlying prediction process—the intuition Casado specifically recommends the talk for.

Martin Casado recommends an MIT talk for an intuitive grasp of LLMs
Research extraction

Decision: Neither item is established by the supplied sources as a sufficiently explicit, noncommercial learning recommendation. The YouTube capture contains generic site navigation, not enough to verify the talk’s exact title, creator, or substance. The RL excerpt identifies and describes a resource, but does not establish its license or noncommercial-use terms.

  • RL resource identity and contents: The resource is the Hugging Face dataset XiaomiMiMo/MiMo-V2.6-RL-oss. Its visible preview includes opensource-code/swe examples, an agent named mimo_swe_agent, and software-engineering task prompts.
  • The page describes “Agentic RL Environments” as RL training environments for LLM agents, listing code/software engineering, cyber/vulnerability reproduction, general/knowledge work, visual/web development, and music/symbolic composition tasks, with their respective verifier or grading types. The excerpt does not verify that the resource is noncommercial; its labels alone do not establish that.
- YouTube Datasets at Hugging Face
clem 🤗
x

MiMo-V2.6-RL-oss (open-source RL-environment dataset/repo on Hugging Face): Clem Delangue praised open-source RL environments and shared a post linking XiaomiMiMo’s repository. The linked post says comparable tasks typically cost hundreds to thousands of dollars per task and describes the repository as worth millions.

Open-source RL envs for the win. On Hugging Face of course! [https://x.com/harveenchadha/status/2103747414775804223](https://x.com/harvee… was looking for a quiet weekend but xiaomi dropped their rl envs repo last night to put in perspective, if you have to buy some tasks lik…
martin_casado
x

Martin Casado calls Vishal Misra’s MIT talk video the best talk on in-context learning and for building an intuitive grasp of how LLMs work. Misra describes it as a first-principles explanation of LLMs that skips attention and transformers and explains SFT/RLHF/RL as reshaping the distribution while the model remains a next-token predictor. Watch the talk.

Still the best talk on in context learning, and getting an intuitive grasp of how LLMs work. [https://x.com/vishalmisra/status/2103954166… Video of my MIT talk on LLMs: a first-principles view of how and why they work, without discussing attention/transformers. SFT/RLHF/RL re…