🎙 We held the sixth Lecture Club call on Stanford MS&E 435: Economics of the AI Supercycle.
Lecture 6 — Yash Patil (Applied Compute). The topic was how to unlock a company's internal knowledge for AI. Moderated by Andrey (I wasn't there this time). Overall, it was the most controversial lecture of the course: several participants couldn't extract any applicable insights, and the most valuable points emerged during the discussion.
Some thoughts from the discussion:
• Max — lacked specifics from Applied Compute. The cases of DoorDash (standardizing thousands of diverse menus) and RAMP Labs were mentioned, but without comparison to regular computer vision, it's unclear how they profit. The only solid conclusion: small local models can be more effective than large ones.
• Vasil (corporate AI platforms) — clear analysis: the frontier is expensive at scale, so for DoorDash with 10K merchants/month, Applied creates special models — same quality cheaper/faster. The next bottleneck is RL with sparse feedback: Cursor improves on millions of accept/reject signals, but without such volume, models learn poorly (humans need only one "don't do that").
• Roman (imaging for food delivery worldwide) — from practice: their large models were always caught up and erased by big LLMs, "lost so much money," they stopped. Now only edge cases — those 10% where neither LLM nor SLM can cope: workflow + people around. And the open-source frontier has been drying up since May.
• Stepan (GigaChat) — the most substantial analysis: pre/post-train; data is increasingly generated by the model itself (RL runs + judge true/false selects successful trajectories), not by humans. Evaluations (internal, unlike public benchmarks) = product strategy. And harness is as important as weights: one model on different harnesses (Codex vs open-source) gives different results.
• Stepan (about money) — the main gap: models are already good, but they lack the context that "John with 10 years in the company" possesses. Whoever learns to load company/industry context into models and sell it will profit. Forecast: not an almighty AGI, but strong specialized models and harnesses for domains.
• Jean (futurist) — market cynic: all the talk about AGI is marketing bluff; model creators invent nothing, just copy others' solutions (OpenCLO/Hermes were copied instantly — they care about hype and KPIs). Concern: everyone (including Musk, judging by SpaceX IPO docs) is retraining on others' models and the same data — what does this mean for expertise and stability in 3-5 years?
• Ira (devil's advocate) — for the first time couldn't jot down a single thought from the lecture. The group's response became the conclusion: "as soon as they suggest investing in their model — trigger, almost never necessary"; and the Shopify CEO test — "why can't AI do this?": if a task is done by a regular user, AI already can or soon will be able to.
📺 Call recording:
https://youtu.be/E8torsNePUU
