🎙 We held the fourth Lecture Club call on Stanford MS&E 435: Economics of the AI Supercycle.
Lecture 4 — Ali Ghodsi (CEO of Databricks). The topic was what AI is doing to the software industry: SaaS, agents, and why 'software is eating itself.' The general feeling of the group was that it was the weakest lecture of the first four, but the discussion turned out to be one of the best.
Some thoughts from the discussion:
• Nikita (me, lead) — The most striking moment of the lecture was the manipulative voting trick 'Is AGI already here?' ~10% always vote 'yes' — regardless of the definition. My definition of AGI: (1) superhuman in most abilities, (2) self-improvement without human intervention. By these criteria, we are still in the 'enhanced calculator' phase, not AGI.
• Ira — 'Every company has its own John, who knows things that no model does.' From her experience in a payment service: even within one company, context is divided by departments for security reasons. The best solution for the client is sometimes an unconventional legal workaround that only John knows. AI cannot connect to this context.
• Stepan — Context = the main bottleneck. Two paths: connectors (Databricks, Palantir AIP) and harness around the model (tools, search, agents). For Russia — a separate stack: 1C, Bitrix, AmoCRM, FNS, public services, banks. And an admission: 'Claude Opus already does code review better than me.'
• Mars — Thesaurus is important: in the 2000s AGI = reasoning + patterns (LLM fits). Over 20 years, autonomous goals and continuous learning have been added — by this definition, it's not here yet. And separately: Ali has a direct commercial incentive to claim AGI is already here — it lowers the sales threshold for his product.
• Alexander — Disagrees: for his use case, AGI is already here, with a record of 4 hours of continuous dialogue with Claude. Most of his AI implementation ideas are not hindered by technology, but by context. The future is many small specialized agents, not one super-agent. And a thought experiment: if AGI is 'sufficient' for SMB → training compute is over-invested and will flow into inference, prices will collapse.
• Boris — Tension between 'AI as a commodity' and safety: will frontier models be widely available as risks grow (Mythos found 26 vulnerabilities in Safari, bio-arms as a real concern)? AI as a US-China 'nuclear race' is also a factor. Parallel with the 2000s: multicast seemed critical — became irrelevant. Perhaps bottlenecks in chips/data centers will be resolved similarly.
• Alina — Practical case: in a year and a half distilled the expertise of investment fund teams into a multi-agent system. The teams are no more — what took a week now takes 30 minutes. Two consequences: (1) the bottleneck shifts from analysis to market reaction speed, (2) two parallel economies are forming — AI-driven and human-dependent. Implementing this, she felt like a 'slave trader.'
• Jean (futurist) — Two AGIs: (1) 'marketing' like 'nanometers' in chips — by this, Deep Blue was AGI back in 1997, (2) 'real' — requires consciousness, which computers definitely don't have. And even with 'AGI' today — 500 people in Sberbank can't agree on processes. The real task is not 'when AGI,' but how to enrich context between businesses without losing privacy.
• Nikita (closing) — Pessimism about unifying context. RTB in advertising started as an open system — ended with walled gardens of Google/Facebook. Transferring playlists from Spotify to iTunes is impossible. Claude literally two weeks ago could automate Facebook — now requires confirmation for every click because FB pressured Anthropic. The utopia of 'one AI knows everything about me' will not happen.
📺 Call recording:
https://youtu.be/9RO_wOTL6ug
