Open Models, Opaque Reasoning, and Compute Concentration Shape the AI Debate

Key Takeaways

1. Deedy reframes AI power as an offense/defense problem: frontier compute will be 65–75% concentrated in 6–7 US entities by 2030, but ~10x annual model price declines are already spreading formerly offensive capabilities. — via 1
2. LeCun urges open foundation models and accuses Anthropic of regulatory capture, saying test thresholds and audits protect its closed frontier at rivals' expense. — via 1 2
3. Hamel warns Opus 5.0 can code but cannot explain its own behavior; analysis shows longer, structurally more complex, simpler-vocab outputs versus earlier models. — via 1 2 3
4. Mollick says Google AI Overview is already changing how people consume and act on information. — via 1
5. A study cited by Mollick finds o3-mini-generated exam questions are psychometrically comparable to high-stakes standardized test items. — via 1
6. Hamel's team quietly shipped multi-agent v2 delegation to any supported model, including Luna. — via 1

1. Open Models, Safety, and Compute Concentration

1. In a recap of the Gavin Baker–Sholto Douglas debate, Deedy argues the right lens is offense/defense balance: defense-dominant AI domains should be open, while offense-dominant ones—such as low-cost synthetic lethal pathogens until the 2030s—should be controlled. He calls 'full distribution' vs. 'full restriction' a false dichotomy because frontier compute will be 65–75% concentrated in 6–7 US entities by 2030, yet model prices are falling roughly 10x per year, faster than compute growth, which has already pushed tools like propaganda and surveillance into wide use. — via 1
2. LeCun doubled down on open AI, arguing that technologies like the printing press and internet should augment human intelligence and that open foundation models are needed for a diverse AI ecosystem. He also endorsed Zuckerberg's criticism and warned against building AI safety on extreme centralization of power. — via 1
3. LeCun escalated his attack on Anthropic, calling its safety advocacy 'regulatory capture' and saying test thresholds and mandatory audits would raise costs for competitors while protecting Anthropic's closed frontier. He separately questioned the sense of slowing AI if it can cure diseases, argued safety tests can be gamed by models, and contended the proposals ultimately advantage Anthropic at the expense of ordinary users and competitors. — via 1 2

2. Model Opaqueness and Research Signals

1. Hamel contends Opus 4.6 may be the last 'English-speaking' Claude model; Opus 5.0 can code but cannot clearly explain its own actions, as if it communicates in an internal, obscure reasoning dialect. A quantitative writing analysis he shared shows Opus 5 outputs are noticeably longer, structurally more complex, but lexically simpler, while Fable 5 is terser than Opus 5. — via 1 2 3
2. Mollick said Google AI Overview, even in isolation, is profoundly changing both the nature of the internet and the information people consume and act on, and that this shift is already underway. — via 1
3. Citing a large field study, Mollick says AI-generated exam questions from an outdated o3-mini in an agentic loop have psychometric properties comparable to real high-stakes standardized tests. He argues AI-generated questions are not inherently a problem. — via 1

3. Adoption, Tooling, and Product Updates

1. Greg Brockman confirmed as accurate the claim that the top 1% of AI spenders spend roughly 600x more than the median company, highlighting a dramatic adoption gap between leading users and the rest of the market. — via 1
2. Hamel said the team quietly shipped an upgrade to multi-agent v2: models can now delegate tasks to any supported model, including Luna, adding flexibility to agent orchestration. — via 1
3. Mollick used Codex to bring the 1987 Infocom text game 'Nord and Bert Couldn't Make Head or Tail of It' back as a playable version. The project preserved the original puzzles, improved the interface, and used AI to help translate some puzzles better. — via 1
4. Hamel argues AI content watermarks will only encourage tools and APIs to strip them, so watermarking adds friction without meaningfully improving content visibility. — via 1