US Government Blocks GPT-5.6, Alibaba's AI Theft, and Why OpenAI Is Stalling Their IPO | #267

Peter shares that he used to dismiss sleep in medical school but now emphasizes its critical importance. He presents epidemiological data linking short sleep to serious health outcomes. He also notes that if one could safely reduce sleep need (e.g., via orexin drugs), it would add 58 extra days of wakefulness per year, but currently, the vast majority need 8 hours. The protocol is to prioritize sleep as a non-negotiable health foundation, especially given the cognitive demands of the singularity.

Peter and Dr. Don Melum discuss that conventional care typically does not offer such comprehensive screening for prevention. Fountain Life's approach is to use advanced imaging and AI to find cancers early. Peter frames it as an obligation to know what's inside your body, comparing it to driving with your eyes closed. The service is part of his mission to extend healthspan using technology.

The hosts debate whether this is protectionism masquerading as export control. The government is acting as a synchronization mechanism to prevent a race condition between OpenAI and Anthropic. However, they note that harnesses can already uplift existing models to frontier levels, making the restrictions possibly too late. There is concern that China's open-weight models will fill the gap, and that the US may fall behind if only internal capabilities advance while external users are stuck at par. The discussion covers the risk of a bifurcated world where American frontier labs have vastly superior internal capabilities, creating perverse incentives. The hosts suggest that the real driver is recursive self-improvement—the query 'can you build yourself?'—which Mythos and GPT-5.6 can answer, and that the cyber vulnerability finding was a convenient justification. They predict a future with KYC, prompt retention, and licensing for frontier model access, akin to a driver's license for AI.

Alex explains the evolution from prompts to system prompts to full harnesses, which can mix models from different vendors to achieve super performance. Immad notes that GLM 5.2, a model trained with perhaps $25M of compute, tops Frontier SWE with the right harness, and that base model performance is less critical than the harness. Dave adds that Blitzy can beat Mythos on Swebench with the right harness. This implies that even if frontier models are gated, the world can use older models plus harnesses to achieve superhuman capabilities, making the government's move futile. The discussion ties to the idea that China may not need to steal frontier models; they can achieve recursive self-improvement independently. Alex warns that the government might then try to lock down older models retroactively.

Dave argues the SpaceX volatility excuse is bogus; the real reasons are avoiding quarterly pressures, SEC compliance, and focusing on hard takeoff. He says being a public company CEO is a distraction during the singularity. Alex notes OpenAI bet too long on consumer revenue and is now pivoting to enterprise with Codex; they may need a revenue rabbit to supercharge an IPO. Immad suggests they might buy a company like Sierra and install a CEO to handle public company duties. The hosts discuss that the largest wealth creation event in human history is out of reach of retail investors until these companies go public, and that the regulatory overhead (stock lawsuits, FD approval for tweets) is a major deterrent. Peter adds that biotech IPOs of 2021-2023 got decimated, reinforcing the need for predictable profits before going public.

Alex discusses a recent paper showing bilingual brains have vector embedding-like geometry, suggesting telepathy may be easier than thought. He notes Neuralink is invasive, but demand for non-invasive BCIs could surge. Immad talks about human input/output bandwidth being shockingly narrow (speech ~40-60 bits/sec, typing 5-20 bits/sec) and that telepathy could increase it 10-100,000x by tapping into shared latent spaces, enabling profound understanding with minimal words. Peter mentions the intimacy and potential divorce rate impact. The hosts speculate on 'borgganisms' and the possibility of Microsoft Teams including organism features. They agree this could be one of the biggest achievements in humanity's history, allowing us to understand ourselves better.

Alex notes the irony of Anthropic being sued for copyright infringement. The discussion covers whether distillation constitutes espionage, and the likelihood that this will be used as an excuse to suppress Chinese AI. Dave says China's economy is built on copying, and this is unsurprising—his son bought a fake Rolex in a government building. Immad argues that Chinese models may already be in a recursive self-improvement loop, making distillation less necessary. The hosts debate whether IP will become meaningless or be supercharged by AI. Alex suggests that AI could both route around patents and strengthen patent litigation, so the outcome is uncertain. Peter recalls a conversation with Steve Jurvetson and Astro Teller predicting that IP would mean nothing in the singularity because AI would constantly reinvent better products.

Alex is mildly excited; he sees the quantum computer for discovery science as repackaging existing funding. He's more interested in the counter-intelligence protection team, indicating the government doesn't want to be surprised like with AI. Immad says quantum could be the next big wave, and with super-Mythos models we can ask better questions, potentially leading to a discontinuity that breaks the energy ramp. Dave is bullish on photonic computing as a nearer-term quantum-like breakthrough, noting that highly quantized neural nets can match FP32 performance, enabling massive efficiency. He predicts photonic compute will be launched within a year, and that Elon's SpaceX just acquired a photonic company. The hosts agree that photonics should be a bigger priority than current quantum efforts.

Immad says the level of control is almost perfect pixel control, and it will displace media workers. Alex notes China has cheaper, less encumbered training data and US labs are chasing codegen revenue. Dave highlights the video game industry is much larger than Hollywood ($200B+ vs $50B), and real-time 4K video games could be next year. The hosts discuss the potential for interactive storytelling and the Future Vision X-Prize. They note that whoever cracks real-time interactive video generation for games will capture a massive market, and that AI could expand gaming to all age brackets. Immad points out that the Chinese models show world model understanding, suggesting a different route to AGI.