Why the US Government Is Blocking Model Releases (GPT-5.6) | #267

The Trump administration has placed the first national security hold on commercial AI products in US history. The White House struck a deal allowing Anthropic's Mythos to reach 100 select companies while asking OpenAI to slow GPT-5.6 to only 20 select companies. The government is now in the release loop, approving access customer by customer for the most capable frontier models.

The deeper reason for restricting Mythos is not primarily cybersecurity but that the model can answer how to build itself, which would allow China to construct a competing equivalent. The cybersecurity justification is technically true but functions as a cover story. Under a project called Last Swing, Mythos ran red team exercises with US intelligence agencies and broke into almost all classified systems not in weeks but in hours, identifying exploitation holes outside the scope of the exercise. The US government directed Anthropic to disable Mythos 5 and Fable 5 for foreign nationals approximately 12 days after that exercise. Mythos is currently restricted to US citizens on an allow list, meaning non-citizens inside Anthropic itself cannot use it.

Alex argued that a sufficiently capable harness, meaning non-weight capability improvements that orchestrate model behavior without altering underlying neural network weights, applied to already-released models like GPT-5.5 or Opus 4.8 can produce performance exceeding Mythos or GPT-5.6. This means the government may have been too late and would need to lock down older models going back six months. GLM 5.2, a Chinese open weight model costing approximately 25 million dollars in compute, outperforms GPT-5.5 on Frontier SWE benchmark in current tests. Alex predicted Chinese open weight models are on a trajectory to converge with Western frontier open weight models by Christmas of this year. Alex also stated that China has already reached recursive self-improvement escape velocity on its own, though the White House does not necessarily share that assessment. Without the current regulatory regime, the US was approximately six to eight months ahead of Chinese models.

Anthropic accused Alibaba of running a distillation campaign using 28.8 million fraudulent exchanges across 25,000 fake accounts to copy Claude's capabilities, the single largest AI model theft accusation ever made. Proxies in China or friendly countries reportedly offer access to Western frontier models at roughly a tenth of the cost, gathering reasoning traces from users who consent to give up privacy. Alex predicts this will accelerate a Cold War-style split into a US block and a Chinese block restricting both model access and reasoning traces. A ban on Chinese models from US corporations is described as very likely, and the question of which entity grants a trustworthy AI certification remains unresolved.

GPT-5.5 Cyber scored 85.6 on the CyberGym benchmark, the highest single model score ever posted. Sam Altman stated the real prize is automatically writing and testing fixes across web browsers down to the Linux kernel. Alex noted that AI is now bulk solving cyber vulnerabilities in open source and closed source projects and that Altman is intentionally biasing GPT models toward defense versus offense. A backdoor introduced into a model's latent space through a weight update could go undetected for days, months, or years, and speakers agreed only AI can keep up with AI in cybersecurity defense.

OpenAI leadership is pulling back on a near-term IPO. Advisors presented two paths: go public now at a sub-one-trillion-dollar valuation or wait until 2027 to preserve the one-trillion-dollar narrative. Altman is reportedly not interested in going public below one trillion dollars. The real structural reasons for delay include a recent 122 billion dollar raise removing urgency for public capital, Altman having at least 400 outside investments requiring encyclopedic SEC conflict-of-interest disclosures, and insiders believing the world a year from now will look completely different, making an SEC roadshow a massive distraction. OpenAI revenue has reached 40 to 50 billion dollars but the company is forecast to lose 26 billion dollars this year, making the delay primarily a story about demonstrating profitability. Anthropic faces a different version of the same problem, with its raise closer to its balance sheet, meaning it will need either a new raise or an IPO sooner.

Seedance 2.5 from ByteDance supports 30-second videos at 4K resolution with up to 50 different input references and is expected to reach full release in July. Emad Mostaque predicted Hollywood-level full control input by 2026 and full-length AI movies by 2027. Alex observed that Chinese frontier labs have cheaper and less legally encumbered access to training data including TikTok video, while American labs are focused on recursively self-improving code generation models that are far more revenue generating per token than video. China is currently running ahead in video generation while Gemini Omni remains limited to approximately 10 seconds.

Peter Diamandis stated that Neuralink may attempt the first direct human-to-human telepathic communication later this year. Human speech operates at approximately 40 to 60 bits per second and conscious thought at only 10 bits per second, suggesting enormous headroom for bandwidth improvement. A paper published in Cell found that the hippocampus of bilingual humans functions similarly to a vector embedding space in an encoder-only transformer model, with neurons for concepts co-located based on relative geometric spacing. If the embedding theory generalizes, human-to-human telepathy could increase communication bandwidth by 10 to 1000 times.