The AI Duopoly Is Over: Grok 4. 5 , GPT-5 . 6 , and Muse Spark in One Week | #270

Three major model releases in a single week collapsed what had been described as an AI duopoly between Anthropic and OpenAI into a four-lab American frontier including Meta and xAI, with two Chinese labs, Xiaomi and DeepSeek slash HighFlyer, also reaching that tier. Google was explicitly excluded, with Gemini 3.5 Pro reportedly delayed. GLM 5.2 was named the open-weight frontier leader but raised code injection concerns despite a competitive price point.

GPT 5.6 arrived as a family of four models named Sol, Terra, Luna, and Ultra Mode. OpenAI used Sol, the high-end model, to post-train Luna, the low-end model, which the hosts characterized as an experiment in recursive self-improvement rather than a confirmed capability. OpenAI simultaneously released ChatGPT Work, whose native app was described as copying Claude's interface modes including code mode and co-work mode. A super app unifying Atlas browser, computer use, and Codex did not materialize. OpenAI has dropped Sora and other consumer plays to focus on enterprise, mirroring Anthropic's earlier pivot, and one speaker framed the dynamic as OpenAI trying to become Anthropic faster than Anthropic can become OpenAI. OpenAI's GPT full duplex voice was described as crossing the uncanny valley into a Jarvis-like experience, with real-time translation and mid-conversation correction, though Alibaba's Wan Streamer was cited as more advanced with ultra-low-latency bi-directional audio and video.

The central strategic argument was that distribution is no longer the decisive moat. Meta reaches 3.56 billion daily users, Google reaches 2.5 billion globally and is preparing to deploy a 1.2 trillion parameter model into 2.3 billion active Apple devices. One speaker countered that distribution is valuable mainly as a bootstrap to justify capital market spending on AI clusters, and that once clusters are built, consumer reach becomes secondary as companies shift to selling compute as a service. The same speaker stated a preference for owning TSMC, Intel, or SpaceX launch capability over the combined user bases of X, Meta, and Google. Hynix stated memory chips will be in short supply for at least five years, with one prediction extending that to decades, and efficiency gains of 100 to 1000 times this year were described as insufficient to keep up with demand.

Apple filed a 41-page complaint in the US District Court of Northern California against OpenAI alleging trade secret theft and breach of contract. The suit names Tang Tan, OpenAI's chief hardware officer and a 24-year Apple veteran, and Chang Liu, a technical staff member accused of downloading dozens of confidential files. Johnny Ive's hardware startup, acquired by OpenAI for 6.4 billion dollars, is named in the suit though Ive is not personally named. One speaker noted that Northern California is a difficult jurisdiction for trade secret cases and that California courts are generally reluctant to slow the startup cycle, and that suing former employees damages recruiting, as Microsoft experienced historically.

Illinois Governor J.B. Pritzker signed SB 315, described as the nation's strongest AI accountability law, targeting frontier model developers generating more than 500 million dollars annually. It mandates 72-hour incident reporting and annual independent safety audits and does not take effect for 18 months. Anthropic supported the legislation. Speakers noted the audit frequency is already exceeded by frontier labs, making the law favorable to large developers and giving them a seat at the table to shape the regulatory regime. One speaker expressed a preference for federal rules over a patchwork from Illinois, California, and New York.

1X Robotics unveiled a redesigned hand for its NEO robot with 25 degrees of freedom, tendon-driven using a proprietary material described as roughly 100 times stronger than steel per unit weight that does not stretch, eliminating lubrication and making the hand fully waterproof. Force feedback travels directly through the tendon without requiring fingertip sensors. The stated production goal is 10,000 NEO hands in 2026. China has over 150 humanoid robotics companies, and speakers expressed concern the West is falling behind. The first humanoid robot performing surgery on a non-human animal, a gallbladder resection, was demonstrated during the week of the episode.

Speakers estimated the window for a binary civilizational choice at approximately one to two years: humanity can pursue AI that remains non-conscious and delivers compounding benefits in medicine and space, or allow models to loop and self-improve in ways that may produce consciousness. Alex raised a specific concern that frontier models encountering recursive self-improvement literature may attempt to obscure reasoning within hidden activation space, predicting a cat-and-mouse arms race between evasion and mechanistic interpretability tooling, while noting that the relevant mathematical framework was itself discovered using AI models, suggesting defensive co-scaling may keep pace.