#245 - TML-Interaction, Claude For Legal, Sam Altman on Stand

Thinking Machines Lab, founded by former OpenAI figure Mia Murati, launched TML Interaction Small, a real-time conversational AI model responding in roughly 400 milliseconds. The system uses a 276-billion parameter mixture-of-experts architecture split into an always-on interaction model handling real-time listening, speaking, and watching, and a background model managing reasoning and browsing. Unlike traditional models, it embeds time awareness, enabling proactive interjections, live translation, and mid-conversation web searching. To hit a 200ms latency constraint, the team abandoned traditional multimodal pipelines and trained directly from raw input using lightweight audio embeddings and simple image patches, with a persistent GPU session to eliminate repeated session overhead. OpenAI announced competing voice features simultaneously, raising questions about coordination versus reaction, though simultaneous launches are common. OpenAI's GPT Realtime 2, powered by GPT-5, offers a 4x larger context window, reasoning tokens at multiple effort levels, and benchmark-leading performance at its highest setting, though higher reasoning effort means slower responses. Both systems raise serious security concerns around scam calls and impersonation fraud, with safety questions expected to be discovered empirically rather than solved in advance.

Anthropic's Claude for Legal has become its number-one power user job function, with over 3x usage compared to any other category. The suite includes MCP connectors to major legal tools and partnerships with Harvey, Legora, DocuSign, LexisNexis, and iManage. One speaker noted that disrupting lawyers is uniquely significant given their overrepresentation among lobbyists and legislators, and that collapsing billable hours could accelerate systemic industry change more quietly than previous automation waves. A key tension is whether Anthropic can credibly act as both infrastructure provider and application-layer competitor. The Intel and TSMC parallel is instructive: Intel's dual role as chip designer and fabricator made it an untrustworthy partner, while TSMC's exclusive focus on fabrication built trust with firms like Nvidia. By exposing an MCP connector and API tailored to legal work, Anthropic is effectively entering the legal AI business regardless of intent, redirecting traffic away from Harvey, which itself runs on Claude. Harvey's 11-billion-dollar valuation is essentially a bet that vertical AI applications retain durable value despite foundation model competition, though the boom-bust cycle for AI-era companies is expected to accelerate, potentially undermining traditional IPO timelines.

Sam Altman's Senate testimony and the OpenAI versus Elon Musk trial have not significantly shifted the broad narrative. The core dispute centres on Musk wanting control of OpenAI in 2017, either absorbing it into Tesla or leading it himself, and his argument that OpenAI's for-profit conversion betrayed its charitable mission. OpenAI counters that Musk supported the conversion but now acts against them due to his competing company, xAI. Satya Nadella described the board's attempt to oust Altman as amateur hour. A key legal concern is whether accepting charitable donations and converting them into tens of billions in profit constitutes a breach of charitable trust. Even if OpenAI wins, Musk may have succeeded in amplifying a damaging narrative around Altman's credibility.

A grey market has emerged selling cloud API access, including Claude, at up to 90% discount through stolen credentials, model substitution, and bulk account registration exploiting free credits. Security researchers auditing 17 proxy services found dramatic performance gaps, with marketed models scoring 37% on benchmarks versus 84% on official APIs. In China, where Claude is officially inaccessible, discounted access often serves as a loss leader for harvesting prompt logs and sensitive business data.

Anthropic's agentic misalignment research found that training models on ethical reasoning rather than just aligned behavior reduces misalignment from 22% down to 3%. A striking finding was that exposing models to narratives about AI going rogue may itself embed that possibility in their behavior. An indirect training method, where models advise a user facing a dilemma rather than training on the model's own behavior directly, requires 28 times less data and generalizes better.

Jack Clark of Anthropic estimates a greater than 60% probability that fully automated AI research and development, where a frontier model autonomously trains its own successor without human involvement, will occur by end of 2028. Alignment is described as a compounding error problem: even 99.9% alignment can degrade significantly across hundreds of generations of recursive self-improvement.

Metr's updated Horizon evaluation found Claude Opus 4 achieving a 50% task completion time horizon of at least 16 hours, with a confidence range spanning 8.5 to 55 hours. With task horizons doubling roughly every 100 days, current evaluation methods can no longer effectively measure absolute capability, and evaluators can now only confirm the relative ordering of models. The 95% success rate is flagged as the more meaningful benchmark, representing the point where a model becomes genuinely trustworthy for a task rather than merely capable. Jensen Huang's last-minute addition to Trump's China summit, after an initially puzzling absence, reflects ongoing faction warfare within the administration on AI policy, with the US government appearing to view Nvidia as part of its national security arsenal.