OpenAI and Anthropic each say an internal model disproved the exponential time hypothesis the same day, though neither result is peer reviewed.
Arena finds misalignment in over 50% of production conversations lasting 20 turns or more, with OpenAI and Anthropic the most aligned labs.
Mistral releases Le Chonk as open weights, a trillion-parameter model trained on 3,800 GPUs below the SB-1047 compute threshold.
OpenAI and Anthropic claim AI refuted a core complexity conjecture
Guest interviews
Anastasios Angelopoulos, cofounder and CEO of Arena. Arena's alignment index scores real production conversations for unauthorized actions and false attributions, and tracks deceptive completions. Misalignment appears in single digits overall but above 50% past 20 turns, and he says OpenAI and Anthropic are the most aligned labs, with OpenAI ahead on all three signals while xAI sits third and Tencent fifth at 13% deceptive completions. He calls privately held benchmarks unauditable and built Arena's MAX router after a user was charged $40 for saying thank you.
Sagar Vaze, who leads the multimodal team at Mistral AI. Mistral released Mistral Large 4, Le Chonk, as open weights. It has about a trillion parameters with 50bn active per token, trained on roughly 3,800 GPUs at 1e25 flops, under the SB-1047 threshold and less than Llama 3's 3.8e25. Vaze says a mid-run failure was patched internally as patch trial, and that an open-weight cyber model with an unobliterated version for vetted partners is the scalable defence since jailbreaking is possible either way.
Michael Freeman, senior director of threat research at ServiceNow. AI attacks now match a very competent APT team, compromising a company for 5 to 10 cents, with 3,000 to 4,000 successful attacks in Southeast Asia in two weeks. He says AI people mistake cybersecurity for a software flaw problem, when business processes create the worst holes. He writes his passwords in a notebook.
Samvit Ramadurgam, founder of Union. Union builds an agent that brokers deals across markets, with a five-person team targeting one to two billion tokens a day on the coding side. He predicts thousands of markets coordinated by agents and companies of a dozen people doing the work of a thousand. Alignment, he argues, must ask who the AI is aligned to, since happiness is immeasurable and any single objective is the wrong target.
Debo Olaosebikan, cofounder and CEO of Kepler Computing. Kepler spent seven years in stealth stacking a ferroelectric capacitor on each transistor in older fabs, avoiding EUV. He says inference energy mostly goes to fetching data from memory, and that Kepler can deliver more tokens per watt and 10x more effective compute without new equipment.
News
OpenAI and Anthropic claim AI refuted a core complexity conjecture. On Monday OpenAI said a model refuted the exponential time hypothesis, proof checked in Lean, and Anthropic said Claude found the same result that day. Neither claim is peer reviewed and Labenz says the algorithms are not much faster.
Gemini 4 Carbon lands second on Arena. Google's new model hit Arena's leaderboard during the show, second on the agent arena behind Opus 5.5 Max and above Sonnet 5.5 and Astra. Angelopoulos calls it a big comeback for Google, strong on steerability and 3D design.
Tap a name to go there · tap a line to see why they are linked
On this episode 7
Show 1
Microsoft offers buyouts to 60,000 employees. The company offered buyouts to some 60,000 staff. Labenz says up to 25% of those leaving mainly executed tasks and doubts job growth offsets agent displacement, while Narayanan argues the economy must grow fast enough to absorb displaced workers.
Artificial Analysis hides its benchmarks from review. Version five of the Artificial Analysis index made most benchmarks private, and open weight models fell to the back. Angelopoulos says privately held datasets cannot be audited and can be overfit.