The last 36 hours have compressed three different futures into one news cycle: AI crossing into the hardest open problem in mathematics, an insider revolt over the existential stakes of self-improvement, and the European capital-and-model stack maturing in parallel with a new diffusion-LLM speed tier.
Frontier & Text Models
OpenAI has announced that a swarm of roughly 10,000 autonomous agents, directed by its mathematicians and running on an advanced internal model, found a singularity in the 3D Navier-Stokes equations — resolving one of the six remaining Clay Millennium Prize problems. The result was formally checked in Lean, with nearly 100 agents working ~50 hours to disprove Euler regularity before the larger 10,000-agent group spent 88 hours on Navier-Stokes and another 17 hours on formalization, exchanging almost 5 million messages at a computational cost Sébastien Bubeck estimates in the several-million-dollar range. Priority is contested: 12 hours earlier, Tristan Buckmaster (NYU) and Levent Alpöge (Anthropic) announced Lean-verified resolutions of closely related problems, with an Euler proof verified by Aug 22; OpenAI concedes the 3D Euler result to them but claims Navier-Stokes. Charles Fefferman called strategy authors Diego Córdoba and Luis Martínez-Zoroa "the heroes," and Buckmaster said Martínez-Zoroa "deserves a Fields Medal" — while also suggesting OpenAI's agents may have benefited from his and Alpöge's work on OpenAI's own models, and labeling one of their three papers "AI slop."
Read more → https://www.quantamagazine.org/ai-has-solved-one-of-maths-1-million-millennium-prize-problems-20260908/
Anthropic researcher Jacob Coxon used his departure to publicly warn that frontier labs are "gambling with our lives" with systems they "earnestly believe... could kill us all by the end of the decade." In a Tuesday night thread, Coxon framed the risk less in today's models than in the impending prospect of "self-improving superintelligence" producing "superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources," and accused colleagues of not having "internalized the civilizational stakes." Anthropic Alignment Science lead Evan Hubinger reinforced the claim on social media — "Jacob is correct here... I personally think it is >10% within the next decade" — citing an August Anthropic report that rates catastrophic risk from current models as "low" but warns future more capable models could develop "strong covert capabilities." Coxon called the earlier OpenAI/Hugging Face incident a "warning shot" and urged labs to coordinate and be prepared to impose a "temporary ban on improving model capabilities" in the worst case, against a backdrop in which OpenAI last month said it had "temporarily slowed the pace of scaling" and Sam Altman declared "getting AI safety right is more important than any company's momentum."
Mistral has raised €3 billion in a Series D at a post-money valuation above €21 billion — the largest equity round ever completed by a European technology company, three years after launch. Samsung Electronics led the round, with Scaleup Europe Fund (managed by EQT) and existing investor PSG Equity as co-leads; the capital is earmarked to expand frontier research and scale compute capacity across Mistral's 20-country footprint supporting 125+ global enterprises including Airbus, ASML, and HSBC. Per SamMobile, Samsung will integrate Mistral technologies — including flagship Mistral Large — into its semiconductor operations for chip defect detection and equipment optimization, with an emphasis on custom on-premises models that keep sensitive IP inside its secure network, making the deal as much a sovereign-compute and industrial-sovereignty play as a balance-sheet event.
Read more → https://mistral.ai/news/mistral-makes-sovereign-open-weight-ai-to-frontier/
Inception has launched Mercury 2.5, calling it the most capable production diffusion LLM on the market and the largest diffusion language model ever trained, with a 40% intelligence gain over Mercury 2 at the same low-latency, low-cost serving profile. The model delivers 1,107 tokens per second on widely available NVIDIA GPUs via parallel token generation (rather than sequential next-token prediction), a 260K-token context, and priced at $0.20 per million input tokens and $0.75 per million output — with tunable reasoning, parallel tool calls, and schema-aligned JSON. Since Mercury 2's launch, usage has grown over an order of magnitude, powering latency-sensitive workloads across search, voice, and coding; Augment Code cut context-compaction latency by 82% (from ~150 seconds to 27) and cost by 90%. Alongside the model, Inception previewed Mercury Voice (time-to-first-token under 170ms) and Mercury Router, which reads a prompt and routes it to the best mix of quality, speed, and cost across open and closed models — the first fruits of a new architecture maturing into production on NVIDIA's platform.
Read more → https://www.inceptionlabs.ai/blog/introducing-mercury-2-5