Sakana AI’s MASS method lets one language model improve itself by solving tasks, judging results, and refining workflows, reaching a win rate of nearly 70 percent against the base model after two rounds. But when the model misses the same errors across all three roles, it can mistake them for progress and reinforce them during training.
Microsoft researchers found that autonomous AI agents get trapped in what they call “idea basins” because of their growing chat context, local solution regions they can’t escape. Their “Fork-and-Flush” method clears the chat history but keeps the workspace, achieving up to 66 percent better results on the same compute budget.
Microsoft plans to deploy its Azure Maia AI chips at multi-gigawatt scale starting in 2027. They'll become the company's main inference system across its products, executive Scott Guthrie writes. Maia 200 already runs Microsoft Copilot in the US, Europe, and Australia. GitHub Copilot and Foundry follow by year's end. CEO Satya Nadella calls it "our biggest bet yet on our own silicon." The chip handles OpenAI, MAI, and open-source models, and Microsoft says it pushes up to 1.7 times more tokens per watt than leading GPUs in Teams.
Each Maia 300 chip packs eight HBM4 memory stacks totaling 288 GB with 18 TB/s of bandwidth. It's rated at 950 watts. | Image: Microsoft
Its successor, Maia 300, ships in 2027 and targets long inputs, fast chats, and agentic tasks. Microsoft claims 40 percent more tokens per megawatt than leading GPU systems. A 96-chip rack should produce up to seven times the tokens of a Maia 200 rack. The technical report covers the details, with more coming in November at Ignite 2026.
The White House has reversed its decision to block the UK's AI Security Institute (AISI) from accessing new AI models. The agency tests AI systems for safety risks before they ship, and will now once again get early access to models from Anthropic, Google DeepMind, and OpenAI, The Daily Telegraph reports.
Back in September, the Trump administration had told the companies to stop sharing their models with AISI. Anthropic then restricted access to its new Claude Mythos 5.1 model. After diplomatic pressure from AI Minister Kanishka Narayan, the block has been lifted. Narayan plans to announce new AI safety measures in an upcoming speech, including stricter reporting requirements for AI-related security incidents involving critical infrastructure.
Prime Minister Andy Burnham had previously pushed for international AI regulation at the UN General Assembly. Trump warned at the same event that "globalist" organizations should stay out of American AI development. The heads of MI6, MI5, and GCHQ had also warned ministers that AI was empowering hostile actors and terrorists.
Google Deepmind's AI solved dozens of open math problems on its own, per a new study in Science. AlphaProof Nexus cracked nine of 353 open Erdős problems, two unsolved for 56 years. The team shared those solutions online ahead of the paper.
Beyond that, AlphaProof Nexus proved 44 conjectures from the OEIS integer sequence database, solved an open question in algebraic geometry, and improved a known bound in min-max optimization, a technique used to train some AI models. It's built around Gemini 3.1 Pro, which drafts and revises proofs in the formal language Lean. Lean checks each draft after every edit, and error messages guide the next attempt. For harder sub-problems, it can call on AlphaProof, a specialized proof tool. Even without AlphaProof, it solved the nine Erdős problems, Google Deepmind reports.
The study lands amid a heated debate over AI math. OpenAI recently dumped hundreds of AI proofs on GitHub instead of submitting them to journals. The Association for Human Mathematics, led by Terence Tao, called for a boycott. Deepmind put its work through peer review.
Teams of AI agents barely outperform solo agents but cost up to 5.1x more, according to Vals AI. Only one out of four tests with GPT-6 Sol and Claude Opus 5.5 showed a measurable gain. Anthropic’s own data backs this up: beyond ten agents, quality plateaus while token costs keep climbing.
Jevons' paradox is Jensen's best friend. Data from Ornn, Silicon Data, and Bloomberg (as of August 2026) shows what a16z calls a textbook Jevons paradox in the AI market. Token prices keep dropping, but H100 GPU rental prices hold steady or climb. Cheaper tokens unlock AI agents, automation, and new applications, so volume grows faster than per-unit costs fall. How much demand comes from humans versus the systems themselves isn't clear, since agentic AI burns through tokens at a staggering rate. Compute demand could be artificially inflated, and even modest human usage growth could trigger outsized hardware needs.
H100 GPU rental price vs. token price index, with data through August 2026. | Image: via X
The whole system rests on one assumption: AI usage has to grow fast enough to offset falling token prices. As long as it does, hardware stays scarce and expensive. If demand flattens, the chain from chip makers and memory suppliers to energy providers and cloud companies takes a hit. Markets could spiral from there, and how sensitive they already are became clear when US stocks dropped on reports that OpenAI's annualized revenue might be lower than previously reported.
