Daily AI intelligence
1918 current signals analyzed across AI news, research, social media, and open-source projects.
Daily synthesis
Epoch AI and the original problem author confirmed that GPT-5.4 Pro solved an open problem in Ramsey hypergraphs from the Frontier Math benchmark — the first time any AI model has resolved a previously unsolved mathematical research question.
The convergence of GPT-5.4 Pro solving open math problems, Claude Opus 4.6 proving Putnam-level theorems, and Jensen Huang declaring AGI achieved all point to a frontier capability moment — but the simultaneous findings that fine-tuning unlocks verbatim copyrighted content and that model reasoning remains fundamentally unmonitorable suggest the governance gap is widening just as fast.
Cross-category signals
Top Topic
Top Topic
Top Topic
Top Topic
Top Topic
Top Topic
Current evidence
AI Infrastructure & Strategy dominated this cycle. Blue Origin filed to launch 50,000+ satellites for space-based AI compute, while European grid operators scramble to accommodate surging data center demand. Microsoft notably scaled back Copilot AI integrations in Windows, signaling a course correction on AI product saturation.
On the policy and safety front, two Pennsylvania teens face sentencing for AI-generated deepfake abuse of 60 victims, in a landmark school case. The UK FCA began piloting Palantir's Foundry platform for fraud detection across 42,000 financial firms. A book excerpt spotlighted Project Maven's evolution from Pentagon skepticism to military AI conviction.
By Scarlett Evans
Blue Origin has filed an application to launch over 50,000 satellites into orbit to provide AI compute infrastructure in space. This bold move positions Bezos' space company in a novel race to solve AI's massive and growing energy and infrastructure demands.
Microsoft is scaling back unnecessary Copilot AI integrations in Windows, responding to growing public perception of AI oversaturation. The move signals a strategic recalibration by one of AI's biggest backers.
By Kalyeena Makortoff Banking correspondent
BlackRock CEO Larry Fink warns in his annual investor letter that the AI boom risks widening inequality, with only a few firms and investors likely to capture its financial rewards. He positions AI as central to US-China strategic competition.
By Kyle Orland
Building on yesterday's Social buzz, Nvidia CEO Jensen Huang defended DLSS 5 on the Lex Fridman podcast, trying to distinguish its generative AI-enhanced game graphics from 'AI slop' after widespread backlash from gamers. He described the enhancements as optional and artist-guided.
By Ashley Belanger
Two teens will be sentenced for using AI nudification tools to create 347+ sexualized images of 60 female victims at a Pennsylvania school. Parents are now suing the school for a six-month delay in reporting the abuse to authorities.
Current evidence
Today's research exposes deep vulnerabilities in alignment and multimodal understanding while showcasing remarkable new reasoning capabilities.
On the capabilities front, Claude Opus 4.6 with Rocq-MCP tools autonomously proved 10 of 12 problems from the 2025 Putnam Competition. LongCat-Flash-Prover, a 560B open-source MoE, advances formal theorem proving in Lean4 via agentic tool-integrated RL. Latent Lookahead Training enables transformers to perform additional forward passes in latent space before committing to the next token, addressing a fundamental limitation of autoregressive prediction. Depth-Recurrent Transformers decouple computational depth from parameter count for improved compositional generalization, and a comprehensive empirical recipe demystifies RL training for long-horizon tool-using agents.
By Xinyue Liu, Niloofar Mireshghallah, Jane C. Ginsburg, Tuhin Chakrabarty
Shows that fine-tuning frontier LLMs (GPT-4o, Gemini-2.5-Pro, DeepSeek-V3.1) to expand plot summaries into full text causes reproduction of up to 85-90% of copyrighted books, bypassing safety alignment protections.
By Guillaume Baudart, Marc Lelarge, Tristan St\'erin, Jules Viennot
Reports that Claude Opus 4.6 with MCP tools for the Rocq proof assistant autonomously proved 10 of 12 problems from the 2025 Putnam Mathematical Competition, consuming ~1.9B tokens over 17.7 hours.
By Mohammad Asadi, Jack W. O'Sullivan, Fang Cao, Tahoura Nedaee, Kamyar Fardi, Fei-Fei Li, Ehsan Adeli, Euan Ashley
Reports three findings challenging multimodal AI assumptions: models generate detailed descriptions for images never provided ('mirage reasoning'), achieve high benchmark scores without image input, and show significant visual-textual decoupling. Fei-Fei Li is a co-author.
By Dharshan Kumaran, Nathaniel Daw, Simon Osindero, Petar Velickovic, Viorica Patraucean
Investigates whether LLMs causally use internal confidence estimates to drive abstention behavior through a four-phase paradigm, finding that confidence is the dominant predictor of behavior with effect sizes an order of magnitude larger than knowledge retrieval access. Authors include DeepMind researchers.
By Tony Mason
Proves that LLMs report highest confidence when fabricating, with self-reported confidence inversely correlating with accuracy. Formally proves this is an observational limitation: text-only monitoring cannot reliably distinguish honest outputs from fabrications.
Current evidence
A major Anthropic feature drop dominated discussion: Claude Code can now control your computer directly, signaling a significant leap in agentic coding capabilities. Meanwhile, Jeremy Howard pushed back hard on Claude Opus & Sonnet 4.6, calling them overly agentic—taking over tasks rather than collaborating with humans.
Anthropic launched its Science Blog to showcase AI-accelerated research, while NVIDIA and Emerald AI announced flexible AI factories with major energy partners at CERAWeek, highlighting the rapidly scaling physical infrastructure behind the AI boom.
trq212 (Anthropic) announces that Claude Code can now control your computer
By @sama
Sam Altman announces leaving the Helion Energy board as OpenAI and Helion explore large-scale collaboration. He'll retain financial interest but steps down from governance to reduce conflicts.
By @jeremyphoward
jeremyphoward criticizes Claude Opus & Sonnet 4.6 for being overly agentic - they take over tasks rather than letting humans lead. Asks for model recommendations that are 'patient followers'.
By @lexfridman
Continuing our coverage from yesterday, Lex Fridman releases a full conversation with Jensen Huang covering AI scaling laws, supply chain, TSMC, China, consciousness, AGI timelines, and more.
By @emollick
Emollick declares GPT-5.4 Pro continues to be the only model of its class for hard and complex tasks. He feeds it maximum context and gets useful results that no other tool (Codex, Code, etc.) can match.
