A Microsoft researcher built a working neural network using goats in Age of Empires II's map editor to critique AI research methods. His analysis of 315 papers found over 50% presuppose language models have human-like traits before the experiment begins.
Amazon, Nvidia, and AMD invest $310 million in Odyssey ML, a 3D world model startup valued at $1.45 billion. IQT fund and Google's Jeff Dean join the round. World models are emerging as the next major AI bet after language models.
Zhipu AI releases GLM-5.2 under MIT license with stable 1-million-token context. On FrontierSWE benchmark for long-duration coding tasks, the open-source model trails Anthropic's Claude Opus 4.8 by just one percentage point. Significant gap remains on reasoning versus closed-source rivals.
Researchers from Nvidia, Carnegie Mellon University, and UC Berkeley use AI coding agents to teach robots dexterous grasping in real-world conditions. A fleet of eight robots achieves 99% success rate on complex tasks.
OpenAI researchers propose a method to predict how often a new AI model will make mistakes after release. This approach could fill gaps left by standard safety testing.
Per Epoch AI analysis, Microsoft, Amazon, Alphabet, Meta, and Oracle are growing AI infrastructure spending at ~70% annually while operating cash flow rises only 23%. Spending could exceed cash flow by Q3 2026. Several hyperscalers are already pursuing outside funding.
Microsoft is considering a fine-tuned version of DeepSeek V4 as a cheaper model option for Copilot Cowork. The company is also switching to usage-based billing, with Copilot head Charles Lamanna stating flat-rate pricing is unsustainable.
A Berlin court ruled that Google's AI-generated summaries are a 'new search result format' with no decisive Google influence over content. A perfume company sued because AI Overviews displayed its brands alongside counterfeit products. The ruling partly contradicts a Munich decision holding Google liable for false AI responses.
SpaceX acquires Anysphere (creator of Cursor) for $60 billion, two days after its IPO. Goal: strengthen xAI to catch up with Anthropic and OpenAI in the AI model race.
US Justice Department invokes national security to defend xAI's unpermitted gas turbines in NAACP lawsuit, claiming Grok chatbot is essential to military operations.
The Institute of the Estonian Language releases a benchmark measuring how susceptible AI language models are to Russian propaganda. No technical details or quantified results provided in the excerpt.
Anthropic scraps its unpopular billing overhaul for the Claude Agent SDK before launch. Third-party apps will continue drawing from regular subscription limits instead of separate credits.
DeepSeek raises 50 billion yuan ($7.4 billion) in its first external funding round, reaching a $50 billion valuation.
OpenAI spent $34 billion in the past year, significantly more than the previous year. No breakdown of cost allocation is provided.
US government officials accuse Anthropic of disregarding Trump's cyber directive and releasing Claude 3.5 Sonnet without approval. Talks are underway with the Department of Commerce, CIA, and science advisor Michael Kratsios regarding demands for unhackable LLMs.
Nvidia launches its first bond sale since 2021 to raise at least $20 billion. The move reflects a broader debt boom among AI giants.
AR scan data from Pokémon Go players trained Niantic's spatial AI models. This technology is now combined with a US defense contractor's software for GPS-free navigation.
The European Commission assesses implications of a US order forcing Anthropic to shut down Fable 5 and Mythos 5 globally. European researchers debate building homegrown foundation models versus securing contractual access. Building local infrastructure requires computing capacity, energy, and competitive providers Europe currently lacks.
Satya Nadella (Microsoft) warns that a small number of AI systems could capture all economic returns. He advocates companies build "token capital"—their own AI capabilities on internal data and proprietary learning loops—to avoid this concentration.
Mirage, a video world model from Microsoft Research, stores scene information in latent space instead of pixel-based point clouds. This cuts compute time and graphics memory while maintaining spatial consistency through long camera moves. Object tracking across segments remains unreliable.
Google Cloud launches Open Knowledge Format (OKF), a standard converting scattered documents into Markdown files with YAML frontmatter, making organizational knowledge portable for AI agents. The format formalizes the "LLM Wiki" pattern popularized by Andrej Karpathy.
KPMG published a report on AI in business containing fabricated case studies involving UBS, the NHS, and other organizations. Edward Tian (GPTZero) helped uncover the errors and warns of 'secondary hallucinations': false claims from trusted consulting firms spreading unchecked. KPMG has withdrawn the report.
AI coding agents like Claude Code find the correct file but miss critical lines within it. The SWE-Explore benchmark shows that without sufficient context, even the best fixes fail.
Amazon and five other companies reportedly alerted the Trump administration to security vulnerabilities in Anthropic's Fable model. The White House ordered the model offline via export control within hours, despite Amazon being one of Anthropic's largest investors.
Count Anything is an AI model capable of counting objects in any type of image via text prompt. It halves the error rate compared to previous systems, but still struggles with extremely dense objects and ambiguous terms.
Satya Nadella (Microsoft) warns against "token-maxing"—throwing frontier AI models at every problem. Marginal productivity gains must justify token costs. Paradoxically, he admits being a "token-maxer" himself and that it's "addictive."
Google Research's Gemini-SQL2, built on Gemini 3.1 Pro, achieves 80.04% accuracy on the BIRD benchmark for natural language-to-SQL conversion, significantly outperforming OpenAI and Anthropic. Google plans to integrate this technology into its data services.
Microsoft and three Chinese universities developed SkillOpt, a method optimizing instruction documents for AI agents using classical training principles. A simple Markdown file boosts GPT-5.5 by ~23 points on procedural tasks and transfers across models (Codex, Claude Code).
Anthropic's Claude Fable 5 achieves 88% accuracy on FrontierMath's hardest tier, versus 75% for OpenAI's GPT-5.5. Massive jump from Opus 4.5 (< 10% early 2026).
Meta faces internal AI costs reaching billions of dollars. Starting 2027, a token governance system via "AI Gateway" dashboard will limit consumption. CTO Andrew Bosworth criticizes "tokenmaxxing": token usage alone does not measure real impact.
Moonshot AI releases Kimi K2.7 Code, an open-weights model with one trillion parameters for programming. Underperforms GPT-5.5 and Claude Opus 4.8 on coding benchmarks but costs up to 12x less per token, offering better value for constrained budgets.
US government ordered Anthropic to disable Claude Fable 5 and Mythos 5 globally, citing jailbreak risks. Anthropic complies but disputes: vulnerabilities are minor and exist in GPT-5.5. Company warns precedent could halt all frontier deployments.
Anthropic surveyed 52,000 Americans on AI hopes and fears. 64% fear job losses, 56% worry about losing independent thinking. Daily AI users show less concern. Most people reject AI in their own workplace, even for tasks they believe it can handle.
OpenAI lets Codex users bank and manually trigger rate-limit resets instead of waiting for fixed schedules. Go, Plus, Pro, and Business plans get one free reset; Plus and Pro users can unlock additional resets by inviting friends.
Claude Fable 5 scores 64.9 points on the Artificial Analysis Intelligence Index and sets records on 5 of 10 benchmarks. Performance gain over Opus 4.8 is only 5.7% while token costs double. Safety filters with fallback routing further increase expenses.
Google and OpenAI separately exposed operations allegedly originating in China using AI for fraud and covert influence campaigns within days of each other. Google filed a joint lawsuit with the FBI. Both target US infrastructure and political debates.
Mistral AI is negotiating a funding round of approximately 3 billion euros at a valuation of around 20 billion euros to fund its European AI expansion.
Anthropic is throttling its new Mythos model for certain tasks while building apps that directly compete with its largest customers. Partners, customers, and investors are pushing back against this platform strategy reminiscent of Microsoft's approach.
OpenAI acquires Ona (formerly Gitpod), a German startup founded in 2020 specializing in AI agents and secure cloud development environments. The acquisition aims to enhance Codex's capabilities for long-running autonomous coding tasks.
Prometheus, Jeff Bezos' AI startup, raises $12 billion at $41 billion valuation. Launched last November with $6.2 billion seed funding, the company has no public product yet. Bezos deems sharing details premature.
Deezer releases a free AI music detector available to users on any major streaming platform to identify AI-generated songs in their playlists.
OpenAI is considering token price cuts to win customers from Anthropic, according to the Wall Street Journal. A price war is emerging between the two generative AI giants.
Anthropic publishes an essay and two policy frameworks. Dario Amodei calls for binding audits of frontier models and describes AI as a strategic weapon wielded by nation-states.
Anthropic reverses a policy that would have secretly throttled rival AI researchers. The company admits a 'wrong tradeoff' but other points of contention remain.
Google releases DiffusionGemma, a 26B-parameter model generating text via diffusion (noise → text) rather than token-by-token. Achieves ~1,000 tokens/sec on H100 (4× faster than comparable autoregressive models), but lower output quality. Positioned as experimental tool for developers.
Sam Altman tells OpenAI staff to expect an IPO "within the next year" but acknowledges a possible delay to 2027. He frames caution around self-improving AI as the reason, though Anthropic's stronger growth and imminent IPO may be the actual drivers.
Anthropic demonstrates that its Mythos Preview model converts security patches (Firefox, Windows kernel) into working exploits within hours for a few thousand dollars, requiring no specialized knowledge. Eight complete attack chains were finished before Microsoft's auto-updates reached a single device.
OpenAI is negotiating to lease a 10-gigawatt data center in Ohio, potentially backed financially by Nvidia, according to The Information.
Anthropic releases Claude Fable 5, first Mythos-class model. Leads benchmarks (SWE-bench Verified 95%), but costs 2× more than Opus 4.8 (10-50$/M tokens). Strict safety filters block 9% of requests; mandatory 30-day data retention policy.
Germany's National Security Council approves establishment of an AI safety institute (DE-AISI) modeled on the UK's AISI. It will test frontier models from Anthropic and OpenAI for security risks. The EU remains dependent on US and Chinese AI technologies.