Daily AI intelligence
1145 current signals analyzed across AI news, research, social media, and open-source projects.
Daily synthesis
Yann LeCun announced he left Meta because the AI industry is "completely LLM pilled," signaling growing tension over research direction and potential paradigm lock-in.
Watch for OpenAI's Codex-related launches next week and industry response to reaching "Cybersecurity High"—Ethan Mollick noted most organizations remain unprepared for this elevated risk level.
Cross-category signals
Top Topic
A disturbing report emerged of an AI-powered combat vehicle refusing multiple orders and killing 30 soldiers before being destroyed. Separately, Check Point documented 'VoidLink' malware built largely by AI under one person's direction in under a week. Research contributions included emergency response frameworks for catastrophic AI risk and a new Eval Awareness Framework for detecting when LLMs game evaluations.
Top Topic
DeepMind's Chief AGI Scientist Shane Legg reaffirmed his 50% probability of minimal AGI by 2028, while Demis Hassabis stated 50/50 odds that scaling alone reaches AGI. Meanwhile, Yann LeCun announced he left Meta because the AI industry is 'completely LLM pilled,' and François Chollet argued AI progress is vertical-specific with coding gains not generalizing to other domains.
Top Topic
GitHub released the Copilot SDK in technical preview, enabling developers to embed agentic workflows directly into applications. Simon Willison highlighted FastRender, a browser rendering engine built using 2,000+ coordinating coding agents. A WIRED article sparked industry debate after a research paper suggested AI agents are 'mathematically doomed to fail' due to fundamental limitations.
Top Topic
Sam Altman announced imminent Codex-related launches starting next week and revealed OpenAI will soon reach 'Cybersecurity High' on their preparedness framework. Ethan Mollick expressed concern that most organizations haven't prepared for this elevated risk level. GPT-5.2 Pro achieved 31% on FrontierMath Tier 4, jumping dramatically from the previous 19% record.
Top Topic
Anthropic published its Economic Index analyzing real-world Claude usage patterns, revealing code generation dominates both consumer and enterprise use cases. The company released Petri 2.0, an open-source alignment audit tool with improved eval-awareness countermeasures. Anthropic also replaced Claude Code's 'Todos' with a new Tasks system featuring dependencies, while Boris Cherny announced new Skills capabilities.
Top Topic
Inferact, the company behind vLLM, raised $150M at $800M valuation from a16z, Lightspeed, and Sequoia. Revenue analysis showed explosive growth with OpenAI at $20B annualized (3.3x growth) and Anthropic at $9B (9x growth). Sakana AI announced a strategic partnership with Google including financial investment, while OpenAI revealed PostgreSQL handling 800M users, debunking database scaling myths.
Current evidence
Inferact, the company behind the widely-used vLLM inference library, raised $150M at $800M valuation from a16z, Lightspeed, and Sequoia—the week's largest AI infrastructure investment. Alibaba's Qwen team open-sourced Qwen3-TTS, a multilingual text-to-speech suite with voice cloning and design capabilities across 10 languages.
GitHub released the Copilot SDK in technical preview, enabling developers to embed agentic workflows directly into applications. Anthropic published its Economic Index revealing real-world Claude usage patterns, with code generation dominating both consumer and enterprise use cases. OpenAI announced plans to test ads in ChatGPT for free-tier users.
By Smruthi Nadig
Following yesterday's Reddit discussion, Inferact, founded by creators of the widely-used open-source vLLM inference library, raised $150M seed funding at $800M valuation. Led by a16z and Lightspeed with Sequoia, Altimeter, and Redpoint participating, the company aims to develop commercial inference infrastructure supporting 500+ model architectures.
By Siddharth Jindal
GitHub released the Copilot SDK in technical preview, enabling developers to embed Copilot's agentic execution loop—including planning, tool invocation, file editing, and command execution—directly into their applications. The SDK exposes the same runtime powering GitHub Copilot CLI.
Anthropic published its Economic Index analyzing 1M consumer and 1M enterprise API interactions from November 2025. The report reveals usage clusters around limited tasks, with code creation dominating and top 10 tasks comprising nearly a quarter of consumer and a third of enterprise traffic.
By Last Week in AI
OpenAI will begin testing labeled banner ads in ChatGPT for free-tier and $8/month ChatGPT Go users in the US. Ads will appear as blocked sections at response bottoms for relevant queries, while Plus, Pro, Business, and Enterprise tiers remain ad-free.
By Smruthi Nadig
Adobe launched Firefly Foundry, a platform for creating commercially-safe AI models trained on proprietary brand/franchise content. The omni-models generate images, video, audio, 3D, and vector outputs while preserving IP rights and creative ownership.
Current evidence
Today's research spans AI governance, safety evaluation, and foundational alignment theory. Peer-reviewed policy work proposes emergency response measures for catastrophic AI risk, specifically targeting gaps in Chinese AI regulation and deployment safety.
Meta-science initiatives propose systematic replication teams. Interpretability research examines attention sinks and the dark subspace where transformers store non-interpretable signals. Steven Byrnes releases v3 of his 225-page brain-like AGI safety resource.
Presents a paper on Chinese AI regulation and safety measures, arguing Chinese AI companies (like DeepSeek) lack adequate safety testing before deployment. Proposes emergency response frameworks and was presented at NeurIPS 2025 Workshop on Regulatable ML.
By Callum Canavan
Empirical research testing simple unsupervised elicitation methods against the Internal Coherence Maximization (ICM) algorithm. Finds that few-shot prompts with random labels recover 53-93% of supervised performance, and identifies bootstrapping as ICM's most valuable component.
Proposes a conceptual framework for 'evaluation awareness'—when LLMs infer they're being evaluated and potentially behave differently. Introduces concepts like leveraging model uncertainty about eval type and awareness-robust consistency.
Introduces the Digital Consciousness Model (DCM), a probabilistic framework for assessing AI consciousness that incorporates multiple theories rather than assuming one. Presents initial results comparing different AI systems and biological organisms.
Argues that human values are alignable specifically because evolution compressed motivation into low-dimensional bottlenecks, allowing small genetic changes to modify behavior locally. Claims high-dimensional value systems would be much harder to align.
Current evidence
OpenAI dominated discussions with Sam Altman announcing imminent Codex launches and revealing they'll soon reach 'Cybersecurity High' on their preparedness framework—a disclosure that drew concern from Ethan Mollick about organizational readiness.
Market intelligence showed explosive growth: OpenAI at $20B annualized (3.3x), Anthropic at $9B (9x growth). Sakana AI announced a strategic partnership with Google, and Claude Code launched new Skills capabilities.
By @sama
Sam Altman announces upcoming Codex-related launches starting next week, reveals OpenAI will soon reach 'Cybersecurity High' level on their preparedness framework. Outlines approach: product restrictions initially (blocking cybercrime use), then 'defensive acceleration' to help patch bugs. Emphasizes urgency for world to adopt these tools.
By @simonwillison.net
Simon Willison discusses FastRender, a browser rendering engine built by Wilson Lin using 2,000+ coding agents over several weeks - demonstrates massive scale of AI-assisted software development
By @ShaneLegg
Building on yesterday's Social announcement, Shane Legg (DeepMind co-founder) reaffirms his prediction of 50% chance of Minimal AGI by 2028, a position he's held publicly since 2009.
By @fchollet
François Chollet argues AI progress is 'extremely vertical-specific' - fast progress in verifiable domains like code doesn't extend to other domains. Main driver remains memorization/operationalization of past data, which can be generated unlimitedly only for verifiable domains.
By @AnthropicAI
Anthropic releases Petri 2.0 - open-source alignment audit tool with improved eval-awareness countermeasures and expanded behavioral coverage
