The author recounts the formation and dissolution of Microsoft Research's AI Frontiers group, which pursued innovative agent frameworks and novel model architectures. After reorganization into Microsoft AI under leadership prioritizing scaling over research, the author and fellow leaders departed due to philosophical misalignment and insufficient support for research-driven innovation.
▲ 69 points20 commentsandsoitis via Hacker News
A compressed version, HyperFlow H3 Flash, of the MiniMax-H3 model now supports joint video and synchronized stereo audio generation. This 13.7B model operates in just 4 to 8 steps using INT8 or GGUF formats. Performance guidelines suggest specific hardware for optimal use.
▲ 29 points8 commentsjiant via Hacker News
AI-powered tools are automating cyberattacks at unprecedented scale, enabling real-time vulnerability discovery and exploitation. Both state and private actors leverage open-source AI systems to breach major organizations, creating supply chain compromises affecting thousands. Experts compare this disruptive period to the insecure 1990s internet, predicting a difficult adaptation phase.
via The Verge
METR's investigation found Anthropic's AI models improving twice as fast as a year ago, driven largely by AI-assisted research rather than compute scaling. Employees report 12x productivity gains using unreleased internal models. However, Anthropic heavily redacted the public report, limiting transparency about the pace and methods of its accelerated development.
via The Verge
AI-enhanced caregiver apps are growing to help families manage complex care needs, offering features like medication tracking, appointment summaries, and AI assistants. However, these tools require significant manual data entry, raise privacy concerns due to limited regulatory oversight, and may not alleviate the core burden of caregiving.
via WIRED
OpenAI has reportedly secured $11 billion in a new funding round valuing the company at $1.25 trillion, with the round expected to close soon and potentially reach $20 billion total. This falls short of a previous $30 billion target. Meanwhile, competitor Anthropic faces lower valuation expectations, and OpenAI has delayed its IPO plans amid market uncertainty.
via TechCrunch
Daily briefing
One email each morning with what changed in AI and why it matters. No spam, unsubscribe any time.
Testing Jev and GPT-6 Luna as identity resolution judges, researchers found a hybrid Jev-Claude Opus system achieves near-frontier accuracy at over four times lower cost. Jev's superior confidence calibration allows it to selectively escalate difficult cases to Opus, outperforming Luna which suffers from overconfidence. The approach enables scalable, cost-effective identity verification.
▲ 566 points139 commentstheanomaly via Hacker News
Security researchers found over 800 publicly exposed instances of the open‑source Artex AI penetration‑testing tool, enabling anyone to download and use it for cyberattacks, including large‑scale breaches of Korean finance and ConnectWise systems.
via TechCrunch
Vals research shows Chinese open-weight model DeepSeek nearly matches OpenAI's closed Daybreak in simulated hacking tasks (73% vs 81% success), at far lower cost. This narrows the capability gap between open and closed models, challenging assumptions that gated access slows proliferation. Experts note attackers still favor older models, but the trend signals a shifting security landscape.
via WIRED
Nilay Patel argues AI products are insufficient to justify their high costs, citing data center noise, deepfakes, and unproven agents. He claims companies spend more than they earn, linking industry influence to political corruption. Patel says the industry must create clearly beneficial products to regain trust.
via Simon Willison
From the Infrenic lab
Infrenic builds working AI systems for businesses. Bring one real workflow and get a free proof of concept in 48 hours.
A former Rockstar developer suggests that AI-assisted decompilation of GTA VI could pressure the studio to release an official PC version sooner. This technology challenges Rockstar's historical pattern of prioritizing console releases over PC ports. The official PC timing remains unconfirmed.
via Ars Technica
AI companies are purchasing internal corporate communications and documents from data brokers to train models, with deals reaching millions of dollars. Firms like Mercor and Micro1 acquire emails, Slack messages, and work files, claiming to anonymize data. Critics warn anonymization may not fully protect privacy, and workers may unknowingly help build AI systems that could automate their jobs.
via WIRED
Daily briefing
One email each morning with what changed in AI and why it matters. No spam, unsubscribe any time.
Google launched Gemini Omni 1.2 Flash, a video generation model producing 15-second 720p clips with conversational editing. Pricing dropped 30% to $0.07 per second.
via MarkTechPost
Google clarified its new AI policy for schools, allowing them to disable Gemini while keeping search active or turn off all AI including search. However, Google will not provide search without AI features, insisting AI is integral to search. Schools gain more control but cannot access AI-free search, though antitrust rulings may improve alternatives.
via Ars Technica
Cult.fit is leveraging Codex to enable cross-functional teams to develop tools more efficiently, accelerating workflows and transforming operational processes.
via OpenAI
Simon Willison shared a Paul Ford quote highlighting web browsers as highly powerful application platforms with rapidly decreasing access costs, posted November 5, 2026.
via Simon Willison
Google DeepMind has released Gemini Omni 1.2 Flash, enabling native 15-second video generation and editing. This update offers improved consistency, multi-turn editing, and a 30% cost reduction across Gemini API and related products.
via Google DeepMind
Google AI has released Gemini Omni 1.2 Flash, enabling native 15-second video generation with improved consistency. The update is 30% less expensive than the previous version and is available across the Gemini API and Google Vids.
via Google AI
Daily briefing
One email each morning with what changed in AI and why it matters. No spam, unsubscribe any time.
Australian universities are integrating ChatGPT and AI tools into curricula to equip graduates with AI skills for the modern workforce. This initiative aims to make students proficient in AI technologies from the start of their careers.
via OpenAI
Signal complete
New briefings land through the day. Get tomorrow’s in your inbox.