01Meta Ships Muse Spark Open: an 816B-Parameter Open-Weight Model Built for Macs [AI Models & Research]Meta calls it the most capable open-weight model in the US: 816B parameters, a 1M-token context window, natively multimodal and tuned for coding and agentic work. The catch for local runs: about 1TB of memory, Mac Studio M5 Ultra territory. Available now via API platforms. → source
02Anthropic Opens the Door: METR Gets Employee-Level Access — and the Company Can't Veto Its Reports [AI Models & Research]Six METR researchers get Anthropic laptops, office space and internal-tools access on par with Anthropic's own risk teams, modeled on banking's resident examiners. Anthropic can demand narrow redactions but cannot block publication — the top tier of a new three-tier evaluator system. → source
03Scaling Indic Pretraining: 2.8 Trillion Tokens Through an Open Data Pipeline [AI Models & Research]NVIDIA and partner teams detail how 2.8 trillion tokens of Indic-language training data were generated through an open, reproducible pipeline — a template for bringing large-scale pretraining to languages the frontier has historically skipped. → source
04LLMs Prove They Can Design State-of-the-Art Quantum Circuits [AI Models & Research]A new paper from a Harvard/MIT team reports large language models discovering quantum circuits at or beyond the current state of the art — concrete evidence that coding agents now operate in one of physics' hardest design spaces, not just software. → source
05Bluesky's Attie Turns Vibe Coding Into a Group Project — Sites and Games on the AT Protocol [AI Tools & Ecosystem]Bluesky expanded Attie so anyone can build websites, quizzes, games and polls that live on the same AT Protocol as the network — in CEO Toni Schneider's words, vibe coding for your community — with Bluesky's built-in distribution answering the 'who's using what you build' problem. → source
06FLUX.3 Image Klein Lands in Diffusers: 5.6B Open Weights With Editing and Layout Control [AI Tools & Ecosystem]Black Forest Labs' latest open image model is live in Diffusers — 5.6B parameters covering text-to-image, image editing and bounding-box layout control, with quantized variants and LoRA finetuning sized for consumer GPUs. → source
07Choosing a Chunking Strategy for Your RAG Stack: a Decision Tree With Failure Signals [AI Tools & Ecosystem]A six-question decision tree maps your documents and retrieval task to a concrete chunking strategy — plus the failure signals that tell you the current one is quietly losing recall. → source
08Anthropic Told Investors: $16.8B Quarterly Revenue, Up Nearly 50%, IPO in Sight [AI Applications & Industry]The FT's readout from Anthropic's investor day: Q1 $4.8B, Q2 $11.5B, Q3 $16.8B, with $1.70 of revenue per dollar of compute and an 11% adjusted operating margin. Distillation — rivals training on Claude's outputs — was cut back to 5–7% of revenue in Q3. → source
09Targeted Ads Delivered 0-Click Mercenary Spyware to Al Jazeera's Devices [AI Applications & Industry]Citizen Lab's AdFection investigation appears to be the first confirmed case of mercenary spyware arriving through laced ads on iOS and Android: 45 devices at Al Jazeera infected, brokered by Israel's Discover-Tech across 70,000 apps and 120,000 websites. → source
10AI Agents Ride Into Rivian's Cars — Anthropic and OpenAI Models Are Next [AI Applications & Industry]RJ Scaringe says Rivian's in-house agentic framework — a 'traffic light' that routes requests between an on-vehicle and a cloud model — will open up to frontier agents from Anthropic and OpenAI. Brakes and steering stay off-limits; details at December's AI & Autonomy Day. → source
11Amazon's Bee Doubles Down: Two All-Day AI Note-Taking Wearables From $75 [AI Applications & Industry]The $75 Bee Frontier and $99.99 Bee Eclipse turn everyday conversation into to-dos, calendar invites and summaries — raw audio processed and deleted, encryption keys held on-device. Amazon's devices chief frames AI gadgets like these as what eventually replaces apps. → source
12DeepMind Sends Antigravity to the Lab Bench — a New Antibiotic Candidate Out the Other Side [AI Applications & Industry]DeepMind's walkthrough shows biologists using the Antigravity agentic platform to collapse a fragmented antimicrobial-peptide search into a single session, scanning millions of uncharacterized UniProt proteins with ProtNLM. The lead candidate is now in wet-lab validation. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.