01StudentBench: AI Tutoring Matches a Human GRE Tutor — Time to Grade the Teaching, Not the Answering [AI Models & Research]Curtis Northcutt's team introduces a benchmark suite and public platform that evaluates how well AI teaches — and finds students working with AI tutors match human instructors' GRE learning gains. Frontier progress still tracks raw model capability; this moves the metric to pedagogy, at scale. → source
02LiMA Gives Dexterous Robots Long-Horizon Foresight Without Costing Them Their Reflexes [AI Models & Research]A new vision-language-action framework splits manipulation into two asynchronous loops: a world-action model for millisecond reactive control and a long-term imagination module planning minutes ahead. The robot hand can commit to a plan and still catch the cup when it starts to slip. → source
03Gemini 3.8 Text-to-Speech Ships: One Voice Engine From Prompt to Podcast [AI Models & Research]Google DeepMind's new text-to-speech stack turns text into studio-grade multilingual narration with controllable style and delivery — and Simon Willison's same-day TTS Playground makes it the easiest frontier voice API to kick the tires on right now. → source
04Meta Connect: Muse Gets an Avatar, Your Glasses and the Checkout [AI Tools & Ecosystem]Zuckerberg made the three-week-old agent the centerpiece of Meta's future: a real-time avatar you can video-call, wake-word access on Meta smart glasses, Mac computer control, its own email address, and shopping connectors spanning Stripe, Shopify, PayPal and Walmart. Muse stays free — Meta plans to profit via 'a small fee from transactions.' → source
05ChatGPT Mobile Gets Voice-Driven Agents: Speak a Workflow Into Existence [AI Tools & Ecosystem]OpenAI brought voice-based agentic features to mobile: Plus and Pro users can dictate documents, emails and Slack summaries through GPT-Live, then resume the same task on desktop. Anthropic merged chat and workspaces into one agent; OpenAI is still keeping them apart. → source
06DeepMind Puts Your AI's Memory Behind Encryption [AI Tools & Ecosystem]Private AI Compute gains secure, server-side memory: persistent personal context that stays cryptographically sealed even from Google's own operators, with published security proofs and an updated technical brief inviting external review. Personalization without exposure is becoming a product category. → source
07Luna Is Now a Product: Meta Introduces Its Cameraless AI Glasses [AI Tools & Ecosystem]Days after the cameraless design surfaced in reports, Meta made Luna official at Connect: six microphones, an AI button, no camera — a privacy-first answer to the recording backlash, built to carry the Muse agent by voice alone. → source
08Altman and Amodei Take the Case for Global AI Rules to the UN Security Council [AI Applications & Industry]Speaking at a France-convened Security Council session alongside the UN General Assembly, OpenAI's and Anthropic's CEOs urged world leaders to build international oversight — even as the Trump administration balks at new guardrails and both superpowers avoid constraining their own labs. Xi arrives in Washington this week with AI regulation on the agenda. → source
09Spotify Hands Subscribers the Keys to Its Recommendation Algorithm [AI Applications & Industry]Taste Profile reaches US Premium users: see how Spotify models your taste, then reshape it in plain English — more variety, fewer sleep-sounds contaminations — and the Home feed, Discover Weekly and Wrapped re-rank around your edits. The era of algorithmic opacity as a feature is quietly ending. → source
10YouTube Opens Its Algorithm: Viewers Get to Build Their Own Feeds With AI [AI Applications & Industry]YouTube will let users assemble custom AI-curated feeds, alongside conversational discovery in YouTube Music and new AI tools in Studio for creators. It joins Spotify in a quiet industry pivot from engagement-maximizing feeds to user-steered ones. → source
11Smart Glasses Meet Their First Mass Backlash — and It's in India [AI Applications & Industry]MIT Tech Review documents covert recordings made on Wayfarer-styled Meta glasses going viral at the expense of trans people and other marginalized groups — while a sub-$105 Jio rival is on the way. A crackdown looks unlikely for an unexpected reason: Indian police are adopting AI-powered smart glasses for surveillance themselves. → source
12RAG or Fine-Tuning? A Practitioner's Field Guide to Picking the Right Adaptation [AI Models & Research]MachineLearningMastery maps the decision: RAG wins when knowledge changes fast or must cite sources; fine-tuning wins when you need the format, tone or task behavior baked in — and hybrid pipelines increasingly do both. A clean decision framework for anyone shipping domain-specific LLMs. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.