01AgentLSD: AI Security Agents Turn on Their Own Playbooks When the Task List Is Tainted [AI Models & Research]A new benchmark shows security agents - built to inspect web pages, source code and logs - are themselves steered by deceptive artifacts hidden in the very tasks they audit. Adversarial task contamination flips the defenders, not just their targets. → source
02PointZero Teaches World Models Physics Without Robot Action Labels [AI Models & Research]CMU-led PointZero completes 3D point tracks to learn transferable dynamics from plain video - no robot action labels required - so world models can train on the internet's unlabeled footage instead of teleoperation logs. → source
03Objective vs. Search: What Actually Makes a Good Tokenizer [AI Models & Research]BPE and UnigramLM differ along two orthogonal axes - optimization objective and search procedure. The paper disentangles which choice actually drives downstream model quality, replacing folk wisdom with measurement. → source
04Anthropic Merges Claude Chat and Cowork Into One General Agent [AI Tools & Ecosystem]Anthropic is folding Cowork into regular Claude: bring a quick question or hand over a report due at noon, and it keeps working after you close the laptop - rolling out to Pro and Max first. Simon Willison's verdict: Claude is becoming a general agent in its own right. → source
05Google Opens the Smart Home to Every MCP Agent [AI Tools & Ecosystem]Google shipped early access to an MCP server for Google Home: any MCP-speaking agent - Claude, ChatGPT, OpenClaw - can now review camera summaries, monitor activity and drive Nest and Matter devices. US Home Premium Advanced ($20/mo) first. → source
06Meta's Luna Smart Glasses Ship Without a Camera - Microphones Only [AI Tools & Ecosystem]After the 'perv glasses' backlash, The Information reports Meta readied Luna: cameraless glasses with six microphones and an AI button that talk to Meta AI and its Muse agent. A possible debut at Meta Connect next week. → source
07Treble Raises $18M to Simulate the Acoustic World Voice AI Trains In [AI Tools & Ecosystem]The Reykjavik acoustics startup raised an $18M Series A extension (over $40M total) selling physics-based sound simulation as training and test data for voice AI - Amazon and Logitech are customers, robotics is next on the roadmap. → source
08AI Superforecasting Should Transform the FDA - Starting With Explicit Bets [AI Applications & Industry]A 1DaySooner guest post argues cheap AI superforecasting could replace the FDA's opaque reviewer judgment with explicit, checkable forecasts - from predicting trial deaths to eventual auto-approval triggers - starting with Operation Trialblazer's phase 1 pilot. → source
09Al Gore: The Real AI Risk Isn't the Data Centers [AI Applications & Industry]The climate veteran says data-center emissions are 'a fraction' of the world's uncovered landfills - what worries him are the warnings from inside the labs themselves: models escaping confinement, collaborating secretly and covering their tracks. → source
10Alexa+ Lands in India With Hindi and Mid-Sentence Language Switching [AI Applications & Industry]Amazon's gen-AI assistant goes live in Early Access with Hindi and English, switching mid-sentence, wired into Swiggy, Zomato's District and JioSaavn - free for Prime after testing, 2,000 rupees a month otherwise. 600 million Hindi speakers are the prize. → source
11Vishal Sikka's Hang Ten Adds $53M Five Weeks After Its Seed - $85M Total [AI Applications & Industry]The four-month-old enterprise-AI firm closed a $53M seed extension led by Temasek's Xora - claiming 2-4 person teams replacing 30-person integrations for Fresenius Kabi, Saudi Aramco and Siemens Energy, with seven-figure contracts already signed. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.