01Language Models Hit a Formal Limit: What Text Alone Can Never Teach Them [AI Models & Research]An information-theoretic proof from Emily Cheng and Ryan Cotterell shows a listener given only text — including any LLM — cannot fully recover what a speaker means: meaning that never shapes the form of an utterance is provably out of reach, no matter how large the corpus. → source
02When Robots Mishear Us: Voice Commands Become a Safety Risk for Embodied AI [AI Models & Research]Researchers simulated ASR errors on voice-controlled embodied agents and found harmful instructions being accepted and executed — the safety chain is only as strong as its noisiest link, and home robots will run on misheard speech. → source
03Chatbots Out-Debunk Search Engines on State Propaganda, NPR-NewsGuard Test Finds [AI Models & Research]Six major chatbots debunked false narratives from Russia, China and Iran about 75% of the time in a 30-question audit — beating every search engine and most AI search summaries; Bing's AI answers failed more often than they challenged the premise. → source
04Simon Willison Decodes ChatGPT Work: Two Products, One Very Confusing Agent Platform [AI Tools & Ecosystem]ChatGPT Work is two products — a cloud agent and a local Codex successor — and the cloud one can run code with open internet access, drive a full headless Chrome, build and deploy sites on Cloudflare Workers, and read a persistent /workspace filesystem across sessions. → source
05OpenAI Bought Tens of Thousands of Macs for RL — and Nvidia Calls Apple Its Top Local-AI Rival [AI Tools & Ecosystem]The Information reports OpenAI quietly bought tens of thousands of Mac minis and Studios for reinforcement-learning and computer-use training while Anthropic rents them via AWS. Apple's unified memory became an accidental AI win, and Nvidia's RTX Spark chips sold out before launch. → source
06Infostealer Malware Is Hijacking Claude Sessions to Drain Paid Usage, Anthropic Warns [AI Tools & Ecosystem]Anthropic says Vidar, LummaC2, StealC, RedLine and Atomic Stealer are lifting authenticated Claude session cookies and burning through paid usage on hijacked accounts. Sessions are revoked and charges refunded — but signing out does not remove the malware. → source
07Inside Meta's Push to Put Robots to Work in Its Data Centers [AI Applications & Industry]Meta is testing Kinova and ABB robot arms to power-cycle servers and swap network cables; one worker estimates a bot could replace 80% of some technicians' workloads. The simplest 'robot' in trials: a motorized finger that presses Mac mini power buttons on remote command. → source
08Caterpillar Brings Its Mining Autonomy Playbook to Enterprise AI Deployment [AI Applications & Industry]CTO Jaime Mineart says decades of autonomous mining translate directly to AI rollout: 1.6 million connected assets, 16 petabytes of proprietary data, a $100M workforce-training pledge — and record $20.5B quarterly revenue driven by data-center power demand. → source
09SpaceX's Secret Turbine-Blade Foundry Aims to Speed Up AI's Power Bottleneck — Pollution Questions Follow [AI Applications & Industry]Musk confirmed a Bastrop, TX foundry will cast single-crystal turbine blades — a part only four companies worldwide can manufacture — to bring gas turbines online up to 18 months faster for data centers, as Memphis pollution litigation over unpermitted turbines continues. → source
10Washington Walls Off Chinese Drones and Robots, but Scale Doesn't Need the U.S. Market [AI Applications & Industry]New FCC restrictions and sweeping drone tariffs target Chinese hardware, yet the world's five largest humanoid makers — all Chinese — shipped 86% of global units in H1 2026. Analysts' verdict: you can't sanction your way around a cost curve; expect a fragmented two-ecosystem market. → source
11$130 Billion in Data-Center Projects Stalled as Community Opposition Goes Mainstream [AI Applications & Industry]CNBC counts roughly $130B of data-center projects delayed or cancelled in Q1 alone — nearly matching all of 2025 — as more than half of Americans now express concern over excitement about AI, turning the backlash into a bipartisan midterm-election issue. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.