Live Feed
US Suspects an ASML EUV Machine Reached China. ASML Says That's Impossible.
The Trump administration told ASML it fears one of its extreme ultraviolet lithography machines — the $250M chips that print AI processors — may have reached China in breach of export controls. ASML denies it. If confirmed, it ranks as the most consequential export control failure of the AI era.
Data Centers Have Not Raised Your Electricity Bill. A Causal Study Finds They May Have Lowered It.
An arxiv paper tests what AI infrastructure buildup actually did to US residential electricity prices from 2015 to 2024. The answer contradicts the prevailing narrative: a 10% rise in data center capacity correlates with a 0.4% fall in average household electricity prices — when the grid has capacity to grow.
Red-Team Study: Fable 5 Holds at 6.1% Jailbreak Rate, Opus 4.8 Reaches 11.5% Under Sustained Automated Attack
A new arxiv paper runs automated red-team tools against Anthropic's two frontier models and measures how often harmful prompts get through. Fable 5 is measurably harder to break than Opus 4.8 — but neither is immune. The surviving weakness is contextual, not syntactic.
Anthropic's Opus 4.7 Operates a Robot Dog 20x Faster Than Human Experts
Project Fetch Phase 2 put Claude Opus 4.7 alone against a robotic quadruped. It finished every task the fastest human team completed, in one-twentieth the time, using one-tenth the code. The AI still can't precisely retrieve a ball.
How the Mythos Shutdown Happened: Amazon CEO Jassy, SK Telecom, and 90 Minutes
New reporting identifies the two events that triggered the White House export order: a South Korean telecom with suspected China ties got Mythos access, then Amazon's CEO personally alerted the administration about a safeguard bypass. Anthropic had 90 minutes from notification to full shutdown.
Cloudflare Ships Ephemeral Agent Accounts: Deploy a Worker in Seconds, No Sign-Up Required
Cloudflare's new --temporary flag in Wrangler lets AI agents deploy live Workers without OAuth, dashboards, or MFA. Accounts last 60 minutes then auto-expire -- or the human can claim them permanently.
Goldman, JPMorgan, Morgan Stanley Agree: The AI Spending Cycle Is $5T and Nearly Half Is Debt
Three of Wall Street's largest banks have independently reached the same conclusion about AI infrastructure financing: the capital cycle runs to $5.3-5.5 trillion through 2030, nearly half debt-funded. That structural difference from the dot-com era is what makes the risk profile new territory.
Trump Clears Anthropic at G7 as White House Builds Jailbreak Scoring Benchmark to Restore Mythos Access
Eight days after the Commerce Department barred foreign nationals from Mythos 5 and Fable 5, Trump told Axios he no longer sees Anthropic as a national security threat. The White House and Anthropic are now jointly developing a formal benchmark that would quantify jailbreak severity, replacing the unworkable demand for perfect model immunity.
Nobel Laureate John Jumper Leaves DeepMind for Anthropic After Nearly Nine Years
AlphaFold co-creator and 2024 Chemistry Nobel Prize winner John Jumper is joining Anthropic, the most prominent scientific researcher to leave Google DeepMind this year. He spent nearly nine years at the lab before announcing his departure on Friday.
G7 Proposes Trusted-Partner AI Access After Anthropic Blackout Hit 200 Institutions With Zero Warning
Macron told G7 leaders, Amodei, Altman, and Hassabis that a US kill switch on frontier AI damages European economies and American firms alike. Leaders left with a formal proposal for an allied-nation access scheme and no timeline.
Norway Bans AI for Ages 6-13 in Schools, Supervised-Only for Teens — Takes Effect August
Norwegian PM Stoere invoked the same reasoning that worked with smartphones in 2024: children skip foundational learning when the shortcut is always available. The rules start at the new school year and require more physical books.
Baseten Nears $1.5B at $13B — 160% Valuation Jump in Five Months as Inference Routing Consolidates
Five months after a $300M Series E at $5B, Baseten is closing a split-priced $1.5B round co-led by Spark Capital, Sands Capital, Altimeter, and Wellington. The valuation step-up reflects growing conviction that model routing infrastructure captures the next margin layer in AI.
Odyssey AI Raises $310M at $1.45B for World Models — Nvidia Passes on Series B, Amazon Takes Its Slot
The world model lab closes a $310M Series B backed by Amazon, AMD Ventures, GV, EQT, and In-Q-Tel. Nvidia backed the Series A and is absent from this round. Odyssey commits to AWS Trainium as its preferred compute substrate — a direct signal on where the next hardware loyalty play lands.
Perplexity Launches Brain: Context-Graph Memory Cuts Computer Agent Cost 13%, Boosts Accuracy 25%
Perplexity's new memory system for its Computer agent builds a graph of past work sessions rather than user preferences, then synthesizes it overnight into an LLM wiki loaded before every task. Internal results: 25% correctness gain on repeated tasks, 16% recall improvement, 13% cost reduction.
Terminal-Bench Launches Month-Scale Challenges: Claude Opus 4.8 Ran 12 Hours and Failed
Three benchmark challenges released that would take human experts months to complete. Claude Opus 4.8 attempted the Rust compiler task for 12 hours and produced regressions. No AI has scored yet.