In partnership with
☀️ TRENDING AI NEWS 🤖 Anthropic Speedup: METR says Anthropic's models are improving roughly 2x faster than a year ago, with little sign of slowing. 🏢 Nvidia Chips: A New York Times probe found H200-approved firm New Matrix operating from an empty office in China. 🛠️ Free Models: DeepSeek and GLM came close to OpenAI's non-public cyber model in simulated network break-ins. 🚨 Deepfake Fraud: Three men were charged with using AI to impersonate CEOs and steal over $100M. |
An empty office with wires dangling from the ceiling where the lights should be. That is apparently where some of the most sought-after AI chips on Earth were cleared to go. Today's issue starts with that image, and it only gets stranger from there.
| 🛠️ Tool of the day | |
ElevenLabs - an AI audio platform that turns text into natural-sounding speech and also makes sound effects and music. It helps creators, teams and developers add voice to videos and apps, and it has a free plan.
| ⚡ Quick hits | |
After the midterms, tech firms face tougher data center rules, and OpenAI has started calling its data centers "compute campuses". Read more →
| 🤓 AI Trivia | |
In a mixture-of-experts model, what is the job of the 'router'?
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
| 🤖 METR Says Anthropic's AI-Building-AI Loop Is Working | |
An outside audit confirms the acceleration
The AI safety research group METR published a report on Thursday finding that Anthropic's models are improving roughly twice as fast as a year ago, with little indication of slowing down. The reason: Anthropic is using AI to build more powerful AI, and it appears to be working remarkably well.
This is the independent follow-up to the company's own claim that self-building AI is close. We covered that warning earlier this week, and now an outside lab is putting numbers on it, which is why researchers are alarmed.
The bottom line
When the speed of progress itself is accelerating, planning around today's capabilities gets riskier every quarter. Treat any 'AI can't do that yet' assumption as having a short shelf life.
285 Free Prompts to Save Time
Your inbox isn't the problem. Neither are your meetings. It's the hours you spend on work AI could do in minutes.
285 copy-paste prompts: built for email, meetings, admin, and delegation. Each one has a clear role, a specific task, and a defined output.
No setup. No prompt engineering. Just copy, paste, and get your time back. Free when you subscribe to Mindstream.
| 🏢 The US Let Nvidia Sell Chips to an Empty Office | |
Export vetting looks shaky
According to The New York Times, New Matrix sits on the list of Chinese firms approved to handle Nvidia's H200 AI chips, yet its offices are empty, "a hollow gray shell with wires dangling from the ceiling where lights should be." Questions are now being raised over whether the Commerce Department or Nvidia is doing real due diligence.
It lands alongside a separate report that China's Ministry of State Security recruited Megaspeed executive Alice Huang to spy on Nvidia CEO Jensen Huang and potentially push for the most advanced chips to be made available to China. Megaspeed is a customer of Nvidia's.
The bottom line
Chip export controls are only as strong as the vetting behind them. Expect pressure on both Washington and Nvidia to tighten how approved buyers are checked.
| ⚠️ Free Chinese Models Are Getting Scary Good at Hacking | |
Open weights meet offensive security
In an experiment where AI models were asked to break into simulated networks, the hacking skills of DeepSeek and GLM were not far behind OpenAI's most advanced cyber model, which isn't public. These are free models anyone can use.
The same week, researchers found hundreds of servers running the Artex penetration-testing tool exposed to the open internet, handing anyone access to the AI hacking tool. Security experts on The Verge's podcast went further, arguing the multibillion-dollar cybersecurity industry as we've known it is essentially broken.
The bottom line
Restricting the best closed cyber models does little when free alternatives are close behind. Defenders should assume attackers already have near-frontier tools.
| 🏢 'AI Clog': Chatbot Complaints Swamp UK Public Bodies | |
A trickle in 2024, an onslaught now
MPs, local authorities, lawyers and schools are receiving floods of AI-generated grievances, requests and legal challenges, often long and complex to read and respond to. A senior caseworker in a Labour MP's office says it started as a trickle in 2024 and is now an onslaught.
The asymmetry is the problem: writing a dense complaint now takes seconds, while answering it still takes a human hours. Schools are asking the same questions, and The Guardian is gathering teachers' and parents' experiences of AI-written emails.
The bottom line
Cheap AI-generated input is overwhelming institutions that can't scale their responses the same way. The next fix is likely AI on the receiving end too.
| 🚨 AI Impersonation Fraud Gets Indicted | |
Cloned CEOs, real money
Three foreign nationals were arrested and charged with using AI to impersonate corporate CEOs and attorneys, allegedly stealing over $100M, according to an unsealed Department of Justice indictment. The men allegedly mimicked the voices and images of executives to convince employees to send large payments. It is a stark example of how deepfakes have moved from novelty to organized crime.
The bottom line
A familiar voice on a call is no longer proof of identity. If you handle payments, add a second channel of verification before moving money.
| 🔬 MoE Routing, Explained | |
A new MarkTechPost guide walks through 7 MoE routing strategies, from Mixtral's top-2 softmax to DeepSeek-V4's hash routing, covering which models use each. Routing is what makes huge models affordable to run, and it's worth knowing if you follow open-source AI releases. Speaking of building: if you need a site to showcase your own AI project, our sponsor 60sec.site is an AI website builder that gets you live in about a minute.
| 🌎 Trivia Reveal | |
The answer is: it decides which experts process each token! The router sends each token to a small group of experts, which is why a huge mixture-of-experts model can activate only a small slice of its parameters per token. Mixtral uses top-2 softmax routing, one of the strategies in today's guide.
| 💬 Quick Question | |
Have you received an email, complaint or message lately that you suspect was written by AI? Hit reply and tell me how you spotted it. I read every response!
That's it for today - see you tomorrow! Visit dailyinference.com for our daily AI newsletter.
| 🎁 Share Daily Inference | |
Know someone who wants to keep up with AI in 5 minutes a day? Share your link. 1 referral unlocks the AI Tools Starter Kit. 5 referrals unlock the AI Insider Briefing.