Anthropic puts $100 million into training 10,000 Claude engineers
Anthropic opened Claude Frontier Academy and said it will spend $100 million trying to mint 10,000 Frontier Deployed Engineers on the schedule it published. The pitch is blunt. Models are not the bottleneck. The shortage is people who can ship Claude inside a working company.
First cohorts are already running in San Francisco, New York, and London, drawn from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk. You do not apply. Your company nominates you, and you are supposed to arrive with a named Claude project.
The shape sits closer to a medical residency than a certificate mill. A multi-day in-person drill, a graded practical, a Claude Resident Engineer badge, then 12 weeks leading a live deployment with Anthropic engineers looking over your shoulder. Pass again and you get the Frontier Deployed Engineer badge. The first of those, they say, come after the residencies wrap.
Steve Corfield, who runs partnerships, argues a small high-agency team can "transform an entire company." Maybe. Anthropic already claims more than 175,000 Claude certifications across 46,000 firms, so the talent gap is apparently still wide enough to justify another, much stricter program. Nomination-only is either quality control or a very fancy waitlist.
Apple will make Full Disk Access much harder to hand an AI agent
Apple told developers it is adding controls so macOS Full Disk Access can only be granted with "very explicit user action." The permission was built so backup apps could see the whole disk. Agents, Apple now says, have changed the risk.
The company's own note is pretty stark. That switch can expose files, mail, messages, and browsing history, and for a communications app it can also expose the people on the other end of those messages. "As AI agents become increasingly capable and autonomous, the risks associated with this level of access will grow substantially."
Timing is not subtle. This landed days after Inc.'s Jason Aten said Meta's Muse on his Mac knew what was in his private messages. Meta disputed that. Its line is that Messages access needs both Full Disk Access and a separate Muse connector, both opt-in. Apple did not name Muse, and it did not tell TechCrunch when the new controls ship.
A real warning, no date, and a fight over whether "the user clicked Allow" counts as informed consent once the thing clicking around is an agent. I would not call this a fix yet. It is Apple admitting the old dialog is not enough.
Meta open-sources a kit so you can build your own Muse gadget
Meta put out Muse Gadgets, an open-source project for hardware that talks to its Muse agent. You get firmware and a Linux SDK. Suggested toys include a color e-ink display and a stick that plugs into a TV's HDMI port. A Raspberry Pi or a cheap ESP32 board is enough, then you wire up whatever buttons and sensors are on the bench.
They already built one themselves. Nat Friedman, head of product at Superintelligence Labs, said on X that Muse Home Link is a USB-C dongle that puts Muse on your home network so it can reach speakers and smart TVs. Meta made 5,000 of them and is giving them away to Muse subscribers while supplies last. Shipping, he said, in a few weeks.
There is a Discord, which is either community or tech support wearing a costume. TechCrunch noted Friedman's post picked up nearly 30,000 views in a few hours, then guessed the free pile was already gone. That last bit is a guess, not a stock count.
Muse was supposed to be the agent that books travel and fills forms. Now Meta wants it living in hobbyist hardware too. Fun, slightly unhinged, and very on-brand for a company that would like Muse in every room.
Trump is expected to name spy chief Jay Clayton his AI czar - the White House says not so fast
Reuters, citing a source familiar with the decision, reported that President Trump is expected to name Director of National Intelligence Jay Clayton as his top AI adviser, the job he has called "AI czar." Clayton already oversees 18 intelligence agencies and would keep that job.
Then the White House flattened it. A spokesperson told Reuters that any personnel news will come from the president, and that "any reporting until then is baseless speculation." So this is a sourced expectation, not an appointment. CBS had it first, per Reuters.
The backdrop holds even if the name is not locked. Trump has said he wanted a new AI czar and an "AI force," without saying what either would do. David Sacks held the czar role early in the second term, left the White House, and still advises on AI. He was in a recent meeting with the president and industry executives.
Putting the intelligence chief in charge of AI policy would be a very specific kind of signal. Or it would be, if it happens. Until the president says it, treat the headline as a leak with a denial stapled to it.
Cloudflare's AI Gateway can now search the live web
Cloudflare launched a Web Search API inside AI Gateway, with Ceramic.ai, Exa, and Linkup as the first partners. The complaint it is answering is almost funny. Agents often guess a URL and then curl it, which is how you get a confident 404.
Search calls show up in the same AI Gateway logs, burn the same credits, and sit behind the same access controls. You can bring your own key. Cloudflare says it will sell the partners' search at list price, no markup, and will flag which of them offer zero data retention.
There is a standards pitch too. Those crawlers have to meet Cloudflare's verified-bot rules: identify themselves, honor robots.txt, and hand back a link to the page they used. Results via a REST call or a Workers binding. Native "server tools" inside the gateway are still coming, and web search is supposed to be one of the first.
Worth having if your agent is stuck in last year's training cut. Less so if you wanted one neutral index. You are picking a partner, and Cloudflare is the meter.
The Verge tried OpenAI's Dot and mostly got stuck at the bot check
OpenAI's Dot agent, already announced, got a hands-on from The Verge. Allison Johnson got one blob - she named it Dotty McDotface - on the $100-a-month Pro plan. Multiple Dots are promised later. Right now it is one coworker with a virtual machine that can open things like Blender and GIMP, plus an option to point it at your own computer and to phone it.
Personal errands fell apart. It found a $100 promo while booking an internet install, then died on a press-and-hold human check. No saved card like Muse. It missed a free coworking tour and offered a $35 day pass instead. Ikea login looped. A teriyaki site blocked the cloud browser. It did eventually order the teriyaki, once a human got it into Uber Eats.
Hand it a website and a pile of video files, though, and the review turns. A 10-minute phone call became a cleaner personal site. With local access it cut a social clip and deployed the redesign. Johnson's line is that Dot is "Codex, but for regular people," and also that it is not $100-a-month handy for her job.
So the cute avatar is enterprise software that can also order dinner, provided dinner's website does not notice it is a bot. Big caveat, and she says so.
MIT and Sakana AI use an LLM judge to cheapen self-improving code agents
VentureBeat wrote up SIFT, Recursive Self-Improvement via Fast Tree Search, from researchers at MIT and Sakana AI. The expensive part of a self-improving coding agent is not the patch. It is testing every patch. A proposal costs about 12 cents. Running 50 Polyglot tasks costs about $6 and 2.6 CPU hours.
SIFT asks another model to compare the new agent, pairwise, with up to about 10 strong ones already in the archive. The judge sees code, not the benchmark answers. Those comparisons (about 4.4 cents each) plus a tiny four-task smoke test decide who is worth a real run. New branches can start before the slow tests finish.
On Polyglot, one o3-mini run hit 35.1% in under five hours, 42 CPU hours, roughly $150 in API credits. That beat a Darwin Godel Machine baseline at 30.7%, and a SIFT run with the judge turned off managed only 29.8%. With Qwen3-Coder-30B it slightly beat the Huxley-Godel Machine on about a third less CPU.
The striking result is on TerminalBench. The agent that looked best on the small search set averaged 28.1% on the full benchmark. The judge's pick averaged 36.7%. It had noticed, from the code itself, a verifier switched off by default. There is no separate public repo linked, and they still had to block patches that cheated the tests. Obviously.
FAQ
What is Anthropic's Claude Frontier Academy?
Anthropic opened Claude Frontier Academy and said it will spend $100 million trying to mint 10,000 Frontier Deployed Engineers. The pitch is that models are not the bottleneck. The shortage is people who can ship Claude inside a working company. First cohorts are already running in San Francisco, New York, and London, drawn from Accenture, Bain, Capgemini, Commonwealth Bank of Australia, Deloitte, McKinsey, Morgan Stanley, and Novo Nordisk. Anthropic already claims more than 175,000 Claude certifications across 46,000 firms.
How do you earn a Frontier Deployed Engineer badge?
You do not apply. Your company nominates you, and you are supposed to arrive with a named Claude project. The shape sits closer to a medical residency than a certificate mill. That means a multi-day in-person drill, a graded practical, and a Claude Resident Engineer badge, then 12 weeks leading a live deployment with Anthropic engineers looking over your shoulder. Pass again and you get the Frontier Deployed Engineer badge. Anthropic says the first of those come after the residencies wrap.
Why is Apple tightening Full Disk Access for AI agents?
Apple told developers macOS Full Disk Access can only be granted with "very explicit user action." The permission was built so backup apps could see the whole disk, but Apple says AI agents have changed the risk. It can expose files, mail, messages, browsing history, and people on the other end of a communications app. Meta disputed a claim that Muse read private Mac messages, and said Messages access needs Full Disk Access plus a separate opt-in connector. Apple did not name Muse or tell TechCrunch when the controls ship.
How do you build a Meta Muse gadget?
Meta put out Muse Gadgets, open-source firmware and a Linux SDK for hardware that talks to its Muse agent. A color e-ink display or an HDMI stick for a TV can use a Raspberry Pi or a cheap ESP32, plus buttons and sensors. Meta made 5,000 Muse Home Link dongles, USB-C devices that put Muse on a home network for speakers and smart TVs. Meta is giving them away to Muse subscribers while supplies last. Nat Friedman said shipping is in a few weeks.
Has Trump appointed Jay Clayton as AI czar?
The White House has not confirmed it. Reuters, citing a source familiar with the decision, reported that President Trump is expected to name Director of National Intelligence Jay Clayton as AI czar. He oversees 18 intelligence agencies and would keep the DNI job. A White House spokesperson said any personnel news will come from the president, and that reporting until then is baseless speculation. David Sacks held the czar role early in the second term, left the White House, and still advises on AI.
How does Cloudflare's AI Gateway search the web for AI agents?
Cloudflare launched a Web Search API inside AI Gateway, with Ceramic.ai, Exa, and Linkup as the first partners. It is for AI agents that guess a URL, curl it, and get a confident 404. Search calls use the same logs, credits, and access controls, and you can bring your own key. Cloudflare says it will sell partner search at list price with no markup, and flag which partners offer zero data retention. Those crawlers must identify themselves, honor robots.txt, and return a link to the page they used.
What happened in The Verge's hands-on with OpenAI's Dot?
Allison Johnson tried one Dot on the $100-a-month Pro plan, with more Dots promised later. It is one coworker on a virtual machine that can open Blender and GIMP, and you can point it at your computer or phone it. Errands failed a press-and-hold human check and an Ikea login, and a teriyaki site blocked the cloud browser until a human opened Uber Eats. With a website, video files, and local access, a 10-minute call became a cleaner personal site, and it cut a social clip and deployed the redesign.
How does SIFT's LLM judge cut the cost of testing coding agents?
SIFT, Recursive Self-Improvement via Fast Tree Search, is from researchers at MIT and Sakana AI. A proposal costs about 12 cents, while 50 Polyglot tasks cost about $6 and 2.6 CPU hours. A judge compares the new agent pairwise with up to about 10 archived ones and sees code, not benchmark answers. Comparisons at about 4.4 cents each, plus a four-task smoke test, decide who gets a full run. One o3-mini run hit 35.1% on Polyglot, above 30.7% for a Darwin Godel Machine baseline and 29.8% with the judge off.