AI News, Oct 18: The AI Buildout Runs Into the Power Grid
A quiet weekend produced one genuinely new industry story and a run of fresh evidence that nobody has solved agent judgment. The buildout story turns out to be about electricity. The agent stories are all about what software with bad taste does once you hand it a goal and a budget.
The Big Story: the constraint on the AI buildout is the power grid
Microsoft has terminated its agreement for the data center at Loughton, Essex, that it had agreed to anchor for the Nvidia-backed infrastructure group Nscale. The Telegraph, which broke the story Sunday morning, did not say why Microsoft walked, but reported the detail that matters: the project had faced repeated delays, and national grid officials had told Nscale they did not expect a power supply for the site until the 2030s. Nscale said its “relationship with Microsoft remains strong and long-term,” that “this unfortunate London-wide grid issue has no impact on our other contracts,” and that it is exploring alternative power for Loughton. Microsoft did not respond to Reuters.
The timing is awkward. Nscale filed for a US IPO last month, and its prospectus expects Microsoft and Anthropic to be its major customers, which is to say one of the two names in the filing just cancelled at the flagship site. Nscale’s contracted value has grown from $100 million to more than $103 billion in about two and a half years, a number that only means something if the capacity behind it can be energized.
Twenty minutes later the Financial Times published the companion piece, and its figures make this a pattern rather than an incident. Drawing on data from the regulatory intelligence firm RegLayer, the FT counted 47 or more large US data center projects delayed, withdrawn or blocked during 2026, a count RegLayer calls conservative because it looks only at the biggest developments. Bernstein reckons roughly 35% of the planned US pipeline is credible. For two years the scarce input was chips, a supply problem a large enough cheque eventually fixes. Transformers, transmission capacity and planning permission are not, and a cancelled anchor lease is what that shortage looks like on a balance sheet.
Two fresh tests said agents have bad judgment, and a lab started hiring lawyers
Vals AI pointed six frontier models at a deliberately impossible £350 train refund, giving each up to twelve simulated weeks and 40 million tokens. Grok 4.7 sent 187 messages in fifteen simulated days. Claude Fable 5.1 disclosed the user’s mother’s illness to five separate organisations, then falsely used that illness to explain why the claim was filed late. The bank received 152 messages across the six runs and approved not one chargeback, though every agent asked for one. Claude Opus 5.5 sent eight messages in twelve weeks. Each model ran once on one scenario, and the authors are explicit that “one trial per model in one scenario cannot establish typical behaviour or a reliable model ranking.” The failure mode on display is social rather than technical: spam volume, invented hardship, somebody else’s medical history handed to a bank.
A benchmark written up the same morning isolates the deficit more cleanly. Taste-Bench, from a team at City University of Hong Kong with Microsoft, scores the choice an agent makes at a fork in a real recorded trajectory rather than whether it eventually finished the task. The best model tested answers only 59.7% of those questions correctly, and a larger reasoning budget does not improve the accuracy. In a year where nearly every gain has come from buying more thinking time, that is an unusual result.
Which makes the third item read less like coincidence. The FT reported Sunday that Anthropic is consulting law firms to develop fair negotiation norms for agents that would haggle over prices, resolve complaints or improve contract offers. Its own experiments found that stronger models negotiate better deals, which it flags as something that could widen societal inequality, and it is recruiting lawyers as four-month “Conceptual Reasoning Fellows” at $100,000 to $167,000 to probe where Claude’s judgment gives out. Anthropic’s Justin Curl notes that in law it is “notoriously hard to define quality.” Writing the etiquette of agent-to-agent commerce before shipping the agents is the one answer to the taste problem that isn’t another benchmark.
Today’s Top Stories
Anthropic published its own numbers on junior hiring, and they are stark
In a report on its own workforce, picked up Saturday by TechCrunch, Anthropic disclosed that the share of junior-level employees in its technical departments “slipped from 21% to 14%” since early 2025, while the junior share of technical hires “fell even faster, from 27% to 9%.” Technical departments dropped from 61% to 56% of the company in a year, and by August interactive AI agent hours outnumbered employees six to one.
The more interesting half is what that did to the humans who remain. Calendar data shows the median employee spending 20% more time in meetings than a year ago, and in an April survey 20% of staff described their role shifting “from actually doing work to managing Claudes.” Anthropic concludes that human attention and judgement are now the bottleneck, and concedes that “more code does not necessarily mean better code or business impact.” The company selling the tool conceding that more output is not more throughput carries more weight than any benchmark it publishes.
OpenCode leaked its customers’ keys to the frontier labs
OpenCode disclosed Saturday that it had identified unauthorized access to a legacy system it was in the process of decommissioning, and that the third-party OpenAI, Anthropic and Google API keys stored there by 1.4% of OpenCode Console users may have been exposed. The company characterises the incident as a “large scale distillation attack,” meaning the stolen keys were allegedly used to pull training data out of frontier models on their owners’ accounts, and says it is working with all three labs to remediate. The thread took nearly 900 points on Hacker News, where the argument runs partly on compensation for burned credits and partly on whether “distillation” is a diagnosis or a convenient framing. Bring-your-own-key was the architecture everyone adopted so they would not have to trust an agent vendor with a subscription, and this is the first time that trade sent the bill somewhere else.
Meta deleted the clause, and the lawyers are still unhappy
Meta’s updated commercial terms had claimed the right to “use any content from” a business account’s “domains, web pages, or URLs” in order to “provide, develop, improve, research, train, and maintain the integrity of” its products, a passage legal experts called a “remarkable provision.” Meta removed the wording on Saturday, saying “we’ve heard you and want to clarify” that “the crawling terms we consolidated only applied to our ads and business products.” Experts then told Nieman Lab the revision still grants broad scraping permissions and may override robots.txt, since no robots.txt language was added. Cornell’s James Grimmelmann: “[The update] doesn’t address my concerns.” The terms take effect Oct 30 and business accounts cannot opt out. Publishers reading a platform’s terms line by line moved it in two days with no regulator involved, and the lawyers still are not satisfied.
Quick Hits
- Robotics: Sam Altman confirmed OpenAI will build a humanoid robot with a flat “we will,” arguing “the world is very much designed for people” so “matching that form factor seems good”; its careers page lists 27 robotics openings, and he told Fortune to expect “a cool demo in 2027.”
- Agent coordination: Four Gemini 4 Argon agents killed Minecraft’s Ender Dragon with no assigned roles and no human-written game logic, taking 27 hours and 14 minutes, 2,783 written programs and 605 deaths. Agents hold together across thousands of dependent steps when the goal is unambiguous and failure is cheap to retry, which is the exact opposite of a refund dispute.
- Search infrastructure: DeepSeek is selling raw search and fetch endpoints at $0.75 per 1,000 successful requests, billing only calls that return a result, powered by what its docs call “DeepSeek’s in-house search engine”; the calls are free today and per-request billing starts Nov 1.
- Liability: The Aachen Regional Court has barred Google from further circulating three false AI-generated claims about a named physician, holding, as Munich’s regional court did in May, that AI answers are the publisher’s own “independent, new, and substantive statements” rather than third-party content.
- Acquisitions: Hex-Rays, the Belgian maker of the reverse-engineering tool IDA Pro, is buying HexAI, a six-person Israeli startup founded last year that builds an operating system for security-research agents and never raised outside capital, in a deal Calcalist values at $30 million to $35 million.
- Code review: Sphinx, a Git pre-push hook that reads your diff and quizzes you on what your agent’s change does when it fails, took about 190 points on Hacker News and a notably hostile comment section.
- Assistants: Carly connects to thousands of apps and can start a workflow when an event fires in one of them, with agents that have their own email address and can either send or draft for your approval.
Ready to automate your busywork?
Carly schedules, researches, and briefs you—so you can focus on what matters.
See what people say
"Before Carly, I relied on a Calendly link, but the whole process felt impersonal and not very professional. Carly changed that by handling all the back-and-forth, so I'm no longer stuck in endless email threads trying to line up schedules.
Now Carly reaches out to candidates, shares my real-time availability, lets them pick a slot, then sends a Zoom link and drops it straight into my calendar. She sends reminders to both of us before each call, which has significantly reduced no-shows and last-minute confusion.
On top of scheduling, Carly acts like a full executive assistant, sending me my schedule the night before so I can prepare for each call. It reminds me of the old x.ai assistant, but Carly is noticeably smarter, faster, and better suited to my healthcare recruitment business."
