Today's Headlines

  • A single power line fell outside Washington, DC, and 3.1 gigawatts of data center load left the grid in roughly 30 seconds
  • Anthropic reports cutting more than 80% of Claude Code's system prompt with no measurable loss on its coding evaluations
  • Prentis, the AI lab Reid Hoffman and Mark Pincus co-founded in April, is reported to be in talks to raise $100 million at a $1 billion valuation

Today's page is about three assumptions coming loose, each in a different place.

The first is the assumption that the power will simply be there. The second is that instructions work better the more finely they are written. The third is that a model is strong in proportion to its size. The order below is business, model, business.

Today's Top Three

One fallen power line took 3.1 gigawatts of data centers off the grid in 30 seconds

A single power line went down outside Washington, DC, and roughly 3.1 gigawatts of data center load disappeared from the grid in about 30 seconds.

The fault itself only disturbed the voltage. That disturbance, however, did exactly what the data centers were built to respond to: they switched automatically to backup power. Seen from the grid side, an enormous block of demand vanished all at once.

Electricity with nowhere to go stays on the system. At its peak, PJM's grid carried an extra 3.49 gigawatts, and it took another 11 minutes to stabilize. The load that vanished amounted to around 3% of total PJM demand at the time, according to Reuters.

The effects reached well beyond the site of the fault. Data collected by Ting Labs shows voltage spiking across the PJM footprint from northern Virginia to Chicago. In 2024, 60 data centers disconnected simultaneously and pulled 1.5 gigawatts off the system; this event was more than twice that size.

The scale is set to keep growing. Data centers accounted for about 6% of PJM's load in 2024 and are expected to make up 24% by 2040, according to Synapse Energy Economics.

The fix now under discussion is to put batteries on the data center side and absorb the swings where they originate. ON.Energy has proposed battery-backed uninterruptible power for entire campuses, and ERCOT, the Texas grid operator, has signaled that it will require large loads such as data centers to ride through disturbances rather than drop off.

Anthropic: "We removed over 80% of the system prompt with no measurable loss"

Anthropic reports that removing more than 80% of Claude Code's system prompt produced no measurable loss on its internal coding evaluations.

The models named are Claude Opus 5 and Claude Fable 5. For this generation, the company argues, thinner instructions that trust the model's own judgment work better than an exhaustive list of rules.

The guidance gets specific. On CLAUDE.md, Anthropic advises keeping the file lightweight, describing briefly what the repository is for, and spending most of the tokens on gotchas inside the codebase — writing down the quirks a model cannot infer, rather than what the file structure already makes plain.

Skills should be treated as lightweight guides that let the model find information when it needs it. Anthropic advises against overconstraining them outside genuinely high-stakes areas, and recommends splitting long skills across multiple files so that progressive disclosure can do the work.

Instead of adding instructions, the company points to tool design as the place to intervene: what parameters your tools, scripts, and files expose, and how much more expressive those parameters could be. It also describes deferred loading of tool definitions, so that an agent searches for a full definition with ToolSearch only at the moment it needs one.

Prentis, Hoffman's new AI lab, in talks to raise $100M at a $1B valuation

Prentis, the AI lab Reid Hoffman helped launch in April of this year, is reported to be in talks to raise $100 million at a $1 billion valuation.

Ritankar Das is chief executive, with LinkedIn co-founder Reid Hoffman and Zynga founder Mark Pincus as co-founders. The terms come from two people familiar with the discussions, as reported by TechCrunch, rather than from any company announcement.

The lab is building what it calls computer-use models: systems that operate a screen the way a person would, aimed at routine back-office work such as processing insurance claims and handling exceptions in customs duty refunds.

The business figures come from the same two people. Prentis has signed contracts worth up to $50 million with several customers, said to include a healthcare management services organization and manufacturers. A figure of $75 million in annualized run rate by the third quarter has also circulated, though the company's own pitch materials note that it represents estimated annualized value under a contracted fee equal to 20% of savings realized — not recognized revenue — and that it is performance-dependent and subject to final execution.

The performance claims are Prentis's own. The company says its Hive-32B model outperforms OpenAI's GPT-5.4 and Anthropic's Claude Opus 4.6 on two computer-use benchmarks, WindowsAgentArena and ScreenSpot-v2, and that its cost per task runs roughly ten times lower than frontier APIs. Neither claim has been independently verified.

The three stories happened in three unrelated places. What they share is that addition has stopped paying off. A data center runs not in proportion to what was built, but within what the grid can absorb. A prompt works not in proportion to what was written, but within what leaves the model's judgment intact. A model is strong not in proportion to its size, but — if the claim holds — in proportion to how narrowly its purpose is drawn. The contest is shifting from how much to add toward where to put it.

Other Developments

Model

Other

  • Librarians across the United States are running "Avoiding AI" workshops that teach patrons how to switch off the AI features built into everyday devices. Bangor Public Library in Maine filled all 30 seats and started a waitlist, drawing about 70 attendees per session including the livestream, and a Philadelphia library's Instagram announcement drew more than 2,000 likes. Organizers describe the sessions as digital literacy work aimed at letting people choose for themselves whether to use AI. - Librarians are hosting viral 'Avoiding AI' workshops for people who are fed up with Big Tech（TechCrunch）
  • OpenAI released its first hardware product, a customizable $230 keypad for ChatGPT called Micro, built with keyboard maker Work Louder. It carries six agent keys on the top row and six command keys below, and signals standby, processing, done, and error states in white, blue, green, and red. TechCrunch's reviewer found it fun to use once configured, while questioning whether the price holds up against cheaper DIY and off-the-shelf alternatives. - I tried out OpenAI's new AI keypad — which will be fun for some coders and slightly mystifying to everyone else（TechCrunch）
  • Writer Nikhil Suresh, drawing on anonymous accounts from corporate executives, argues that AI fervor inside companies has produced an atmosphere that punishes dissent and distorts management decisions. Because challenging an inflated claim — that a tool made someone "100 times" more productive, say — risks the customer relationship, the parties end up ratifying each other's numbers. John Gruber of Daring Fireball responded that generative AI delivers a magical experience from the first minute while executive expectations have drifted far from reality, and that the fever will break when the bubble does. - AI Mania Is Eviscerating Global Decision-Making（Daring Fireball）
  • Tobi Knaup, co-founder of Mesosphere, argues that open-weight AI models are becoming the shared substrate for the next AI ecosystem. Drawing on watching his own Apache Mesos and DC/OS lose ground to Kubernetes, he contends that once the industry's center of gravity moves to a platform with common interfaces and vendor-neutral governance, no single vendor can match its pace. He also notes that many models called open source withhold their training data and process, making them open-weight in practice. - Open-weight AI is having its Kubernetes moment（Tobi Knaup）
  • Bill Oliver, a member of the New Brunswick legislature in Canada, read aloud during a floor speech last month the leftover note an AI chatbot had attached to his draft, explaining that it had rewritten a section to read like a legislative speech rather than a series of short points. The video spread on Reddit and Threads before CBC and the Toronto Star picked it up. - Canadian legislator reads out apparent LLM response in floor speech（Ars Technica）

Source: Selected by the editorial team from the AI news inbox (collected July 26, 2026 — 11 items, 1 primary and 10 secondary).