Today's Headlines

  • Opposition to AI data center construction is surging across the US — active groups roughly doubled from 396 at the end of 2025 to 833 by the end of Q1 2026, spanning 49 states
  • Apple's failed self-driving car program is reported to be the source of its Neural Engine and M7-generation chips
  • Programmer George Hotz (geohot) writes that he loves LLMs but hates the AI industry's hype

It is a day when questions, not events, hold the front page.

There were no big announcements. Instead, the writing gathered around the frictions the AI boom keeps running into — residents pushing back, the assets a failed hardware bet left behind, and the gap between the marketing and the reality. Rather than the momentum forward, the day was about checking the ground beneath it.

Today's Top Three

The Backlash Against AI Data Centers Spreads Nationwide — Opposition Groups Double in Half a Year

Grassroots opposition to building AI data centers is spreading rapidly across the United States.

According to Data Center Watch, a research project backed by the AI security firm 10a Labs, the number of active opposition groups more than doubled from 396 at the end of 2025 to 833 by the end of Q1 2026, now spanning 49 states. The project also says more than 235,000 petition signatures were gathered in that quarter alone. From January through March, it reports, protesters blocked or delayed at least 75 projects valued at $130 billion in total.

There is movement at the federal level as well. Senator Bernie Sanders (D-VT) and Representative Alexandria Ocasio-Cortez (D-NY) introduced a bill that would pause construction of new AI data centers until Congress passes laws to prevent the facilities from raising utility prices or harming the environment. Alongside this AI Data Center Moratorium Act, a Ratepayer Protection Act — which would codify an agreement for big tech companies to pay their own data-center energy costs — is drawing bipartisan support.

States and localities are regulating in their own ways too. Per a tally from the policy outlet Tech Policy Press cited by The Verge, both Democrat- and Republican-led states have enacted 28 laws related to AI data centers: Florida introduced rules to stop cost pass-throughs to residents, Idaho placed restrictions on water usage, and Washington State removed a tax break for the companies operating the facilities.

Even so, the response remains a patchwork of individual projects and local rules, and the federal bills have yet to become law. The Verge notes that this patchwork alone is not enough to rein in the buildout, leaving many communities to fend for themselves.

Apple's Failed Self-Driving Car Leaves a Legacy of Powerful AI Chips

Apple's self-driving car program, which never really came together, is reported to be the source of the AI performance in the company's signature chips.

According to The Verge, drawing on Bloomberg's Mark Gurman in his Power On newsletter, Apple realized early in developing its self-driving platform that it would need powerful on-device AI processing done in real time. The car processor itself was never finished, but that research led to the development of the Neural Engine, the backbone of Apple's on-device AI.

The Neural Engine made its debut with the iPhone X and the A11 Bionic, used at first mainly for computer vision. Apple later extended it to the desktop through its M-series chips, establishing itself as an early mover in on-device AI.

By Gurman's account, Apple is skipping the Pro, Max, and Ultra versions of its upcoming M6 chip and accelerating the M7 instead. The M7 should arrive in the first half of 2027 with significant Neural Engine upgrades, and the higher-end M7 Ultra is expected to be the basis for a new server product supporting up to 1.5TB of RAM. The technical assets of a failed hardware bet, in other words, have been redeployed into the performance of Apple's core AI products.

geohot: "I Love LLMs, I Hate Hype" — Taking Aim at the Industry's Overselling

Programmer George Hotz (geohot) wrote on his blog that he loves LLMs as practical tools while hating the AI industry's hype.

He first praises LLMs as down-to-earth productivity tools — like find-and-replace, Stack Overflow, or all the regexes he never bothered to learn. He writes candidly that he is genuinely excited about progress in self-improvement and coding agents.

He then names two things he dislikes. One is hype built to make you anxious and to make you feel bad about yourself — the "you're falling behind," "the window is closing" variety.

The other is the strawman leap from LLMs as "fancy autocomplete" or "a smarter compiler" straight to a world-changing superintelligence. That, he says flatly, will not happen.

At the core of his argument is the view that AI's progress is happening mostly because of Moore's law and general advances in computing, not because of anything the frontier AI labs are doing. Those companies, he contends, have an incentive to keep you from realizing this — because if you did, you "might not want to give them billions of dollars." That, he says, is why he is skeptical of their valuations. He adds that much of the code AI produces is still slop, and that the wave of new software those productivity gains should imply has yet to appear.

More to Note

Research

  • Reinforcement learning pioneer Richard Sutton argues that one-step world models — widely used in POMDPs, Bayesian inference, and elsewhere — are broken in practice despite their theoretical appeal. One-step errors compound into large long-term prediction errors, and in stochastic environments the future branches exponentially and becomes computationally intractable; as a fix, he proposes temporally abstract models built on options and generalized value functions (GVFs). - The One-Step Trap (In AI Research) (incompleteideas.net)

Models

  • Ploy, an AI marketing automation platform that autonomously builds websites, reports migrating its production agent from Claude Opus 4.8 to GPT-5.6, cutting build time from 8 minutes to 3 minutes 42 seconds and reducing cost per build by 27%, according to its own blog. The post also describes migration work such as handling incomplete tool-call parameters and restructuring its prompt-caching strategy. - Migrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper (Ploy)

Other

  • A technical analysis comparing coding agents' token usage finds that Claude Code sends about 33,000 tokens on its first turn from its system prompt and 27 tool definitions alone, versus about 7,000 for OpenCode's 10 tool definitions — a gap the author attributes mainly to the size of the tool definitions. - Claude Code sends 33k tokens before reading the prompt; OpenCode sends 7k (Systima)
  • An essayist, drawing on Folk Computer — a physical computing environment programmed by manipulating paper and gestures — argues that paradigm-shifting research needs stubbornly useless inquiry funded by patient patrons rather than market pressure. Once a prototype becomes useful, the writer says, its creators are pulled into serving existing users and lose the freedom to explore new futures. - Against Usefulness (Motive Notes)

Source: Selected by the editorial team from the AI news inbox (collected July 13, 2026 — 10 items: 0 primary, 10 secondary).