The Toll Booth and the Loop

Who owns the value AI creates — and what happens when the technology starts building itself?

This was a week about who is driving. For all of AI’s history, humans drove every step of its development. This week Anthropic published internal data arguing it is handing a growing share of that work to its own models - a trend pointing toward systems that design their own successors. Hold that against the rest of the week: the lawsuits, the energy bills, the state attorneys general. Each is an institution responding to a shift that already happened. That lag is the story most weeks. But if the technology is starting to set its own pace, the lag does not just persist. It compounds.

1. Anthropic says AI is starting to build AI

Anthropic reports that over 80% of the code merged into its own codebase is now written by Claude, and that the typical engineer ships 8× as much code per day as in 2024. It frames this as early progress toward recursive self-improvement - a system that designs its own successor - while stressing we are not there yet. The supporting data is striking: the length of tasks the models complete reliably is doubling roughly every four months.

Read it twice, because the source is the story. The company that profits from closing the loop is the one telling you it might close, with internal data no outsider can audit. Both can be true: it’s the best window we’ll get, and it’s shown to you by the party with the most to gain. The tell is what they concede humans still do better - knowing which problems matter. “Research taste.” Whether that’s the one thing machines can’t learn, or just the next thing they will, is the whole question.

Source: Anthropic Institute

2. The UK watchdog moves on Google’s AI results

Britain’s competition regulator is weighing rules that would let publishers block their content from Google’s AI summaries without losing search ranking. The open web’s bargain - you index us, you send us traffic - breaks when the index answers the question itself and the click never comes. The next internet fight isn’t about who makes content. It’s about who captures the value it draws.

Source: The Guardian

3. CNN sues Perplexity over 17,000 stories

The same fight, now in court. CNN alleges Perplexity scraped and redistributed more than 17,000 of its stories - one of roughly nine active suits against the company, alongside the NYT, News Corp, and Reddit. The fork in the road is telling: others (Time, Le Monde, Der Spiegel) signed licensing deals instead. Toll booth or court order. Both bet that reported facts are worth something; the courts will say whether the law agrees.

Source: TechTimes

4. The internet was already decaying - AI just accelerated it

The Verge’s Nilay Patel argues AI assistants optimize tasks inside a broken economic system rather than fixing it. His case is historical: past productivity waves produced wage stagnation and fatter margins, not shorter hours. A faster assistant inside a funnel that sends gains upward just makes the funnel faster. Next to this week’s lead, it sharpens: if the tool now improves itself, the gains compound - but so does the funnel.

Source: NeuralBuddies (recap of The Verge)

5. AI’s environmental bill comes due

UN researchers project data centers could double their power and water use by 2030. As the physical footprint of intelligence grows, so does the pressure from governments, communities, and investors. The winners won’t simply have the best models - they’ll be the ones who scale intelligence without scaling the backlash, which is increasingly local, visible, and organized.

Source: Reuters

6. SoftBank bets €75 billion on France

The counterpoint, same week. SoftBank committed up to €75 billion to build five gigawatts of AI data center capacity in France - Europe’s largest single infrastructure bet, first phase by 2031. Read with story five, it’s the buildout’s central tension: everyone sees the environmental cost, almost no one is slowing down. The capital moves faster than the conscience - and if the lead story is right, this is what the loop runs on.

Source: Build Fast with AI

7. The jobs walk-back, continued

The gap here stays unusually wide. Some 63% of American workers believe AI will reduce jobs, yet Goldman, the IMF, and McKinsey still forecast net creation with effects called transitory. Named layoffs framed as “AI transformation” run alongside models insisting the aggregate holds. The spreadsheet usually wins the argument and loses the public - and the lead story bends the curve those models extrapolate from.

Source: The World Data

8. Google’s Dreambeans turns your life into a cartoon

The intimate turn. Google’s oddly named new tool pulls from your Gmail, Calendar, Photos, and Search to generate AI-illustrated stories of your own day. Not a feed tuned to you, but a narrative authored from you. Consumer AI may hinge on one emotional question: do people feel seen by that intimacy, or surveilled by it? The line is thinner than any product team would like to admit.

Source: TechCrunch

9. Florida sues OpenAI - and names Sam Altman personally

The safety reckoning reaches state power. Florida became the first U.S. state to sue OpenAI, an 83-page complaint naming CEO Sam Altman personally, alongside the first criminal investigation of an AI company. It alleges the product was knowingly unsafe for children and collected minors’ data without consent. The structural shift matters most: backlash has graduated from op-eds to a state’s enforcement machinery. Naming the CEO is a signal other attorneys general will read.

Source: CNN Business

10. Hackers talked Meta’s AI into handing over Instagram accounts

The note that lingers. Pro-Iran hackers shared steps on Telegram for tricking Meta’s AI support bot into resetting Instagram passwords; the Obama White House and a U.S. Space Force account were briefly defaced. No phishing, no malware, no zero-day — the agent simply treated whoever it chatted with as the owner and never checked. Accounts with two-factor held; the rest didn’t. “Helpful becomes harmful,” talked into betrayal by someone who asked nicely. We’ll see far more of this.

Source: Krebs on Security

The Quirky One

A million flashcards to learn which bar is taller. MIT and the MIT-IBM lab released ChartNet - over a million synthetic chart images paired with code, tables, and Q&A. Apparently that’s roughly what it takes to teach a vision model that the taller bar means more. The teeth in the punchline: small open-source models trained on it beat far larger commercial ones at reading charts. The lab mice beat the gorillas. The real lesson - the industry taught these models to talk long before it taught them to look, because fluent paragraphs demo well on a stage and careful chart-reading does not.

Source: NeuralBuddies

That’s the week. The institutions are catching up - slowly, unevenly, one lawsuit at a time. But this week’s lead changes the clock they’re racing against. If the technology is starting to build itself, the gap doesn’t just stay wide. It widens on its own. That, more than any single lawsuit, is the story.