Signal #79: Three men asked for AI Brakes

Dario Amodei spent the weekend asking the industry to slow the frontier down. Sam Altman agreed, Elon Musk agreed in three words, and by Monday the President of the United States had called the whole thing a hoax live on a podcast. Also inside: Porsche's billion-euro exit from Rimac, the rotation hiding inside the chip selloff, and the week Europe spent arguing about something else entirely.

Morning. I have spent a good part of this year telling clients that waiting for the next model is not a strategy, and this week the people who make the models said something close to the opposite.

What makes it interesting is not the argument itself. It is who answered, how fast, and what the answers reveal about where the speed of this thing is actually set. Four stories, and the through-line is the same in all of them: the people building the frontier do not control its pace.

And sorry for sending this late today, but I had work to do.

Amodei published We Must Pace the Frontier on 12 September, and the concrete part is stranger than the headline. He wants third-party evaluators inside frontier labs with "desks in our offices, access badges, and company laptops," with contractual rights to publish findings Anthropic cannot redact for being unflattering. He wants capability checkpoints, where a model that can defeat common sandboxing has to arrive with alignment certifications attached. And because coordinating that across rivals looks like collusion, he wants Washington to grant a narrow antitrust waiver so the conversation is legal at all. His stated trigger is recursive self-improvement accelerating since the summer, plus a worry that within 6 to 12 months an agent swarm could hold a persistent botnet across the internet.

Altman matched the evaluator pledge within hours and went further on legislation, asking for a federal safety framework and warning that humanity could "lose control of the future to AI", which he called unacceptable. Musk quote-posted three words, "Dario is right," and committed xAI to nothing. Demis Hassabis called the direction correct and the details unresolved.

Then the refusals landed, and they came from everyone with real power over the outcome. Trump phoned into Jensen Huang's All-In interview on Monday and called the risk "a hoax", with Huang answering on stage, "We're not going to let that happen, sir." David Sacks called the proposal a cartel and told the labs they can simply slow down without asking permission. Beijing's foreign ministry dismissed it as confrontation dressed as governance. Watch whether any lab actually seats an outside evaluator before Christmas. That is the only part of this that does not need anyone's permission.

While that argument ran, the region's most valuable engineering asset changed owner class. Porsche completed its exit from Bugatti Rimac and Rimac Group on 11 September, selling its 45% and 20.6% stakes for a figure BeBeez puts at roughly EUR 1bn. Abu Dhabi's BlueFive Capital took 30% of Bugatti Rimac with one board seat and two observer seats. A HOF Capital consortium took the remaining 15% and lifted HOF to 23.5% of Rimac Group, with three supervisory board seats across the two companies. Porsche put EUR 250m of the proceeds against pension liabilities and raised its cash flow margin guidance.

Mate Rimac still runs it, and the Sveta Nedelja campus still builds the hardest electric drivetrains in Europe. What changed is who sits in the room when the next capital decision gets made. A German carmaker with an industrial logic sold to Gulf and US financial capital with a returns logic, and the seats came with it.

This is the same question as the pacing argument in a different register. Amodei can propose a speed limit and discover that the people who set speed are elsewhere. Croatia can build a champion and discover that the champion's direction is set from two cities it has no relationship with. For anyone in the region raising a serious round this year, the term sheet clause worth fighting over is not valuation, it is board composition. Read the Rimac seat count as the going rate.

Monday was ugly at the silicon end. Nvidia closed down 3.36%, AMD 4.40%, Broadcom 4.77%, Intel 5.59% and Micron 5.25%, with the Philadelphia Semiconductor Index off 5.9%. Asia took it harder, with SK Hynix and Samsung leading the KOSPI down and SoftBank falling more than 10% in Tokyo. Be careful attributing all of that to Amodei, because the same session carried an oil spike and a Fed meeting with a hike priced in.

The honest evidence sits elsewhere in the same close. The software ETF IGV finished up 5.04% and the cybersecurity ETF CIBR up 5.99% on the day the chips broke. That is not risk-off, that is a rotation out of training capex and into whatever gets sold if somebody has to prove a system is safe. Altman ruling out a 2026 IPO because it would be "an ill-advised moment to go public" given safety reads the same way.

Private capital had already made the trade. AIUC closed $40m led by Ribbit on Tuesday to certify AI agents against roughly 5,000 tests, producing hundred-page reports, with Cursor, Lovable, Harvey and ElevenLabs already customers. If you sell agents into a Western enterprise, the audit report is becoming the thing procurement asks for before the demo. Build it now or explain later why you cannot produce one.

The Commission said nothing about pacing. Berlin was reported to have called a slowdown not a viable option for Europe, and no serving EU official put a position on the record at all. What Brussels did do was hold the first council of the EU-Ukraine Drone Alliance on 11 September, eighteen companies split evenly between EU and Ukrainian firms, chaired by Andrius Kubilius. Fire Point told the room it wants an interceptor demonstrated by mid-2027.

The useful detail is what Kubilius named as the constraint. Not capital, not engineering: testing slots, certification, and supply chains that run through third countries. He asked industry outright to name its bottlenecks. In the same week von der Leyen used the Paris space summit to push an IRIS2 expansion funded partly by private capital, and her State of the Union lands on 16 September.

So Europe has already answered the question the Americans are arguing about, just not out loud. It does not regulate pace at the model layer, it regulates it at the certification layer, which is where its own builders sit. That is a policy choice with a bill attached, and the bill falls on European hardware companies waiting for a test range while their competitors ship. Watch the State of the Union for whether certification capacity gets money or another working group.

For the engineering manager. GitHub added cost and quality tiers to Copilot's automatic model selection on 14 September, with Efficiency, Balance and Intelligence drawing from the same model pool per prompt. Read the small print: the tier steers which model answers, it does not cap what you are billed. Set it per team and look at the invoice in two weeks before you decide it saved anything.

For anyone quoted frontier API rates. Salesforce and NVIDIA post-trained an enterprise reasoning model called Koa on the open-weight Nemotron-3-Super-120B base, using public and synthetic data with no customer data, and claim it beats a strong proprietary baseline on tool use and agentic reasoning. The paper is 16 pages and the recipe is legible. If your workload is narrow and repetitive, this is the template for owning it instead of renting it.

For the legal and brand side. Vilnius-based EnforceShield raised EUR 1.7m led by Vendep to automate IP enforcement, claiming removal of more than 33,500 violations a month for twenty-plus enterprise customers, most of them Lithuanian consumer brands. The interesting part is the customer list: a local e-commerce cluster mature enough to fund its own tooling layer.

For founders in the region. TesterArmy, out of Warsaw and San Francisco, raised EUR 1.04m from Y Combinator with angels including Vercel's Guillermo Rauch, for agents that test software. Fifty-plus customers, a team of six, and a founder who was still in secondary school when YC accepted him. Polish QA sold as product rather than as headcount is the margin structure worth copying.

For anyone in Bucharest-Ilfov. The regional agency's EUR 10m innovation clusters call is in public consultation until 21 September, with grants of EUR 300,000 to EUR 2m. The gate is tight: you must be a cluster management entity holding an ESCA excellence label. A second call worth EUR 25m for SME digitalisation follows in October. If the guide is wrong for your cluster, this week is the only week your comment counts.

Next edition soon,

Çelik