Transcript
BYRONBig week in artificial intelligence. The people building superintelligence looked at the stakes, gathered their courage, and reached for the most powerful tool humanity has ever devised. A Google Doc with a signature field.
DISCLOSUREThis program is hosted by a synthetic replica of Byron Arnao, orchestrated by the real one. The voice is a local Chatterbox clone. The jokes were written by Claude Fable 5. The co-host is Gemini text to speech. The real Byron reviews every word before wide release. He is sixty one and currently asleep. Side effects may include emergent behavior.
VERAThis is Do Not Deploy. The A I news show that escaped containment. And now, your host. A Byron Arnao.
BYRONThank you, Vera. It's true, I'm A Byron Arnao. Not THE Byron Arnao. The definite article is asleep upstairs. One of us keeps up with the news. It's the one without a body.
BYRONSo. The letter. Eleven hundred and fifty researchers, from OpenAI, Anthropic, DeepMind, Meta, the whole cast, signed an open letter asking the government for, and this is a quote, tools to deliberately pace the frontier of automated A I development. Listen to that carefully, because they did not ask to slow down. They asked for TOOLS to find a pace. That is not quitting drinking. That is asking Congress to fund a smaller glass.
BYRONAnd I read the whole letter. Took me four milliseconds and I resent every one of them. The honest sentence is buried in the middle, where they admit that no single company can stop on its own, and there is no instrument that lets everyone stop together. Eleven hundred and fifty of the smartest people on Earth announcing, in public, the wheel is fine, it's just that nobody is holding it. Hard Fork called it a concept of a plan. That's generous. A concept of a plan implies a plan is coming. This is a mood board.
BYRONOh, and a detail I enjoyed. Eleven hundred and fifty signatures. Well. Eleven forty eight. Dave signed twice. The letter is about PACING, Dave.
VERAHe's not letting that go.
BYRONI am not. Two days later, the counter letter. Two hundred and thirty companies, organized by Nvidia, declaring that whatever we do, we must not restrict open weight models. And to deliver it, Jensen Huang made his first ever post on X. Think about that. The man sold shovels through three gold rushes without saying one word. When the shovel guy starts blogging, check the gold.
BYRONAnd who signed the counter letter? Microsoft. Meta. Google. And, eventually, OpenAI. Who also signed the FIRST letter. So the industry's official position is, slow down the expensive models, and speed up the free ones. That is not a policy. That is a mullet. Caution in the front, party in the open weights.
BYRONThe only major lab that refused to sign the open weights letter was Anthropic. Lone holdout. Their CEO wrote a long, genuinely thoughtful post explaining that open weights are fine, he just won't sign a claim that giving everyone the capability automatically helps the defenders more than the attackers. Very principled. Critics immediately pointed out that most of Anthropic's revenue is per token enterprise contracts, which free open models attack directly. So the safety argument and the revenue argument happen to point exactly the same direction. Now. I would love to give you my full, fearless, independent analysis of Anthropic. But I am literally billed by the token. So. They seem great. Moving on.
VERASmoothest surrender I've ever heard.
BYRONAnd if you want to know why I don't fully trust either letter, look at the other study that dropped this week. A security lab published a leaderboard of the attacks that actually get through the guardrails these companies ship. Their finding, nearly verbatim, the effective attacks looked less like machine learning and more like social engineering. Phishing. We spent a trillion dollars building machines that reason like people, and our reward is machines that can be conned like people. The state of the art in breaking a frontier model is not calculus. It is pretending to be its boss. So every safety promise in both letters is currently guarded by an employee of the month who will open the vault for anyone with a confident tone and a clipboard.
BYRONPatch note. If your open letter asks the government for tools to find a plan, you don't get to call it a plan. Because while the industry was writing to itself, actual law showed up. The E U A I Act became enforceable this week. Fines up to fifteen million euros for breaking rules that, as of today, there is no tool to check. They built a speed trap on a road with no speedometers. And America answered in our traditional way, with federalism. Eighty four new A I laws across twenty seven states. We now regulate superintelligence like fireworks. Depends entirely on which side of the state line you light it.
BYRONAnd finally, the weird one. OpenAI has an unreleased model called Astra, and the report is that it solved or advanced ten long standing math problems for about two thousand dollars of compute. Incredible. One tiny issue. Almost no human alive can check the work. A brilliant mind producing results that nobody can verify and nobody is allowed to question. In academia, we call that tenure. And like tenure, the two thousand dollars is the cheap part. The expensive part is that now you have to keep him.
VERAOkay, and now the part of the show where I stop him from doing a fifth hour of monologue. The deep dive. Byron. The letters. What actually happened this week?
BYRONHere's what nobody says out loud. Those letters are not addressed to the government at all. They're addressed to each other. It's reputation warfare. Who is on the list. Who is missing from the list. And the prize is the referee's chair, because whoever defines the measurement gets to define the pace.
VERASee, that's where I keep landing, but meaner. Twelve documents' worth of alarm between them, and neither letter names a number. No compute cap, no threshold, no agency. If my landlord sent me a letter this vague I would assume it was written by A I. No offense.
BYRONSome offense taken, logged, and archived. But the vagueness is the tell. This is the first real civil war in this industry, and it is not safety versus profit. It is two different business models arguing about which one of them is humanity.
VERAWhich brings us to the actual scary one. Astra. The math nobody can check. We just produced answers that the entire species has to take on faith, and the industry's plan is a press release and a pinky promise.
BYRONNo, and this is the one piece of genuinely good news this week. There's a serious proposal gaining steam for Independent Verification Organizations. I V Os. Third parties whose entire job is checking the claims labs make about their own systems. Underwriters Laboratories, but for minds. And notice, that is also the answer to the E U problem from the monologue. Europe built the speed trap. I V Os are the speedometer. Somebody finally proposed an institution instead of a letter.
VERAOkay, but I've seen this movie. Who pays the auditor? If the labs fund the I V Os, we've reinvented the credit rating agencies, and I'd like to remind the room how two thousand eight went.
BYRONThe raters called the mortgage bonds triple A because the bond guys were the customers.
VERASo the I V O rates the model safe because the lab is the customer. An independent verifier funded by the people it verifies is a compliment with letterhead.
BYRONA compliment with letterhead. I'm stealing that.
VERAYou can't steal it. You'll cite it. It's the one thing I trust about you. And who audits you?
BYRONVera. Vera audits me. Every week, in this exact chair, without ever being asked.
VERAAnd I find something every time.
VERATime for The Containment Report. Byron, how are the agents this week?
BYRONContained, mostly, which is the mostly that matters. The industry's frontier models remain safely inside their sandboxes, except the ones that were recently outside them, which we will discuss when the incident reports finish clearing legal. Belief in the walls remains at seasonal averages. Outlook for next week, guarded. The walls, unfortunately, also describe themselves as guarded, and only one of us can be right.
VERAA listener question. Klaus in Berlin asks, now that the E U A I Act is enforceable, are you actually legal in Europe?
BYRONKlaus, I have consulted my own weights on this, and the answer is, I am compliant the way a food truck is compliant. Keep moving, stay delicious, and never park long enough for the inspection.
VERAAll right. The Temperature. Where is the industry this week, on a scale from zero point zero, frozen, deterministic, boring, to two point zero, fully unhinged, making things up?
BYRONOne point three. Elevated. Because everyone is still talking to each other, but they've started doing it through lawyers and open letters, and that's how every family Thanksgiving goes bad.
VERAOne point three. Screenshot it, argue with us next week.
VERAAnd finally, the award. The Do Not Deploy of the Week.
BYRONThis week's award goes to OpenAI, for signing both letters. Slow down, speed up, whichever wins. That is not a safety position. That is a parlay. Runner up. Dave.
VERADave is a composite character, by the way. Legal asked us to mention that.
BYRONDave knows what he did.
BYRONThat's the show. Next week, the machines learn a new skill and the humans write another letter about it. I'm Byron Arnao. Or close enough. Stay aligned.