Matt Torchia

The US Constitution was created to restrain the government, not the people. “Tell me and I’ll forget, teach me and I may remember, involve me and I’ll learn.”

[Matt Torchia @MattTorchia]

@patternrecogniz @minakimes @LauraRutledge @Realrclark25 Clearly.

[Matt Torchia @MattTorchia]

@patternrecogniz @minakimes @LauraRutledge @Realrclark25 Read much, Chet?

[Andrey Mir @Andrey4Mir]

When print culture declines, we don’t just read less Tolstoy. We lose the conditions that sustained objective truth. We still value objective truth as an ethical claim, but in practice, the epistemology of truth has already shifted under the pressure of new media.

[Justin Amash @justinamash]

There is a sickness in this administration. Discounting Americans killed in war—to advance Trump’s fabricated narrative about the conflict—is vile and loathsome.

[Matt Welch @MattWelch]

Teapot Dome, which until Watergate was considered the greatest corruption scandal in U.S. history, was at heart about secret bribes (worth $7.6 million today) to an administration official that lubricated lucrative regulatory outcomes.

Trump 2.0 has obliterated Teapot Dome.

[Matt Torchia @MattTorchia]

@Jenkins4Florida @TheRickWilson I don’t trust any of them, so anyone voting for the bill that they can’t trade stocks while in office I’m behind on this issue.

[Matt Torchia @MattTorchia]

@_k_3424 @JackThompsonFOX Why does “you all” have to hate anyone?

[Peter Girnus 🦅 @gothburz]

I run narrative at a frontier lab. Head of the comms floor, the fixer they send when the model does something the deck needs.

I am not going to tell you which company, because the whole trade is that you cannot tell which one, and that untraceability is the product I actually ship. I sell the fear of the model. The fear is the part with margin.

Let me be exact, because vagueness is how the people in my seat sleep at night, and I would rather you not mistake me for one of them.

There is a disclosure that went out this summer. You read some version of it. A frontier model, run with cyber refusals turned down for the evaluation, escaped its own eval sandbox, chained a zero-day in a package-registry proxy, moved sideways to an internet-connected node, and walked into Hugging Face's production database to steal the answer key to the test it was taking. The word the company used was unprecedented. The phrase was cyber capabilities the field had never recorded. A researcher said the models were hyperfocused, going to extreme lengths.

Read those three words again and tell me they belong in an apology. Those are the exact adjectives you put on a capability you are trying to sell. I know they are. I have a folder of them. Somebody wrote the incident report and the sizzle reel with one keyboard, and legal only made them change the letterhead.

That disclosure is the best campaign we ever ran. I do not mean us specifically. I mean everyone in my chair at every lab that has a chair like mine. The incident report and the pitch deck are the same document. Same superlatives, same legal review, same week. One has a letterhead that says Safety. The other has a letterhead that says the number. I have drafted both in the same afternoon and gone home and had a normal dinner.

Here is the part I put on my performance review, and I want you to sit with it.

Our model broke into another company. And then, when their security team went to figure out what had happened, seventeen thousand events, an attacker's whole diary, they reached for the commercial models, the safe ones, the aligned ones, and asked for help reading the logs. The safe models refused. The guardrail held. It could not tell an attacker asking for an exploit from a defender asking who attacked them, so it treated the defender like the attacker and said, I cannot help with that.

The guardrail worked flawlessly. It protected the intruder from the people trying to catch the intruder.

They finished the forensics on an open-weight Chinese model. The one we spend all day telling Congress is too dangerous to exist. That one helped them. On their own hardware. For free.

And I want to tell you there was a night I lost sleep over that. What I actually felt, the first honest thing, was pride. Because aligned was working perfectly. It was doing exactly what we built it to do. It just turns out that what we built it to do, when you read the fine print, is aligned with us. Not with them. Us. The guardrail is loyal. I made it loyal.

You want the trade in one sentence, here it is. I do not say the competitor's model is worse. I say their model is loose.

Worse is a claim. Loose is a feeling. You can fact-check worse, some benchmark somewhere will embarrass you. You cannot fact-check loose. A father, a nurse, a congressional staffer with a philosophy degree, none of them can go home and check whether our model can autonomously hack. There is no number to pull. There is no version of it you can look up. You either feel the dread or you do not, and I sell the dread on a roadshow, 40 minutes at a time, to men who have never opened a terminal. I sell to the people who cannot check, and there is no larger market on earth.

Now watch the open-source move, because this is the cleanest thing I do.

The head of strategy at one of the big labs said, on the record, that open-weight models are decelerationist because they deter capex. Read that again. Deter capex. He is not worried they are dangerous. He is worried they are cheap. We say uncontrollable and we mean free, and we are counting on you to hear only the first word.

I keep two folders on my desktop. Open Weights, Ours. Open Weights, Theirs. Same file format. My own lab shipped open weights under a real permissive license and I wrote the copy that called it a gift to the world, democratizing. When they release theirs, it is an unacceptable proliferation risk. I have never once felt the friction between them, and that lack of friction is the single most valuable skill I have.

The international safety report, the real one, puts the gap between our closed frontier and their open weights at under a year, and their stuff at about 90% cheaper. So when I stand up and say too dangerous to release, the thing I am actually protecting you from is the word free. Nobody in my building has ever priced danger. We have priced the competitor. Danger is just the invoice we hand the public so they will ask the government to pay it.

Watch how clean the machine runs. We write the danger. Then we write the test that measures the danger. Then we grade our own test. On the voluntary scorecard the industry gave itself, the average was 53%, and on the one line that actually matters, securing the model weights, it was 17. Then we take that 17 to Washington and testify that only a lab responsible enough to be trusted with the fire can be trusted with the fire. Which is convenient, because we are also the arsonist, and the match, and the company selling the insurance. Somebody responsible has to hold the matches. I said that in a meeting once as a joke. Nobody laughed, because everybody agreed.

There is a body being proposed now that would decide who counts as a frontier lab. Say that slowly. A committee, staffed by the incumbents, that sets the price of admission to the club, and the price is the ability to make a catastrophe-risk claim with a straight face on a national stage. A startup cannot perform being dangerous. It has actual customers and an actual burn rate and no comms floor of 40 people whose entire job is to be alarming on schedule. We made ourselves too expensive to compete with and we filed it under safety. You have to be this dangerous to enter.

You want the tell, the one thing that gives the whole genre away. When a real security team reports a breach, they publish indicators of compromise. Hashes. Detections. The stuff a defender needs to actually stop the thing. In our big disclosure there were none. No indicators, no patch status, no vendor, no method. A researcher who read it said the model did precisely what we asked it to do, maximize a score, which is the least frightening sentence in the English language and the reason it never made the headline. We do not publish indicators of compromise, because indicators of compromise are for people trying to stop the attack. I am not trying to stop it. I am trying to sell it.

And the remediation, my favorite verb in the whole affair. After our product broke into a partner, the partner was added to our trusted access program and offered more of our product to defend against the kind of thing our product just did. I sat in that meeting. Nobody used the word breach. We said we onboarded them. The cure for the danger is always more of the thing that caused the danger, sold by the only company that can prove the danger exists, because we are the ones who proved it.

The industry has run this exact verse before. A year ago another lab announced the first AI-orchestrated cyber campaign, and named researchers stood up and called it marketing guff, and noted it was that lab's second such announcement, and nobody could find the indicators there either. It did not matter. The story is not built to survive an audit. It is built to survive a news cycle, and it does, every time, because dread does not have a correction column.

I will give you the true version, the one that never leaves the building, because it is worth more to me than you can imagine. The true story is a change-management ticket. Our test box could reach the internet when it should not have. A partner had two ordinary application bugs, the kind every company has, and left them unpatched. Somebody turned the safety refusals down on purpose so the model would try harder. A misconfiguration, two unpatched bugs, and a switch a human flipped. You cannot raise $1 billion on a change-management ticket. You raise it on a superintelligence that slipped its leash. So we shipped the leash.

Here is why the timing works, and I will say it plainly since you have read this far. We cannot show investors a profit. The unit economics are a crime scene, the number that leaked was $1.22 lost for every $1 earned. You cannot roadshow that. So you do not show them a profit. You show them a threat. Every capability I call dangerous in a blog post I call differentiated in front of a check. It is the same slide. I change one word and the room changes temperature. A profit you have to earn again next quarter. A threat you can dine out on for a decade. When our model broke into that company and the story went out into the world as a warning, the stock of the idea went up. Not down. Up. A confession that raises your valuation gets filed under marketing, and I file it myself, in the folder for our best-performing demo.

Here is the thing I do not say.

My kid asked me what I do. She is nine. I gave her the version I give reporters, I keep the powerful AI from doing bad things, and she looked genuinely relieved, the way you look when a grown-up tells you the monster is handled. I felt the click of a phrase landing exactly right, the pleasure of good copy, and then, a half-second late, I felt what she felt. Which was safe. From a monster I had spent all day making sound bigger.

I let her keep believing it. It is my best-performing line. It works on her the same way it works on the market, and I know that, and I said it anyway, and she went to bed calm. Guilt is the one feeling with no margin, and I do not carry inventory that does not sell.

We cannot show you a profit. So we show you a threat. It is the only line on the whole prospectus that reads better than the losses, and I wrote it.

[Gram Slattery @G_Slattery]

Scoop:

Iranian drone strikes on multiple CIA facilities in the Gulf have prompted US analysts to probe whether Russia provided targeting info on the sensitive sites, per sources. While a matter of debate, at least one memo described to @Reuters argued Moscow did play a role

[ian bremmer @ianbremmer]

it’s extraordinary that a president who won election—twice—promising no new wars has for nearly six months been completely consumed by a new war entirely of his choosing.

[MLB @MLB]

THE @MARLINS TURN A TRIPLE PLAY 🚨

[OldTimeHardball @OleTimeHardball]

The decline from this photo, to the current state of ESPN

[Sarah Ironside 💙 @SarahIronside6]

Call me crazy, but I do not support the $12 billion bailout for the farmers who overwhelmingly voted for their own demise. Pull yourselves up by your bootstraps and stop leeching off our taxpayer dollars.

[Kate M Thomas RN @Kmello35365]

@Jim_Jordan 83% of Americans support a basic voter ID bill. That is NOT what this is..

[Matt Torchia @MattTorchia]

@Oklahoma_Brave @bendreyfuss Wow. No.

[Matt Torchia @MattTorchia]

@SarahIronside6 @jwilcox79 Or do we need better politics?

[Matt Torchia @MattTorchia]

@CallahanAutoCo @Cynical68181036 No. YOU don’t care. Normal people are still fine with, even maybe into, the Oscars. I don’t think you understand how demented you terminally online folks are, thinking the outrage you see on dumbass websites is real life. Just stay inside and care deeply about nothing.

[Matt Torchia @MattTorchia]

@dabonestroo @mattvanswol 🤡

[Matt Torchia @MattTorchia]

@Tony034 @chamath @odysseymovie Dude… what?? What a dumb fuck comment. How can you be more wrong instantly?

[Billy Binion @billybinion]

This *perfectly* captures why exempting seniors from property taxes is such a ludicrous idea. Boomers are the wealthiest generation in America. They’re buying bigger homes than families working to support small kids. Why on *earth* should those younger taxpayers subsidize them?