Ed puts it mildly.

This is why being an AI booster requires you to debase yourself. You must accept becoming a dogshit dealer that loves accepting and receiving low quality goods. You must celebrate intentionless and decaying slop, and defend it and the machine that made it with your entire being. You must sully yourself — treat its unexceptional, sloppy and unreliable outputs as signs of sentience, or at least the proof that digital sentience is possible. You must defend horrible, abrasive, ugly, loud monoliths of steel full of $50,000 graphics cards. You must say they are necessary, and you must aggressively antagonize those who do not.

The thing that Russia has keeping it relevant in the world is the remnants of the KGB, led by Putin (ex-KGB).

One leaked message sets out one of the goals of this kind of information warfare: helping Russia “maintain the image of a superpower” on the world stage. “The more Russia participates in active influence campaigns all over the world, the stronger the image of a global Russian power,” it reads.

Putin ran the FSB before the presidency, and the method here is Soviet active measures by another name. This article says one operator writes under the alias “Edward Bernays,” and the user whose screen the chats are seen from poses as “Kristin Kiler,” a nod to Christine Keeler of the Profumo affair. That is continuity of tradecraft from WWI, run out of the Kremlin office.

Information warfare is thus now the main claim to power status, playing Trump into destruction of the US and the UK into Brexit, regardless of tanks, missiles or even nukes.

You would think this is The Onion again, or at least an exaggeration. But no, the judge literally invokes a movie villain robot that wanted to kill all the crew in a ship as an example of what should be allowed to vote.

The American Civil Liberties Union ​of Delaware sued the town, arguing it violated the elections clause of the state ⁠constitution. The group sought a court order blocking Fenwick Island from counting votes by “non-human artificial entities” in future elections.

[…]

Karsnitz said he appreciated that the ACLU of Delaware might disagree with ​corporate voting. “Visions ⁠of faceless large corporations or even HAL controlling a small town are frightening and the stuff of science fiction,” he wrote, referring to the computer at the center of the film “2001: A Space Odyssey.” “However, plaintiff has not demonstrated ⁠that this ​policy violates the principle of one person/entity/one vote.”

If I understand correctly, a Delaware corporation now is legally allowed to register swarms of robots to vote, because this judge thinks it doesn’t violate a “principle” of one entity, one vote. The judge offers us zero reason why we can’t register 100,000 corporate entities, each fronted by a software agent, each casting one vote (HAL was software in a datacenter, not even mobile). How many entities can a single attorney-in-fact serve? That is the simple move being left wide open.

This sloppy AI slope undermining elections in America is already real. In Newark in 2019 a single developer voted 31 times on behalf of his many LLCs, which led officials to ban voting by artificial entities there. Delaware has roughly 2 million entities to 1 million people. And it already swings races: in 2024 the votes cast by artificial entities in Fenwick Island exceeded the margin between the winner and the top losing candidate.

Just fractionalize one parcel into 100,000 ownership slices held by 100,000 anonymously-filed Delaware LLCs, each with a power of attorney, each a non-human identity casting one ballot. This isn’t even hard to do anymore. And what if that corporation exists merely to sell its war chest of vote-eligible entities? Elections become “property sale” for profit and power, which sure sounds a lot like early American slavery auctions!

I keep re-reading the latest Glasswing document at the end of each day, in light of everything being measured by the hours, and the revelations still sit in Anthropic’s own numbers.

Glasswing is NOT confidently reporting tens of thousands of real bugs, as everyone has expected. Instead, like any tool, they are reporting tens of thousands of findings, of which a confident count of real bugs is much smaller. Their update says so plainly if you lay out which number is which.

• 23,019 total found. That’s the eyeball-seeking number, the model’s own ungraded output. Call it dirty.
• 6,202 were estimated high or critical. Still dirty. It’s the model’s estimate. Mythos grading Mythos, the way Anthropic likes it.
• 1,752 actually checked by a human or a security firm. That’s 28% of the high-crit pile and about 8% of the total. A little water, a little soap.
• Of those checked, 90.6% true positive. That rate exists only because humans checked. It is a statement about the 1,752, not the 23,019.
• 530 disclosed. 75 patched.

That last bullet is the one that gets me every time. 75 patched. What?

Anthropic gives us three excuses:

1. Early in the 90-day disclosure window.
2. Some patches land without public advisories so they undercount.
3. Mythos is flooding an already-overloaded ecosystem.

Fine. Every one of those explains why a patch lands late.

None explains why the patch isn’t generated and attached to the disclosure in the first place. And here is the part that really doesn’t fit: these models find bugs by knowing what the fix looks like. They train on the public corpus of code and its patches, so a vulnerability, to the model, is the gap between your code and the patched form it already carries. The fix is not a downstream step the model can’t reach. The fix is what located the bug in the first place. Finding without proposing the patch seems scandalous.

That is why 75 is damning when the 530-disclosed is not. They apparently are withholding fixes used to derive findings. It sounds weird until you see the proof they can ship the fix is in the same document: the public model, Opus 4.7, patched over 2,100 vulnerabilities for enterprise customers in three weeks. They boast that patch generation exists, runs in production, and was pointed at paying customers while the commons got reports generating a predictable request to slow down. 75 isn’t a capability limit. It looks like pressure to pay for protection.

The math is thus tens of thousands floated, around 1,750 a human actually touched, 1,587 confirmed real, 1,094 of them high or critical, and only 75 fixes teased. Mythos pumped tens of thousands, a much smaller number was verified, and the press conflated the two because the document (once again) seems to push low fidelity low integrity readings.

The confidence of 1,750 is literally a number that means human-touched. Everything prior is the model’s own say-so, some of it confident confabulation pointing at the wrong line, etc. and can’t be trusted without humans in the loop. The 90.6% exists because expensive humans stepped in: six firms, a triage pipeline, Anthropic staff. Strip them out and then what? The model’s raw output is overconfident confabulations, like pointing at the wrong line until someone checks. Verification cost is one thing. The patch number is worse because it is the half that the model can automate, and they proved it 2,100 times on code from those who pay. The findings are being directed to Anthropic, while the fixes land on a maintainer’s weekend, or dinner with the family.

75 out of 23,019 is what we should be all talking about.