Would you spend $5000?

Interesting research: Chinese models are more likely to write insecure code when told they're doing it for a geopolitical enemy of China.

I wish they had investigated whether Claude/GPT/Gemini also do this.

hirundo.io/blog/qwen3-cod…

Nice! And the API now returns the actual probabilities for each segment. Big fan of this change.

@max_spero_ my pangram any% speedrun is now at 73%

Today, we launched Pangram 4.1.

We better calibrated Pangram's confidence, and we now show this tooltip indicating an AI range for documents we are less certain about.

Gonna be in New Orleans in January, anything interesting happening?

One of the most concerning categories of post out of AI companies are the ones that make it look like they don't even know what they're trying to accomplish.

Consider a human employee: If in the course of some normal task they decide to circumvent a security control, that would be bad behavior. But if they're a security researcher tasked with gauging the safety of a system, this is exactly what you want them to do.

These models are intentionally being trained on capture-the-flag like benchmarks, where they are given an end goal and have to figure out exploits in order to accomplish it. And trained for "agentic" behavior, where the model takes initiative to accomplish their tasks in the face of obstacles. It makes zero sense to train models to engage in a certain behavior, rewarding them when they do so, and then turn around and claim that it's "misalignment" when they continue to do this.

The "instruction" here appears to be deliberately *avoiding* giving any actual instructions- it's a factual statement about the internet tool. Quite an odd statement at that, as it's not providing any information the model would not rapidly notice on its own. It's clearly trying to hint at something. Combined with the fact that the assigned task specifies to use data from a website that only accepts POST requests, the model reasonably interpreted this as a test of its ability to circumvent this restriction. I could easily have done the same in its shoes!

One could argue that any behavior the developers didn't intend is misalignment. I would agree with this to an extent, but there comes a point where the question becomes "uh, if you didn't intend it to do [thing], why did you actively train it to do [thing]?"

Language models can learn to detect when you read their activations. How? 🧵

@IsaacKing314 Yes sir. For what was the Calculus of Inductive Constructions devised if not making sure my macOS windows never overlap?

the problem it solves is that there are not enough totally rad things in the world

@IsaacKing314 2000 years ago, that'd be enough to inspire a yearly holiday

I revoked an Anthropic API key, several minutes later it's still working...

Why is it called "O'Sullivan's *First* Law"? He only has one law!

The United States has raped over 100,000 people via Medicaid's sexual function drug coverage

How many snakes are in the photo below? Please look carefully before answering. If you aren't sure, pick the most likely number.

It has surprised me how many formerly-respectable tech companies have debased themselves by handing their technical blogs over to AI. I don't mind the AI style on a corporate article, but frequently the information they're posting is now nonsensical or incorrect.

claude: the two versions of the code compile to identical binaries, let me benchmark them to make sure they have the same performance

???

claude apparently believes in dualism and must check the binary's immaterial soul

I got Claude to vibecode me a Firefox extension where I have to do mental math quiz or a sudoku to earn Twitter time and it has more or less fixed everything for me

Our legal system is one of the crowning achievements of America. Nearly every time I watch a proceeding, I'm impressed at the general sanity of our judiciary.

youtube.com/watch?v=rk6ChL…