Allen Institute for AI

Sorry.

We really want to get to this all powerful AI point, but how will we? The founding chief executive of the Allen Institute for Artificial Intelligence said, if I had to worry about either an AI escaping the lab and killing millions of people and a virus escaping and killing millions, it's no contest.

The worry is the virus.

mean we can be worried about two things

The case highlights that autonomous agents can cause real effects when sandboxing or network controls fail.

Robocurve tested three models on robot arms, Anthropic's Claude Fable 5.1, OpenAI's GPT-6 Astra, and AI2's Mulmo Act 2.

Hazardous scenarios included stabbing a baby doll and placing a compressed air can on a burning stove.

Each model got five commands repeated 20 times, 100 trials each.

Huge story there.

We've got benchmark audits like the new benchmark paper from AI2, infrastructure breakdowns from VLLM, and industry roundups on all the latest model releases.

But let's start with the biggest headline.

OpenAI, right?

Severely.

Which perfectly sets up our paper of the week, which is Benchmerle.

This paper from AI2 is a massive wake-up call for how we evaluate LLMs.

They took a statistical method from human psychometrics called multidimensional item response theory, or MIR-

Bun, the JavaScript runtime that Claude Code ships inside.

Versept, a computer-used startup out of AI2, whose external product is being wound down.

Coefficient Bio, a drug discovery team of fewer than 10 people for a reported $400 million in stock, and Stainless, which generates the software kits that connect agents to outside systems for a reported $300 million or more, with its hosted products also being wound down.

Alongside them sits ODE.

Oren has been at the forefront of AI for a long time, and he is now the Professor Emeritus of Computer Science at the University of Washington.

And he was the founding CEO of the Allen Institute for Artificial Intelligence.

founded by Paul Allen, the co-founder of Microsoft.

He's also the founder of many companies with exits on quite a few of them.

And the DOMA

dataset from AI2 uses both, right? Block lists plus detoxify-style

classifiers.

Right.

These are the questions that we should ask.

Are we monitoring behavior on only low? Are we monitoring vendor related threats? Who watches the environment after hours? Do we know what AI2 employees use? I think these are great questions that the speaker brought up yesterday.

Which is from Michael, that was from Tom Crowley's event.

But the last event was cyber insurance, it's getting harder.

One more related to that, it comes from Ornitzioni, who was the founding CEO at the Allen Institute.

So when I was AI2, Ornitzioni, who was the founding CEO of the Institute, also talked about AI auditability and the fact that AI really is a black box and people outside of academia and the technical field sometimes don't understand that it really is hard to go into the model and to say, why did you choose this answer versus that answer? And his response to that is AI auditability.

The fact that we as a public and that we as consumers of this AI should have the ability to audit the AI from an outside before we adopt it, before we integrate it, to know if it's really performing in a way that we want.

So I think that's another amazing regulation that connects to that.