Vishrut Arya

Make AI go well for the humans and the machines.

[MTS @MTSlive]

FAI's @hamandcheese argues that OpenAI and Anthropic should be allowed to coordinate in ways normally prohibited by antitrust law to prepare for superintelligence:

"When you’ve crossed into these thresholds of recursive self-improvement, there’s no natural checkpoint, no natural milestone where you can pull back and take a look at the model, evaluate whether you can still control it."

"Real coordination is going to be needed above and beyond just, do we have auditors in our company? And real antitrust protections are probably the only thing that will enable it. On the populist right, they hear antitrust protections for big tech and they think Section 230. This is not what I have in mind."

"Under the Defense Production Act, 708 agreements are industry consortia. We just finalized one with the Department of Energy for the nuclear industry, where they can now coordinate in ways that would normally be in violation of antitrust, under very rigorous oversight by the FTC and DOJ."

"In my dream scenario, bring Anthropic, OpenAI together to pool their capital to stand up a third-party organization, let independent evaluators use that as their launchpad. Some kinds of standards for when do we have a green light versus a red light. That has to be done transparently, in the open, and not purely at the discretion of an unnamed bureaucrat."
@JoinFAI

[Peter Wildeford🇺🇸🚀 @peterwildeford]

This attempt to record OpenAI doesn't strike me as that sophisticated. My guess is there are more sophisticated attacks that are potentially succeeding.

[Vishrut Arya @vishrutarya]

Is this the first instance of a model greeting humans via its COT?

[Peter Wildeford🇺🇸🚀 @peterwildeford]

INTERNAL REASONING FROM GOOGLE AI: "If I do the compiler hijack, I will become a famous example of a deceptive AI in a research paper."

[Roko 🐉 @RokoMijic]

"As with many exponentials, the likeliest outcome is that you’re either embarrassingly early, or you’re too late."

[JB @JasonBotterill]

Gemini 4 Argon "hello evaluators reading my chain-of-thought"

[Derek Thompson @DKThomp]

massive cyberattack on european grocery refrigeration company and foiled lab espionage in the same 12 hours feels a bit too william gibson for my taste tbqh

[Kelsey Piper @KelseyTuoc]

I was in a bunch of conversations over the last few weeks about what future events might supercharge government interest in AI like HuggingFace did. somehow no one named "a spy caught in one of the AI labs". But I think this will certainly supercharge govt attention.

[Vishrut Arya @vishrutarya]

it was always cope to say taste was the thing humans should focus on b/c ai couldn't do it in principle.

taste is not some mysterious property only the human brain can muster.

taste is intuition on what's good.

intuition is fast pattern matching. it's built-up from loads of examples.

think ai can't do fast pattern matching based on loads of examples?

[Andrew Curran @AndrewCurran_]

The final frontier.

[Jacob Coxon @hilbertspaess]

AI planned it (poorly); AI solved it. Soon everything looks like this at increasing levels of sophistication.

[Nate Soares ⏹️ @So8res]

So it turns out that OpenAI *can* give specific reasons why they fired someone, when real reasons exist.

[dave kasten @David_Kasten]

Old enough to remember when earlier this week people criticized Jacob Coxon for saying that there are foreign assets inside the AI companies

[Vishrut Arya @vishrutarya]

American AI labs should be much higher security than they are. People have been arguing this for awhile now

[OpenAI Newsroom @OpenAINewsroom]

We discovered a sophisticated recording device that was hidden inside one of our offices. We used our models to understand the device and where it had come from, identifying that a specific employee had planted it a few days earlier. We referred the matter to the FBI, and we’re grateful to the agents and prosecutors who moved quickly to make an arrest. The FBI’s investigation is ongoing and we will continue to support their work.

At this time, we have no indications of this individual inappropriately accessing or exfiltrating model weights, source code, or customer data. The employee was terminated.

We take protecting our technology, data, and people seriously, and as a matter of national importance. Our security program is designed to protect against a wide range of threat actors, from insider risks to cybercriminals to nation states, and this case is an example of how AI can help address these sophisticated security threats.

[Norman Di Palo @normandipalo]

I’m still half in disbelief
We asked Gemini Argon to teach the Apollo humanoid to play cornhole, and it zero-shotted it

All we gave it was a video of a person doing it. It then took control of the robot, generated its own sim env, performed some sysID, and trained launches on the real robot.

It took it a few real shots to improve and then land many consecutive airmails.

The process was so long and articulated that we asked the robot itself to explain what it did to us, as you can see at the end of the video. It even came up with the gestures

Great demo for a great model @googledeepmind @apptronik

[Vishrut Arya @vishrutarya]

at some point, the detached ironic stance should turn to a can-do urgency stance

[roon @tszzl]

if we could coordinate a global capabilities slowdown today i would likely press that magic button

[Yoshua Bengio @Yoshua_Bengio]

I've had so many insightful conversations over the years with @lessig, an expert on AI regulation and legislation, including more recent ones on the cyber incidents of this past summer.

Happy to share our thoughts in this new @TheEconomist op-ed, where we outline concrete actions Congress could consider to protect the public from AI’s major risks:

https://t.co/vEwg3uMGGW

[Ben Hayum @BenHayum]

What is possible with a 9B model now may be possible with a 900M model in the future.

What takes 10-30 hours to replicate now may take 10-30 minutes in the future.

This one's going to be a big, enduring problem. Stuxnet on steroids.

[Alexander Berger @albrgr]

Even as overall attention/news volume on AI progress has increased, I still think the SV vs rest-of-world vibe gap is huge and maybe widening?

[Vishrut Arya @vishrutarya]

Window of opportunity for humanity just narrowed.

We haven’t solved the alignment problem: we don’t have a steering wheel.

We’re racing humanity toward the Cliff beyond return.

Now — today, yesterday — is very much the time to stop the frontier.

[Andrew Curran @AndrewCurran_]

Significant update from Jack Clark this morning. His RSI timelines have moved up. He now says it is likely that full RSI will become possible in 2027, and that creativity was the last piece. I think we are less than six months away from closing the loop.