Andrew Critch (🤖🩺🚀)

Executive Director @ @ Researcher @ Berkeley.
Views my own.

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

This new eval shows how you can use AI models from different providers to boost performance by 20%+ on a suite of technical problems involving logic, numeracy, geometry, calculus, statistics, and coding.

Simply put, the models can catch each other making mistakes.
1/🧵

[dominic dyer @domdyer70]

Good news

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

Don't play along with calling extinction risks "improvement".

Also, Recursive Self Improvement (RSI) is too narrow of a term, being too imprecisely used, for what's happening right now. There is:

1) Recursive Self-Modification (RSM) := AIs modifying their next versions, including their self-modification tendencies. E.g.,

→ an AI that modifies itself or its next version to become increasingly powerful, that eventually kills all humans, is definitely engaged in RSM.

2) Recursive Self-Improvement (RSI) := RSM where the modifications follow a trend of improvement, for some notion of "improvement". What constitutes "improvement" can be debated, and varies importantly based on who is building the AI and what they want.

→ Extinctive RSM is not "improvement" from the perspective of most humans. I think we should not cooperate with branding decisions that allow people to call it "improvement".

3) Runaway RSM := RSM that gets out of the control of humans in a manner unexpected to humans. I think this is uncontroversially extremely dangerous.

4) Attempts at RSI can fail to yield morally-speaking "improvements" in AI and yield runaway RSM instead. That does not mean RSM and RSI are the same thing, and does not warrant calling extinctive AI developments "improvement" or even "self-improvement".

5) Objecting when someone says "RSI could lead to human extinction" takes some courage. One option is to say: "Sorry, that's not an improvement, and I'm not going to play along with branding games that call it that."

5a) If you generally dislike AI companies and want to blame them for the 'improvement' language, you can add:
"It's wrong for companies to call extinctive risks 'improvements', and they should not go unchallenged in that."
or
"Attempts at RSI are likely to yield runaway self-modification instead, which is not what basically anyone wants, and is actually a human extinction risk."

5b) If you generally like the idea of RSI, you can add:
"Actual improvement, under our values, precludes extinction threats."

Either way, don't play along with calling human extinction risks "improvement".

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

This ban needs to be expanded to "Recursively Self-Modifying", to avoid unfortunate debates about whether killing or driving humans to extinction counts as "Improvement". (Answer: it doesn't, but many people disturbingly claim it does.)

[MTS @MTSlive]

SITUATION DETECTED: Bloomington, Indiana's city council voted unanimously this week to ban any recursively self-improving entity capable of greatly outperforming all humans on all cognitive tasks from inside city limits, citing existential hazards and loss of human control.

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

Dear skeptics of multi-provider AI workflows: the world is about to start looking very strange to you.

Specifically, you'll start seeing a lot of math results coming out from independent researchers and groups using models from two or more frontier providers collaborating on the same project, because you can get much better performance that way, compared to using a single provider.

After math, other fields will follow, starting with basic physics and materials science. For each field, first you'll see hype from single-agent and multi-agent teams from a single provider, like a swarm of Claudes or a swarm of GPTs. But the then you'll start seeing better results, at least per dollar, coming from independent teams using models from multiple provders in combination.

These multi-provider breakthroughs won't be advertised as heavily, because no single frontier lab will pay for the marketing. But if you watch for who outside frontier labs is able to crank out impressive results, *without* billion dollar budgets, you'll start to see the pattern.

To many, this multi-provider advantage is obvious in advance, e.g., if you read and grokked my ARCHES report from 2020. Sorry if this post was boring for you!

But if this is *not* yet obvious to you, and especially if you're skeptical, you should probably start updating how you think about intelligence.

[Elliot Glazer @ElliotGlazer]

hello there ℤ is Diophantine in ℚ, thanx to my close friends Astra and Claude

The polynomial characterizing ℤ isn’t quite Tweet-sized; it has m=292,890,231,905 variables, is of degree d < 2^^7, and all its coefficients have absolute value < 2^^8.

[Voters For Animal Rights @theanimalvoters]

ELECTION NIGHT WAS A WIN FOR ANIMALS 🗳️

✅ 40+ animal-friendly candidates won in New York State
✅ Washington, D.C. banned foie gras
✅ Denver, CO also prohibited foie gras
✅ Lemon Grove, CA prohibited the sale of birds in pet stores
❌ Colorado REJECTED an amendment that would have made hunting and fishing a constitutional right
✅ Ashley Hinson and Roger Marshall, supporters of the dangerous Save Our Bacon Act that would put farmed animals back in cages, lose their seats

This is why animal advocates need to be politically active!

When we show up at the ballot box, animals win. 💪

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

Oh, and cage-free eggs.

I'm not assuming these products today are net-positive for animal welfare, but I'm hoping that purchasing them en masse can send a meaningful market signal that meat-eaters are willing to pay money to treat animals better.

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

I'm organizing an event where some people want to order meat products. I want to insist on using caterers who serve *relatively* ethical animal products, like humanely pasture-raised beef and/or wild-caught fish.

Can anyone make recommendations for this in the Berkeley area? Ideally Zerocater vendors, but other catering options would be great.

(I personally eat animal products, but am very picky about the treatment of the animals, and want to help spread that sort of discernment.)

[Lewis Bollard @Lewis_Bollard]

Some great news for factory-farmed pigs last night.

In Iowa and Kansas, voters rejected the lead sponsors of the Save Our Bacon Act in both the House and Senate: Rep. Ashley Hinson and Sen. Roger Marshall.

These were the politicians who did the pork lobby’s bidding and tried to wipe out every state ban on selling pork from crated pigs.

Yet even in two of the most pro-agriculture states in the country, it cost them more votes than it won them.

The politicians who follow the factory farming lobby aren’t as safe as they thought.

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

The event will have around 50 to 100 people.

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

@ESRogs Nope! E.g., if it reveals one model weight, that's nonzero knowledge, and no one cares.

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

Nice! For years I've said SNARKs would help prove AI outputs. Here Anthropic reference zk-SNARKS, which impose an unnecessary zero knowledge constraint, needlessly increasing compute costs. I wonder if they know that? Probably Claude will tell them soon, if they ask :)

[Anthropic @AnthropicAI]

New Anthropic research: progress on provable inference for large language models.

Here, we introduce FlashProve. Building on established cryptographic techniques, FlashProve allows developers to verify whether a given response came from a given AI model. It’s substantially more efficient than prior open-source methods for doing so.

These proofs have a range of applications. They can, for instance, make it harder for attackers to hijack a chat, or enable safety evaluators to verify that the model they’re testing is the same one being served to customers.

For our full write-up, see here: https://t.co/OIZWPb6QZa

Further progress in provable inference will likely be an important part of improving trust in AI deployment. To that end, we're open-sourcing FlashProve’s code, here: https://t.co/7ky3Ir8V06

We’re grateful to our collaborators: Alessandro Chiesa, @GiacomoFenzi, @gwberg95, @miranda__christ, Rahul Dalal (https://t.co/pjWKLHWXyz) and Ben Mayer (@transmissions11).

[Benonwine @benonwine]

Marina Sirtis gives an incredibly emotional, powerful tribute to Sir Patrick Stewart 💔

"He's gone. My friend. My Captain. My fellow football fan." 😢

You can tell it really comes from the heart. You can see how much he meant to all of them, especially the Next Generation crew.

Rest easy, Captain. 🖖❤️🙏

[Ian McKellen @IanMcKellen]

SIR PATRICK STEWART (1940-2026)

Before everything else, Patrick was defined by being born and bred in Yorkshire, the proud county in the north of England which he loved to the end. You could hear it in his accented voice and in the stern honesty with which he told you what he thought and felt. Yorkshiremen have no truck with half-truths or avoidance of big issues. He wore his heart on his sleeve, which made him a man to rely on, to admire and in my case, love.

We first met as ambitious young actors with the Royal Shakespeare Company half a century ago but we weren’t in the same productions. My first strong memory was a few years later, meeting him in the street in central London when he told me about an offer to take over a leading part in 'Star Trek'. “But Patrick think of your stage career in the UK. Do you really want to risk—getting lost in Hollywood for years and years?”. Well, he took the risk and thank goodness for us all that he did for not just his devoted Trekkies but film and theatre audiences, who marvelled at his using international celebrity to star in a string of cinema successes ('X-Men' of course) and in great classical plays by Shakespeare, Ibsen, Arthur Miller, Peter Shaffer, Pinter and Beckett at Stratford-upon-Avon, in London’s West End and on Broadway.

America was his adopted home, where he met his adored Sunny who has cared for him through sickness and health, through thick and thin. But he never forsook his origins: and the proudest achievement of the most respected and beloved actor of his generation on either side of the Atlantic, was his appointment as Chancellor of Huddersfield University in Yorkshire.

His farewell performance in 'Avengers: Doomsday' is yet to be seen.

Ian McKellen 29 October 2026

Photo: University of Huddersfield

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

Meeting Sir Patrick Stewart in person was one of the happiest days of my life. His portrayal of integrity, dignity, bravery, and so many qualities I look up to, was truly inspirational.

When we met, I thanked him and shook his hand. He was kind.

I was so happy I got to thank him like that.

Thank you again for everything, Sir Patrick. Today forward, I will try to be a better person in your honor, to help shape the future into something you would be proud of us for creating.

I hope many others will as well.

We will miss you.

[Patrick Stewart @SirPatStew]

In lieu of flowers, memorial contributions may be made to:
Johns Hopkins Alzheimer’s Disease Research Center Fund   
https://t.co/luC5CdcWDy
Alzheimer’s Foundation of America Caregiver Resources    
https://t.co/zvNptfyuNd

Photo Credit: Tommy Garcia

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

@ElliotGlazer Some of the disproofs might involve weird Gödelian Quine-like stuff that's hard to put into Lean though, so I'm not making Lean-specifically my only win route unless you give me better odds.

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

@Michael_Druggan I'm getting 20:1 from others, so you're not the best deal in town ;)

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

There's too much groupthink assuming unproven hypotheses to be true by inertia. My relatively dubious credences:

p(Hodge conjecture)≈50%
p(Riemann hypothesis)≈60%
p(ZFC is consistent)≈75%
p(P ≠ NP)≈80%
p(Goldbach Conjecture)≈85%
p(Collatz conjecture)≈90%

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

@ElliotGlazer Then we'd also be betting on whether Lean is how we do formal proofs in 2036 ;)

So no deal if the disproof has to be in Lean specifically. Unless you want to give me 100:1 odds!

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

@ElliotGlazer For the disproof, there's a question of the proof system. Would you accept "Any one of the top three automated proof verification systems available at the time"?

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

@ElliotGlazer Could we do all four? Either for $1k each or $250 each from me, against $20k each or $5k each from you? Each bet would settle at 10 years or at first disproof. Positive proofs wouldn't settle the bet(s) because ZFC is in question.

[Andrew Critch (🤖🩺🚀) @AndrewCritchPhD]

1. What's your probability that Zermelo–Fraenkel set theory with Choice (ZFC) is actually self-contradictory?
a) ≤5%
b) >5%

2. Are you a mathematician?
a) Yes
b) No