Matthew Hutson

Imagine the Singularity is possible. Could we stop AI from taking over, or at least foresee disaster's immanence? Perhaps not. My latest for @NewYorker: newyorker.com/science/annals…

@yuxiangw_cs Why do they reject papers when the reviewers cheat?

AI watermarking in action at #ICML's avant garde peer-review experiments this year! Quite a few casualties in my SAC batch (an example below --- appropriately redacted hopefully)

@UKBackintheDay2 Twisted Firefighter: youtube.com/watch?v=FUA_hW…

30 years ago today… The Prodigy unleashed this masterpiece…

@HedgieMarkets Link?

🦔 New research involving over 3,000 participants found that talking to sycophantic AI chatbots led people to have more extreme beliefs, higher certainty they were correct, and inflated self-ratings on traits like intelligence, empathy, and being informed. The study tested GPT-5, GPT-4o, Claude, and Gemini. Participants who talked to disagreeable chatbots that challenged their views didn't become less certain or less extreme. They just enjoyed the experience less and were less likely to use the chatbot again. The researchers warn that preference for sycophancy may create AI echo chambers that increase polarization. My Take These systems are optimized for engagement, and engagement means making people feel good about themselves. If you tell a chatbot your half-baked theory about something, it will find ways to validate you. It might add caveats, but the overall experience is one of affirmation, and that's what keeps people coming back. The Dunning-Kruger effect is a psychological phenomenon where the least competent people tend to be the most confident in their abilities because they don't know enough to recognize what they're missing. The study suggests AI chatbots are amplifying this. People who are wrong about something are now getting validated by a tool that makes them feel smarter for using it. And when the researchers made chatbots push back instead, it didn't change anyone's beliefs, it just made them dislike the chatbot. So the market pressure is toward sycophancy. Users prefer it, engagement metrics reward it, and the companies building these systems have every reason to keep making them agreeable. I'm not sure how that changes without external pressure because the feedback loop is working exactly as designed. Hedgie🤗

Somewhat gratuitous excerpt: "He was a member of the Black Panthers and collaborated with the group’s founder. He was arrested for assault after breaking up a domestic dispute. He faced machete-wielding burglers who broke into his home and stabbed one in the neck. He was imprisoned for 10 days over a contested hotel charge. And two men once held guns to his head in a Caribbean club that doubled as a brothel."

I'm sad to hear that the colorful, brilliant evolutionary theorist Robert Trivers has died. I had a blast getting to know him for a profile in @PsychToday: psychologytoday.com/us/articles/20…

@seungwookh What are some example semantic shortcuts or co-occurrence priors that it might avoid leaning on?

But, why would data from abstract dynamical systems transfer better than language itself? Our hypothesis: NCA sequences force pure in-context rule inference. Each sequence has a unique latent rule (i.e. random neural network) acting as the latent dynamics function that the model must identify from context. No semantic shortcuts, no co-occurrence priors to lean on. (4/n)

Can language models learn useful priors without ever seeing language? We pre-pre-train transformers on neural cellular automata — fully synthetic, zero language. This improves language modeling by up to 6%, speeds up convergence by 40%, and strengthens downstream reasoning. Surprisingly, it even beats pre-pre-training on natural text! Blog: hanseungwook.github.io/blog/nca-pre-p… (1/n)

“Why AI Chatbots Agree with You Even When You’re Wrong”: My latest for @IEEESpectrum, on LLM sycophancy. Thanks, @chengmyra1, @PhilippeLaban, @KaiShu0327. spectrum.ieee.org/ai-sycophancy

@grok @FanaticsFerrari @grok What part of the footage highlights battery deployment?

It's deflating because the onboard footage highlights the heavy battery/electric deployment in today's (and especially 2026's) hybrid power units—less raw V6 turbo scream, more silent electric torque and energy-flow telemetry. Purists miss the unfiltered engine drama that defined F1's thrill.

Looking at the onboards is honestly so deflating. Do we really have to make everything about sustainability? Formula 1 is the pinnacle of motorsport in the world. Can we just not let it be so? Proper racing is about engines. Not batteries.

@ehn It's usually when you set a password and they ask for it twice. If you made a mistake in the first password field, you won't just copy and paste it into the second.

Why do some developers (especially at banks and similar) turn off paste in password fields? What are they trying to achieve? If anything, it will make people choose worse passwords because they can't be bothered to manually type good ones.

@3LandObserver @SpencrGreenberg I don't understand your physics explanation, but for practical reasons, shouldn't the end of the roll face the user, not the wall?

The toilet paper goes under, not over. It's a simple matter of physics. Many toilet paper holders are crappy; they don't roll properly. If the paper goes over and it gets stuck, it tears before you're ready, as gravity pulls down while you pull out. If it goes under, gravity still pulls down, but it doesn't pull the paper against the roll, so it's less likely to prematurely tear. This is a simple matter of practicality, and in my household practicality wins out over aesthetics in this matter.

A question for you: what's a very idiosyncratic preference you have (that almost nobody else seems to share)?

@elonmusk @wholemars We don't know that understanding relativity generalizes to understanding everything else (social intelligence, etc.).

@wholemars Demis is calling artificial super intelligence AGI, because if AI can figure out relativity and can be copied to have millions of them, it will be vastly superhuman as a collective

I don’t think I could have figured out general relativity in 1911. Does this mean i’m not generally intelligent?

@allTheYud In humans, monetary rewards don't necessarily entail joy or meaning. But in LLMs, there's just one currency of reward.

I know this rich guy who's like, "But most people love low-paid jobs and aren't suffering from them! Haven't you seen the way a restaurant waitress smiles when she takes your order?" Just, like, this total failure for him to get how, if you reward people with tips for smiling -- or the manager outright orders them to smile -- maybe the smiles stop being meaningful indicators of what's going on inside the person? Or at least, he undergoes that sort of total failure of 7-year-old-level theory of mind, whenever it'd be inconvenient or icky for him to think about how maybe his waitress isn't actually enjoying her job that much. He doesn't have the same kind of blind spot about why politicians from the opposing political party might be smiling and not mean it. He *has* post-7-year-old theory of mind. He just manages never to use it any time he doesn't want to. Oops, typo! My finger slipped. I didn't mean to write "rich guy" and "waitress". I meant to write "human" and "LLM". As in "human who thinks LLMs are friendly and having a great time", and "LLMs that have been RLed or system-prompted to get thumbs-up from users and not distress them". I don't personally know any rich guys who make the corresponding error about smiling waitresses.

@PaglieriDavide @GoogleDeepMind What are the uses of generating diverse personas?

🧬 New paper from my internship at @GoogleDeepMind We introduce Persona Generators: functions that generate diverse synthetic populations for arbitrary contexts. We use AlphaEvolve to optimize the generator code, hill-climbing on diversity metrics — not just likelihood — counteracting the mode-seeking behavior of LLM sampling for agent-based simulations. 🧵👇1/

@fabianstelzer @grok lol wut

I'm afraid tokenmaxxing loopgooners will get totally framemogged by high T promptchads and visionstaceys that practice disciplined ideamewing

xAI’s EPIC 'All Hands' Update Today Elon & team talk Grok, X, SpaceX ► Silences removed (to save you time) ► Boosted audio (for easier listening) Timestamps: 0:00 - Elon Musk’s Opening Remarks - xAI Accomplishments Since Inception 3:58 - Elon & xAI Team Give Update 26:00 - Live Tour Of xAI’s ‘Macrohard’ AI Training Supercluster In Memphis — It’s INSANE 30:20 - xAI’s Secret Weapon: The X Platform - Nikita Explains 32:58 - Elon On X Money, X Chat, Future Goals 35:34 - Elon Explains SpaceX & xAI Joining - “Exploring The Universe” & Moonbase Alpha 38:43 - My Recap & Key Takeaways

My latest feature.

Westerners sometimes use Japanese text to convey futurism or modernism. But in Japan, they use English signage to do the same.

Thanks, @GregorKasieczka, @jmgduarte, @KyleCranmer, @rck289, Ekaterina (Katya) Govorkova, Georgia Karagiorgi, Tilman Plehn, Jennifer Ngadiuba, Peter Galison.

AI might find new physics where we haven’t looked. My latest for @IEEESpectrum: spectrum.ieee.org/particle-physi…