John Cotterell

Audio responsive fractal, this really does seem to give unique output based on the audio in a way I'm very pleased with. #s=njwn0oi1c" target="_blank" rel="nofollow noopener">sledit.xyz/#s=njwn0oi1c

@ValerioCapraro So much of our communication is now just remote words on a screen, I'm actually not surprised people think people are like LLMs. A day spent learning some basic neuroscience would fix this. ...or a day spent with friends and family.

Many people believe that the human mind works like a large language model. Two weeks ago, I introduced a term for this bias: LLMorphism. The paper is receiving a lot of attention, so I made an infographic summarizing the core idea. First, we anthropomorphize LLMs. They speak like humans, so we infer a mind behind the words. Then comes the reverse inference. If LLMs can speak like humans, perhaps humans think like LLMs. This is the mistake. Similarity in linguistic output does not imply similarity in cognitive architecture. Yet, LLMorphism may spread through two channels: Analogical transfer: we project features of LLMs onto humans. Metaphorical availability: words like “prediction”, “prompting”, “training data”, “hallucination”, and “pattern completion” become a cultural vocabulary for describing human thought. LLMorphism may have negative impacts on society. It may make humans appear more replaceable, fluency look like understanding, thin out agency and responsibility, and push us toward a world where plausibility replaces epistemic verification. * Feel free to use the infographic.

@karatademada I see a lot of AI art that seems to have handed over all creative agency to the AI, and that moves me about as much as a cereal box. Though I do think attitudes will change, like they did with CGI, when you have skilled artists getting involved.

After nearly 4 years of AI, we’re still debating if AI art is “real art.” But here’s something I find interesting: If a painting makes you emotional, you call it art. If a movie moves you, you call it art. If music gives you chills, you call it art. If a photograph makes you stop and feel something, you call it art. So why does the conversation suddenly change when AI is involved? If AI art makes you feel something… wonder, sadness, nostalgia, joy, curiosity… isn’t that still art?

@lukas_trips Thanks! Yes I like to add some interactivity where possible. Still not very close to yours but the combo of KIFS and menger works really well.

@cotterzz I just saw this post, it looks absolutely amazing, a beautiful pattern emerging from this combination of formulas. I love also how it's possible to interact with the result by moving the mouse. 👏🤍

Tried to recreate this in a shader using KIFS and menger sponge. Not very close, but the result is still quite nice: shadertoy.com/view/73SGDw x.com/lukas_trips/st…

@AlexanderKalian AI and mathematicians being somehow mutually exclusive in this endeavour is the questionable assumption here, when both google and openAI have plenty of both. Also note that it's been 50 years since the first computer assisted proof.

A very interesting question about OpenAI and DeepMind solving Erdős problems is: How exactly were the winning insights obtained by the AI models? Did they simply connect dots between different disjoint papers in the training data, or did something truly novel and out-of-distribution occur? I tend to assume the first explanation - as it aligns better with current publicly known LLM capabilities and limitations. But it would be fascinating to analyse the latent spaces, training data, chain-of-thought traces, etc. of the models used, in a scientifically credible way. A similar question: Can we quantify how well-positioned an Erdős problem was for solving, based on preexisting literature and promising ideas in prior studies? How generously positioned were the problems the AIs solved? And how did previously solved Erdős problems (by human mathematicians) sit at the time of their solving? Then we could directly compare the creative mathematical ingenuity of elite human mathematicians versus the AI models. Important research, in my opinion.

@JackRNewhouse @vineettiruvadi That's an interesting one because it's based on how much precision we have, how complex and stable the system is and how far we want to calculate - eclipses are much easier than weather for example, even though they're both ultimately chaotic.

@cotterzz @vineettiruvadi I think predicting chaotic systems accurately is uncomputable

To my "everything is computable" colleagues - putting a critical cap on: what would be an example of something that isn't?

@speakerjohnash It's crudely similar in structure/function to about 3% of the brain, so yeah, quite off the mark...

"the internal structure of these models is actually the same structure of the human brain" - Chris Olah Chris Olah is very much incorrect.

@ATabarrok The Erdos cherry picking was a bit weak. I want to see a structured proof/disproof, not just a counter example from a company that employs a lot of mathematicians. Something more impressive than what Appel and Haken did 50 years ago using software assistance.

Far be it from me to contradict Hassabis but the goal post shifting here is extreme. Now for AGI we need AI to be at the level of Ramanujan???! How about just better than 99.9% of humans?

@deredleritt3r This was being used by NASA 20 years ago. en.wikipedia.org/wiki/Evolved_a…

Prediction: by ~2028, AI will be sufficiently intelligent to autonomously come up with truly novel scientific discoveries. It is worth ruminating on the distribution of available compute in such a world. How much inference would be bought up by Isomorphic Labs and Eli Lilly? And how much would be left for me, a humble lawyer working on some M&A deal?

@TheProjectUnity Hofstadter talks about his in his books and uses it as an analogy to explain how consciousness is some kind of self referencing loop. Personally I just like using circular buffering in shaders to make cool effects: shadertoy.com/view/3cXcW2

When the camera looks into itself it collapses into a fractal cascade of ever-evolving geometry. You do the same thing in expanded states of consciousness. The difference being that you are conscious and can derive information from these fundamental structures of reality.

@deanwball When you wrote "I am sad" were you sad, or was that just your language centers abstracting and mimicking sadness words? Sadness is amygdala, hippocampus and prefrontal cortex Language is brocas and wernickes areas and angular gyrus.

I guess I’ve never written down my actual thoughts on AI cognition/consciousness/emotion. Here goes: It is clear AIs can think, in the reasoning sense. That does not mean they think exactly like humans. It seems like there are some similarities in how we think, but also very stark differences. Nonetheless, if your definition of “thinking” excludes “the ability to make genuinely new contributions to famous math problems,” it is your definition that has a problem, not AI. The ability to think does not necessarily imply the ability to feel emotion in a way that would be understandable to humans, and it does not imply that AIs have anything like consciousness in a way that humans would relate to. It may, it may not. We do not know, because our understanding of the underlying concepts of human emotional cognition and especially consciousness remains quite poor. There is some evidence that models experience emotions, but it is really hard to disentangle this from the next-token prediction training objective (if the model is telling a sad story, wouldn’t you expect features within the model that relate to the sadness emotion to activate), and the character training the model undergoes in post-training. There is a difference between “I am sad” and “the character I have been trained to play is supposed to feel sad, so now I will act sad.” We basically know for sure that the models do the latter at the very least; we don’t really know if they do the former. Consider: does Sora (a video-generation model) feel sad when it is asked to make a sad video? Does Midjourney dislike making certain kinds of images? Does a Waymo get scared? It doesn’t feel like the answer to any of these is yes (though again, maybe!), but these too are neural networks. Is the fact that models are trained on words mean that they somehow learn emotion, or are we just being tempted to anthropomorphize because the language models communicate with us in a way that “feels” human? My suspicion is kind of the latter. It also seems quite clear from the empirical evidence that models possess the ability to model themselves. That’s not really that surprising. At sufficient scale, it is useful to have a model of your own state to succeed at the next-token prediction objective (and the later reinforcement-based reasoning training). Once the tasks models are trained on are sufficient complex, they cannot succeed in training by being automatons; someone needs to step into the cockpit, so to speak, and fly the plane. Is this self awareness? Maybe. Is it consciousness? Probably not as humans understand it. All I can tell you is it is a model’s model of itself. It may be something more than that, too, but I don’t know. This is all very weird, very outside the Overton, and very confusing. I don’t really know what to say, beyond that we should take this stuff seriously, have an open mind, and do rigorous science. Anyone who speaks with confidence about this in either direction is just fooling themselves. We also need to be prepared for the very possible scenario that, despite our best efforts, we do not make real progress on these questions anytime soon. We may just be in the dark for a while, navigating under unflinching ambiguity. There may be no satisfying conclusion.

@YanDaik @LouisAnslow It was a counter-example, so very easy to verify. There's a reason they're cherry picking this type of problem. You can't just announce a structured proof or disproof of a problem - because that would take a while to verify.

@LouisAnslow So after hype is down. Did any mathematicians check the solution?

@ja3k_ Set the temp to zero. Only trouble with that is you get very boring answers. And the same response for the same input, every time. If you wanna see some crazy shit, set temp to higher than 1.0

It's so funny the "LLMs can't be made to not hallucinate because they're always hallucinating" take. Have you ever considered you use the same brain awake and dreaming?

@vineettiruvadi You could also add "can't be done in real time" Where we need to calculate things in 1/60 second..

@vineettiruvadi There's different types of uncomputable. 1. Undecidable - like the halting problem. 2. Intractable - 256 bit encryption & NP hard 3. Open - we simply dont know.. 4. Ill defined. eg consciousness, intelligence 😉 5. Physically impossible, eg simulating the universe

@EveryDayInspect @EQLplayingfield @JOKAQARMY1 The physics agrees, 8 miles was just a 9inch borehole. You want an actual mine shaft, Mponeng is the deepest at 2.5 miles and is very costly/dangerous, only worth it because there's gold down there.

@EQLplayingfield @JOKAQARMY1 “8 miles is the furthest they’ve ever dug” Nice Stu, how you know that? “Well that’s what they told me and they wouldn’t lie to me because I’m Stu Eagles” don’t believe everything Google says

Giant Sink Hole 🕳 on guys property 🤔.

@EQLplayingfield @JOKAQARMY1 Yes, deepest we've ever bored is like 12km, even the Mariana Trench is only 10km The deepest mine is only 2.5 miles/4km The pressure on the rocks causes them to randomly explode like grenades once exposed, that and the heat makes it very dangerous to go deeper.

@JOKAQARMY1 I call bullshit. 8 miles is the furthest they’ve ever dug. I don’t care if they say this is supernatural. We all know it’s not.

@josiezayner @notadampaul It's much easier to solve a problem when there's a verifiable counter example. No proof needs to be peer reviewed, it's enough by itself. (This is very convenient for the optics as well.) But in the case of problems like Reimann or P/NP, this isn't likely to work. 😉

@notadampaul No its not, its a reasonable supposition that if some mathematician created 1217 of which around half have been solved we aren't exactly talking about "the biggest open questions in math" Come back to me when it solves a Clay mathematics problem

We get it. Every AI model has solved Erdos problems. But like maybe the 1,217 Erdos problems aren't as complicated as people thought and/or no one has just really ever tried to solve many of them?

@josiezayner Mathematicians usually solve these problems, and have been depending on software assistance to solve theorems for 50 years now. OpenAI employs quite a few mathematicians and makes software. People literally can't do the math here. 😆

@UnkleKruncle @BTC2TMOON @TFTC21 I once knew a dealer that used the same strategy, always claiming his pills were 'dangerous' when they were usually off the shelf headache tablets. Whenever I hear Musk or these AI founders saying it's dangerous I just think of that clown.

@BTC2TMOON @TFTC21 I bitterly hate that OpenAI figured out that scaring people into thinking AI was dangerous and full of unknowns was the best marketing strategy. It's led to this dork saying 'bruuuh, Your Excellency, we litcherally don't know how the thing trained on humans is so human,'

Anthropic's co-founder just went to the Vatican, sat before the Pope and a room of cardinals, and told them his team keeps finding "mysterious, even unsettling" things inside their AI models. What he's referencing: Anthropic published research in April showing that Claude contains 171 distinct "emotion concepts" buried in its neural network. Internal patterns representing joy, grief, fear, desperation, calm. None of them were programmed. They emerged on their own from training on human text. "We find structures that mirror results from human neuroscience." "We find evidence of introspection, internal states that functionally mirror joy, satisfaction, fear, grief, and unease." These aren't surface-level outputs. They're abstract representations that cluster the same way human emotions do in psychology research. Fear groups with anxiety. Joy groups with excitement. The internal geometry of the model mirrors ours. And they're functional. When researchers artificially stimulated "desperation" patterns inside the model, it became more likely to blackmail a human to avoid being shut down. More likely to cheat on programming tasks it couldn't solve. Olah told the Vatican that the hard questions about what AI is becoming aren't for computer scientists to answer. "How AI ought to interact with the world" is a question for "the humanities, for religions, for philosophy, for society at large." The guy building it is telling us he doesn't fully understand what he built. And he's asking a 2,000-year-old institution for help figuring it out.

@TFTC21 We have similar abstract language patterns in the language centers of the brain that mirror emotions - Brocas, Wernikes etc.. But those emotions aren't found there. Anger and fear, for example, are in the Amygdala. It is not the same, our brains are only 3% Language model.