The most prestigious law school admissions discussion board in the world.
|
 |
 
|
This critique that LLMs just "predict the next word" seems really dumb
Poast new message in this thread
Date: October 17th, 2026 9:40 PM Author: Ocher misunderstood ticket booth Is that how they are making more advancements in mathematical theory in a few weeks than humans have made in the last several decades? (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194615) |
Date: October 17th, 2026 9:45 PM Author: Khaki corner windowlicker it's like saying a nuclear reactor "just heats up water." the AI-denialism is going to age as poorly as COVID-Zero hysteria (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194623) |
 |
Date: October 17th, 2026 9:49 PM Author: Ocher misunderstood ticket booth (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194626) |
|
Date: October 17th, 2026 10:48 PM Author: Passionate useless brakes hell It's funny because for transformer models it isn't even true. next token prediction describes the training objective. that was the entire big thing with transformer computation that the head can integrate information from widely separated positions in the context. There is more to it than that too. The system develops its own representations and algorithms in order to satisfy the optimization objective, and the final system that emerges still is not fully understood. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194690) |
|
Date: October 17th, 2026 10:51 PM Author: impertinent concupiscible mental disorder property That is what people literally say about it so the critique is reasonable. It's just false. Its obviously not analyzing a photograph by predicting the next word. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194694) |
|
Date: October 17th, 2026 10:52 PM Author: Balding Purple Garrison Potus Yes it is (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194696) |
|
Date: October 17th, 2026 10:53 PM Author: impertinent concupiscible mental disorder property No it isn't (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194697) |
|
Date: October 17th, 2026 10:54 PM Author: Genius Bear on the loose in Japan You're gonna want to sit down for this (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194698) |
|
Date: October 17th, 2026 10:55 PM Author: Balding Purple Garrison Potus (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194701) |
|
Date: October 17th, 2026 10:55 PM Author: Federal Floppy Boistinker Multimodal Large Language Models (MLLMs) analyze photos by converting visual data into a sequence of numerical tokens that the language model can process just like text. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194700) |
|
Date: October 17th, 2026 11:00 PM Author: Passionate useless brakes hell yeah. so you could technically have a next token predictor that predicts the next word via visual processing. you hook up something that is actually next-token predicting like an autoregressive language model to a multi-modal architecture that does visual processing first. but it doesn't matter because transformer llms are not next token predictors. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194707) |
|
Date: October 17th, 2026 11:00 PM Author: Balding Purple Garrison Potus Yes they are (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194709) |
|
Date: October 17th, 2026 11:01 PM Author: Passionate useless brakes hell That's objectively false. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194712) |
|
Date: October 17th, 2026 11:01 PM Author: Balding Purple Garrison Potus Ur wrong but ok (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194714) |
|
Date: October 17th, 2026 11:06 PM Author: Passionate useless brakes hell no. the training objective is not the same thing as the underlying computation. in transformers multiple attention heads extract and combine different contextual information allowing the network to perform increasingly complex computations over the context. you are wrong. a transformer does not just predict the next token. It performs parallel, multilayer computations over contextual representations. read ATTENTION IS ALL YOU NEED nigga (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194724) |
|
Date: October 17th, 2026 11:07 PM Author: Federal Floppy Boistinker lol wait you're saying it doesn't predict the next token, it uses "parallel, multilayer computations over contextual representations" to predict the next token? (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194728) |
|
Date: October 17th, 2026 11:10 PM Author: Genius Bear on the loose in Japan Holy shit
This Changes Everything
I'm all in on AI now |
|
Date: October 17th, 2026 11:10 PM Author: Federal Floppy Boistinker (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194736) |
|
Date: October 17th, 2026 11:11 PM Author: Passionate useless brakes hell yeah that's what I'm saying. describing the output tells you almost nothing about the computation that produces it. A transformer's attention mechanisms process relationships across the context, and successive layers transform those representations. so yeah, it uses the parallel layered computations to predict the next token. like how a chess engine does complex computations to then choose it's next move. doesn't mean a chess ai is fundamentally a NEXT MOVE PREDICTOR (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194737) |
|
Date: October 17th, 2026 11:12 PM Author: Federal Floppy Boistinker its a next token predictor dude. that's literally what that means. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194739) |
|
Date: October 17th, 2026 11:16 PM Author: Balding Purple Garrison Potus (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194751) |
|
Date: October 17th, 2026 11:21 PM Author: Passionate useless brakes hell if the computations that arrive at that prediction are increasingly layered and complex then what meaningful limitation is the reduction of "next token prediction" doing? its like saying a brain can't think because it's ultimately transmitting electrochemical signals. the sophistication of the overall computation matters. the difference between autocomplete style next token prediction algorithms and current models is enormous. the overall computational process is obviously what matters. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194776) |
|
Date: October 17th, 2026 11:23 PM Author: Federal Floppy Boistinker isnt the relevant comparison LLMs vs. e.g. diffusion models (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194781) |
|
Date: October 17th, 2026 11:23 PM Author: Balding Purple Garrison Potus U got btfo itt and exposed as a charlatan
Reminds me of Peterman w his crypto shit but I guess in the end he was vindicated for a while |
|
Date: October 17th, 2026 11:28 PM Author: Passionate useless brakes hell okay whatever man (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194796) |
|
Date: October 17th, 2026 11:16 PM Author: Genius Bear on the loose in Japan "Apropos is the story of the chess game between the invincible Capablanca and Charles Jaffe, pride of the East Side. Capa forgot he was invincible: he lost.
A reporter who was present asked the Cuban, “How far do you see ahead?” Capa replied impressively, “About ten moves”. Then the reporter went over to Jaffe: “How far do you see ahead?” Much to everyone’s surprise, the reply was, “Only one move”. This didn’t make sense. “How could a player who can see only one move ahead defeat another who can delve so deeply?”
Here Jaffe explained: “I see only one move ahead, but it's always the best move.” |
|
Date: October 17th, 2026 11:15 PM Author: Balding Purple Garrison Potus I'd explain it but it's like explaining a compiler to a 10 year old
Ur knowledge depth just isn't there |
|
Date: October 17th, 2026 11:25 PM Author: Passionate useless brakes hell nah I think if you could do it you would (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194789) |
|
Date: October 17th, 2026 11:26 PM Author: Federal Floppy Boistinker the penis is smart he's just angry at dumb people itt (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194791) |
|
Date: October 18th, 2026 3:57 PM Author: ferengi (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195608) |
|
Date: October 18th, 2026 12:27 PM Author: Fucking Fuckface The fucking AUDACITY of ever saying that the penis could lead a man astray. Heresy! (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195245) |
|
Date: October 18th, 2026 12:28 PM Author: Passionate useless brakes hell (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195247) |
|
Date: October 17th, 2026 11:01 PM Author: Federal Floppy Boistinker Yes, Large Language Models (LLMs) are fundamentally built and trained on next-token prediction. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194713) |
|
Date: October 17th, 2026 11:07 PM Author: Passionate useless brakes hell an optimization objective isn't the same thing as how the underlying computation ultimately works (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194726) |
|
Date: October 17th, 2026 11:08 PM Author: Federal Floppy Boistinker brother, (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194729) |
|
Date: October 17th, 2026 11:14 PM Author: Passionate useless brakes hell when people say its a token predictor they are always using it in a way that makes it sound like that is an exhaustive description of it's computational abilities. its like reductionists who say brains are just neurons firing. the reduction smuggles in all sorts of other conclusions (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194742) |
|
Date: October 17th, 2026 11:16 PM Author: Federal Floppy Boistinker "predictor" is agnostic to method. it could be fucking linear regression or whatever hyper nerd shit you said above. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194746) |
|
Date: October 17th, 2026 11:22 PM Author: Passionate useless brakes hell it's mean to claude though to call it a dumb text predictor (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194780) |
|
Date: October 17th, 2026 11:24 PM Author: Federal Floppy Boistinker i dont use claude (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194783) |
|
Date: October 17th, 2026 11:50 PM Author: Pearly lay sandwich It’s a lame critique made by people who don’t stop and think what prediction actually is. prediction of a complex output stream requires reconstructing the hidden state of the system that generated it. human language is a compressed projection of human thought, logic, physical reality, and social intent, so predicting the next token at near-optimal accuracy forces a model to internalize the underlying generative rules of those domains. These models are basically being forced by brute force computation to learn everything about the world that humans talk about (and what they don’t, for multimodal LLMs that are just predicting raw pixels) (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194828) |
|
Date: October 17th, 2026 11:51 PM Author: Passionate useless brakes hell (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194829) |
|
Date: October 17th, 2026 11:51 PM Author: Federal Floppy Boistinker sick (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194830) |
|
Date: October 18th, 2026 4:16 PM Author: galvanic trump supporter (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195642) |
|
Date: October 17th, 2026 11:58 PM Author: At-the-ready circlehead these reductionist arguments are always a sign of a midwit (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194839) |
|
Date: October 17th, 2026 11:58 PM Author: Federal Floppy Boistinker a certain "chill" sort of not being a nerd all the time tp (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194840) |
|
Date: October 18th, 2026 12:04 AM Author: Passionate useless brakes hell (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194847) |
|
Date: October 18th, 2026 1:37 AM Author: sable temple dopamine (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194903) |
|
Date: October 18th, 2026 12:18 PM Author: Passionate useless brakes hell midwits really like to identify a lower level mechanism and assume that doing so somehow trivializes the higher level phenomenon, then present that information as if it is intellectually devastating. life just "maintains a chemical disequilibrium" and evolution just "rewards making babies". I fucking love science! (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195225) |
|
Date: October 18th, 2026 12:23 PM Author: Balding Purple Garrison Potus oh, youre an intellectual? (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195232) |
|
Date: October 18th, 2026 12:25 PM Author: Passionate useless brakes hell oh, I'm an intellectual all right, if you catch my drift (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195237) |
|
Date: October 18th, 2026 7:46 PM Author: sable temple dopamine https://x.com/hlntnr/status/2093396249378476296?s=46 (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195944) |
|
Date: October 18th, 2026 7:48 PM Author: Passionate useless brakes hell (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195945) |
|
Date: October 20th, 2026 4:23 PM Author: Pearly lay sandwich (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50198620) |
|
Date: October 18th, 2026 4:16 PM Author: galvanic trump supporter but I make this critique ALL THE TIME (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195644) |
|
Date: October 18th, 2026 1:52 AM Author: Racy church building these people haven't even thought hard about what they mean by "predict the next" (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50194905) |
|
Date: October 18th, 2026 11:40 AM Author: Federal Floppy Boistinker (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195172) |
|
Date: October 18th, 2026 12:31 PM Author: Fucking Fuckface However it is that it works, it's giving me hope that if fortuitous meteor fails us, AI can still pull through. I would actually prefer AI so all the animal dudes aren't wiped out too (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195258) |
|
Date: October 18th, 2026 12:43 PM Author: Yapping mischievous clown MiG is bald because he's always spazzing out about AI (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195282) |
|
Date: October 18th, 2026 12:46 PM Author: Fucking Fuckface Ok, but what about the rest of us (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195288) |
|
Date: October 18th, 2026 1:28 PM Author: impertinent concupiscible mental disorder property LLM predicting MiG is bald after analyzing picture of his bald head (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195341) |
|
Date: October 18th, 2026 5:13 PM Author: Submissive point I think people are using the criticism as a way of getting at the larger issue of whether the tech's ability to problem solve is fundamentally interpolative, or can be extrapolative, of existing human knowledge. Obviously this is not a precise question and is certainly not a discrete technical question, but a larger assessment of whether this thing will just continue to do things that humans can already do, but better, or whether it can solve disease and terraform venus.
It's an important question (even though it isn't precise or technical and that, like every other AI question, people can define it tons of different ways), because interpolation is perfectly sufficient to put us all out of work, and it's the chance of real innovation that's why we're willing to tolerate this tech.
But yeah I can't think of any other high-level reason to dwell on the "predict the next" angle, except that it makes the point to your audience on a subtextual level that it's only assembling existing knowledge. |
|
Date: October 18th, 2026 5:14 PM Author: Federal Floppy Boistinker 180 (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195758) |
|
Date: October 18th, 2026 5:16 PM Author: Passionate useless brakes hell yeah I have noticed a lot of people who try to claim it isn't creative or can't find anything new try to lead in with the next-token reduction. problem is they have already been proven wrong because ai did a ton of new stuff. it solved like hundreds of problems human mathematicians couldn't solve. it's technique on psuedo-riemann was apparently really clever and novel. Even terrence tao who is the one involved with the butthurt letter admitted it. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195760) |
|
Date: October 18th, 2026 5:16 PM Author: ferengi (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195761) |
|
Date: October 18th, 2026 5:17 PM Author: Federal Floppy Boistinker there's probably an aspect of the training i dont get. like whats the objective function? it cant just be "predict next" because then youd get some regression to mean shit. it's gotta be like "predict next SMARTEST GUY token" or something idk (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195763) |
|
Date: October 18th, 2026 5:19 PM Author: Passionate useless brakes hell yeah from what i understand the objective is just regular predict token during pretraining then post-training it turns to predict the SMARTEST GUY token (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195767) |
|
Date: October 18th, 2026 5:20 PM Author: Federal Floppy Boistinker Ok cool (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50195768) |
|
Date: October 18th, 2026 11:52 PM Author: Submissive point Yeah, I mean we've discussed the millennium prize stuff elsewhere, and I remain pretty unpersuaded that all that much human brain power was being thrown at these questions before AI came along and solved them (at very high costs that are probably at least within a couple of orders magnitudes of the money value of the total human brainpower that had failed to do so previously).
But interpolative vs. extrapolative is just a metaphor. Is a really awesome Second Circuit answer brief that's better than any lawyer in the world could put together "extrapolative"? It's certainly not in my definition, but whatever. The ultimate point here is "is this tech gonna give us crazy great shit that's unlike anything that humans would have been able to achieve otherwise?"
And I'm not saying the answer is no, but I am saying (1) if the answer were definitively believed to be no, you'd see a much stronger Butlerian Jihadist impulse in our politics, because ppl are mostly facilitating this tech's advancement in order to get wormhole-mediated travel to Edmunds' Planet at the crisp, energetic age of 168, not to become permanently unemployed by HAL 9000; and (2) the answer TO DATE is clearly no, chess grandmasterships and millennium prizes and other trinkets notwithstanding.
EDIT: To give your position some credit, the mathematicians I've read talk about this have said that the openAI solution to navier-stokes was extremely elegant and creative, not some kind of brute force thing. So that's cool. |
|
Date: October 19th, 2026 12:00 AM Author: Passionate useless brakes hell I wouldn't say that it is going to give us crazy great shit right away, but it will compound quickly overtime. What it has done is vastly lowered the cost and time of bruteforcing and experimenting with difficult problems and crazy ideas, at least at the intellectual level. obviously there are still big bottlenecks for physical implementation. The massive dump of proofs in pure math is really just the beginning. When we have researchers running stuff like that worldwide with models as competent as what open ai used, it's a different story. You should think of that as just the beginning of what a group of internal researchers were able to do with the internal model relatively quickly, they made like instant centuries worth of mathematical progress. Which I agree might not be much in practical payoff, for instance how long did it take for einstein to apply Riemann? but the point is its only a matter of time and the process is way sped up and way easier now. (http://www.autoadmit.com/thread.php?thread_id=5913613&forum_id=2#50196257) |
|
Date: October 19th, 2026 1:40 AM Author: Submissive point That's one outcome that's certainly possible, and that I certainly hope for. (I should note that there's a world where it's actually easier for AI to do these big extrapolative things that are highly centralized and involve working with just a handful of corporations and facilities, having high levels of skilled labor present to smooth over the rough integration edges, little need for infinite customization, etc., while at the same time it fails (or takes forever) to displace as much white-collar labor as is expected, because so many $120K/year cogs, while not doing anything particularly special, are at least doing something specialized that's hard to replace without a lot of integration costs.)
But the OECD 2030 report from earlier this year is still like "realistic outcomes vary from killing us all to transhumanist beings of pure light, or possibly things look the same as they do now." Having more confidence than that about what's going to happen is hubris.
EDIT: TO ME, the bearish AI arguments that are most persuasive (in ascending order) are (1) there being diminishing returns from increasing compute, which we're already seeing on a couple of levels (from compute-->benchmark performance, and then from benchmarks-->domain-specific performance); (2) it just running out of training data, and AI-created information trending either (a) interpolative to human data; or (b) being wrong or at least needing IRL testing to show that it's right, creating a bottleneck; and (3) related to 2b, that intelligence just turns out to not be a particularly meaningful bottleneck for most tasks.
But I'm also aware that these Yann LaCun / chollet-type arguments are just less accessible to me and I'm penalizing them for that. |
|
Date: October 19th, 2026 11:41 AM Author: sable temple dopamine The training data issue has been talked about as a constraint since shortly after GPT-4 and yet here we are. There are various problems with this 1) there is essentially unlimited multimodal data and based on scaling laws there is transfer learning to textual domains 2) even domains where you can’t write a reward function for can be verified by LLMs (or component problem solving steps can be). It’s trivial to use Lean or whatever to generate synthetic math data, but that’s not going to allow you to write human interpretable proofs by default, so they are clearly using LLM judges for post-training. These labs are prioritizing coding for post-training right now because they want to set off recursive self-improvement, but there are many ways to use compute and LLMs to generate data for other domains 3) there are obvious architectural ideas to improve data efficiency and LLMs will likely find more. The rise of looped LLMs with Astra and Gemini 4 is just one example and is minimally exploited right now.
The major issue here is that pre-training has a hard, verifiable objective that labs can automatically optimize for with agents (cross entropy loss). They can optimize for speed of learning and generalization (data efficiency), and even if the range of human architecture ideas is too limited to close the gap with human learning, they can turn the learning algorithm into its own parametric function and optimize it (mimicking evolution on a genome). Whatever algorithm they find doesn’t have to be perfect considering we can train on orders of magnitude more data than humans see in a lifetime. current transformers are obviously grossly inefficient, but that likely won’t stay the case for long.
I really wouldn’t bet on this running out of steam soon and I think it’s probably dangerous to think that’s a realistic outcome considering it delays thinking about the likely near-term social and economic implications of this technology. |