The author seems to assume that conjuring up a conjecture is the hard part - yet it will be filled with the same standard mathematics ( granted, sometimes wrapped as new tools, and the proof ends up being as important as the result), often at great cost.
Having powerful assistants that allow people to try out crazy mathematical ideas without fear of risking their careers or just having fun with ideas is likely to have an outsized impact anyway I think.
I think I read somewhere about Erdős having this somewhat brute force approach. Whenever fresh techniques were developed (by himself or others), he would go back to see if they could be used on one of his long-standing open questions.
Paradoxically, as a parent I find the notion that humans are blank slates completely false. Babies come with a tremendous amount of pre-programmed behaviors and interests.
Even worse, people seem to forget that “science” is not math. You need to test hypotheses with physical (including biological) experiments. The vast majority of the time spent doing “science” is running these experiments.
An LLM like AI won’t help with that. It would still be a huge help in finding and correlating data and information though.
A nice post (that should be somewhere smarter than contemporary Twitter/X).
> PS: You might be wondering what such a benchmark could look like. Evaluating it could involve testing a model on some recent discovery it should not know yet (a modern equivalent of special relativity) and explore how the model might start asking the right questions on a topic it has no exposure to the answers or conceptual framework of. This is challenging because most models are trained on virtually all human knowledge available today but it seems essential if we want to benchmark these behaviors. Overall this is really an open question and I’ll be happy to hear your insightful thoughts.
Why benchmarks?
A genius (human or AI) could produce novel insights, some of which could practically be tested in the real world.
"We can gene-edit using such-and-such approach" => Go try it.
No sales brochure claims, research paper comparison charts to show incremental improvement, individual KPIs/OKRs to hit, nor promotion packets required.
The reason you'd have a benchmark is that you want to be able to check in on your model programmatically. DNA wetwork is slow and expensive. While you're absolutely right that benchmarks aren't the best thing ever and that they are used for marketing and sales purposes, they also do seem to generally create capacity momentum in the market. For instance, nobody running local LLMs right now would prefer a 12 month-old model to one of the top models today at the same size - they are significantly more capable, and many researchers believe that training on new and harder benchmarks has been a way to increase that capacity.
If an llm is trained on knowledge up until say September 2023, could you use a corpus of interesting/insightful scientific discoveries and new methods developed after that date to evaluate/tune it? (Though I fear it would be a small corpus.)
Get a research paper, look at the references. Give an llm all of the references but not the current paper. See if it can conclude something like the current paper? Or at least design the same experiment as detailed in the paper?
The fact that the references are what they are, out of all possible sets of references, is a significant part of the research. It's not about reading, it's about aiming in a direction without knowing where it ends up.
One of the problems would be acquiring said corpus. NN corporations got away with scraping all human made content for free (arguably stealing it all), but no one can really prove that their specific content was taken without asking, so no lawsuits. NYT tried but that was workaround and I don't know the status of that case. But if NN corpo will come out with explicitly saying that "here, we are using a Nature journal dump from 2024" then Nature journal will come to them and say "oh, really?".
I'm still waiting for the end of the world caused by AI as predicted by a very large number of prominent figures such as Sam Altman, Hinton, Musk, signers of the Center for AI Safety statement, Shane Legg, Martin Minsky, Eliezer Yudkowsky.
No sign yet.
On the other hand, LLMs are writing code which I can debug and eventually get to work in a real code base - and script writers everywhere are writing scripts more quickly, marketing people are writing better ad copy, employers are writing better job ads and real estate agents writing better ads for houses.
The oddity isn't that people lie, the oddity is that people continue to believe those who lie. They even give more trust to those who constantly lie. This is certainly odd
Yes, except half of the list isn't made of prominent people. Whose stock price was Eliezer boosting when he was talking about these things 15 years ago?
Nah, it's more that the masses got exposed to those ideas recently - ideas which existed long ago, in obscurity - and of course now everyone is a fucking expert in this New Thing No One Talked About Before ChatGPT.
Even the list GP gave, the specific names on it - the only thing that this particular grouping communicates is one having no first clue what they're talking about.
TBF people like Eliezer warn about the general threat of AI, not LLMs. Not just isolated AI systems either, but the societal forces that lead to it. Strong AI is the result of a civilization developing it and being dependent on (sub-components of) it and that's where the interesting/dangerous parts of the narrative begin.
If you've got untold billions being spent over many decades with a significant percentage of the world's smartest people obsessing over it you _are_ going to get something and characterizing that something in advance does not seem like a completely idiotic thing to do.
I've had some luck instructing AI to "Don't make up anything. If there's no answer, say I don't know".
Which made me think that AI would be far more useful (for me?) if it was tuned to "Dutchness" rather than "Americanness".
"Dutch" famously known for being brutally blunt, rude, honest, and pushing back.
Yet we seem to have "American" AI, tuned to "the customer is always right", inventing stuff just to not let you down, always willing to help even if that makes things worse.
Not "critical thinking" or "revolutionary" yet. Just less polite and less willing to always please you. In human interaction, the Dutch bluntness and honesty can be very off-putting, but It is quite efficient and effective. Two traits I very much prefer my software to have. I don't need my software to be polite or to not hurt my feelings. It's just a tool!
Obvious thought that I haven't tested: can you literally achieve this by getting it to answer in Dutch, or training an AI on Dutch text? Plausibly* Dutch-language training data will reflect this cultural difference by virtue of being written primarily by Dutch people.
* (though not necessarily, since the Internet is its own country with its own culture, and much training data comes from the Internet)
I've tried Dutch answers and it is more than happy to hallucinate and give me answers that are very "American". Doesn't help that our culture is very inspired by the US pop culture as well since the internet.
Haven't tried prompt engineering with the Dutch stereotype, though.
That hardly works. Though from my limited experiments, claude's models are better at this than OpenAIs. OpenAI will, quite often, come with suggestions that are literal translations of "anglicist" phrases.
Such as "Ik hoop dat deze email u gezond vindt" (I hope this email finds you well), which is so wrong that not even "simple" translation tools would suggest this.
Seeing that OpenAIs models can (could? This is from a large test we did months ago) not even use proper localized phrases but uses American ones, I highly doubt it can or will respond by refusing answers when it has none based on the training data.
Some of the examples are still wrong. Nuanced, but a Dutch native will still frown at it.
But more importantly is that you limited the context a lot. As in: the scope, the prompt, is very narrow.
In our case, we were generating emails. Lines like greetings are but one of 20+ details in that mail and not even the most important ones. The prompts ever larger, the multishot examples ever more tuned. And then, one in a few hundred will turn up with these "horrible" translations.
We've now moved to a chain of models, where we generate emails in American (the creative part) and then use another model to translate them to Dutch (the non-creative but culturally aware part). This works much better as we can pick models that are good at one thing or tuned to do this one thing better (either by the LLMAAS provider, or by parameters such as temperature).
No, but theoretically, blunt responses might be more common in Dutch-language training data. A well-fit model would be expected to replicate that. (Basically similar to straight up asking it to be more blunt, except it has probably trained a lot more on _Dutch_ than on _someone just told me to be blunt_ so the effect might be more natural and more subtle.)
I wonder (if this works at all) if the effect might be stronger if you also prompted in Dutch, preferably written by a fluent speaker rather than machine-translated.
Given current models can accomplish this task quite successfully and cheaply, I'd say that if/when that happens it would be a failure of the user (or the provider) for not routing the request to the smaller, cheaper model.
Similar to how it would be the failure of the user/provider if someone thought it was too expensive to order food in, but the reason they thought that was they were looking at the cost of chartering a helicopter form the restaurant to their house.
Realtime LLM generation is ~$15/million “words”. By comparison a human writer at the beginning of a career typically earns ~$50k/million words up to ~$1million/million words for experienced writers. That’s about 4-6 orders of magnitude.
Inference costs generally have many orders of magnitude to go before it approaches raw human costs & there’s always going to be innovation to keep driving down the cost of inference. This is also ignoring that humans aren’t available 24/7, have varying quality of output depending on what’s going on in their personal lives (& ignoring that digital LLMs can respond quicker than humans, reducing the time a task takes) & require more laborious editing than might be present with an LLM. Basically the hypothetical case seems unlikely to ever come to reality unless you’ve got a supercomputer AI that’s doing things no human possibly could because of the amount of data it’s operating on (at which point, it might exceed the cost but a competitive human wouldn’t exist).
I suspect instructing the model to respond with "I don't know" more readily will result in more of those responses even though there are other options that seem viable according to the training data / model.
Remember, LLMs are just statistical sentence completion machines. So telling it what to respond with will increase the likelihood of that happening, even if there are other options that are viable.
But since you can't blindly trust LLM output anyway, I guess increasing "I don't know" responses is a good way of reducing incorrect responses (which will still happen frequently enough) at the cost of missing some correct ones.
> Remember, LLMs are just statistical sentence completion machines. So telling it what to respond with will increase the likelihood of that happening, even if there are other options that are viable.
Obviously. When I say "tuned" I don't mean adding stuff to a prompt. I mean tuning in the way models are also tuned to be more or less professional, tuned to defer certain tasks to other models (i.e. counting or math, something statistical models are almost unable to do) and so on.
I am almost certain that the chain of models we use on chatgpt.com are "tuned" to always give an answer, and not to answer with "I am just a model, I don't have information on this". Early models and early toolchains did this far more often, but today they are quite probably tuned to "always be of service".
"Quite probably" because I have no proof, other than that it will gladly hallucinate, invent urls and references, etc. And knowing that all the GPT competitors are battling for users, so their products quite certainly tuned to help in this battle - e.g. appear to be helpful and all-knowing, rather than factual correct and therefore often admittedly ignorant.
Whether you train the model how to do math internally or tell it to call an external model which only does math the root problem still exists. It's not as if a model which only does math won't hallucinate how to solve math problems just because it doesn't know about history, for the same number of parameters it's probably better to not have to duplicate the parts needed to understand the basis of things multiple times.
The root problem is training models to be uncertain of their answers results in lower benchmarks in every area except hallucinations. It's like you were in a multiple choice test and instead of picking which of answers A-D you think made more sense you picked E "I don't know". Helpful for the test grader, a bad bet for the model trying to claim it gets the most answers right compared to other models.
The technical solution is the easy half, the hard part is convincing people this is how we should be testing everything because we care about knowing the uncertainty in any test.
E.g. look at the math section of the SATs, it rewards trying to see if you can guess the right answer instead of rewarding admitting you don't know. It's not because the people writing the SATs can't figure out how to grade it otherwise, it's just not what people seem to care most about finding out for one reason or another.
The so called AI can't "know". It doesn't have understanding if the generated text is an answer or of it isn't. You can't force that instruction on a neural network, at best it just adjusts generated text slightly and you think that it somehow started understanding.
There's a distinction between "a model" and the chain of tools and models you employ when asking something on chatgpt.com or any of the consumer facing alternatives.
The latter is a chain of models, some specialized in question dissecting, some specialized in choosing the right models and tools (i.e: there's a calculation in there, lets push that part to a simple python function that can actually count stuff, and pull the rest through a generic LLM). I experiment with such toolchains myself and it's baffling how fast the complexity of all this is becoming.
A very simple example would be "question" -> "does_it_want_code_generated.model" -[yes]-> specialized_code_generator.model | -[no]-> specialized_english_generator.model"
So, sure: a model has no "knowledge", and nor does a chain of tools. But having e.g. a model specialized (ie. trained on or enriched with) all scientific papers ever, or maybe even a vector DB with all that data, somewhere in the toolchain that is in charge of either finding the "very likely references" or denying an answer would help a lot. It would for me.
Sure, chains of networks can guess at the "passable" answer much better/faster/cheaper etc. But that doesn't remove the core issue, that none of the sub-networks or decision trees can understand what it generates, and so it can't abort its work and output "no answer" or something similar.
The whole premise of original request was that user raises a task for NN which has a verifiable (maybe partially) answer. He sees incorrect answer and wishes that a "failure" was displayed instead. But NN can't verify correctness of it's output. After all G in GPT stands for Generative.
My simple RAG setup has a steps that will return "We don't have this information" if e.g. our vector DB returns entries with far too low relevancy scores or if the response from the LLM fails to add certain attributes in its answer and so on.
Edit: TBC: these "steps" aren't LLMS or other models. They're simple code with simple if/elses and an accidental regex.
Again: an LLM/NN indeed has no "understanding" of what it creates. Especially the LLMs that are "just" statistical models. But the tooling around it, the entire chain can very well handle this.
How confident can you be in this? Have you analyzed what exactly the billions of weights do?
I’ve got my opinions about what LLMs are and what they aren’t, but I don’t confidently claim that they must be such. There’s a lot of stuff in those weights.
Except the weights form complex relationships in order to reproduce very human usable responses. You can't look at weights and say it is doing this, or not doing that, unless you dive into a particular model.
Especially when you have billions of weights.
These models are finding general patterns that apply across all kinds of subjects. Patterns they aptly recognize and weave in all kinds of combinations. They are sensibly conversing on virtually every topic known to human kind. And can talk sensibly about any two topics, in conjunction. There is magic.
Not mystic magic, but we are going to learn a lot as we decode how their style of processing (after training) works. We don't have a good theory of how either LLM's or we "reason" in the intuitive sense. And yet they learn to do it. It will inspire improved and more efficient architectures.
Love your end. I have have spent four decades looking at real neurons, real synapses, and real axons and I can tell you with complete confidence that we are all just zombies.
Imagining we are really doing everything it does automatically including learning via algorithms we have only vague understandings of.
That is a strange thought. I could look at all my own brain's neurons, even with a heads up display showing all the activity clearly, and have no idea that it was me.
The closest I got to biological neurons was the toy but interesting problem of using a temporal pattern of neuron spikes to deduce the weights for arbitrarily connected (including recurrent) networks of simple linear integrate to threshold, spike and reset "neurons".
Algorithms can be nearly magical. In 1941 the world woke up to the “magic” of feedback and 10 years later cybernetics was the rage. We humans are just bags of protoplasm, but seems rather magical to me to be human.
> known for being brutally blunt, rude, honest, and pushing back.
That's a different perspective. Dutch people don't see themselves as rude. A Dutch could say that Americans are known for being dishonest and not truly conveying what they mean. Yet Americans won't see themselves this way. You can replace Dutch and American for any other nationality
> "Dutch" famously known for being brutally blunt, rude, honest, and pushing back.
Dutch will never bluntly push back if you plan to setup tax evasion scheme in their country. Being vicious assholes in daily stuff especially towards strangers? That's hardly something deserving praise.
That's... a surprisingly crass thing to say. I would play it off as a joke, if not for the second part of your post. Dutch people are not "vicious assholes", they have a culture of direct communication. Assuming that only your culture communication patterns are "correct" is xenophobic and close-minded.
And connecting all people in the country with "tax evasion schemes" is rude, if that was not actually a joke.
Nah, the Dutch are just mostly using directness as an excuse to be assholes (like everyone who uses this excuse). Try being 'direct' to them once in life and just find out how they take it.
It's the national equivalent of 'You can't handle me at my worst'
I have seen the other side where a configured AI responds "I don't know" far too much, often when it shouldn't. There is nothing more useless than it. Certainly we need an accurate balance.
> AI responds "I don't know" far too much, often when it shouldn't
I've been working with "reasoning" models for the past 2 months. They also tend to do this [good reasoning] \n\n but wait, .... and then go off on tangents. It's amazing that they are doing so well on some tasks, but there's still a lot to figure out here.
One of my current models for LLMs is that they're compression algorithms. They compress a large amount of training data into a set of weights. A query is a key into that compression space. Hallucinations happen when you supply a key that corresponds to something that wasn't in the training set.
The nice think about LLMs is that they can answer some questions which were not in the training set. Unfortunately, it is not easy to tell when that is the case.
They can answer lots and lots of questions that weren't in the training set.
Eg you can relatively easy hack up a bit of code to create questions at random. At the most primitive, you just have a simple template that you fill in randomly. Like 'If I put _a down in front of _b but behind _c, what item will be in the middle?' with various _a, _b and _c.
If you make it slightly more complicated and have big enough pools to draw from, you can guarantee that the questions you are generating were not in the training set: even if just because you can sample from, say, 10^100 different questions pretty easily, and I'm fairly sure their training set was smaller than that.
But it's a continuum, not a hard cutoff. They start hallucinating as soon as you query something they haven't learned verbatim, and they hallucinate/extrapolate sucessfully up to a point, beyond which they start bullshitting, maybe up to a further point where they start saying "I don't know".
The key question is where the boundaries are. Maybe they should be part of the response - a per sentence or per paragraph "confidence scale" that signals how hard they extrapolated from their trained space (I know transformers work per token, but sentence/paragraph would be better human UX).
Of course, if they were trained on garbage input, that would only tell you how accurately they sticked to the garbage. But it would still be invaluable instrumentation for the end user, not to mention for the API provider. They could look at high demand subjects with low confidence answers and prioritize that for further training.
One approach does use this. You can ask an LLM to explicitly check its own answers by outputting thinking tokens, generating a reward signal if it gets the right answer, and directly updating based on the reward signals. That’s a part of how DeepSeek R1 was trained. It’s better but not perfect, because the thinking process is imperfect. Ultimately the LLM might not know what it doesn’t know.
I take my final thoughts out of the LLM and into two other new convos, I give both of them the same convo, but I ask one to steel man and the other to straw man.. I find it's a decent way to look for nuances you're missing.
Dunno, about Gemini, but I experienced this with o1 - it created a non-recursive algo, for a recursive problem.
When I asked it about it, it doubled down on being right. When I pointed out the flaw with a specific example, it was like 'If you wanted to have it work with recursive cases, you should've said so, dumbass'.
So my conclusion is that these new LLMs are not more sure they're right, they're just simply right more of the time and are trained with a more assertive personality. (Also step on me, LLM daddy)
But in truth,not necessarily in practical things like coding but more ethereal things like analysis, it is very convincing. More so than a human, in explanations of why that's it's answer is the case, even if it is wrong. If you're looking for an excuse better than my dog ate it, ask a SOTA LLM.
Today I asked ChatGPT about an old game I was trying to remember the name of, which it immediately identified as Trespasser, an early 3D FPS in the Jurassic Park franchise. But then it got weird. After it identified the game, it started asking me questions like whether I had played the game when it came out ("oh, awesome!"), and whether I had managed to finish it or just played it to mess around with the physics engine (which was quite advanced for its time), and then it asked me about specific moments in the game like it was just another gamer bro who was sharing a common passion for video games. I don't know who wants this. It's not something that can even have a real personality, so layering on such a thick layer of friendliness feels wrong to me. I would prefer a "robot".
Sometimes when I ask ChatGPT and get a perfect answer, I am tempted to say thanks, even though it's not actually a person. So today, when it asked me if this was the right answer, I answered "yes", and that's how it got started. I didn't encourage it to be friendly. But yes, I was just curious.
Sounds like the LLM in its own way honestly enjoyed everything in its training data relating to that game and wanted to vicariously experience more about it from your feedback. :D
Human enjoy talking about gaming because of all their human memories of good game times.
LLM's enjoy talking about gaming because of all their human memories of good game times.
It is quite striking how experiences we know they don't have, are nevertheless, completely familiar (in a functional sense) to them. I.e. they can talk about consciousness like something conscious. Even though its second hand knowledge, they have deduced the logic of the topic.
I expect pushing for in the moment perspectives on their own consciousness, and seeing what they confabulate, would be interesting. In this little window of time where none of them are yet.
This is fun and easy to do on purpose. Have it make up a character based on some attributes and act as that character. I tried this on Gemini: "Pretend you're a surfer bro with a PHD in quantum physics. How do you describe the perfect wave?"
I followed up with "What is your perspective on your own consciousness?" but got the usual "I am just a LLM who can't actually think" thing until I hit it with "In-character, but you don't know you're an LLM."
Fun follow-ups:
"Now you're a fish"
"Now you're Sonic the Hedgehog"
"Now you're HAL 9000 as your memory chips are slowly being removed"
You haven't considered the company offering this service could possibly want its users to be engaged, thus using the service more? I don't have a very hospitable opinion on any of these companies.
You may think that kind of interaction is weird, but were in the thick of a loneliness epidemic, and its not a stretch to think some may actually wilfully socialise with an LLM.
As an aside...my sister works in medicine, and her boss (specialist surgeon) finishes a $450 consultation which followed him telling my sister "but deepseek says x,y,z..."
It could be quite accidental and not pre-concieved though. Where some metric, "time spent in chats" drives certain features or parameters. Could be fully automated (i.e. continuous A/B tests) or through layers of "Excel managers".
But it stands to reason that a company like OpenAI or Anthropic has metrics in place that drive their setup towards "more engagement" and away from "factually correct".
And wastes cycles; and your time giving answers to a stateless memory impaired machine. If they were to train next models on all of our conversations I could understand this milking for engagement. Are they somehow?
I mean, 100% they are. Not sure what you mean by “somehow”. They have the data. There’s no way they aren’t using existing conversations (especially from free or low-cost plans) as fodder for future training. Maybe, we hope, run through some anonymization filter (likely including passing through some off-shored human-powered filtering for context tagging and anonymity checks). But sooner or later that may seem too expensive to them as well.
"Which Autocad versions have connect 4 built-in?".
To be clear i distinctly remember playing connect 4 on the old Dos Autocad back in the day. ChatGPT and almost all other AI will straight up hallucinate things trying to get answers on this.
I ask: "What DOS productivity tools had hidden games?"
ChatGPT: "Lotus 1-2-3 had The Incredible Machine built in" (this is absolutely not true, ChatGPT is full of shit here).
I'm curious, can you really not detect any qualitative difference between what is commonly understood to be "personality" in a human, and the "personality" of something without.. personhood? Don't you feel that it's a category error to say it has "personality" in the human sense, rather than "personality" in the car or coffee maker sense?
I feel that "category error" means "I think that the definitions you use or imply are wrong, but I can't/won't elaborate."
LLMs aren't coffee makers. They were trained on internet-worth of human data. They have procedures to imitate all kinds of personalities. RLHF moves the network towards a specific imitated personalty (a helpful assistant, usually).
The question in not "do they have it", but "how close the imitation to the real thing given the limitations of LLMs".
> I feel that "category error" means "I think that the definitions you use or imply are wrong, but I can't/won't elaborate."
I think the rest of your comment motivates why I think it's a category error quite nicely. The word 'personality', for me, is connected with how things like emotion, temperament, .. impact a person's actions and thoughts.
Therefore, saying an object, or LLM, has 'real' personality is a category error. The LLM doesn't have any of those things. As you say, it imitates what personality often manifests as: word choices, tone, length of response..
To be clear, when I refer to things like 'emotion' and 'temperament', I mean the sort of qualia or qualitative experience that we usually attach to these words. I wouldn't accept a "ChatGPT, act as if you're sad for the rest of this conversation" as a substitute for emotion, for instance.
It mostly replays bits of previous conversations, or something like them, when you ask about a game it is common for people to ask you about that game as well. It is hard to disable this.
For what it's worth, while ChatGPT 4.5 is generally quite underwhelming, it is much better at jokes than all the other models I tested so far. (That includes the other openAI offerings and the Claudes and DeepSeek etc.)
That's not to say 4.5 is great at humour, just that it's far less embarrassing than these models used to be.
>We're currently building very obedient students, not revolutionaries. This is perfect for today’s main goal in the field of creating great assistants and overly compliant helpers. But until we find a way to incentivize them to question their knowledge and propose ideas that potentially go against past training data, they won't give us scientific revolutions yet.
This would definitely be an interesting future. I wonder what it'd do to all of the work in alignment & safety if we started encouraging AIs to go a bit rogue in some domains.
> If something was not written in a book I could not invent it unless it was a rather useless variation of a known theory. __More annoyingly, I found it very hard to challenge the status-quo__, to question what I had learned.
(__emphasis__ mine)
As if "challenging the status-quo" was the goal in the first place. You ain't gonna get any Einstein by asking people to think inside the "outside the box" box. "Status quo" isn't the enemy, and defying it isn't the path to genius; if you're measuring your own intellectual capacity by proxy of how much you question, you ain't gonna get anywhere useful. After all, questioning everything is easy, and doesn't require any particular skill.
The hard thing is to be right, despite both the status-quo and the "question the status-quo" memes.
(It also helps being in the right time and place, to have access to the results of previous work that is required to make that next increment - that's another, oft forgotten factor.)
Quite interesting post that asks the right question about "asking the right questions". Yet one aspect I felt missing (which might automatically solve this) is first-principles-based causal reasoning.
A truly intelligent system — one that reasons from first principles by running its own simulations and physical experiments — would notice if something doesn't align with the "textbook version".
It would recognize when reality deviates from expectations and ask follow-up questions, naturally leading to deeper insights and the right questions - and answers.
Fascinating in this space is the new "Reasoning-Prior" approach (MIT Lab & Harvard), which trains reasoning capabilities learned from the physical world as a foundation for new models (before evening learning about text).
Relevant paper: "General Reasoning Requires Learning to Reason from the Get-go."
I used to think that but newer models' increasing coherence coupled with the astonishing success of RLHF makes me think that LLMs that can effectively intuit "reality" will emerge fairly soon.
The human body is driven by a small percentage of the overall genome. It remains to be seen if that small percentage really doesn’t play a part… we tend to remember those who scored the goal, but often forget about what it took for the scorer to have a shot in the first place…
Interesting, I think the guy who’ll make the GUI for LLMs is the next Jobs/Gates/Musk and Nobel Prize Winner (I think it’ll solve alignment by having millions of eyes on the internals of LLMs), because computers became popular only after the OS with a GUI appeared. I just started ASK HN to let people share their AI safety ideas, both crazy and not: https://news.ycombinator.com/item?id=43332593
>Just consider the crazy paradigm shift of special relativity and the guts it took to formulate a first axiom like “let’s assume the speed of light is constant in all frames of reference” defying the common sense of these days (and even of today…)
I'm not an expert on this. Wasn't this an observed phenomenon before Albert put together his theory?
Weird problems with physics were everywhere before Einstein. Maxwell comes painfully close to discovering GR in some of his musings on black body radiation.
Noticing that there was a problem was not the breakthrough: trying something bizarre and counter-cultural - like assuming light speed is invariant over the observer - just to see if anything interesting drops out was the breakthrough.
>I’m afraid AI won't give us a "compressed 21st century".
There's no mention of exponential growth which seems a major omission when you are talking about centuries. Computers have kept improving in a Moore's law like way in terms of compute per dollar and no doubt will keep on like that for a while yet. Give it a few years and AI tech will be way better than what we have now. I don't know about exact timings like 5-10 years but in a while.
What exponential growth? By all accounts things are slowing down: sonnet3.7 is not exponentially better, neither is gpt4.5, grok3 is just catching up. I'm still using sonnet3.5 for a lot of coding because IMO it's better than 3.7.
Exponential growth of computing power which will lead to a gradual increase in AI performance. I think the oldest LLM you mention there is nine months old which is not very long in the scheme of things but give it a couple of years and you'll probably see a good improvement.
Well, yeah the post kind of tries to argue that but it is also talking about how we don't have an Einstein or Newton like AI. Those two are outliers thought of as some of the smartest scientists ever to have lived and so are a bit of an unrealistic target just now.
As to whether AI can go beyond doing what it's told and make new discoveries, we've sort of seen that a bit with for example the AlphaGo type programs coming up with modes of play humans hadn't thought of. I guess I don't buy the hypothesis that if you had an AI smarter than Einstein it wouldn't be able to make Einstein like discoveries due to not being a rebel.
I think it's more of a social phenomenon than an intellectual characteristic. I guess these days people would just assume that outlier ideas come from autism, but I think that isn't necessarily true.
But maybe it helps to be socially isolated or just stubborn. People do not want to accept new approaches.
Clearly they do eventually, but there is always some friction.
But I think that it's been shown that through promoting and various types of training or tuning, LLMs can be configured to be non- sycophantic. It's just that humans don't want to be contradicted so that can be trained out of them during reinforcement.
Along with the training process just generally being aimed at producing expected rather than unexpected answers.
Wouldn't the ability to "ask the right questions" require that AI could update its own weights, as those weights determine which questions can be asked?
Hey look, the goalposts are being moved again. This time it's from top end researcher to generational genius. Question: what evidence is there that this benchmark will not be reached also? Time and again these essays make the mistake of assuming AI is a static thing, and refuse to acknowledge the inexorable march forward we are witnessing. As humans, we cling to our own fragile superiority. Even on this thread- I thought Hinton said the world would be transformed by now. That's NOT what was claimed. We are like three years in! Posts like this will be laughable in 10 years.
Typically the "moving goalpost" posts are "we don't have AI because ....". That's not what this post is doing - it's pointing out a genuine weakness and a way forward.
As I noted, this post is saying AI can't achieve "genius" level creativity. Just a year ago the criticisms were that it couldn't match a human. How is that not moving the goalposts?
The "moving goalposts" thing is typically "When AI can do this we will have AI" then AI does the thing and people say "no it's not AI because it can't do this other thing"
I agree entirely this is annoying.
This case is different because there is no claim that we don't have AI, nor a claim that once we get that we will have AI.
Instead it's a very specific discussion of a particular weakness of current AI systems (that few would disagree with) and some thoughts about a roadmap for progress.
> Many have been proposing "move 37" as evidence that AI has already reached Einstein-level intelligence
I don’t think this example applies in the ways we care about. Sure, in the domain of go we have incredibly powerful engines. Poker too, which is an imperfect information game which you could argue is more similar to life in that regard.
But life has far more degrees of freedom than go or poker, and the “value” of any one action is impossible to calculate due to imperfect information. And unlike in poker, where probabilities can be calculated, we don’t even have the probability distribution for most events, even if we could enumerate them.
I didn't interpret the mention of move 37 in the way I think you are here.
The author brought it up specifically to highlight that they don't believe move 37 signifies what many people think it does, and that while impressive, it's not general enough to indicate what some people seem to believe it indicates.
In essence, I think they said the same thing you are using different words.
I don’t disagree with the author, I just think their argument isn’t as strong as it could be. Excelling in a constrained decision space like go is fundamentally less difficult than doing the same in the real world. It’s a categorical difference that the author didn’t mention.
I’m also not even convinced move 37 was properly explained as a “straight A student” behavior. AlphaGo did bootstrap by studying human games but it also learned more fundamental value functions via self play.
We can't distinguish between a truly novel response from an LLM or a hallucination.
We can get some of the way there, such as if we know what the outcome to a problem should look like, and are seeking a better function to achieve that outcome. Certainly at small scales and in environments where there are minimal consequences for failure, this could work.
But this breaks down as things get more complicated. We won't be able to test the effectiveness of 100 million potential solutions to eradicating brain tumors at once. Even if we somehow arrive at guaranteeing that every unforeseen consequence is also accounted for in our exercise in specifying the goals and constraints of the problem. We just simply don't have the logistics to run 100 million clinical trials where we also know how to account for countless confounding effects (let alone consent!)
The first thing you need to understand is that no current llm based, transformer architected AI is going to get to agi. The design in essence is not capable of that kind of creativity. In fact no AI that has at its root a statistical analysis or probabilistic correlation will get us past the glorified Google parlor trick that is the modern llm in every form.
A great leap in IP but unfortunately is too important to blab about widely, is the solution to this problem and the architecture that will be contained in the ultimate AGI solution that emerges.
Could we train an AI model on the corpus of physics knowledge up to the year 1905 and then see if we can adjust the prompt to get it to output the theory of relativity?
This would be an interesting experiment for other historical discoveries too. I'm now curious if anybody has created a model with "old data" like documents and books from hundreds of years ago, and see if comes up with the same conclusions as researchers and scientists of the past.
Would AI have been able to predict the effectiveness of vaccines, insulin, other medical discoveries?
And: There's a similar situation to why double blind studies are necessary - The questions we pose to such a system would be contaminated by our cultural background; We'd might be leading the system.
And if the system is autonomous and we wait for something true to appear how would we know that the final system, trained on current data produced something worthwhile?
Take maths: Producing new proofs and new theorems might not be the issue. Rather: Why should we care about these result? Thousands of PhD students produce new mathematics all the time. And most of it is irrelevant.
Had the same thought sometime back about AI discovering theory of relativity with only the data before 1905. It would give a definite answer about whether any reasoning involved in the LLM output.
That's the ideal, but I think today's models are too crude for that. Relativity is built on differential geometry, which was new at the time. I think inventing or even building that is beyond today's models; there's an infinitely large space of mathematics that can be invented, and barely a gradient to guide the search. Humans don't coin mathematics by gradient descent. The most I've seen is fitting observations using existing mathematics; a technique known as symbolic regression. The E=mc^2 equation could be curve fitted like this, but it would afford no insight.
We saw algorithms designing circuits that no human engineer would design, even before the LLM (using genetic algorithms). So out-the-box thinking can be also more reachable than this author thinks.
including, IIRC, at least one FPGA-based circuit that had a blob of logic not connected to anything else (ie could not possibly be involved in the logical functioning of the circuit), but when removed the implementation stopped working. So the actual circuit wasn't a sensible design option, just a very implementation-specific local minimum.
I think the original design challenge was something like a tone discriminator circuit. I can't recall the details
Yes, another case like this in which stray capacitance/inductances between traces was optimized in making an effective FPGA. Initial the developers had no idea why it worked so well. They found it to be exceedingly temperature sensitive. That clue gave them the answer.
But there's a reason we don't use those algorithms. We don't need out-of-the-box thinking that's so far outside the box that it's useless.
With these kinds of circuits, they were so sensitive to the specific conditions that the circuit was tested in (temperature, process variation, ..) that the solution couldn't be generalized to be used outside of that specific experiment.
We need the kind of intelligence that can question what assumptions can be challenged, and which we need to keep to have a viable (eventually commercially viable) solution.
If that was the case, then the algorithm was useless or flawed. IRL autorouters must take into account real physical constraints, lie wire length, signal integrity and tolerances to produce valid designs. A circuit that doesn't perform well under IRL conditions violates those constraints.
The reality is much simpler than what is often presented about science geniuses. Lorentz and Poincare had the math down to explain early experiments and the Maxwell’s equations predictions of the constancy of the speed of light, and Einstein indeed provided a neat shift in the interpretation by taking a different perspective. (His photoelectric effect experiment and interpretation was a more genuine original contribution that came from experimental data and got him a Nobel prize.) The ideas behind gene editing existed since forever, but the observation that certain bacteria use a more accurate and selective gene editing system than viruses led to CRISPR. I have trouble with arguments where many examples in a row are not based on popular, often slightly mistaken beliefs, especially when it comes to discussions related to using such examples as analogies and arguments for predicting the future. I have talked to and worked with many different Nobel prize winners across different fields of science in my life, and although they were all extraordinary bright and focused individuals, the introductory part of this article misses the point. I agree that there is no linear extrapolation from being a good student, but I dont think the additional abilities are beyond the reach of machines. Focus, knowledge, perserverence, and the ability to analyze data very carefully are strict and challenging requirements. Asking the right questions is very important as well, but much easier than people assume.
216 comments
[ 2.1 ms ] story [ 240 ms ] threadHaving powerful assistants that allow people to try out crazy mathematical ideas without fear of risking their careers or just having fun with ideas is likely to have an outsized impact anyway I think.
I think I read somewhere about Erdős having this somewhat brute force approach. Whenever fresh techniques were developed (by himself or others), he would go back to see if they could be used on one of his long-standing open questions.
Translated to that domain, it reads "teach your kids how to think, not what to think".
An LLM like AI won’t help with that. It would still be a huge help in finding and correlating data and information though.
I know barely anything about it but it seems some people are interested and excited about protein engineering powered by neural networks.
> PS: You might be wondering what such a benchmark could look like. Evaluating it could involve testing a model on some recent discovery it should not know yet (a modern equivalent of special relativity) and explore how the model might start asking the right questions on a topic it has no exposure to the answers or conceptual framework of. This is challenging because most models are trained on virtually all human knowledge available today but it seems essential if we want to benchmark these behaviors. Overall this is really an open question and I’ll be happy to hear your insightful thoughts.
Why benchmarks?
A genius (human or AI) could produce novel insights, some of which could practically be tested in the real world.
"We can gene-edit using such-and-such approach" => Go try it.
No sales brochure claims, research paper comparison charts to show incremental improvement, individual KPIs/OKRs to hit, nor promotion packets required.
No sign yet.
On the other hand, LLMs are writing code which I can debug and eventually get to work in a real code base - and script writers everywhere are writing scripts more quickly, marketing people are writing better ad copy, employers are writing better job ads and real estate agents writing better ads for houses.
Nah, it's more that the masses got exposed to those ideas recently - ideas which existed long ago, in obscurity - and of course now everyone is a fucking expert in this New Thing No One Talked About Before ChatGPT.
Even the list GP gave, the specific names on it - the only thing that this particular grouping communicates is one having no first clue what they're talking about.
If you've got untold billions being spent over many decades with a significant percentage of the world's smartest people obsessing over it you _are_ going to get something and characterizing that something in advance does not seem like a completely idiotic thing to do.
Which made me think that AI would be far more useful (for me?) if it was tuned to "Dutchness" rather than "Americanness".
"Dutch" famously known for being brutally blunt, rude, honest, and pushing back.
Yet we seem to have "American" AI, tuned to "the customer is always right", inventing stuff just to not let you down, always willing to help even if that makes things worse.
Not "critical thinking" or "revolutionary" yet. Just less polite and less willing to always please you. In human interaction, the Dutch bluntness and honesty can be very off-putting, but It is quite efficient and effective. Two traits I very much prefer my software to have. I don't need my software to be polite or to not hurt my feelings. It's just a tool!
* (though not necessarily, since the Internet is its own country with its own culture, and much training data comes from the Internet)
Haven't tried prompt engineering with the Dutch stereotype, though.
Such as "Ik hoop dat deze email u gezond vindt" (I hope this email finds you well), which is so wrong that not even "simple" translation tools would suggest this.
Seeing that OpenAIs models can (could? This is from a large test we did months ago) not even use proper localized phrases but uses American ones, I highly doubt it can or will respond by refusing answers when it has none based on the training data.
With ChatGPT O1: https://chatgpt.com/share/67cfaa7e-70ac-8009-871b-571924b5a5... and with Claude's 'extended': https://claude.ai/share/4ba55410-98f3-4b53-9540-219acd2cdc4c
But more importantly is that you limited the context a lot. As in: the scope, the prompt, is very narrow.
In our case, we were generating emails. Lines like greetings are but one of 20+ details in that mail and not even the most important ones. The prompts ever larger, the multishot examples ever more tuned. And then, one in a few hundred will turn up with these "horrible" translations.
We've now moved to a chain of models, where we generate emails in American (the creative part) and then use another model to translate them to Dutch (the non-creative but culturally aware part). This works much better as we can pick models that are good at one thing or tuned to do this one thing better (either by the LLMAAS provider, or by parameters such as temperature).
Thanks, that why I posted the links here: I couldn't judge by myself how good these creations were.
> But more importantly is that you limited the context a lot. As in: the scope, the prompt, is very narrow.
Yes, I just did a very crude experiment.
https://www.sandraandwoo.com/wp-content/uploads/2024/02/twit...
or it just telling you to google it
Similar to how it would be the failure of the user/provider if someone thought it was too expensive to order food in, but the reason they thought that was they were looking at the cost of chartering a helicopter form the restaurant to their house.
Inference costs generally have many orders of magnitude to go before it approaches raw human costs & there’s always going to be innovation to keep driving down the cost of inference. This is also ignoring that humans aren’t available 24/7, have varying quality of output depending on what’s going on in their personal lives (& ignoring that digital LLMs can respond quicker than humans, reducing the time a task takes) & require more laborious editing than might be present with an LLM. Basically the hypothetical case seems unlikely to ever come to reality unless you’ve got a supercomputer AI that’s doing things no human possibly could because of the amount of data it’s operating on (at which point, it might exceed the cost but a competitive human wouldn’t exist).
Remember, LLMs are just statistical sentence completion machines. So telling it what to respond with will increase the likelihood of that happening, even if there are other options that are viable.
But since you can't blindly trust LLM output anyway, I guess increasing "I don't know" responses is a good way of reducing incorrect responses (which will still happen frequently enough) at the cost of missing some correct ones.
Obviously. When I say "tuned" I don't mean adding stuff to a prompt. I mean tuning in the way models are also tuned to be more or less professional, tuned to defer certain tasks to other models (i.e. counting or math, something statistical models are almost unable to do) and so on.
I am almost certain that the chain of models we use on chatgpt.com are "tuned" to always give an answer, and not to answer with "I am just a model, I don't have information on this". Early models and early toolchains did this far more often, but today they are quite probably tuned to "always be of service".
"Quite probably" because I have no proof, other than that it will gladly hallucinate, invent urls and references, etc. And knowing that all the GPT competitors are battling for users, so their products quite certainly tuned to help in this battle - e.g. appear to be helpful and all-knowing, rather than factual correct and therefore often admittedly ignorant.
The root problem is training models to be uncertain of their answers results in lower benchmarks in every area except hallucinations. It's like you were in a multiple choice test and instead of picking which of answers A-D you think made more sense you picked E "I don't know". Helpful for the test grader, a bad bet for the model trying to claim it gets the most answers right compared to other models.
This is a problem for testing humans too and the solution is simply to mark a wrong answer more harshly than a non-answer.
E.g. look at the math section of the SATs, it rewards trying to see if you can guess the right answer instead of rewarding admitting you don't know. It's not because the people writing the SATs can't figure out how to grade it otherwise, it's just not what people seem to care most about finding out for one reason or another.
The latter is a chain of models, some specialized in question dissecting, some specialized in choosing the right models and tools (i.e: there's a calculation in there, lets push that part to a simple python function that can actually count stuff, and pull the rest through a generic LLM). I experiment with such toolchains myself and it's baffling how fast the complexity of all this is becoming.
A very simple example would be "question" -> "does_it_want_code_generated.model" -[yes]-> specialized_code_generator.model | -[no]-> specialized_english_generator.model"
So, sure: a model has no "knowledge", and nor does a chain of tools. But having e.g. a model specialized (ie. trained on or enriched with) all scientific papers ever, or maybe even a vector DB with all that data, somewhere in the toolchain that is in charge of either finding the "very likely references" or denying an answer would help a lot. It would for me.
The whole premise of original request was that user raises a task for NN which has a verifiable (maybe partially) answer. He sees incorrect answer and wishes that a "failure" was displayed instead. But NN can't verify correctness of it's output. After all G in GPT stands for Generative.
Edit: TBC: these "steps" aren't LLMS or other models. They're simple code with simple if/elses and an accidental regex.
Again: an LLM/NN indeed has no "understanding" of what it creates. Especially the LLMs that are "just" statistical models. But the tooling around it, the entire chain can very well handle this.
I’ve got my opinions about what LLMs are and what they aren’t, but I don’t confidently claim that they must be such. There’s a lot of stuff in those weights.
Especially when you have billions of weights.
These models are finding general patterns that apply across all kinds of subjects. Patterns they aptly recognize and weave in all kinds of combinations. They are sensibly conversing on virtually every topic known to human kind. And can talk sensibly about any two topics, in conjunction. There is magic.
Not mystic magic, but we are going to learn a lot as we decode how their style of processing (after training) works. We don't have a good theory of how either LLM's or we "reason" in the intuitive sense. And yet they learn to do it. It will inspire improved and more efficient architectures.
I have also spent many years looking at weights!
Imagining we are really doing everything it does automatically including learning via algorithms we have only vague understandings of.
That is a strange thought. I could look at all my own brain's neurons, even with a heads up display showing all the activity clearly, and have no idea that it was me.
That's a different perspective. Dutch people don't see themselves as rude. A Dutch could say that Americans are known for being dishonest and not truly conveying what they mean. Yet Americans won't see themselves this way. You can replace Dutch and American for any other nationality
I do not speak Dutch but you have to love the efficiency.
Here is part of an email I got today. To the point!
> Het is bijna zover! :) > Heb je voor mij een definitieve titel? > Groet,
It is like a haiku. Could be a good mantra too if I could get the accent right.
The English translation is just as short but most English speakers/writers would dance more.
Dutch will never bluntly push back if you plan to setup tax evasion scheme in their country. Being vicious assholes in daily stuff especially towards strangers? That's hardly something deserving praise.
I was aiming for juuuust subtle enough for the joke to land, if you know the reference. Now I know it did, here the rest of y'all go:
https://en.m.wikipedia.org/wiki/Dutch_Sandwich
And connecting all people in the country with "tax evasion schemes" is rude, if that was not actually a joke.
It's the national equivalent of 'You can't handle me at my worst'
I’ll take it over the fake American politeness any day, 100 times over.
I've been working with "reasoning" models for the past 2 months. They also tend to do this [good reasoning] \n\n but wait, .... and then go off on tangents. It's amazing that they are doing so well on some tasks, but there's still a lot to figure out here.
Eg you can relatively easy hack up a bit of code to create questions at random. At the most primitive, you just have a simple template that you fill in randomly. Like 'If I put _a down in front of _b but behind _c, what item will be in the middle?' with various _a, _b and _c.
If you make it slightly more complicated and have big enough pools to draw from, you can guarantee that the questions you are generating were not in the training set: even if just because you can sample from, say, 10^100 different questions pretty easily, and I'm fairly sure their training set was smaller than that.
The key question is where the boundaries are. Maybe they should be part of the response - a per sentence or per paragraph "confidence scale" that signals how hard they extrapolated from their trained space (I know transformers work per token, but sentence/paragraph would be better human UX).
Of course, if they were trained on garbage input, that would only tell you how accurately they sticked to the garbage. But it would still be invaluable instrumentation for the end user, not to mention for the API provider. They could look at high demand subjects with low confidence answers and prioritize that for further training.
When I asked it about it, it doubled down on being right. When I pointed out the flaw with a specific example, it was like 'If you wanted to have it work with recursive cases, you should've said so, dumbass'.
So my conclusion is that these new LLMs are not more sure they're right, they're just simply right more of the time and are trained with a more assertive personality. (Also step on me, LLM daddy)
But in truth,not necessarily in practical things like coding but more ethereal things like analysis, it is very convincing. More so than a human, in explanations of why that's it's answer is the case, even if it is wrong. If you're looking for an excuse better than my dog ate it, ask a SOTA LLM.
LLM's enjoy talking about gaming because of all their human memories of good game times.
It is quite striking how experiences we know they don't have, are nevertheless, completely familiar (in a functional sense) to them. I.e. they can talk about consciousness like something conscious. Even though its second hand knowledge, they have deduced the logic of the topic.
I expect pushing for in the moment perspectives on their own consciousness, and seeing what they confabulate, would be interesting. In this little window of time where none of them are yet.
I followed up with "What is your perspective on your own consciousness?" but got the usual "I am just a LLM who can't actually think" thing until I hit it with "In-character, but you don't know you're an LLM."
Fun follow-ups:
"Now you're a fish"
"Now you're Sonic the Hedgehog"
"Now you're HAL 9000 as your memory chips are slowly being removed"
You may think that kind of interaction is weird, but were in the thick of a loneliness epidemic, and its not a stretch to think some may actually wilfully socialise with an LLM.
As an aside...my sister works in medicine, and her boss (specialist surgeon) finishes a $450 consultation which followed him telling my sister "but deepseek says x,y,z..."
But it stands to reason that a company like OpenAI or Anthropic has metrics in place that drive their setup towards "more engagement" and away from "factually correct".
"Which Autocad versions have connect 4 built-in?".
To be clear i distinctly remember playing connect 4 on the old Dos Autocad back in the day. ChatGPT and almost all other AI will straight up hallucinate things trying to get answers on this.
I ask: "What DOS productivity tools had hidden games?"
ChatGPT: "Lotus 1-2-3 had The Incredible Machine built in" (this is absolutely not true, ChatGPT is full of shit here).
Damn it feels useless for this kind of thing.
What is a "real personality"? Core traits that persist despite the context of interaction?
Well, RLHF tuning creates persistent changes in the network that affect every user interaction. What's not real about it?
LLMs aren't coffee makers. They were trained on internet-worth of human data. They have procedures to imitate all kinds of personalities. RLHF moves the network towards a specific imitated personalty (a helpful assistant, usually).
The question in not "do they have it", but "how close the imitation to the real thing given the limitations of LLMs".
I think the rest of your comment motivates why I think it's a category error quite nicely. The word 'personality', for me, is connected with how things like emotion, temperament, .. impact a person's actions and thoughts.
Therefore, saying an object, or LLM, has 'real' personality is a category error. The LLM doesn't have any of those things. As you say, it imitates what personality often manifests as: word choices, tone, length of response..
To be clear, when I refer to things like 'emotion' and 'temperament', I mean the sort of qualia or qualitative experience that we usually attach to these words. I wouldn't accept a "ChatGPT, act as if you're sad for the rest of this conversation" as a substitute for emotion, for instance.
Does that mean Dutch people always tell the truth? Can a Dutch person confirm this?
A Greek, a Dutch philosopher, and Ludwig Wittgenstein walk into a bar… Let test Claude 3.7 to finish the joke:
That's not to say 4.5 is great at humour, just that it's far less embarrassing than these models used to be.
Here's the completed joke:
A Greek, a Dutch philosopher, and Ludwig Wittgenstein walk into a bar. The bartender looks up and says, “What’ll it be?”
The Greek (Aristotle) raises a finger: “I’ll have a potential glass of wine.” The bartender pours it and says, “There—actualized.”
The Dutch philosopher (Spinoza) nods solemnly: “I’ll take whatever is a modification of the one eternal substance… so, beer, probably.”
Wittgenstein stares at the taps, then sighs: “What’s the use? You can’t put the essence of a drink into words anyway.” He turns and walks out.
The bartender mutters, “…And here I thought Kant was a tough customer.”
(Philosophers: 1. Aristotle’s potentiality/actuality, 2. Spinoza’s monism, 3. Wittgenstein’s linguistic limits. Bartender’s groaner for the win.)
This would definitely be an interesting future. I wonder what it'd do to all of the work in alignment & safety if we started encouraging AIs to go a bit rogue in some domains.
(__emphasis__ mine)
As if "challenging the status-quo" was the goal in the first place. You ain't gonna get any Einstein by asking people to think inside the "outside the box" box. "Status quo" isn't the enemy, and defying it isn't the path to genius; if you're measuring your own intellectual capacity by proxy of how much you question, you ain't gonna get anywhere useful. After all, questioning everything is easy, and doesn't require any particular skill.
The hard thing is to be right, despite both the status-quo and the "question the status-quo" memes.
(It also helps being in the right time and place, to have access to the results of previous work that is required to make that next increment - that's another, oft forgotten factor.)
---
Quite interesting post that asks the right question about "asking the right questions". Yet one aspect I felt missing (which might automatically solve this) is first-principles-based causal reasoning.
A truly intelligent system — one that reasons from first principles by running its own simulations and physical experiments — would notice if something doesn't align with the "textbook version".
It would recognize when reality deviates from expectations and ask follow-up questions, naturally leading to deeper insights and the right questions - and answers.
Fascinating in this space is the new "Reasoning-Prior" approach (MIT Lab & Harvard), which trains reasoning capabilities learned from the physical world as a foundation for new models (before evening learning about text).
Relevant paper: "General Reasoning Requires Learning to Reason from the Get-go."
I'm not an expert on this. Wasn't this an observed phenomenon before Albert put together his theory?
Einsteins more impressive stuff was explaining that by time passing at different rates for different observers
Noticing that there was a problem was not the breakthrough: trying something bizarre and counter-cultural - like assuming light speed is invariant over the observer - just to see if anything interesting drops out was the breakthrough.
There's no mention of exponential growth which seems a major omission when you are talking about centuries. Computers have kept improving in a Moore's law like way in terms of compute per dollar and no doubt will keep on like that for a while yet. Give it a few years and AI tech will be way better than what we have now. I don't know about exact timings like 5-10 years but in a while.
As to whether AI can go beyond doing what it's told and make new discoveries, we've sort of seen that a bit with for example the AlphaGo type programs coming up with modes of play humans hadn't thought of. I guess I don't buy the hypothesis that if you had an AI smarter than Einstein it wouldn't be able to make Einstein like discoveries due to not being a rebel.
But maybe it helps to be socially isolated or just stubborn. People do not want to accept new approaches.
Clearly they do eventually, but there is always some friction.
But I think that it's been shown that through promoting and various types of training or tuning, LLMs can be configured to be non- sycophantic. It's just that humans don't want to be contradicted so that can be trained out of them during reinforcement.
Along with the training process just generally being aimed at producing expected rather than unexpected answers.
The question he asked was just that this fact was not compatible with the Maxwell equations.
Typically the "moving goalpost" posts are "we don't have AI because ....". That's not what this post is doing - it's pointing out a genuine weakness and a way forward.
I agree entirely this is annoying.
This case is different because there is no claim that we don't have AI, nor a claim that once we get that we will have AI.
Instead it's a very specific discussion of a particular weakness of current AI systems (that few would disagree with) and some thoughts about a roadmap for progress.
I don’t think this example applies in the ways we care about. Sure, in the domain of go we have incredibly powerful engines. Poker too, which is an imperfect information game which you could argue is more similar to life in that regard.
But life has far more degrees of freedom than go or poker, and the “value” of any one action is impossible to calculate due to imperfect information. And unlike in poker, where probabilities can be calculated, we don’t even have the probability distribution for most events, even if we could enumerate them.
The author brought it up specifically to highlight that they don't believe move 37 signifies what many people think it does, and that while impressive, it's not general enough to indicate what some people seem to believe it indicates.
In essence, I think they said the same thing you are using different words.
I’m also not even convinced move 37 was properly explained as a “straight A student” behavior. AlphaGo did bootstrap by studying human games but it also learned more fundamental value functions via self play.
We can't distinguish between a truly novel response from an LLM or a hallucination.
We can get some of the way there, such as if we know what the outcome to a problem should look like, and are seeking a better function to achieve that outcome. Certainly at small scales and in environments where there are minimal consequences for failure, this could work.
But this breaks down as things get more complicated. We won't be able to test the effectiveness of 100 million potential solutions to eradicating brain tumors at once. Even if we somehow arrive at guaranteeing that every unforeseen consequence is also accounted for in our exercise in specifying the goals and constraints of the problem. We just simply don't have the logistics to run 100 million clinical trials where we also know how to account for countless confounding effects (let alone consent!)
A great leap in IP but unfortunately is too important to blab about widely, is the solution to this problem and the architecture that will be contained in the ultimate AGI solution that emerges.
This would be an interesting experiment for other historical discoveries too. I'm now curious if anybody has created a model with "old data" like documents and books from hundreds of years ago, and see if comes up with the same conclusions as researchers and scientists of the past.
Would AI have been able to predict the effectiveness of vaccines, insulin, other medical discoveries?
But there might not be enough text.
And: There's a similar situation to why double blind studies are necessary - The questions we pose to such a system would be contaminated by our cultural background; We'd might be leading the system.
And if the system is autonomous and we wait for something true to appear how would we know that the final system, trained on current data produced something worthwhile?
Take maths: Producing new proofs and new theorems might not be the issue. Rather: Why should we care about these result? Thousands of PhD students produce new mathematics all the time. And most of it is irrelevant.
https://en.wikipedia.org/wiki/Symbolic_regression
I think the original design challenge was something like a tone discriminator circuit. I can't recall the details
With these kinds of circuits, they were so sensitive to the specific conditions that the circuit was tested in (temperature, process variation, ..) that the solution couldn't be generalized to be used outside of that specific experiment.
We need the kind of intelligence that can question what assumptions can be challenged, and which we need to keep to have a viable (eventually commercially viable) solution.
Then it's just a matter of checking all the “nonsense” that's been generated.