Eliezer Yudkowsky on large language model economics
twitter.com
twitter.com
What he's describing is common knowledge. This technique is called knowledge distillation (there are some small differences, but the idea is the same). A large model (here, openai's davinci) is used as a teacher model to train a smaller student model (llama). This is used practically everywhere in industry to improve inference times.
Perhaps I am a little behind the times, but what are reasons/achievements/credentials behind his huge, prophet like status within the cult of rationalism? Beyond writing a popular Harry Potter fanfic, I mean.
The rationalists are an interesting group. I like a lot of the vibe of the whole thing, but there is also some very weird navel gazing that has cropped up. It can be hard to differentiate from the good and the bad.
I am overall a little surprised by the vitriol towards EY in this thread though. While I agree that the guy is somewhat shrill and his ideas are pretty out there, he really doesn't seem like a con artist from my standpoint. It seems to me that he is genuinely extremely concerned about AI alignment/risk, and his actions seem to make sense in that framework without needing to ascribe a bunch of weird and sinister motivations to him.
Not professing agreement or disagreement with him on any particular thing, just my two cents.
As somewhat of a cynical person, I also find it funny that
> While I agree that the guy is somewhat shrill and his ideas are pretty out there, he really doesn't seem like a con artist from my standpoint.
I've heard this exact same sentiment about another popular Effective Altruist in the past year. I will let the reader guess which one.
The Objectivists I know hate hypotheticals and refuge to engage with them with the same dismissive tone used in your comment.
I may be missing something, as I don't really follow this "community" besides reading the occasional Slate Star Codex post or whatever.
But I don't see much ground for comparison here. SBF (who I assume you're referring to) was in much different position with regards to potential harm as a result of con man antics than EY is. (As far as I can tell).
SBF was managing billions of dollars. EY is not. If SBF turns out to be a scammer, real people lose money. If EY is a scammer... then so what? He publishes a bunch of blog posts crying doom about AI that suddenly look less sincere? To what end? I can't see a good reason at this time for assuming his reasons are anything less than what he claims they are.
I'm not surprised by the vitriol, though. He has got to be one of the most arrogant people alive, and people arrogantly writing as though they are world experts on everything when they are, like, dumb and wrong... is extremely irritating.
Essentially him and his crowd have been trying to create a rationally grounded purely logical non contradictory moral framework for action from axiomatic first principles (he’s not the first or the last to do this), and they believe such a system is essential if AGI comes to be.
I personally think AGI is impossible, his purely logical moral system is impossible, and think what he’s working on is not applicable to current AI (I think he knows this, but his conclusion seems to be that the path to AGI is out of the bag/was made incredibly foolishly and should have been constructed differently with more ability to add guardrails)
But I do think the high level abstract way he thinks is fruitful/very valuable, and a good moderating counterbalance to the frenzied “harder better faster stronger” ML stampede that’s not really thinking about the ramifications of what they’re doing with any sort of discipline.
That’s a very weird statement to make. Do you think you have a soul or something? Otherwise how is your own Intelligence something a computer can’t replicate?
So, I really don’t get why people think AGI is impossible.
I mean, I have hope that it won’t ever happen, but not because I suspect it is impossible.
However I do think it’s quite possible and likely that we’ll create mimics that convince most people they’re intelligent. But I think that’s a weird type of reflection/anything we can make is limited to mirroring our observable past and existing collective thought and can never be truly introspective, because true introspection requires evolutionary context and hidden motivation you can’t transfer.
It doesn't seem very logical to think that because evolution took so long to get us to where we are now, we consequently won't be able to design an intelligent AGI system.
Imagine an amoeba that can only detect light and dark. Its observable universe exists on a simple gradient. But the cells that are used in its perceptual system cannot be described on a gradient. I think that probably scales up indefinitely.
We confuse what we see for all that exists because of our recent history and scientific advances. The triumph of science comes from focusing solely on what we can see and reason about and improving our vision because thats what we have control over, but it doesn’t mean we see everything. I think true intelligence originates from something we can’t see. You can call it a soul, or call it a scaled up amoeba cell, but I personally think the origin of true intelligence and qualia is from some weird very very old evolutionarily created thing we can’t see (and I think other animals have similar invisible perceptual machinery somewhere, I don’t think it’s just a human thing).
I suppose there is a discussion to be had about what "true intelligence" is and whether some or most human beings have it in the first place.
Gödel proved it’s impossible for any deterministic system to determine its own correctness close to a hundred years ago. Whatever we’re doing when we say an output A is “right” and output B is “wrong” is fundamentally different than what a propositional deterministic system does, regardless of how large or chaotic or well fitted to expected output it is.
What qualifies as “true intelligence” is the core issue here, yes, and transformers don’t qualify. That doesn’t mean they aren’t valuable or can’t give correct results to things much faster and better than humans, but I think anything we could deliberately design inevitably ends up being deterministic and subject to the introspective limitations of any such system, no matter how big or all encompassing.
I think we’re going to create better and more comprehensive “checkpoint”/“savepoints” for the results of collective intelligence, but that’s different than the intelligence itself.
> Gödel proved it’s impossible for any deterministic system to determine its own correctness close to a hundred years ago.
I don't think non-deterministic systems fare much better. At the core, completeness is a quantitative problem, not a qualitative one. In order to fully understand and analyze a system of a certain size, there is no way around the simple fact that you need a bigger system. There are more than ten things to know about ten things, so if you only have the ten things to play with, you will never be able to use them to represent a fact that can only be represented using eleven things. Gödel's theorem pushes this to infinity, which is something that we are notoriously poor at conceptualizing. I think this just obfuscates how obvious this limitation is when you only consider finite systems.
Which brings me to the core disagreement that I think we have, which is that you appear to believe that humans can do this, whereas I believe they blatantly cannot. You speak of the grounding problem as if the human brain solves it. It doesn't. Our reasoning is not, has never been, and will never be grounded. We are just pretty damn good at lying to ourselves.
I think that our own capacity for introspection is deeply flawed and that the beliefs we have about ourselves are unreliable. The vast array of contradictory beliefs that are routinely and strongly held by people indicates that the brain reasons paraconsistently: incoherence, contradiction and untruth are not major problems and do not impede the brain's functioning significantly. And so I think we have a bit of a double standard here: why do we care about Gödel and what deterministic systems can or cannot do, in a world where more than half of the population, but let's be honest, it's probably all of us, is rife with unhinged beliefs? Look around us! Where do you see grounded beliefs? Where do you see reliable introspection?
But I'll tell you what. Evolution is mainly concerned with navigating and manipulating the physical world in a reliable and robust way: it is about designing flexible bodies that can heal themselves, immune systems, low-power adaptable neural networks, and so on. And that, strangely enough, is not something AI does well. Why? My theory is that human intelligence and introspection are relatively trivial: they are high impact, for sure, but as far as actual complexity goes, they are orders of magnitude simpler than everything else that we have in common with other animals. Machines will introspect before they can cut carrots, because regardless of what we like to think about ourselves, introspection is a simpler task than cutting carrots.
I think this explains the current trajectory of AI rather well: we are quickly building the top of the pyramid, except that we are laying it directly on the ground. Its failure will not be intelligence, introspection or creativity, it will be mechatronics and the basic understanding of reality that we share with the rest of the animal kingdom. AKA the stuff that is actually complicated.
Our perception of the physical world is itself an evolved construct. The wider thing that perception is fitting is something I don’t think we can fit a machine to. I think we can only fit it to the construct.
I get where you’re coming from/appreciate the argument, and you may be right, but I‘ve come to appreciate the ubiquity of hidden context more and more/lean that way.
Strictly speaking, it is for you to say why you think that a digital computer, which is nothing like a human, can be intelligent like a human.
Btw, I see this quip about souls often on HN, in response to comments that AGI is impossible and it is invariably introduced in the conversation by the person arguing that AGI is possible. I have never, ever seen anyone argue that "digital computers can't be intelligent because they don't have a soul". I don't even know where that "soul" bit comes from, probably from something someone said centuries ago? In any case it's irrelevant to most modern conversations where most people don't believe in the supernatural, and have other reasons to think that AGI may be impossible.
And why wouldn't they? For the time being, we don't have anything anyone is prepared to call "AGI". We also don't have a crystal ball to see into the future and know whether it is possible, or not. For all we know, it might be a theoretical possibility, but a practical impossibility, like stable wormholes, or time travel, or the Alcubierre Drive. We will know when we know, not before.
Until then, invoking "souls", or any other canned reply, only seems to serve to shut down curious conversation and should be avoided.
To say it in very plain English: you don't know if AGI is possible, you only assume it to be, so let's hear the other person say what they assume, too. Their opinion is just as interesting as yours, and just as an opinion.
> [...]
>If the Great Secret of Natural Selection, passed down from Darwin Who Is Not Forgotten, was only ever imparted to you after you paid $2000 and went through a ceremony involving torches and robes and masks and sacrificing an ox, then when you were shown the fossils, and shown the optic cable going through the retina under a microscope, and finally told the Truth, you would say "That's the most brilliant thing ever!" and be satisfied. After that, if some other cult tried to tell you it was actually a bearded man in the sky 6000 years ago, you'd laugh like hell.
>And you know, it might actually be more fun to do things that way. Especially if the initiation required you to put together some of the evidence for yourself—together, or with classmates—before you could tell your Science Sensei you were ready to advance to the next level. It wouldn't be efficient, sure, but it would be fun.
>[...]
>And no, I'm not seriously proposing that we try to reverse the last five hundred years of openness and classify all the science secret.
~ Eliezer Yudkowsky, https://www.lesswrong.com/s/6BFkmEgre7uwhDxDR/p/3diLhMELXxM8...
That's really all there is to Yudkowsky's persona of occultism at the end of the day. It's an intentional affectation.
I think the most impressive feat of GPT/LLMs thus far has been all the ML experts its caused to materialize from thin air.
> You may not ... use the Services to develop foundation models or other large scale models that compete with OpenAI
Also, think this stuff is becoming a commodity real quick.
(2) "Training a smaller model off davinci using the logits" would have been vastly more expensive than what they did, which was to fine-tune another pretrained foundation model, which happened to be smaller, using output of another foundation model, which happened to be larger. The relative size isn't the key idea, it's that they were able to extract the fine-tuning for instruction-following (which is relatively computationally cheap to train, but requires a custom expensive small dataset) and port it from one foundation model to another. This is what has implications for the companies that planned to have this custom fine-tuning as part of their competitive moat.
(3) I understand that ad hominem is against the culture that Hacker News tries to cultivate here. Consider reading or rereading the local guidelines on discussion, as well as reading or rereading an article on knowledge distillation, the Stanford publication of Alpaca, and my tweet.
Here, the Stanford authors are not doing anything clever to enable a smaller model. They're just yoinking the fine-tuning onto what happens to be a smaller model. The destination model being smaller is not the point. The cheap yoinking of just the instruction tuning is the point. They used a small model as the destination because that was cheaper.
> This sequence-level approximation leads to a simple training procedure wherein the student network is trained on a newly generated dataset that is the result of running beam search with the teacher network
2) Fine-tuning is a step in the training process. Language models are first pre-trained, then fine-tuned. This is a pedantic quibble.
3) It is unsurprising that you don't understand ad hominem. Giving background information and pointing out the style of writing is relevant to arguments made.
I threw a null cultural reference exception on that one. Can you please clarify?
Edit: I know the "Masks of Nyarlathotep" and what a shoggoth is; I just don't know what is the "mask of the shoggoth".
Edit 2: Really, the one thing I wanted to be able to do with a language generator has always been to get it to GM a CoC campaign for me. It's still the case that there's no chance of that. Just as an aside.
The implications aren't quite common knowledge yet: a lot of companies are still operating with the "we'll beat the competition by building a better AI than them" mindset.
I don't think the broader market has caught up with the idea that investing in AI gives you no long-term competitive advantage. (eg StabilityAI hasn't gone bankrupt yet)
As a matter of fact, the API/SaaS business for machine learning 'direct' model results have suffered from this since the early days. Now, what may make more sense and survive longer are full data-storage-oriented platforms, and niche products that encapsulate all underlying model calls, including to publicly available models. Sometimes glue code or even devops becomes the true solid business case of AI/ML apps ;)
Right now we're in this weird phase where we have no accurate metrics for how good an LM is, so people just try them with a handful of queries and are subjectively impressed or not. Maybe clones can pass that test, but as the market gets more sophisticated, people will prefer superior models.
People's mouths were watering over the commercial implications of the recent 90% drop in cost for the new ChatGPT model. Now imagine if you can get similar performance on a model that requires <5% of the parameters.
which would be a much easier and more reliable(explainable) way to do it. just find and parrot the facts.
> The AI companies that make profits will be ones that either have a competitive moat not based on the capabilities of their model,
This is the “open source” case, where profitable firm provide ancillary services and support, rather than directly selling the capabilities of the software, because the software is not at all exclusive.
> OR those which don’t expose the underlying inputs and outputs of their model to customers
This is the “internal software” case, where the profitable firm does not provide access to the software at all, but uses it internally in a way which is opaque from the outside to support other products or services it provides, maintaining exclusivity because no one outside actually gets access to the software itself. (Arguably, non-source-available closed source software might fit in here, though its often possible to reverse engineer software that you have access to without the source, from the executable, so I'd argue that even without source available its more the next category.)
> OR can successfully sue any competitor that engages in shoggoth mask cloning.
This is the “proprietary software” case, where the profitable firm protects its exclusivity through legal constraints on people who have access to the software, while giving customers access to it.
Given that the big firms in AI largely either are or are closely associated with big firms in software that have successfully used all three models for different parts of their portfolios, I don’t think that them figuring out AI profitability is particularly challenged by this “new idiom”.
Web hosting will always be less demanding, unless something weird happens and GPU/APU/TPU type hardware (along with HBM and related ancillaries) somehow becomes less expensive than (or subsumes) general purpose CPUs and RAM.
https://news.ycombinator.com/item?id=35113781
> Perhaps we will provide feedback to open source Llama using ChatGPT. The cost to adjust the model is presumably what's hard?
What I am so impressed by is that the state of the art is so democratic now. All the top AI posts on HN are things that I was able to get to myself days before the were posted here. This must mean either that the true SOTA is way ahead of what we see or that anyone can hit the current SOTA. Both are very exciting developments!
TL;DR: If your customers can get text input-output samples from your super-secret LLM, whether via web interface or API, then your customers can easily and cheaply train their own LLM to learn to act like your super-secret one.
For example, if you're OpenAI, your customers can spend ~$100 to get input-output samples from text-davinci-003, plus another ~$500 to fine-tune LLaMA on those samples, and -- voila! -- they get Alpaca.
I was interested in Scientology as a microcosm of society around the time alt.religion.scientology appeared in the early 1990s. Scientology documented everything about how it runs, even management is scriptural, and only Hubbard could write scripture so it is like a specimen that's been fixed to a microscope slide.
Since then renegade "Free Zone" Scientologists have put all of his writings online so it is very accessible.
Dianetics is the Sequences of Scientology in that it is a long, windy, and nonsensical book that selects for people who are perseverant readers who are impervious to critical thinking. (Like one of those Nigerian scam emails.)
Hubbard's hypnotic language is demonstrated in his Philadelphia Doctorate Lectures where he not only mentions the hypnotist Milton Erickson but uses Erickson's techniques extensively. (There is no record of Hubbard meeting Erickson but Hubbard and Erickson were both highly active at the same time in Arizona and I think Hubbard must have sat in on Erickson’s lectures.)
(That is, that weird ramblely kind of writing is by no means innocent but it is misparsed by many people's minds and helps create an altered state of consciousness.)
Hubbard's most satanic (in opposition to conventional Eastern & Western religious ideas) works are Introduction to Scientology Ethics and A New Slant on Life. The first one describes an elaborate system of punishments that are applied to anyone who violates group norms and a system of evaluation that guarantees you will violate those norms by accident. The list of "ethics conditions" is folded so the positive conditions come first and a person who starts reading at the beginning and doesn't go all the way through will miss the batshit crazy stuff. The second describes how a Scientologist is supposed to disconnect themselves from anyone who disagrees with their spiritual journey.
Something Scientology shares with the EY cult (as well as the LaRouche organization and the third international) is the proliferation of front groups. Scientology has Narconon, Criminon, The Way to Happiness Foundation, and many others. The EY cult has "effective altruism", "rationalism", "longtermism", "AI safety" and probably those parties Aella used to run.
It's a typical path that a coercive group makes multiple 360° turns. "Effective Altruism" seems designed to attract morally weak people but seems rational enough at first. I mean, it seemed to make sense when Bill Gates insisted on funding only international NGOs that could prove the value of what they were doing.
Once, however, you are funding "longtermist" organizations that are concerned with imagined problems on the way between here and the glorious future that we've turned all the planetary mass of the Milky Way into Dyson spheres there are no results to measure!
It is like the technique a stage hypnotist uses of putting volunteers through a number of screens that rapidly selects the most suggestible and/or likely to go along subjects. Both in the short term and long term this process expels the trouble makers and leave behind the potential Manchurian Candidates.
It reads like a "Hitler also ate sugar" argument.
It’s true Hubbard divided the world into ‘public’ who would have their bank accounts empty and ‘staff’ who, penniless, would be turned into worker drones.
EY is not building that sort of machine. But Hubbard was operating at peak egalitarianism in the US and EY is operating at a time (and places) that have peak inequality. A Scientology whale gives a few million to the organization, the ideal EY whale is Sam Bankman Fried. Since it is money that is fungible and ‘makes the world go round’ he doesn’t need an army of worked drones when just one rich kid at Stanford or Cambridge can be quietly bled.
Other groups to look at with apocalyptic ideology are the Symbionese Liberation Army (Why make worker drones when you can brainwash Patty Hearst?), People's Temple (...go to San Francisco), Heaven’s Gate, Aum Shinrikyo, etc. It amazes me that one of them hasn’t tried to assassinate an A.I. researcher yet.
>Other groups to look at with apocalyptic ideology are the Symbionese Liberation Army (Why make worker drones when you can brainwash Patty Hearst?), People's Temple (...go to San Francisco), Heaven’s Gate, Aum Shinrikyo, etc. It amazes me that one of them hasn’t tried to assassinate an A.I. researcher yet.
Not obviously bad: have apocalyptic ideology (might actually be a little bit bad, admittedly)
Obviously bad but Eliezer et al. haven't done it: try to assassinate a researcher
It seems EY is now less fashionable, so criticizing him as a crackpot is more acceptable. But I remember a few years ago when he was in full swing, he had a lot of traction even here on HN, and of course in many nerdy circles. Also see that debacle with Roko's Basilisk and all those silly antics. The LessWrong crowd is insufferable.
I can sort of understand why EY had so much pull with internet nerds. He talked the talk, he was confident, and he was a nerd. But I'm also relieved that more people seem to understand now that he is a crackpot, and that claiming one is an expert on AI doesn't make it so, and that writing long essays mixing Bayes, Harry Potter and "rationalism" doesn't really make anyone an expert on anything.
>It's a typical path that a coercive group makes multiple 360° turns.
What is the evidence that Eliezer is running a coercive group? Whom are they coercing into what?
I fully get that LessWrong et al. are obnoxious, use their own terminology, have far-out ideas about the future of AI that are probably wrong, weird moral systems (though any moral system is weird when you look at it hard enough imho)... But you are making the charge that Eliezer is running a cult and doing so on purpose, which is more specific than just running an annoying community around some non-mainstream ideas.
> Some of what appears in it sort of actually makes sense, but when you reproduce it in monosyllables, it turns out to be truisms. It’s perfectly true that when you look at scientists in the West, they’re mostly men, it’s perfectly true that women have had a hard time breaking into the scientific fields, and it’s perfectly true that there are institutional factors determining how science proceeds that reflect power structures. All of this can be described literally in monosyllables, and it turns out to be truisms. On the other hand, you don’t get to be a respected intellectual by presenting truisms in monosyllables.
https://www.openculture.com/2013/07/noam-chomsky-calls-postm...
In contrast, this comment chain was more about EY using large words pretentiously, to suggest expertise or intellectual depth that he doesn't actually have.
And the point is that you don’t need to RLHF as long as you have access to another model that has been trained with RLHF that you can blackbox.
Why does he use 2 different niche analogies to explain the thing that is still quite densely written at the end? Basically I read his point 3 times, and felt it could’ve been distilled into 1 sentence…
“Making profit on AI is basically the same as making profit on other software, except that the sensitive bits include the direct inputs and outputs to the core model, not just the software itself.”
Yes, it could. But then, the attempt to sell the idea this was some kind of radical phase transition that the players in the AI industry (largely, very experienced players in the software industry) simply were fundamentally not prepared for would be a lot harder.
https://www.lesswrong.com/posts/fLRPeXihRaiRo5dyX/the-magnit...
> you're giving away your business crown jewels to competitors that can then nearly-clone your model without all the hard work you did to build up your own fine-tuning dataset.
you're loosing the capacity to continue to earn money for work that is already done. This just means they're not gonna be able to charge rent on the done work for too long.
the work is the training of the model. I chose to reject this to way to frame the issue; i.e. I reject that it is problem that they're giving away that hard work. It's only a problem due to social (market) ideologies of ownership which are necessary only by the logic of trade (or commerce).
As I see it, Eliezer is pointing out hat people who use the "product API"/ "chaptGPT service" can use this in such a way that they can clone it so to stop being a customer.
I see some kind of divine comedy in this (due to my own ideological beliefs).
> OR can successfully sue any competitor that engages in shoggoth mask cloning.
this is just the legal, modern day equivalent of old school "break their knees unless they pay you"
I don't EY is advocating for this, so much as saying "unless AI companies can find a way to establish such a racket, they won't be profitable". That's a factual claim, not a value judgment.
[0] https://en.m.wikipedia.org/wiki/War_Is_a_Racket
[1] https://books.google.co.il/books/about/War_is_a_Racket.html?...
In the same way that proprietary software, or property rights in commercially useful property more generally, are that, yes.
chat history but is that really that important?
It's one api call from an integration stand point.
It's not a marketplace so you don't have network effects much.
the interface isn't all that much.
it's kind of a boring bot. it doesn't even tell dirty jokes.
I really do feel like this is a race to the bottom.