They're not going to go away for the same reason that the social media companies are so hesitant to openly report on the number of bots on their platform. For the same reason that Tesla still calls their assisted driving "AutoPilot": They make too much money on the suckers.
Of course they can be fixed. To say otherwise is nonsense. The questions are:
1. How to do it? My thought is the answer involves building AI's that are much more than just LLM's. You need a system that has knowledge and a world model and common-sense reasoning, etc. When you're doing nothing but "predict the next token based on statistical patterns" then yeah, you're going to get hallucinations.*
2. How much will it cost?
3. How long will it take?
And so on.
* Let me add that that's a very "off the cuff" view of my current thinking on building AI's that are less susceptible to "hallucinations" (or "confabulations" if you prefer). But I'm not saying that is the only way, or that it will ever be possible to construct AI's that never make mistakes at all. After all, our current "gold standard" for intelligence to compare with (eg, human intelligence) is quite prone to making mistakes. So I'm only talking about getting to a level that eliminates the really egregious hallucofabulations (to coin a word).
When people say that something can’t be done, I think the majority of the time they mean (but don’t say) that it can’t be done within unspecified x,y,z constraints.
Those constraints, as you eluded to, are most often things like “within a certain timeframe” or “under a certain cost”.
Saying that something cannot be done without specifying those unspoken constraints - or being explicit that you do actually mean it can’t be done, even on an infinite timescale or with infinite money - is kind of a useless statement to make IMHO.
I’ve never really understood that tendency but maybe working in tech kind of conditions you to think in a more flexible way about what may be possible?
Yeah, like when clients ask me if something is possible. Happens incredibly frequently. What they actually mean to ask is how much would it cost to do. So I usually answer in those terms.
Well, yes, by taking AI to mean something different than the current meaning. Of course if you build something that's not an LLM you may find a way to fix the confabulation* problem, but that isn't fixing what we have now.
The history of 'AI' the term is one of continually redefining it to mean something different, as the winds of technology and business blow. In this case, the assertion is that, under what we currently call 'AI', there's not a fix.
Well, I agree with your general sentiment to a point, but I would say that in the lexicon of the day, "AI" and "LLM" are not synonymous. LLM's are one (current very popular and useful) approach to AI, but only a VERY casual observer would conflate the two terms as being completely equivalent. To me, it goes without saying that the world of AI is bigger than just LLM's.
But since we're talking about an article in Fortune, maybe it's fair to think of this strictly from a lay-person perspective.
Also, to the extent that it matters, I'm not necessarily saying "build something (completely) different than an LLM". I can picture a world with systems that include LLM's that interoperate with other components that do more of the "world model" and "commonsense reasoning" parts, yielding an aggregate system that does more/better than any of the parts could do in isolation.
No argument from me. I see good uses for LLMs, both now, with care, and as part of a toolkit for building towards more general intelligence.
One way to think of LLMs as available now is as sort of thinking partner, much like bouncing ideas off another person, except LLMs use the ideas of entire communities of persons. Of course some of the ideas bounced back at you from thousands or millions of people are going to get things wrong. Don't be like the lawyers who blindly used ChatGPT to write their legal briefs without even checking the citations, and you'll be OK.
From 2008 there was How to Build a Person by John Pollock. I don't recall whether he included some putative "explainer" disciplined by other components.
Heh. Seems to be a common pattern in book titles. There are also How to Build A Brain[1] by Chris Eliasmith and How To Create A Mind[2] by Ray Kurzweil as well.
Without discussion of grammars, I think you're lost in the sauce on this.
> You need a system that has knowledge and a world model and common-sense reasoning, etc. When you're doing nothing but "predict the next token based on statistical patterns" then yeah, you're going to get hallucinations.
Arguably, LLMs already have this. They can do impressive reasoning already. The problem may be that the way they encode all that is untenable. Imagine trying to describe a car in terms of its individual pieces (tires, suspension, engine). Then imagine trying to do that in terms of its molecules. Uh oh. Then its atoms. Even worse.
> How much will it cost?
To describe a car in terms of atoms? Infinite cost. We can barely even get a baby fruit fly's connectome (total brain schematic) read and stored somewhere. A car is much bigger.
It may be that LLMs are structured such that they're trying to encode all that common sense reasoning in terms of atoms, and a fundamentally different grammar is necessary to go up to parts.
Whether this is the case is still being debated. If it is the case, you'll see LLMs near a ceiling of diminishing returns.
> How much will it cost?
To figure out how to do LLMs with more expressive grammars? Unknown. I think that'd take a qualitative breakthrough in AI, and I think LLMs are just the result of a quantitative breakthrough (we're still doing the old, non-clever stuff; we're just doing it faster than ever before thanks to continued growth in GPU compute power; more monkeys on typewriters than ever before!). I don't think we've had much in the way of qualitative breakthroughs since the AI winter.
Arguably, LLMs already have this. They can do impressive reasoning already. The problem may be that the way they encode all that is untenable. Imagine trying to describe a car in terms of its individual pieces (tires, suspension, engine). Then imagine trying to do that in terms of its molecules. Uh oh. Then its atoms. Even worse.
I will freely admit, I'm torn on how to think about this. On the one hand, some of the kinds of mistakes LLM's make lead me to think that they aren't exactly "doing reasoning" as I think of it. But the things they can do hint at some sort of reasoning.
Part of me wants to think of it as "they can simulate reasoning", but there's a fair objection to be raised that "simulating reasoning is just as good as doing reasoning." Part of me wants to reject that, but then I remind myself that I've always claimed to be something like a "behaviorist" when it comes to evaluating AI's. That is, I've always tried to eschew those "it's not real thinking" arguments and focus on the observed behavior, holding to a "if it displays intelligent behavior then it is intelligent" mindset.
By no conception of reasoning I'm familiar with do LLMs reason, at all. They are clever software systems of an advanced sort of autocorrect/predictive text kind.
I'm not sold on this. The ability to often reasonably combine knowledge to derive new solutions to a problem to me implies some kind of model of a world. If it fails it might as well just be a model of a very different, wrong world. The idea that we have "persistent memory" of contextual knowledge is also questionable, relations generally don't seem to be hard-coded in the brain.
In the optimal scenario, it would just be sufficient to massively increase the context space of current LLMs.
Isn’t an AI “hallucination” just the equivalent of going down a wrong path due to closely related or incomplete information in the LLM? I thought that it was possible to “tune” the LLM when this happens?
Maybe “one-size-fits-all” solutions won’t be fixed; but it seems that domain-specific systems should be easily tailored.
I think that's right. Most likely general AI will be some fractal type system where the larger models first outline the boundaries then smaller models refine the pathways.
I'd say 2-3 years the money proposition will be a market of small, efficient models with clear domain bias and you build out by overlapping the domains till you cover the footprint.
These large models just interconnect too much to efficiently produce reliable output unless you don't care about accuracy(eg, art or far right propaganda)
I believe a more accurate term would be AI confabulation.
When one confabulates, one shamelessly inserts words which sound smooth and grammatical and potentially acceptable to a listener (and the speaker), but shamelessly with no hesitation or corrections to achieve actual truth.
I suspect there needs to be some layering on of reflection and inhibition to avoid this.
Perhaps with both 'hallucination' and 'confabulation' there are suggestions of both thought and absence of thought. The former suggestions are likely anthropomorphic. Suggesting some seamless lack of thought might not be?
A meta point as someone who has been working in mathematical signal processing and ML professionally for a while -- at the time of writing, every single comment in this thread displayed a serious (yet unique!) misconception about AI, either as it works today or in general. I understand why it's gotten so buzzy in popular discourse, but it's sadly all but ruined discussion on social media (including HN).
Are there any good, maybe more gatekept forums for discussing AI news? It is great to have a community to discuss with, the field is so large and moving so fast, and there's so much to learn from other practitioners, but the SNR here (where i often find news) has just gotten so low.
That's not how public forums work. Unless you are put on a pedestal (which is effectively what writing your own articles, joining larger organizations, making your own content, etc does), you will be drowned out by misunderstandings and misinformation.
There will always be a much greater effort required tackling bad information than it takes to post that bad information. This is why public forums are a terrible place for debates or discussions. It's also why you're considering their post a lower SNR when it's really not, certainly not in comparison to those just claiming that it's possible with a sort of blind optimism or ranting about bots or what have you. Important to signify here, I don't lean in any which direction on this specific topic, but just highlighting how the low effort posts being viewed as higher SNR to you is exactly what leads to more noise on public forums.
As someone who also works professionally in this field, I'd be curious to know what you find problematic about my comment. I'll grant you that I may be over-simplifying a bit, but it seems appropriate in the context at hand.
Actually you're right, that was my mistake. Current-gen autoregressive token predictor "hallucinations" probably can't be fixed, but yeah, the way forward will involve new architectures and I would place my bets that something like world models will be involved to ground generative output.
Humans cannot 'fix' other humans' hallucinating, embellishing, or outright lying so why would we think we could 'fix' that in a system designed by, and operated by, humans?
This issue goes away if one accept that LLM projects produce opinions not facts. IMHO LLM responses resemble our own initial opinions about an issue. Our initial opinions are often flawed and require review. I wouldn't be surprised if there is actually a deep equivalence between LLM processes and the sub-conscious processes from which we form our opinions.
38 comments
[ 2.9 ms ] story [ 95.0 ms ] thread1. How to do it? My thought is the answer involves building AI's that are much more than just LLM's. You need a system that has knowledge and a world model and common-sense reasoning, etc. When you're doing nothing but "predict the next token based on statistical patterns" then yeah, you're going to get hallucinations.*
2. How much will it cost?
3. How long will it take?
And so on.
* Let me add that that's a very "off the cuff" view of my current thinking on building AI's that are less susceptible to "hallucinations" (or "confabulations" if you prefer). But I'm not saying that is the only way, or that it will ever be possible to construct AI's that never make mistakes at all. After all, our current "gold standard" for intelligence to compare with (eg, human intelligence) is quite prone to making mistakes. So I'm only talking about getting to a level that eliminates the really egregious hallucofabulations (to coin a word).
Those constraints, as you eluded to, are most often things like “within a certain timeframe” or “under a certain cost”.
Saying that something cannot be done without specifying those unspoken constraints - or being explicit that you do actually mean it can’t be done, even on an infinite timescale or with infinite money - is kind of a useless statement to make IMHO.
I’ve never really understood that tendency but maybe working in tech kind of conditions you to think in a more flexible way about what may be possible?
Well, yes, by taking AI to mean something different than the current meaning. Of course if you build something that's not an LLM you may find a way to fix the confabulation* problem, but that isn't fixing what we have now.
The history of 'AI' the term is one of continually redefining it to mean something different, as the winds of technology and business blow. In this case, the assertion is that, under what we currently call 'AI', there's not a fix.
* I prefer the term confabulation over hallucination. See <https://universeodon.com/@siderea/109883198218504351>
But since we're talking about an article in Fortune, maybe it's fair to think of this strictly from a lay-person perspective.
Also, to the extent that it matters, I'm not necessarily saying "build something (completely) different than an LLM". I can picture a world with systems that include LLM's that interoperate with other components that do more of the "world model" and "commonsense reasoning" parts, yielding an aggregate system that does more/better than any of the parts could do in isolation.
One way to think of LLMs as available now is as sort of thinking partner, much like bouncing ideas off another person, except LLMs use the ideas of entire communities of persons. Of course some of the ideas bounced back at you from thousands or millions of people are going to get things wrong. Don't be like the lawyers who blindly used ChatGPT to write their legal briefs without even checking the citations, and you'll be OK.
[1]: https://www.amazon.com/How-Build-Brain-Architecture-Architec...
[2]: https://www.amazon.com/How-Create-Mind-Thought-Revealed/dp/1...
> You need a system that has knowledge and a world model and common-sense reasoning, etc. When you're doing nothing but "predict the next token based on statistical patterns" then yeah, you're going to get hallucinations.
Arguably, LLMs already have this. They can do impressive reasoning already. The problem may be that the way they encode all that is untenable. Imagine trying to describe a car in terms of its individual pieces (tires, suspension, engine). Then imagine trying to do that in terms of its molecules. Uh oh. Then its atoms. Even worse.
> How much will it cost?
To describe a car in terms of atoms? Infinite cost. We can barely even get a baby fruit fly's connectome (total brain schematic) read and stored somewhere. A car is much bigger.
It may be that LLMs are structured such that they're trying to encode all that common sense reasoning in terms of atoms, and a fundamentally different grammar is necessary to go up to parts.
Whether this is the case is still being debated. If it is the case, you'll see LLMs near a ceiling of diminishing returns.
> How much will it cost?
To figure out how to do LLMs with more expressive grammars? Unknown. I think that'd take a qualitative breakthrough in AI, and I think LLMs are just the result of a quantitative breakthrough (we're still doing the old, non-clever stuff; we're just doing it faster than ever before thanks to continued growth in GPU compute power; more monkeys on typewriters than ever before!). I don't think we've had much in the way of qualitative breakthroughs since the AI winter.
I will freely admit, I'm torn on how to think about this. On the one hand, some of the kinds of mistakes LLM's make lead me to think that they aren't exactly "doing reasoning" as I think of it. But the things they can do hint at some sort of reasoning.
Part of me wants to think of it as "they can simulate reasoning", but there's a fair objection to be raised that "simulating reasoning is just as good as doing reasoning." Part of me wants to reject that, but then I remind myself that I've always claimed to be something like a "behaviorist" when it comes to evaluating AI's. That is, I've always tried to eschew those "it's not real thinking" arguments and focus on the observed behavior, holding to a "if it displays intelligent behavior then it is intelligent" mindset.
LLM's have me sort of vacillating on all of this.
By no conception of reasoning I'm familiar with do LLMs reason, at all. They are clever software systems of an advanced sort of autocorrect/predictive text kind.
In the optimal scenario, it would just be sufficient to massively increase the context space of current LLMs.
Maybe “one-size-fits-all” solutions won’t be fixed; but it seems that domain-specific systems should be easily tailored.
https://gptea.beehiiv.com/p/ai-hallucinations-explained
I'd say 2-3 years the money proposition will be a market of small, efficient models with clear domain bias and you build out by overlapping the domains till you cover the footprint.
These large models just interconnect too much to efficiently produce reliable output unless you don't care about accuracy(eg, art or far right propaganda)
When one confabulates, one shamelessly inserts words which sound smooth and grammatical and potentially acceptable to a listener (and the speaker), but shamelessly with no hesitation or corrections to achieve actual truth.
I suspect there needs to be some layering on of reflection and inhibition to avoid this.
The generated text is wrong, false, inaccurate. I don't think we really need a jargon-y term for it.
At least the day wasn't all lost. Learnt a new word.
Are there any good, maybe more gatekept forums for discussing AI news? It is great to have a community to discuss with, the field is so large and moving so fast, and there's so much to learn from other practitioners, but the SNR here (where i often find news) has just gotten so low.
From my perspective, this is easily the lowest SNR comment on this post.
There will always be a much greater effort required tackling bad information than it takes to post that bad information. This is why public forums are a terrible place for debates or discussions. It's also why you're considering their post a lower SNR when it's really not, certainly not in comparison to those just claiming that it's possible with a sort of blind optimism or ranting about bots or what have you. Important to signify here, I don't lean in any which direction on this specific topic, but just highlighting how the low effort posts being viewed as higher SNR to you is exactly what leads to more noise on public forums.
Early bird gets the worm, and all that fun jazz!