AI capability isn't humanness(research.roundtable.ai) |
AI capability isn't humanness(research.roundtable.ai) |
I think it's the other way round: humans have effectively unbounded training data. We can count exactly how much text any given model saw during training. We know exactly how many images or video frames were used to train it, and so on. Can we count the amount of input humans receive?
I can look at my coffee mug from any angle I want, I can feel it in my hands, I can sniff it, lick it and fiddle with it as much as I want. What happens if I move it away from me? Can I turn it this way, can I lift it up? What does it feel like to drink from this cup? What does it feel like when someone else drinks from my cup? The LLM has no idea because it doesn't have access to sensory data and it can't manipulate real-life objects (yet).
That said, LLMs are still trained on significantly more data pretty much no matter how you look at it. E.g. a blind child might hear 10-15 million words by age 6 vs. trillions for LLMs.
A camera hooked up to the baby's head is absolutely not getting all the input data the baby gets. It's not even getting most of it.
I don't know how to count the amount of words a human encounters in their life, but it does seem plausible that LLMs deal with orders of magnitude more words. What I'm saying is that words aren't the whole picture.
Humans get continuous streams of video, audio, smell, location and other sensory data. Plus, you get data about your impact on the world and the world's impact on you: what happens when you move this thing? What happens when you touch some fire? LLMs don't have this yet, they only have abstract symbols (words, tokens).
So when I look at it from this "sensory" perspective, LLMs don't seem to be getting any data at all here.
The acquired knowledge is a lot less uniform than you’re proposing and in fact is full of gaps a human would never make. And more critically, it is not able to peer into all of its vast knowledge at once, so with every prompt what you get is closer to an “instance of a human” than “all of humanity” as you might think of LLMs.
(I train and dissect LLMs for a living and for fun)
We get a lot of high quality data that's relatively the same. We run the same routines every day, doing more or less the same things, which makes us extremely reliable at what we do but not very worldly.
LLMs get the opposite: sparse, relatively low quality, low modality data that's extremely varied, so they have a much wider breadth of knowledge but they're pretty fragile in comparison since they get relatively little experience on each topic and usually no chance to affirm learning with RL.
Intelligence likely doesn’t require that much data, and it may be more a question of evolutionary chance. After all, human intelligence is largely (if not exclusively) the result of natural selection from random mutations, with a generation count that’s likely smaller than the number of training iterations of LLMs. We haven’t found a way yet to artificially develop a digital equivalent effectively, and the way we are training neural networks might actually be a dead end here.
Humans ship with all the priors evolution has managed to cram into them. LLMs have to rediscover all of it from scratch just by looking at an awful lot of data.
This is clearly untrue. All information a human ever receives is through sensory data. Unless your position is that the intelligence of a brain that was grown in a vat with no inputs would be equivalent to that of a normal person.
Now, does rotating a coffee mug and feeling its weight, seeing it from different angles, etc. improve intelligence? Actually, still yes, if your intelligence test happens to include questions like “is this a picture of a mug” or “which of these objects is closest in weight to a mug”.
But as we interact with other people using mostly language, and since the start of internet a lot of those interactions happen in way similar to how we interact with AI, the difference is not so obvious. We are falling into the Turing test in this, mostly because that test is more about language than about intelligence.
Didn't they used to say that 90% of communication is non-verbal?
Look, that was a while ago, when people met IRL. So maybe not as true today.
Before 2022 (most of history), if you had a long seemingly sensible conversation with something, you could safely assume this other party was a real thinking human mind.
it's like a duck call.
edit, i want to add because this is neural net that's trained to output sensible text, language isn't just the interface.
unlike a website there's no separation between anything, with LLM's the back and front end are all one blob.
edit2: seems I have upset the ducks that think the duck call is a real duck .
I think this is incorrect on two accounts: Yes, transformers and individual layers are parallel, but the entire network is not. On a first level, it's obviously sequential over generated tokens - but even generation of a single token is sequential in the number of layers that the information travels through.
Both those constraints are comparable to the way humans think I believe. (The human brain doesn't have neatly organized layers, but it does have "pathways" where certain brain regions project into other brain regions)
But as we saw over the course of recent months or years, AI outputs are becoming more indistinguishable for human output.
I find myself wondering why Google isn't doing as good a job as you guys. In particular, I'm wondering if you think they've been kind of lazy about the problem because they don't see captcha detection as faring particularly poorly, and perhaps once AI agents really start taking off and becoming prevalent, then Google will buckle down, at which point they'll be able to hone their troves of data to build a newer captcha even better than yours. Or is there an additional secret sauce you guys have?
That kind of leads me to my other question-I assume your secret sauce is the cognitive science, stroop-esque approach to bot detection (which I think is brilliant in the abstract). But I'm curious how that scales. Like, do you go one-by-one for each website you have a partnership with and figure out how a human would likely use the site, or do you use general techniques (besides the obvious ones like typing speed and mouse movements, which almost certainly Google could implement too if it decided to buckle down). Like, can you scale some broad stroop task across websites?
By the way, asking all of this out of admiration. You guys are doing really clever work!
I would bet completely against this, models are becoming more human-like, not less, over time.
What's more likely to change (that would cause a difference) is the work itself changing to adapt to areas where models are already super-human, such as being able to read entire novels in seconds with full attention.
Some people are obsessed with chatting with ghosts. It seems like a rational adult couldn't be seriously harmed by chatting with a ghost, but there are news reports showing that some people get possessed.
It's a better metaphor than parrots, anyway.
more:
This is literally their inhumanness.
A car is not just an engine, it's a drivetrain, a transmission, wheels, steering, all of which affect the end-product and its usability. LLMs are no different, and focusing on alignment without even addressing all the scaffolding that intermediates the exchange between the user and the LLM in an assistant use case seems disingenuous.
And I am sorry to be negative but there is so much bad cognitive science in this article that I couldn't take the product seriously.
> LLMs can be scaled almost arbitrarily in ways biological brains cannot: more parameters, more training compute, more depth.
- Capacity of raw compute is irrelevant without mentioning the complexity of computation task at hand. LLM's can scale - not infinitely - but they solve for O(n^2) tasks. It is also amiss to think human compute = a singular human's head. Language itself is both a tool and protocol of distributed compute among humans. You borrow a lot of your symbolic preprocessing from culture! Like said, this is exactly what LLM's piggyback on.
> We are constantly hit with a large, continuous stream of sensory input, but we cannot process or store more than a very small part of it.
- This is called relevance, and we are so frigging good at it! The fact that machine has to deal with a lot more unprioritized data in a relatively flat O(n^2) problem formulation is a shortcoming, not a feature. Visual cortex is such an opinionated accelerator of processing all that massive data that only the relevant bits need to make to your consciousness. And this architecture was trained for hundreds of millions of years, over trillions of experiment arms - that were in parallel experimenting on everything else too.
> Humans often have to act quickly. Deliberation is slow, so many decisions rely on fast, heuristic processing. In many situations (danger, social interaction, physical movement), waiting for more evidence simply isn't an option.
- Again a lot of this equivocates conscious processing to entire cognition. Anyone who plays sports or music knows to respect the implicit, embodied cognition that goes on to achieve complex motor tasks. We are yet to see a non-massively-fast-forwarded household robot do a mundane kitchen cleaning task, and go play table tennis with the same motor "cortex". Motor planning and articulation is a fantastically complex computation; just because it doesn't make it to our consciousness or instrumented exclusively through language doesn't mean it is not.
> Human thinking works in a slow, step-by-step way. We pay attention to only a few things at a time, and our memory is limited.
- Thinking, Fast and Slow by Kahneman is a fantastic way of getting into how much more complex the mechanism is.
The key point here is as limited in their recall, how good humans are at relevance, because it matters, because it is existential. Therefore when you are using a tool to extend your recall, it is important to see its limitations. Google search having indexed billions of pages is not a feature if it can't bring the top results well. If it gets the capability to sell me whatever it brought up was relevant, that still doesn't mean the results are actually relevant. And this is exactly the degradation of relevance we are seeing in our culture.
I don't care if the language terminal is a human or a machine, if the human was convinced by the low relevance crap of the machine it just a legitimacy laundering scheme. Therefore this is not a tech problem, it is a problem of culture; we need to be simultaneously cultivating epistemic humility, including quitting the Cartesian tyranny of worshipping explicit verbal cognition that is assumed to be locked up in a brain; we have to accept that we are also embodied and social beings that depend on a lot of distributed compute to solve for agency.
Which gives us no information on computational complexity of running that algorithm, or on what it does exactly. Only that it's small.
LLMs don't get that algorithm, so they have to discover certain things the hard way.
If it walks like a duck and quacks like a duck, then it's not anything remotely resembling a duck, and is just a bag of statistics doing surface level duck imitation. According to... ducks, mostly.
I think this would depend entirely on how the sensory impairment came about, since most genetic problems are not isolated, but carry a bunch of other related problems (all of which can impact intelligence).
Lose your eye sight in an accident? I would grant there is likely no difference on average.
Otherwise, the null hypothesis is that intelligence (and a whole host of other problems) are likely worse, on average.
The fact that they can pull it off to this extent was a very surprising finding.
They mentioned the training data is much higher for an LLM, LLM's recall not being uniform was never in question.
No one expects compression to be without loss when you scale below knowledge entropy that exists in your training set.
I am not saying LLMs do simple compression but just pointing a mathematical certainity.
(And I think you don't need to be an expert in creating LLMs to understand them, albeit I think a lot of people here have experience with it aswell so I find the additional emphasis on it moot).
Again just my impression from exposure to many LLMs at various states of training (my last sentence was not an appeal to expertise)
Entirely possible - we just don’t know. The closest thing we have to a real world case study is Helen Keller and other people with significant sensory impairments, who are demonstrably unimpaired in a general cognitive sense, and in many cases more cognitively capable than the average unimpaired person.
This seems like a very philosophical standpoint, rather than practical. And I guess that’s fine, but I feel like the implication is that if an LLM is in some way intelligent, then it was exactly as intelligent before training. So we are talking about “potential intelligence“? Does a stack of GPU’s have “intelligence”?
I’m trying to illustrate that the constraints that apply to LLMs don’t necessarily apply to humans. I don’t believe human intelligence is reliant upon sensory input.
The term cannot mean something measurable or concrete when it’s convenient, but be vague and indefinable when it’s not.
By now, I am fully convinced that this denial is "AI effect" in action. Which, in turn, is nothing but cope and seethe driven by human desire to remain Very Special.