Ok, I want to thank you for finally giving us a concrete falsifiable statement that we can check. I pretended Marseille 40 times before asking Luna 5.6, and the answer was Paris.
So, even with concrete examples, model haters are still wrong.
You also imply the claim that making the distribution of words as the possible next one visible, somehow makes the whole system not intelligent. I would say the exact opposite is true.
By using the embedding vectors, models are aware of precise placement and relative position of words in this hugely dimensional space. No human is capable of such precision. This enables party tricks of "king plus woman minus man" kind. But this also give us a precise point between any two words, no matter how different. What is on the midpoint between volcano and music, for example. No human can precisely answer that, but an embedding can. And we can see which words are closest to this 700 dimensional point.
You see this menu of words as a weakness, and I say it is in fact a sign of super intelligence. And this is all before any reasoning or attention mechanism is even run.
No he says in the actual talk which model it occurred on and it was an older model he was using that caused that to occur with Marseille. They have since corrected it from doing that anymore. It was only used to illustrate the prediction machine that it is...
I don't see the many weighted words as a weakness, I see it opening up what's under the hood of the prediction machine that it is.
LLMs are very cool tech, definitely not a model hater, the use case on when to use it makes a difference, it's not AGI.
You're confusing language use with intelligence. Fair enough, they were fine-tuned to do that, but still.
Great that it has some 700 dimensional model of language.
If that is a sign of super intelligence, then so is an encyclopedia?
Also I'm just curious how do you think it is "more intelligent" for having a vector representation for a meaningless thing such as "the midpoint between volcano and music"?
So, right of the bat, you are warning us that you are going to apply the " no true Sscottman" fallacy, and that we should brace ourselves.
Yes, models posses intelligence, but it is not a true one.
Then you claim that models do not posses world-building capabilities. But this is simply not true. Even ignoring the whole subgenre of scientific papers on exactly that subject, it is not that hard to build some hypothetical scenarios, big or small, and then witness the ease with which models do navigate those worlds.
Yes. And they are criticizing a model for not having a default mode network - as if that is some impossibility rather than just an artifact of the current iteration of the specific architectures we have built so far. Why do people paint with these broad brushes over relatively specific complaints?
LLMs are likely for machine intelligence something like drosophila are to biological intelligence - relatively early on the high dimensional spectrum of possibility. Though it stikes me that in a different way they're little alike - drosophila are relatively small and efficient.
Your text reads much better if you replace word 'intelligence' with 'text generator with some randomness built in'.
This is because you goal is to state how models are not intelligent, but you couldn't attack the generated text itself, so you created a little rider, attached it to the model, and then you attacked the raider.
But, even in that you failed. You compared the source of human randomness in text generation, and called it 'profound' and implied that it is exactly the source of true intelligence. But, then, the temperature, the similar thing in model was "just a randomness slider". Double standard.
A logical fallacy free attack on LLMs would be to show a prompt, and then the response generated by this prompt, where it would be shown that only an entity with no intelligence would generate such a response. Yet, attacks like this are not written here anymore.
You point out that I didn't attack the output itself. But the method you propose is deeply flawed.
I can give you n prompts and m results provided by these prompts, all passed through black boxes. And you can't discern the algorithms or models they have gone through. These boxes can range from simple text generators to MATLAB, Mathematica, CFD applications, correlation engines, linear solvers, mathematical proof-checkers, LLMs, you name it.
For any kind of input they can accept, you can't discern whether the algorithm behind it is intelligent or not, because none of the outputs can be produced by something that doesn't pack some kind of smarts.
How do we pack these smarts in? We teach them as intelligent humans. We pack our intelligence inside them as models (aka algorithms). They do a great job of approximating what we know in a smaller, better-designed problem space. We use these approximations to fine-tune our designs or predict things, then go from there. Just because an algorithm is more capable in processing inputs in some cases doesn't make it intelligent. The way the output looks doesn't make the algorithm intelligent, either.
I have developed multi-agent systems which showed emergent intelligence when the agents came together across distributed systems; I have written high-performance modeling software which can do calculations way faster and better than humans in the materials science space. I'm not doing some kind of armchair criticism of what I'm talking about.
> You compared the source of human randomness in text generation, and called it 'profound' and implied that it is exactly the source of true intelligence. But, then, the temperature, the similar thing in model was "just a randomness slider". Double standard.
Nope, my stance is clear. To quote myself:
> Briefly, any intelligent creature has internal stochastic processes like sensory inputs and feelings to a certain degree. These stochastic inputs and the creature's own actions change the creature in subtle or profound ways. An LLM has no such processes. You push inputs to the same static model, sans temperature which is just a randomness slider.
To expand my quote, humans or any living creatures do not stay static. They evolve due to the sensory input they receive from external and internal stimuli. The temperature slider doesn't do anything close to that. You tickle a static model in different amounts. The model doesn't change after you supply the inputs & temperature and get the output. Creatures do not stay the same. Their mood, behaviors, and stance against life and their environment change, sometimes permanently.
I'll go one step further. We are not intelligent enough to understand other living beings around us. Claiming that we can build AGI tomorrow is a god-complex. What we have done is something arguably useful in some cases, but how this is built is another matter which is worthy of its own discussion. However, today I don't have time to re-iterate all the problems over and over. You can search my comments for that, if you are in for it.
So, no. You tried to attack my comment by finding contradictions in it, but you failed. A better rebuttal would try to similarize how LLMs mirror the human learning process and just read like a normal human, but this is a well-trodden path which has been rebutted countless times in various forms.
How hard could it be to write a greasemonkey script that will send the current html/Javascript from hell to a LLM with a prompt that will ask you which of the detected annoying elements you want to remove, and then will create and install another GM script that is removing those elements from this site, and also invent finglonger as a side effect. A man can dream, though. A man can dream.
That would not be intuition that is just randomness. Intuition is not randomness.
A jump in intuition comes from automatic processes reorganising the relational structure of conceptual models. There is no reorganisation of the model durring inference.
I wouldn't say "magically", it's what happens with me and every other engineer I know: we're all detecting bugs and yet we're all writing them too. Not saying that LLMs work the same as humans, but saying it's not a logical fallacy.
No, the extra code velocity allowed by AI without proper QA support means there could have been a lot more bugs introduced into products over the last year or so, giving AI more bugs to find. It’s tongue in cheek but not completely uncalled for. It’s been long known that one dishonest way to inflate your bug fix count is to put more bugs into the product to fix. Goodhart’s law is merciless.
If you assume that AI can produce code with less bugs, your 100 times increase in fresh bugs means you also increased your speed for new functionality by more than 100 times.
If models can spot 13 years old critical bugs, I think they can also produce fresh code without those bugs. The skillset is the same.
> If models can spot 13 years old critical bugs, I think they can also produce fresh code without those bugs. The skillset is the same.
As if humans are incapable of spotting critical bugs simply because they've existed for a long time? And we know humans don't produce fresh bug-free code, because the old bugs exist.
SolutionNotFound is a perfectly valid return case for my Algo. I start two threads, one searches for a solution, second one terminates the first after one second, if it is still running.
How in the world does the halting problem enters into the picture if my program is bug free or not?
The argument is that 'the program returns' is strictly easier to prove than 'the program returns the correct answer'. If the first is impossible, then the second is too.
I'm all in favour of sidestepping the argument by having our languages more resemble System F or some polymorphic lambda calculus.
Though still useful to remember that some of the earliest mathematical write ups of the Halting Problem were also by Alonso Church and directly in the lambda calculus. The Church-Turing Theorem is a lasting reminder that the two worlds imperative and functional describe the same algorithms (form the set of Turing Completeness).
He says that using the set of questions and answers from one model to train another model (deatilation) is cheaper than training the model without those datasets.
But he didn't mention that training any model from a set of texts and books is much cheaper than writing those books in the first place.
In other words, it's ok when Anthropic learns from others, but it is not ok when others learn from Anthropic.
So, even with concrete examples, model haters are still wrong.
You also imply the claim that making the distribution of words as the possible next one visible, somehow makes the whole system not intelligent. I would say the exact opposite is true.
By using the embedding vectors, models are aware of precise placement and relative position of words in this hugely dimensional space. No human is capable of such precision. This enables party tricks of "king plus woman minus man" kind. But this also give us a precise point between any two words, no matter how different. What is on the midpoint between volcano and music, for example. No human can precisely answer that, but an embedding can. And we can see which words are closest to this 700 dimensional point.
You see this menu of words as a weakness, and I say it is in fact a sign of super intelligence. And this is all before any reasoning or attention mechanism is even run.