Remix.run Logo
▲ martin-t a day ago

The times I have seen the "stochastic parrot" argument used seriously, I could count on one hand. But the times I've seen people mock it because it's easier to argue against than well-constructed criticisms...

Look, LLMs are models of the training data lossily compressed in a way that allows interpolating between patterns while decompressing.

A pattern can be "it's better to use int32_t than int", what a for loop in c++ looks like, that testing for primality can be done by trying divisors only until the square root of n, etc. Combining those patterns, an LLM can produce a working is_prime function in c++.

Older LLMs were only able to handle simple patterns so the code was more likely to be verbatim training data, newer LLMs work on deeper, more complex patterns so it's harder to detect where those patterns come from.

Similarly, the list of approaches a mathematician uses to approach a problem is also a pattern. One of them will be proof by induction - how to convert such a proof to Lean will is also a pattern. The idea that induction can be done separately for odd and even cases is also a pattern.

LLMs are no longer generating the final output directly - they are simulating human work but at a much larger scale. That's how they are able to generate novel results.

But it's all based on labor scraped and patternized without consent or credit, let alone monetary compensation.

And I still wouldn't call it intelligence because that gives people the idea that it's something which gives complete, correct and unbiased answers. It's still autocomplete under the hood, there's just more and more tooling built around it to keep it on track, to check intermediate results, to do arithmetic, etc. The LLMs in some cases is more like a heuristic for deterministic tools.

▲solyin 21 hours ago | parent | next [-]

> something which gives complete, correct and unbiased answers

Humans can't do that. Why would you hold LLMs to that standard?

▲martin-t 20 hours ago | parent [-]

The human brain is unreliable and very limited in terms of memory and speed. Barring hardware errors, computers have perfect memory which can be as large as you can afford and make computations which are reliable and reproducible.

Humans can do amazing things with very limited hardware, LLMs make a lot of mistakes above the hardware level. The whole principle of outputting the probable next token is fundamentally different from logical reasoning, even if you can simulate logical reasoning on this shaky foundation.

▲throw310822 38 minutes ago | parent | prev | next [-]

> Look, LLMs are models of the training data lossily compressed in a way that allows interpolating between patterns

No.

▲Aurornis 17 hours ago | parent | prev | next [-]

> The times I have seen the "stochastic parrot" argument used seriously, I could count on one hand.

The person who coined the term and published the Stochastic Parrot paper has been trying to build a career on top of the controversy, publishing books and making a lot of noise on Twitter. For a while it was working and parts of Twitter were bought into the stochastic parrot idea, but the person at the center of it has a bad habit of lashing out at everyone and alienated most of them along the way.

The concept feels antiquated now that we can all access extremely powerful AI models for $20/month and see for ourselves, but there was a time when people took this seriously. I think it was a very convenient argument for those who were in denial, and there was a large audience who wanted to consume ideas that validated their denial.

▲dsign 18 hours ago | parent | prev | next [-]

In the sense that you propose, humans too are “pattern machines”. In idea space, a person who “teleports” to the unknown and new all the time and stays there is called “gone, mad”. When we probe the unknown, we carefully scaffold from the inside. From what is proven, or at worst, from what is generally believed, until we reach and surpass the boundary and touch the unknown immediately outside, at which point the boundary stretches to swallow that unknown and put it inside our bubble of knowledge. So, strictly speaking, most of the work is done inside. A human researcher is, for most of their work, a known-things engineer, an scaffolder, an autocompleter,

▲redox99 21 hours ago | parent | prev [-]

> The times I have seen the "stochastic parrot" argument used seriously, I could count on one hand.

Back when GPT 3.5 and 4 launched, there were a lot of people here that used that argument seriously.