5 ms·

You can use Claude, like the author, to reproduce the result.

by kelseyfrog 4mo ago

justin_dash 4mo ago
Unless if it was done by Fable!
- kelseyfrog 4mo ago
  The 'major insight' described in the article predates Fable's release by two week four days. It would be a complicated timeline.
atrus 4mo ago
somehow I suspect it was a bit more involved than: Claude, please solve Linear A.
- smsm42 4mo ago
  You also have to add "make no mistakes"!
- fragmede 4mo ago
  A little bit more. If you ask ChatGPT to "solve linear a" it thinks you mean linear algebra. If you specify that it's the Minoan translation problem, you get a table similar to the one that we get a glimpse of in the without access to the paper, we can't say how much more work the paper has than my gist. https://gist.github.com/fragmede/bbf277d36a2398065f109484f3473684 https://gist.github.com/fragmede/bbf277d36a2398065f109484f34...
- Kosturdistan 4mo ago
  You are correct!
_verandaguy 4mo ago
This isn't really a reasonable approach, is it? The original prompts aren't provided, nor is the original context; even then, you can't really treat a stochastic system like an LLM as a major component in reproducibility.
- ben_w 4mo ago
  > even then, you can't really treat a stochastic system like an LLM as a major component in reproducibility. If you had the other things, being "stochastic" is not even remotely a show-stopper. Stochastic processes abound and are the reason the mathematics of statistics was developed in the first place, ultimately allowing us to create such things as LLMs. When all the relevant steps gets published, I absolutely expect a lot of people to (attempt to) reproduce this work even though LLMs are stochastic.
  - _verandaguy 4mo ago
    My issue with this is that it's a form of "soft" reproducibility, where it'll work for many (maybe even most!) people, but that depends on the way the original prompt was formulated (read on) and the state of the random noise in the system. On the prompt formulation; prompts with very similar formulations (in terms of both semantics, hamming distance, or both) can lead to _wildly divergent_ outputs in my experience. It's not rigourous, and when that divergence happens, it's extremely difficult (arguably impossible, by nature of the architecture of transformers) to identify why the divergence happened and where.
- fragmede 4mo ago
  Sure it is. We're humans, not robots (well, I think I am, and I presume you are as well, but for all we know, we could be living in a simulation), so if the non-deterministic system decides to generate code that calls the variable foo one day and bar the next, as long as the code still does what's being asked of it, why do I care that the non deterministic system chose to call the variable something different when run on Tuesday? There's the computer science definition of determinism and the engineering result of "does it work", which are at odds. It's like the halting problem. We haven't solved the computer science definition of the halting problem, but give some C code with a loop that won't terminate to Claude, and it'll call that out as not halting.
  - _verandaguy 4mo ago
    All things aside, I think this misses the forest for the trees on the halting problem. It's not about being able to throw claude or codex at a loop and having it evaluate it for halting, it's about being able to do this for arbitrary code. Computer science rigourously defines the halting problem as not computable and undecidable. within the framework of using something akin to static analysis using any deterministic Turing machine. There's not really a question of "solving" the halting problem like there's some as-yet unknown way of generally figuring out if arbitraty code halts. Turing proposed a proof in 1937 in favour of undecidability of what we now know as the halting problem, building on ideas first articulated by Church a few years prior. Frankly, if anything, it's reasonable to say that the halting problem's been solved, just in the direction of undecidability rather than decidability. Anyway, back to LLMs; as code gets more complex, the robot will need a bigger context window, more hardware resources, and more time, all of which will be variable due to the noise inherent in the system. It'll be difficult to put a useful upper and lower bound on how much computing power and time it'll take to figure out if a program ever halts. Which is all a bit moot, frankly, in the context of halting, but useful to keep in mind in the more general context of using these things as analysis tools.

You can use Claude, like the author, to reproduce the result.

justin_dash 4mo ago

Unless if it was done by Fable!

kelseyfrog 4mo ago
The 'major insight' described in the article predates Fable's release by two week four days. It would be a complicated timeline.

atrus 4mo ago

somehow I suspect it was a bit more involved than: Claude, please solve Linear A.

smsm42 4mo ago
You also have to add "make no mistakes"!
fragmede 4mo ago
A little bit more. If you ask ChatGPT to "solve linear a" it thinks you mean linear algebra. If you specify that it's the Minoan translation problem, you get a table similar to the one that we get a glimpse of in the without access to the paper, we can't say how much more work the paper has than my gist. https://gist.github.com/fragmede/bbf277d36a2398065f109484f3473684 https://gist.github.com/fragmede/bbf277d36a2398065f109484f34...
Kosturdistan 4mo ago
You are correct!

_verandaguy 4mo ago

This isn't really a reasonable approach, is it? The original prompts aren't provided, nor is the original context; even then, you can't really treat a stochastic system like an LLM as a major component in reproducibility.

ben_w 4mo ago
> even then, you can't really treat a stochastic system like an LLM as a major component in reproducibility. If you had the other things, being "stochastic" is not even remotely a show-stopper. Stochastic processes abound and are the reason the mathematics of statistics was developed in the first place, ultimately allowing us to create such things as LLMs. When all the relevant steps gets published, I absolutely expect a lot of people to (attempt to) reproduce this work even though LLMs are stochastic.
- _verandaguy 4mo ago
  My issue with this is that it's a form of "soft" reproducibility, where it'll work for many (maybe even most!) people, but that depends on the way the original prompt was formulated (read on) and the state of the random noise in the system. On the prompt formulation; prompts with very similar formulations (in terms of both semantics, hamming distance, or both) can lead to _wildly divergent_ outputs in my experience. It's not rigourous, and when that divergence happens, it's extremely difficult (arguably impossible, by nature of the architecture of transformers) to identify why the divergence happened and where.
fragmede 4mo ago
Sure it is. We're humans, not robots (well, I think I am, and I presume you are as well, but for all we know, we could be living in a simulation), so if the non-deterministic system decides to generate code that calls the variable foo one day and bar the next, as long as the code still does what's being asked of it, why do I care that the non deterministic system chose to call the variable something different when run on Tuesday? There's the computer science definition of determinism and the engineering result of "does it work", which are at odds. It's like the halting problem. We haven't solved the computer science definition of the halting problem, but give some C code with a loop that won't terminate to Claude, and it'll call that out as not halting.
- _verandaguy 4mo ago
  All things aside, I think this misses the forest for the trees on the halting problem. It's not about being able to throw claude or codex at a loop and having it evaluate it for halting, it's about being able to do this for arbitrary code. Computer science rigourously defines the halting problem as not computable and undecidable. within the framework of using something akin to static analysis using any deterministic Turing machine. There's not really a question of "solving" the halting problem like there's some as-yet unknown way of generally figuring out if arbitraty code halts. Turing proposed a proof in 1937 in favour of undecidability of what we now know as the halting problem, building on ideas first articulated by Church a few years prior. Frankly, if anything, it's reasonable to say that the halting problem's been solved, just in the direction of undecidability rather than decidability. Anyway, back to LLMs; as code gets more complex, the robot will need a bigger context window, more hardware resources, and more time, all of which will be variable due to the noise inherent in the system. It'll be difficult to put a useful upper and lower bound on how much computing power and time it'll take to figure out if a program ever halts. Which is all a bit moot, frankly, in the context of halting, but useful to keep in mind in the more general context of using these things as analysis tools.