That's what I was going to say. The natural language version actually claims that it leaves the dog behind unattended in every step, even though the following step continues as though it still has the dog and not whichever vegetable it brought back in the previous step.

Either it's not actually good at natural language processing or some element of the solution isn't surviving the shift from the river_cross() tool to natural language output. Whatever actual state it's tracking internally doesn't track to the output past the headline.

Other funny thing: it only became a fully automatic plagiarism machine when it claimed that it wrote the code (referring to itself by name which is a dead giveaway that the system prompt makes it do that).

I wonder if code is where they will ultimately get nailed to the wall for willful copyright infringement. Code is too brittle for their standard approach, "we sort of blurred a lot of works together so its ours now, transformative use, fuck you, prove that you don't just blur other people's work together, huh?".

But also for a piece of code, you can very easily test if the code has the same "meaning" - you can implement a parser that converts code to an expression graph, and then compare that. Which makes it far easier to output code that is functionally identical to the code they are plagiarizing, but looks very different.

But also I estimate approximately 0% probability that the assholes working on that wouldn't have banter between themselves about copyright laundering.

edit: Another thing is that since it can have no own conception of what "correct" behavior is for a piece of code being plagiarized, it would also plagiarize all the security exploits.

This hasn't been a big problem for the industry, because only short snippets were being cut and pasted (how to make some stupid API call, etc), but with generative AI whole implementations are going to get plagiarized wholesale.

Unlike any other work, code comes with its own built in, essentially irremovable "watermark" in the form of security exploits. In several thousands lines of code, there would be enough "watermark" for identification.

@HedyL @diz I kinda wonder if this would work better if it just was worded the other way round: "must be supervised always"
If I understand correctly, LLMs have difficulties encoding negative correlations (not, un-, ...)

Edit: or maybe not, seeing it did this transformation already in the introduction and still lets the dog escape on the very first turn

Yeah I'm thinking this one may be special cased, perhaps they wrote a generator of river crossing puzzles with corresponding conversion to "is_valid_state" or some such. I should see if I can get it to write something really ridiculous into "is_valid_state".

Other thing is that in real life its like "I need to move 12 golf carts, one has low battery, I probably can't tow more than 3 uphill, I can ask Bob to help but he will be grumpy...", just a tremendous amount of information (most of it irrelevant) with tremendous^tremendous^ possible moves (most of them possible to eliminate by actual thinking).

At the very least, many of them were probably unable to differentiate between "coding problems that have been solved a million times and are therefore in the training data" and "coding problems that are specific to a particular situation". I'm not a software developer myself, but that's my best guess.