[advael]

Dam, I really thought this would be much more interesting than it is

People have been doing some cool stuff for like a decade with giving dogs buttons to use human language, something they can seemingly get decent at communicating effectively with if they can get around the pesky issue of not having the sophisticated vocal machinery needed to produce recognizable phonemes, through the power of a good interface for them, even if the output is discretized to the level of words

I thought maybe this would be about creating a way for a dog to create stuff said dog might actually want or enjoy via the more powerful lever of effective long-context natural language processing that came of a similar tokenization approach - which can even sometimes churn out working code - that we have now

Instead it seems to be an exploration of how the capabilities you can produce from essentially random noise from this technology is less distinguishable from the result of thoughtful input than I might have hoped. Still interesting, but way less so

[krupan]

Are you disappointed that the LLM generated random games, or are you disappointed that the dog didn't actually care about the games?

Neither are that surprising to me, tbh.

[advael]

The LLM being able to generate random games is kinda expected behavior. It's trained on sequence probability distributions that include the code for a great many games and nudged toward doing so by the human user. I'm disappointed that the dog is basically used as a noise generator here. A process driven by the desires of the dog in a meaningful way would be more interesting and at least seems somewhat plausibly technologically feasible, and the article's title kind of implies it. I am especially disappointed because what most excites me about new technologies is applications that were not possible before its invention, which this seemed like it could be an example of