Mastodon Feed
jonny@neuromatch.social ("jonny (nonvenomous)") wrote:
Some context on the top thread behavior, a person might naively think "why not just log every single message surely that's just a line at the end of whatever get messages() function exists" and that turns out to be stupendously wrong. First of all there are a million get messages functions. Second, there is a tension between the thing being able to work at all (needs to roundtrip all the messages to retain context because anthropic can't backend for shit apparently) and appearing as a well behaved consumer product (needs to hide all the intermediate messages and "thinking" that happen before a user visible message). As a result there is lots and lots and lots of code, some of it routing through the model itself, that strips, censors, and filters the messages before they are logged anywhere.
Fable in particular is known to have baffling gibberish intermediate messages, and in general if any of the intermediate "thinking" text was visible it would make it abundantly clear that the only reason these things appear to work is by sheer brute force, just keep generating text until something seems plausible. Thinking text is vastly more voluminous than output text, I don't have numbers open but 10:1 is a decent ballpark. Some of that gets returned to the client, some only happens on the server.
So anyway that's why its possible to drop messages at all, its a byproduct of keeping the scam going.