Regarding the interactions shown in the screenshots:
LLMs are pattern-matching machines. They keep the pattern going. Once "the agent disobeys the human's instructions" has made its way into the context, that is the pattern that it's going to keep matching. No amount of telling it to stop will make it stop.
The only possible solution is excising it from context and replacing it with examples of it doing the right thing. Given that these models have massive context windows now and much of the output is hidden from the user, that's becoming less viable.
It makes the behavior more obvious from simply looking at the file, for one thing, and it means you can just lump it into your next `git add -A` without needing to handle it specially.
I've never been a fan of dual grid, and personally prefer the rpg maker approach of using 5 sample tiles and then chopping up and recombining them to make the 47 tiles needed for what I believe is called "blob tiling".
Yup, email is usually the preferred communication tool of record. In a previous job, our messages on Teams were wiped after 8 days so anything that needed to be recorded had to be in an email or some form of document.
We use Mulesoft where I work, and XML namespaces are a constant issue. We never managed to define an API spec in such a way that the RAML compiler and the APIKit validator would both accept the same payload. In the end we just had to turn off validations in APIkit.
Yeah I strongly feel that the best outcome of all of this would be the end of sponsorships and affiliate links, and a general reduction in price discrimination.
Something I've occasionally wished for is a classic-style Zelda game[1] where partway through the adventure you discover that all the dungeons are actually adjacent to each other, and you can open up passages connecting them turning it all into one big Metroidvania experience.
[1]: i.e. one with 4-8 dungeons and new navigation/combat tools in each, not a sandbox like BotW
We had this problem, 119 services that all got their dependencies from a shared domain. Individual services had to depend on the exact version of libraries provided by the domain. It made updates essentially impossible.
Eventually we were forced by a licensing change to move to containers, which fixed that issue but substantially increased our research usage, we went from 16 GB of RAM for the entire domain to 1.5 GB per service, with similar increases on CPU.
LLMs are pattern-matching machines. They keep the pattern going. Once "the agent disobeys the human's instructions" has made its way into the context, that is the pattern that it's going to keep matching. No amount of telling it to stop will make it stop.
The only possible solution is excising it from context and replacing it with examples of it doing the right thing. Given that these models have massive context windows now and much of the output is hidden from the user, that's becoming less viable.
reply