Some Cynic

[Some Cynic @SomeCynic]

@badlogicgames @FrameworkPuter @antirez Plastic cup shitty robot made an appearance too :)

[Some Cynic @SomeCynic]

@badlogicgames Even the standard seems a bit half-baked. Like, what's the purpose of "allowed-tools"? Skills aren't a subagent, that could limit tools available or determine when it begins or ends. Skills are just some additional info appended to the current context on demand. Not a boundary.

[Some Cynic @SomeCynic]

@badlogicgames It's so much fun too!! I did the same with our 3yo some weeks ago. She created a side scroller (with dinosaurs too), a flappy-birds type thing (underwater-themed with baby shark) and a Pac-Man type game with procedural levels and music. All slop with just html and messy js.

[Some Cynic @SomeCynic]

@ThePrimeagen Plus there isn't even a need for this degree of personification to explain or to steer behavior.
These models are ultimately huge role playing machines.
And they also role play the activation of emotional concepts it seems. Doesn't make them organisms.
https://t.co/nITGlIHNnR

[Some Cynic @SomeCynic]

@badlogicgames But I didn't think that would also expand into AI Gateway. If you have to do specific things per vendor, when working with an abstraction, that defeats the purpose.
I do understand why it may happen though. Dealing with "reasoning" across the varied formats is quite a mess.

[Some Cynic @SomeCynic]

@badlogicgames Yeah, that's the sort of reason why after exploring AI SDK for one of our harnesses we decided to just build our own connector specific for each provider. It seemed like even if you used their abstraction you still had to have a bunch of conditional code for each model provider.

[Some Cynic @SomeCynic]

@ashpreetbedi Even at the price of more complexity to the system, which the agent itself is not aware of.

[Some Cynic @SomeCynic]

@ashpreetbedi It's not because a VFS is the optimal tech solution from the system's point of view. But, it is quite optimal from the LLMs point of view. The agent benefits greatly from a very simple and well known paradigm for dealing with content, through a file system...

[Some Cynic @SomeCynic]

@TheAhmadOsman DeepWiki from Cognition is pretty good. I've only tried it with public repos though.
https://t.co/1Fy6RhafgW

[Some Cynic @SomeCynic]

@mattpocockuk I believe your first two can be accomplished using isolates, through isolated-vm. We do it in our agents that way and works very well.
Not sure I understand the need for the AST analysis though. I'm def missing something.
What sort of use-case would you be guarding against?

[Some Cynic @SomeCynic]

@DamianCatanzaro Si en la casa tenes ya bocas de cable coax en los ambientes podes usar eso directamente con MoCA y tenes opciones hasta de 2.5gbps. Yo lo tengo así, y son re contra confiables y rápidos.

[Some Cynic @SomeCynic]

@karpathy Wrote about this back in November https://t.co/F8QP0MwTxy

[Some Cynic @SomeCynic]

@scott_stouffer @sandhawke @BrianRoemmele No, they are no related. When you index content you embed chunks. And during querying you yourself embed the input query. The LLM only sees the original text chunks. Doesn't deal with the vectors.

[Some Cynic @SomeCynic]

@YorkSwampFox @BrianRoemmele The author you are talking about is Michael Lewis, the first face you see in this 60 minutes interview. It's an old episode, actually.

[Some Cynic @SomeCynic]

@david_firrin Yes, blocked by cloudflare it seems, at least for the API

[Some Cynic @SomeCynic]

@the_moisrex You could have computes.js, gets.js, and inputs.js, and at least you'd have some logical grouping without it being one single huge file.

[Some Cynic @SomeCynic]

@derekcheungsa @FlowiseAI Shouldn't you be using a Temperature closer to 0? 0.9 may make it too random and prone to hallucinations, I would expect.

[Brian Roemmele @BrianRoemmele]

It’s happening…

[Some Cynic @SomeCynic]

@simonw But if you mean a direct output from one as unaltered input to another call, then no, nothing that direct.

[Some Cynic @SomeCynic]

@simonw Another example: Conversation history gets periodically condensed to include a summary of very old messages. That summary built by 3.5 gets included into the running conversation for gpt4.