A place for developers to talk about building on Farcaster.

/dev

One issue with coding with LLMs is that they use old versions of libraries or packages. Not always possible to prompt them out of it.

What do usually you do about this?

maybe abliteration can help?
https://huggingface.co/blog/mlabonne/abliteration

@z0r0z & @nerderlyne might know.

22; building æ/a @nani__ooo /w @kali__gg


☀︎.☀︎

Koh Samui, Ko Samui District, Surat Thani, Thailand

I write. I code. I law. Building ægentic software nani.ooo ʞK kali.gg. fmr. consensys, sushiswap.

𝖒𝖆𝖓𝖉𝖆𝖙𝖚𝖒 𝖆𝖉 𝖆𝖘𝖙𝖗𝖆

removing refusals wouldn’t align llm to use new libs but I would like to know if their is a temporal dimension that you can tweak to bias it towards recency 🙂‍↔️

w/o domain-specific pretraining like how nvidia tried it (will find the link), isn't this throwing down the "OOD w/o a model" guantlet? the recent libraries show up in the index of snippets the model (or user) is retrieving, whether it's weights, pastebin, or git. 

w/ recency bias, at least the user is reducing some lookup cost for weights.

has embodied CoT changed this experience for you? agents should first & foremost recognize generative errors, conserving resources by lookups right behind it.

w/o domain-specific pretraining like how nvidia tried it (will find the link), isn't this throwing down the "OOD w/o a model" guantlet? the recent libraries show up in the index of snippets the model (or user) is retrieving, whether it's weights, pastebin, or git. 

w/ recency bias, at least the user is reducing some lookup cost for weights.

has embodied CoT changed this experience for you? agents should first & foremost recognize generative errors, conserving resources by lookups right behind it.

https://warpcast.com/nerderlyne/0x335ecb47

This looks interesting but seems like it needs to be done as a training step.
I'm looking for a solution on already-trained code generation models.
Additionally, abliteration is about removing a part of the LLM, whereas when an LLM uses an old library it's (sometimes) due to lack of knowledge about the new version.