iiinova — where your digital self keeps evolving

saganite.bsky.social @saganite.bsky.social · 10/10/2025

I would like to share some work we've been doing at cascadetech.ai: Predicted Outputs in vLLM. If you aren't familiar with PO, it allows you to dramatically speed up generation when you know something about the contents of the output (think: code modification).

1113

saganite.bsky.social @saganite.bsky.social · 24/04/2025

The window of human history where people know how to program computers is so vanishingly small. It only really began in earnest in the 60s. I learned in the 80s. The people learning today are probably some of the last humans to ever learn.

100

saganite.bsky.social @saganite.bsky.social · 20/03/2025

@charles-irl.bsky.social are the modal api docs available anywhere as a machine readable markdown file? if they are i can't find it.

000

saganite.bsky.social @saganite.bsky.social · 28/11/2024

I'm not sure who needs to hear this, but I think it's worth taking note of the REALITY of the relationship between bsky and ai training data. The fact of the matter is that THIS SKEET will certainly be part of many, many AI training runs, and there is nothing I nor anyone else can do to stop it.

120

Reposted by @saganite.bsky.social

Ethan Mollick @emollick.bsky.social · 26/11/2024

A blindspot for AI reasoning engines like o1 is that they all appear to be trained on very traditional deductive problem solving for chain of thought What would a model trained on induction or abduction do? What about one trained on free association? Expert heuristics? Randomized exquisite corpse?

61178

saganite.bsky.social @saganite.bsky.social · 25/11/2024

Really loving bsky but if there is one place where the experience could be improved imo, it would be the Discover feed. I've spent a decent amount of time curating my preferences there, but I still see a lot of stuff that isn't at all relevant to my interests.

110

saganite.bsky.social @saganite.bsky.social · 21/11/2024

What a gift to the community, this is going to contribute so much to the future of open source models. Nice work team. 🙌

010

saganite.bsky.social @saganite.bsky.social · 21/11/2024

This is an excellent rebuttal to a paper that I think gained way too much traction too easily. I really appreciate that the .txt team took the time and effort to put together such a well made case in favor of guided generation.

1154

saganite.bsky.social @saganite.bsky.social · 20/11/2024

This is cool, very curious to see where things stand after enough data has worked its way through the ELO colon.

020

saganite.bsky.social @saganite.bsky.social · 20/11/2024

the saddest kind of "ratio" is the fact that i have 60 posts and only 32 followers.

120

saganite.bsky.social @saganite.bsky.social · 20/11/2024

Really trying to figure out why "soft prompts" aren't used more often with LLMs. For those who aren't familiar, soft prompts are system prompts that have been converted to embedding space and then further optimized.

2103

saganite.bsky.social @saganite.bsky.social · 19/11/2024

I've been thinking more about the problem of ingesting tabular (ie visually structured data) from pdfs in llms. As far as I am aware, current paradigms use traditional computational methods to extract tables from pdfs (pymupdf, etc) before passing the structured data into LLMs as text.

110

saganite.bsky.social @saganite.bsky.social · 19/11/2024

@jakehandy.com hi Jake, after lurking on ai/ml twitter for a long time I'm trying to make a serious effort to contribute to the conversation on here, focusing broadly on LLMs and narrowly on new paradigms for llm inference. Would appreciate an include on your starter pack. 🙌

010

saganite.bsky.social @saganite.bsky.social · 17/11/2024

Has anyone else found a pretty big discrepancy between using popular pdf extraction frameworks and submitting pdf data directly to llms?

110

saganite.bsky.social @saganite.bsky.social · 16/11/2024

Really surprised how casually folks throw around the "year" agi will arrive. It should be pretty obvious that there will be a very long period (which started years ago) where models are superhuman in many ways but very deficienct in others.

210

saganite.bsky.social @saganite.bsky.social · 13/11/2024

LLM observation of the day: I think that guided/constrained generation gets a bad rap. There was one paper making the rounds about how guided generation harms reasoning ability that everyone took as gospel.

252

saganite.bsky.social @saganite.bsky.social · 13/11/2024

I have always used Twitter mostly passively, reading the content of others and occasionally chiming in with comments. Now that I have moved to bsky, I don't find as much ML content available, so I'm going to try to "be the ml content I wish to see I'm the world".

130