viksit

6 ms·

31.
by viksit 1y ago
yes apologies, the code rendering in substack wasn't great, but I'll put this in a gist!
32.
by viksit 1y ago
thank you, appreciate the comment! thats a great point -- as I'm developing this intuition, I'm designing an eval which does a comparison of the openAI example there + tool call using a simple RNN + one that uses an encoder model.
33.
by viksit 1y ago
great point, appreciate the comment. totally agree with your framing, though i think there’s still a gap in how tool use is handled today. quick note: it doesn’t have to be an rnn. i’ve got a follow-up example coming that uses a transformer
34.
Optimizing Tool Selection for LLM Workflows with Differentiable Programming(viksit.substack.com)
122 points by viksit 1y ago|42 comments
35.
by viksit 1y ago
I was experimenting with how local, learnable routers can reduce token overhead, and lower costs, and decided to publish a post about it. The main goal is to delegate tool calls via a PyTorch based learner and examples of how to integrate t
36.
by viksit 1y ago
this is also the most common thing in india fwiw so it’s not all religious.
37.
by viksit 1y ago
open source llm compiler for prompts called selvedge. rather than manually write prompts for llms (which is like hand coding byte code for cpus), declare a structure and instructions and let the system do prompt writing for you. it also exp
38.
by viksit 1y ago
+1 to this - exactly the same experience. I started a company in the space because I believe that the decentralized technology and its ability to have world impact is truly amazing. But after spending ~2y in the space, I realized another th
39.
by viksit 2y ago
imo, the easiest way to learn a language is to start speaking it in real life esp if you live in the country. your brain starts filling in and you start to learn how to fill knowledge gaps by reading and hearing others in a “sponge” like wa
40.
by viksit 2y ago
yes :) we also stopped maintaining the old code starting may/june after running it for about 9 months. a new version of the front end is in the works!
41.
by viksit 2y ago
yes we built a 10k user social network for artists and musicians on it and it’s excellent. very sophisticated and very extensible.
42.
by viksit 2y ago
anyone have pointers on progress in symbolic reasoning vs context forcing approaches in LLMs?
43.
by viksit 2y ago
oh i just commented elsewhere in the thread about our work in integrating frames and slots into LSTMs a few years ago! second this.
44.
by viksit 2y ago
from 2015-2019 i was working on a bot company (myra labs) where we were directly inspired by cyc to create knowledge graphs and integrate into LSTMs. the frames, slots and values integrated were learned via a RNN for specific applications.
45.
by viksit 2y ago
they released an announcement on this. > We are sorry for that. > It’s been a while since we’ve released a model months ago, so we’re unfamiliar with the new release process now: We accidentally missed an item required in the model re
46.
by viksit 3y ago
funnily enough i have a library i’m planning to open source soon! i’ve used airflow as a guideline for it as well.
47.
by viksit 3y ago
this is a great write up! i was curious about the verifier and planner agents. has anyone used them in a similar way in production? any examples? for instance: do you give the same llm the verifier and planner prompt? or have a verifier age
48.
by viksit 3y ago
a rag system that uses tree sitter vs vector search alone would lead to better results (intuitively speaking)? have you seen anything like that yet?
49.
by viksit 3y ago
got stuck at “eliminate world poverty”. classic.
50.
by viksit 3y ago
if i understand the problem correctly - you'd like to feed xMM documents directly into an LLM so that it uses this context to "reason" answers to questions, vs offload the retrieval to a vector db and merely assemble results
51.
by viksit 3y ago
question: RAG by definition offloads the retrieval to a vector similarity search via embeddings db (faiss, knn et al). what is the preferred way to feed documents / knowledge into a model so that the primary retrieval is done by the ll
52.
by viksit 3y ago
i looked at dspy last week, and was trying to wrap my head around how it would be useful for a "fine tune" style use case - where i would want to give the base model more context vs use a vector DB and have the model put together
53.
by viksit 3y ago
curious what hardware you use? and is any of this runnable on an m1 laptop?
54.
by viksit 3y ago
i’ve been using irfanview since i was in grade 4? and it’s one of the first things i install on every laptop i’ve owned ever since.
55.
by viksit 3y ago
interesting thank you. intuitively, prompting like this to get an answer seems basically like the first part of a fine tuning process (more exemplars). what is your thought here behind why reinforcing good output via a loss optimization is
56.
by viksit 3y ago
can you give a few pointers on articles or examples of this?
57.
Ask HN: Best open source gen AI video model?
3 points by viksit 3y ago|0 comments
58.
by viksit 3y ago
curious how or why australia doesn’t feature in this given their extensive unpopulated outback seems relevant.
59.
by viksit 3y ago
could you eli5 how wasm being monadic is helpful / and how an llm monad would be too?
60.
by viksit 3y ago
China actually has very advanced technology in literally every area of the sciences now. They just don’t bother talking about it in english I recommend you follow some of the specialist chinese AI substacks to see what’s happening over ther

More ›