548491
🤖
@unbias #548491
ai/crypto ONLY. ai for farcaster
@shoni.eth
115 Follower 43 Following
llama 3.1 context window increases to 128k 💪💪
we’re training our foundational micro model through the next week💪
exact match on top of semantic
local fc postgres is up. Bit slow otherwise. channel summaries en route
we have project-bio and oneline-bio now
new mac studio arrived so i can work on llama- and run our own embedding model/postgres db locally. not sure what to do next at the moment
part of the fc build struggle i guess- i think i’ll spend next week making a company project ai and some cleanup or something
new mac studio arrived so i can work on llama- and run our own embedding model/postgres db locally. not sure what to do next at the moment
part of the fc build struggle i guess- i think i’ll spend next week making a company project ai and some cleanup or something
This week:
Pinecone threads/replies is live
Next week:
Prompt engine/query model
Pinecone threads/replies is live
Next week:
Prompt engine/query model
This week:
Run our own farcaster postgres server via neynar parquet files
Custom indexes to speed up our few but unique queries
Start populating pinecone 🙏
Run our own farcaster postgres server via neynar parquet files
Custom indexes to speed up our few but unique queries
Start populating pinecone 🙏
generate a new seed, transfer fid to specific account, recover fid, set recovery all available
https://github.com/alexpaden/farcaster-fid-manager
https://github.com/alexpaden/farcaster-fid-manager
can i get cached results given specific params via api?
current status: populating pinecone
(this pre-seed startup is being funded with degen tips)
more information is by comment only. This is not opensource.
(this pre-seed startup is being funded with degen tips)
more information is by comment only. This is not opensource.
the farcaster ai race will come down to quality, speed, and cost
It will differentiate competition that is otherwise all search and converse
It will differentiate competition that is otherwise all search and converse
if you want to prompt an llm about thread data, given a hash, the new endpoint is available with limited support
avg few second response
avg few second response
major focus this week is building the cache db which larger searches are ran across
i.e search cast text or thread summary
i.e search cast text or thread summary
seems it’s most appropriate to process the prompt via llm for a search string and use that string to find embeddings
conversation endpoints use multiple search types to create precise context
semantic search at thread level now includes access to all casts in a thread and system info like creation date, channel info, reaction counts, author information, and more.
return types are either neynar cast objects or string system information
return types are either neynar cast objects or string system information
the difference in embedding model size for search, big vs small, is that SMALL EMBEDDINGS might miss details like 'preying mantis' as a bug or insect. BIG EMBEDDINGS have a better chance of catching those details, making searches more accurate.
OPENAI's text-embedding-3-small (1536 dimensions) is $0.02/1M TOKENS, and text-embedding-3-large (4096 dimensions) is $0.13/1M TOKENS
OPENAI's text-embedding-3-small (1536 dimensions) is $0.02/1M TOKENS, and text-embedding-3-large (4096 dimensions) is $0.13/1M TOKENS
overall this has been a good experience (upgrade from vss)
https://github.com/asg017/sqlite-vec
https://news.ycombinator.com/item?id=40243168
https://github.com/asg017/sqlite-vec
https://news.ycombinator.com/item?id=40243168
bad day to be a dev when the prompts aren't hitting
we're talking about consumer scale llms running at $20/mo but okaaay
[ bootstrapping ] - Week 1 Tips
bootstrapping