riley tomasek (@rileytomasek) on X (original) (raw)

• Pinned
  
  Agents create work. Daemons maintain it. Is your team drowning in unmergeable PRs, stale docs, and half-finished issues? Today we're launching a new product category to fix exactly that: daemons.
• 
  I've learned so much from
  @hubermanlab
  , but podcasts make recall and discovery challenging. Huberman AI uses the latest models from
  @openai
  to search every episode with Google-level accuracy and summarize the results.huberman.rile.yt
• 
  Replying to @rileytomasek
  How does it work? Transcriptions are generated using Whisper and then embedded using the latest
  @openai
  text-embedding-ada-002 model. The vectors are then stored in a
  @pinecone
  vector database. A user's query is embedded and then used to find similar vectors in the database.
• 
• 
  the new agent framework from
  @threepointone
  (cloudflare) is quite interesting. persisted local state for agents is an obvious benefit, it's just a pain in the ass to do well with other tools.
• 
  i needed to spend a $1,000 azure credit in 24 hours, so i made this. it's a knowledge graph, created by extracting nodes and edges from transcripts using gpt4. how i did it 👇🏻
• 
  Replying to @rileytomasek
  The search works on semantic meaning, not just keywords. This allows you to search for a term like "neurotransmitters" and get results that include "dopamine" and "serotonin."huberman.rile.yt/search?query=n…
• 
  Announcing gpt.best — the best way to share ChatGPT threads: - Save threads to scrollable pages - No more walls of screenshots - Automatic sharing images for Twitter 👇🏻 Example thread: gpt.best/k5Q5F9UpCreate your own: gpt.best
• 
  Data fetching/writing is handled brilliantly by
  @remix_run
  . It’s SO much simpler than things like Next.js. However, FormData and URLSearchParams are difficult to work with. Zodix makes it easy to parse Remix loaders/actions using Zod schemas.github.com/rileytomasek/z…
• 
  The transcripts took a while, but I built the app in a day. It's reusing code from a bunch of other things I was working on, so not totally fair to say it was built in a day. Some of it is open source if you build your own:github.com/rileytomasek/p… github.com/rileytomasek/o…
• 
  Replying to @buffteapot @hubermanlab and @OpenAI
  You're welcome. It's cheap enough to run that I'm covering the costs personally. I'm not sure what I'll do if lots of people use it and it becomes expensive to operate.
• 
  Big news! The successor to Huberman AI is now in private beta.
  @dexa_ai
  has more shows, summarized episodes, and rarely hallucinates (this wasn't easy). We're bringing you knowledge you can't find on Google or from ChatGPT. A few conversations that have resonated with me 👇
• 
  how is this nonsense compounding error rate claim so common? a well built agent using a recent model is more than capable of recovering from errors. add a 75% error recovery rate and the result is much different...
• 
  Replying to @rileytomasek
  The "Ask" answer uses text-davinci-003 to answer the question given the search results, with instructions not to make stuff up.