Tim

Important win in the lawsuit we filed against Deere for blocking farmers from easily fixing their equipment. Great to see continued momentum towards ensuring that Americans have the freedom to fix their own products or use repair shops of their choice. fingfx.thomsonreuters.com/gfx/legaldocs/…

@waitbutwhy The Diamond Age by Neal Stephenson. Lonesome Dove by Larry McMurtry

My current to-read list. What am I missing?

The Washington state legislature is sending TWO right to repair bills to the governor to sign -- and both passed overwhelmingly (total of 2 no votes across 4 full chamber tallies!) pirg.org/media-center/r…

@arosborough @kwiens @ChattLab repair.org/store

@kwiens @ChattLab Where can I get a Repair.org shirt?

Pitching in today at the @ChattLab repair cafe. What an incredible group of fixers!

@teknium instructor. func calling w/pydantic has been fantastic, and made getting useful outputs out of data extraction tasks dead simple and nearly 100% accurate. Bonus, just found out it works with GPT4-V model too using MD_JSON mode.

Give me every agent's library you've ever heard of or worked with for json mode/func calling/text extraction/agent traces/whatever, please & ty :)

@bhallcpa Paolo Bacigalupi is my favorite in the genre - the Water Knife, Ship Breaker, Windup Girl. Apocolyptic Sci-fi. Don't really follow the "major outage" trope - more like the apocalypse has happened, now what does the world look like.

My guilty pleasure is reading fiction apocalypse books. They always start with a major outage...

Right to Repair Right to Repair 2014 vs 2023

@ArtirKel The quality of the PDF, and complexity of the content, is the biggest factor in RAG performance. If they are well formatted pdfs, not much complexity - most out of the box solutions mentioned will work. Otherwise you need to OCR. Vision models have helped, but v. expensive.

I have around 5000 PDFs and documents that I want to LLM through, what's the best current way? ie I want to be able to ask things like "What do we know about X?" "Show me some documents that discuss Y" etc.

The Polish hackers who heroically fixed a series of passenger trains that the manufacturer had artificially bricked with DRM are now being threatened with lawsuits. OEM demanding the trains be removed from service: 404media.co/polish-hackers…

New from 404 Media: you may have seen that a group of Polish hackers repaired trains the manufacturers had bricked (hackers were successful and helped). Now the train company is threatening to sue them. We spoke to the hackers. 404media.co/polish-hackers…

@levelsio Did OAI lift your rate limits? I'm capping out at 200/day still.

✨ My first 100% ChatGPT startup using GPT Vision ❤️‍🔥 ProfileReview.com Review your dating app: - profile - photos - convos 🚀 On average it can ~10x your matches in volume + quality It uses data from the most popular dating profiles on Tinder and Bumble Free for now!

@dylan522p @AnthropicAI 100k-instant or 100k-1.3?

There's seriously something wrong with @AnthropicAI Claude 100k attention mechanism! Yes you can input huge seqlen, but it seems to ignore the middle part heavily and only barely recognize it. Inference latency is crazy low too. Do they use some sort shortcut attention mechanism?

Huge Right to Repair progress in Minnesota pirg.org/updates/right-…

"Right-to-repair bills are having a moment" gizmodo.com/colorado-farme…

When farmers don't have access to what they need to fix their tractors, serious costs & downtime ensue. In my latest for @uspirg, I calculated that #RightToRepair could save U.S. farmers $4.2 BILLION a year. Here's a few of the key stories and findings from the report:

@Peter_0_0_g @abacaj I’ve had success using agents as simple controllers,deciding the path to continue based on fuzzy context, and letting it go a few levels deep. But letting them run wild any further gets into loops quickly in my experience.

@abacaj Has anyone actually managed to get anything useful out of them?

I think current GPT “agents” are very limited in capacity (you can try using them to see what I mean). Often getting lost the deeper the task goes or just outright failing to make any progress

@Ishaank1999 @langchain @pinecone How have the costs been on pinecone? Been a fan of pinecone so far but struggling to understand their pricing. Have about 250k pdfs ranging in size from 10-150 pages.

@xathis @tito @LPitsoulis @pinecone @langchain It seems to depend heavily on the subject content. Larger chunks for text heavy docs works best, but smaller chunks works better with graph or table heavy docs. No hard data, just experience at this point. Very interested in research on this

@tito @LPitsoulis @pinecone @langchain Do you know of any scientific investigation on this? Does overlap really help? In the open source retrieval gpt plugin OpenAI suggests further embedding calculations. Would love to see this analyzed systematically

Have been playing around with document embeddings on @pinecone and @langchain. Picking the right documents to feed into the LLM context is not quite solved by a vector store search for anything beyond a demo use case. Domain-specific metadata filtering engineering should help

@xathis @tito @LPitsoulis @pinecone @langchain I rarely use overlap. The only time is search across pdf pages. I use it to capture sentence continuation past the page end, but I’m not convinced some arbitrary number, token based overlap is the right method. paragraph / sentence boundary seems better.

@dwhdai @hwchase17 Just openai for now.

@hwchase17 @tim_asp nice! does this work with all chat/completion models?

Lots of great small features added over the past week My favorites: improvements to OpenAI tracking to allow for cost tracking (@tim_asp) + number of successful requests (Jon Page)