This is awesome! Using a CLIP or Dino v2 model to produce image embeddings would probably improve the similarity search a lot - kind of similar to http://same.energy/ .
This is very useful! I think the interfaces around models being used in an async manner will look very different than the synchronous chat UIs we are today. Claude Code is the first real “agent” that is providing true economic value, and there’s so much low hanging fruit in making the interfaces far better.
I have a few private “vibe check” questions and the 4 bit QAT 27B model got them all correctly. I’m kind of shocked at the information density locked in just 13 GB of weights. If anyone at Deepmind is reading this — Gemma 3 27B is the single most impressive open source model I have ever used. Well done!
Wow! I'd love to read a more in-depth blog post describing how to create one of these myself, and maybe even contribute my own splats to a collaborative library for iconic landmarks. I could see interactive splats being added to Wikipedia for popular locations.
This is great! I wonder how hard it would be to use a pen plotter instead of a thermal printer. You could even use a procedurally generated handwriting font and Claude to make it feel like a handwritten letter.
Agreed. I keep my iPhone in my right pocket and my Ricoh GRIIIx in my left pocket whenever I go out. It’s such a fantastic camera given its size, especially with the APSC sensor in such a compact body.
I also have a Sony A7C but I haven’t used it since getting the Ricoh. Being pocketable is a massive factor in how much I use a camera.
This is awesome! Could you share more details on how you’re storing the image embeddings and performing the KNN search? Is there an on-device vector database for iOS?
Last month Lexica served a little over 1 billion images and the Cloudflare bill (I'm using R2 + workers) was a little over $5k. I've since gotten it down to a more reasonable amount after spending some time to re-encode the images to reduce our bandwidth usage significantly. If Lexica were running on AWS/S3 I imagine our first month's bill would be closer to $100k rather than $5k. This is only image serving, so not including costs to run the beefy CPU servers to run CLIP search, frontend, DB, backend, etc.
Great work! If anyone’s planning to use AI generated artwork in their projects, I’ve added an image search API to Lexica, similar to Unsplash. All the images are licensed CC0 and millions more are being added every few weeks.
I’ve used both DALL-E and stable diffusion extensively, and in my experience DALL-E is more forgiving while stable diffusion is more powerful — with a much higher skill ceiling for writing prompts.
Also, stable diffusion’s open source weights are great for self-hosting which can’t be done with DALL-E.
The GPT-3 generated conversations were coherent most of the time, and even interesting! However the generated speech via Google Cloud's API was monotonous and could do with a bit more intonation and excitement.
Ah, I figured out what went wrong. Libgen has two different databases, a fiction and non-fiction one, and right now Alexandria only queries Libgen's non-fiction DB. I'll update it to search both databases later this week. Thanks for bringing this to my attention!