reading "AI News" (previously Smol Talk) is probably the highest-leverage 45 mins I spend everyday on catching up with what's going on in AI. So much alpha, organized hierarchically; an exceptionally well curated (and summarized) daily newslettter. Probably only suited for
@soumithchintala
-
Multimodal LLM Development Reference Resource Released
By
–
This is likely going to become **the** Reference for developing Multimodal LLMs (among other things).
— Soumith Chintala (@soumithchintala) 4 mars 2024
Great resource, and great work from a long list of awesome people! https://t.co/xGKJbvJIJXThis is likely going to become **the** Reference for developing Multimodal LLMs (among other things). Great resource, and great work from a long list of awesome people!
-
Mistral-7B KVCache Optimization for XLA Compilation
By
–
hey @fchollet . I think this is tricky to do in general (like a no free lunch way).
I think the Mistral-7B implementation KerasNLP writes out a KVCache (for compiler optimization) and specifically writes code in a way that is needed for XLA Compilation to work well.
Do let me -
GPT-Fast Integrates Gemma with Optimized Token Performance
By
–
+/-12 lines of code in gpt-fast integrated Gemma from @GoogleDeepMind ; and shows:
* 234 tokens / sec on V100 in int8
* 144 tokens / sec on V100 in float precision https://
github.com/pytorch-labs/g
pt-fast/commit/ef055fc12188eaf80d8ba948ad743ee5583d0f3c
… -
AI Competitive Moat Shifts to Personalization and User Preferences
By
–
As AI models get to some baseline level of intelligence, the moat might transition to human preferences — whether the model is giving you the answer you want, not just the most frequently preferred answer. If this happens, companies who already have personalization models on
-
Generative AI Needs Personalization Through User Embeddings
By
–
Generative AI Products need to go through their "personalization" revolution.
A single "prompt" for the entire world is a single human preference for the entire world. That obviously doesn't work.
You'd want to adjust the prompt embeddings with user embeddings, to give -
Breakthrough in Photorealistic Video Generation Models Debate
By
–
I don't think its a fair characterization. If the average understanding last week was "we have no idea how to get to 1-minute photorealistic video models", and that understanding massively changed with a single work, it is a breakthrough.
The focus on "novelty is only when you -
Stretch Robot V3 Release Improves Speed and Performance
By
–
been enjoying using stretch as a research platform for 3 years now.
— Soumith Chintala (@soumithchintala) 16 février 2024
its biggest strength is its predictable and simple kinematics, and cost.
they just released v3, which looks faster and better! https://t.co/77uS3XHIg4been enjoying using stretch as a research platform for 3 years now.
its biggest strength is its predictable and simple kinematics, and cost. they just released v3, which looks faster and better! -
LLMs exhibit distinct intelligence pathological in unique ways
By
–
LLMs and diffusers exhibit intelligence distinct from animal intelligence, better in many ways worse in many ways, pathological in distinct ways. Unfortunately the use of a single term "intelligence" confuses people; "meme" personalities arbitrage this to fame lol
-
Game Engine and Diffusion Model Hybrid for Video Generation
By
–
i thought it was a hybrid, a game engine creates the camera controlled low-res, and a diffusion model upscales it (and produces the weird stuff).
camera rotations and translations are so notoriously hard in video-gen, that I thought some runtime trick was a great way to shortcut