AI
HuggingGPT: Solving AI Tasks with ChatGPT and its Friends in HuggingFace
ChatGPT chooses models from Huggingface to solve various tasks it’s not usually capable of when used alone.
Tasks: counting objects on images, answering questions about images, generating images, text-to-video and many more.
Eight Things to Know about Large Language Model
LLMs predictably get more capable with increasing investment, even without targeted innovation.
Many important LLM behaviors emerge unpredictably as a byproduct of increasing investment.
LLMs often appear to learn and use representations of the outside world.
There are no reliable techniques for steering the behavior of LLMs.
Experts are not yet able to interpret the inner workings of LLMs.
Human performance on a task isn't an upper bound on LLM performance.
LLMs need not express the values of their creators nor the values encoded in web text.
Brief interactions with LLMs are often misleading.
Core Views on AI Safety: When, Why, What, and How by Anthropic
Our approach to AI safety by OpenAI
Basically “we are working very hard to make AI safe” without much specific commitments
TheZvi’s commentary on Eliezer Yudkowsky's Letter in Time Magazine
Solid case for being less of a doomer (but still a doomer).
Research
Segment Anything by Meta
Powerful segmentation model
Interactive: can point at an object to mask it. Demo: https://segment-anything.com/demo#
New generative model: Kandinsky 2.1
New encoder: MoVQGAN. Beats VQGAN and other encoders.
Uses an image prior model.
Feature: mixing two images.
Available for anyone to try.
StackLLaMA: A hands-on guide to train LLaMA with RLHF by Huggingface
Blogs
Most Technologies Aren't Races
Towards ML-enabled cleaning robots
Behind the curtain: what it feels like to work in AI right now
Tech
Twitter released the source code of their recommendation algorithm. You can see find on github.
Three-step recommendation system: candidate sourcing, ranking, heuristic filtering
Candidates sourced from three different models
Ranked by a neural network with ~48M parameters