An *exclusive* from me about a new startup,
@synth_labs
, which is working to help companies get their AI systems do what they want (and avoid doing what they don't want). via
@technology
Waited a long time to tweet this 😅 but I'm happy to announce that I am a recipient of
@StabilityAI
PhD fellowship that'll be funding my studies at
@BrownCSDept
until 2026.
Excited to announce that I've founded new company! is dedicated to solving the AI alignment challenge–ensuring AIs are robustly aligned with human intentions & values.
Grateful to have
@NathanThinks
@fdesouza
as cofounders & Seed funding from
@M12vc
&
As of today I have left my role at Stability and Carper. I wish them best of luck, and I am excited to join some of my closest friends at
@AiEleuther
along side
@BlancheMinerva
:)
Recent advances with language models have been powered by Reinforcement Learning with Human Feedback (RLHF).
At Carper, we're developing production ready open-source RLHF tools. Blog post by
@natolambert
, myself,
@lvwerra
and
@Dahoas1
Thanks for the retweet! Are you interested in doing RLHF for story generation but using no human annotation data? Our research direction at Carper has got you covered! We explore the limitations of CLIP-like story critic models ala CARP and show various
.
@carperai
just passed a major milestone a few days ago, its now a whole year old. I thought we could take this time to recap the achievements we've accomplished over the last year as well as discuss potential directions. of the lab. A thread 🧵1/N
we're hiring for all roles.
Open science stuff we're working on:
1) RLAIF for pretraining (we're making open source datasets).
2) benchmarks benchmarks benchmarks.
3) collaborating with
@AiEleuther
on some awesome projects.
Work with us.
@geoffreyhinton
Yes, if t can’t explain how it works then how do you know who is to be held liable if something goes wrong? How do you know if it’s reasoning was flawless but it just so happen to be unlucky circumstances?
doing an open source RLHF/mech interp meet up. Capacity is a bit limited. I'll be accepting people with substantial open source work in this space first and foremost.
Releasing the demo notebook!
You can try CARP-L here, just click "Run all"
You can edit the stories and the critiques. It'll perform prompt softening under the hood using Pegasus.
Cut the CARP: Fishing for zero-shot story evaluation
abs:
a scalable, efficient method for performing qualitatively superior, zero-shot evaluation of stories and a new corpora composed of 1.3M aligned story-critique pairs derived from over 80,000 stories
Zero shot semantic style transfer for editing faces. Original, Spock, James T. Kirk designed by Bauhaus, Low poly Hikaru Sulu.
Edits are performed using CLIP system guided by L-grammars and Faces VQGAN.
Shame on
@WriteSonic
for using Stable Diffusion on their website while (illegally) not mentioning the RAIL license or crediting
@StabilityAI
. I really hope you guys didn't try to pass that model as your own to VCs ;)
Thanks to everyone who came out to the third new england RLHF hackers hackathon. We're going to be doing the next RLHF hackathon at NeurIPS 2023. An exact time and date has yet to be decided, but you can join the discord for more information.
Carp seq2seq (based off of GPT-J) is on the eye. You provide stories and it critiques them. Carp eval, for automated story evaluation, will be out within the coming weeks.
License is MIT. Paper soon.
@colinraffel
“Oh cool some hard math! The work must be right then” it’s so hard to get feedback on my mathematics if no one in my field fully understands what I’m actually doing! Everyone gets it at an intuitive level sure, but I still don’t know if I’m using the right abstractions.
In an event to mimic the success of
@KordingLab
's recent neuro event, I decided to host my own pure math in deep learning event. Here is the calendar event URL. Please RSVP if you intend to attend. The meeting is tomorrow night (Tuesday) at 10PM EST.
@arankomatsuzaki
Absolutely shameless plug:
I wrote about this exact relationship back in April. I think my theory took it in a different direction than theirs as I borrowed more deeply from the cognitive neuroscience side of things. I should contact the authors.
Trying to figure out interest, who would be interested in a
@StabilityAI
launch event in Boston on October 16th? A dinner party in Cambridge, hosted by yours truly.
@omarsar0
Incredibly misleading. We have no fast how ChatGPT trained or if it was based on a chinchilla-like approach. Similarly its only 15x faster because inference fits on a single gpu, not because of anything the repository does. This is a thin DS wrapper around a naive PPO implt.
Anyone have a paper to cite that all RLHF is implicitly model based RL? Seems super obvious but I can't recall any papers making this claim, just a few lesswrong posts.
@kchonyc
@GoogleDeepMind
This is misinformation. The model does not know this information and people will take this as evidence that the numbers are correct. Please delete it and refrain from encouraging people to ask models about themselves
Now that we can write Tiny Papers
@iclr_conf
, what should we write about?
I'd like to invite all established researchers to contribute Tiny Ideas as inspirations, seeds for discussions & future collaborations!
#TinyIdeasForTinyPapers
I'll start. Note: bad ideas == good starts.
MUSIQ is a new approach for image quality assessment (IQA) that uses a patch-based multi-scale transformer architecture, bypassing the typical fixed input size constraint of CNN-based architectures, to achieve state-of-the-art IQA performance. Read more →
In the coming weeks and months, we'll be putting out blog posts outlining our new direction. We want to make RLHF accessible to everyone, no matter how advanced you are. We want to lower the barrier of entry until some person in their basement with a GPU can make ChatGPT.13/N
Passed part of my quals yesterday, got a new road bike to celebrate (have wanted a new one since last summer anyway and had been saving up for it since then). Never liked drop bars but for some reason on the Kona they're so incredibly comfortable I'm kinda shocked haha
@_joaogui1
@rodrigfnogueira
@carperai
Didn't wanna reply from the CarperAI account since this is my opinion and not my org's opinion but we have some preliminary results already confirming that RL makes a noticeable difference in terms of capabilities.
1/2
Hi 👋 if you are interested in:
🪿 Goose explainers, discussion, and resources
📃 Sharing the top goose related news
🔬 Learning more about geese
😂 Goose-related memes
Consider following me. ✔️
I have lots of great content planned for the new year! Stay tuned!🎉
Hi 👋 if you are interested in:
🤖 AI explainers, discussion, and resources
📃 Sharing the top AI papers every week
🔬 Learning more about medical AI research
😂 AI-related memes
Consider following me. ✔️
I have lots of great content planned for the new year! Stay tuned!🎉
New CARP model just dropped!
Model:
Example script:
Half the number of parameters but matches CARP-L performance across the board. Also perhaps subjective but qualitatively the results it produces are much more satisfying ;)
Releasing the demo notebook!
You can try CARP-L here, just click "Run all"
You can edit the stories and the critiques. It'll perform prompt softening under the hood using Pegasus.
@pixlpa
Fabula vs syuzhet. The idea that a written story is different than the actual raw story (Eg a story need not be written to express identical information)
@gabrielpeyre
@jwkritchie
Bill Cook did a whole course on integer programming and solving TSP using DNNs. I recommend you read it! It’s super cool. A lot of people are copying his work now adays and not crediting him, it really sucks. He did combinatorial deep learning back in 2012ish
“Towards a Model-theoretic View of Narratives”
@BlancheMinerva
@recardona
@DavidThue
and myself describe a foundation of narratology in the vocabulary of model and information theory. 1/N
@arankomatsuzaki
1) algorithm distillation
2) evolution through chain of thought to build out extensive behavior cloning datasets
3) Offline RL, results in better reward/TFLOPS
4) synthetic preferences
5) various (many) ways to improve generation diversity and prevent modal collapse.