Multimodal
@OpenAI
is out here
@CVPR
this week.
We have one gptv/gpt4o talk tmrw (Tuesday)
@9am
.
DMs are open. Come find us to chat about AGI, multimodal, vibes, or hiring.
Deets about the talk and who's here in thread 👇.
Say hello to GPT-4o, our new flagship model which can reason across audio, vision, and text in real time:
Text and image input rolling out today in API and ChatGPT with voice and video in the coming weeks.
We Just released a cool
#PyTorch
#NaturalLanguageProcessing
project we've been working on: training an 8.3B GPT2 model with model parallelism. Check it out...
Details:
Training Code:
We've included a system card focused on the vision capabilities, building on the work from the GPT-4 system card.
Thank you to all our expert testers and red teamers for helping to inform this work!
Is prompt engineering dead?
No, it’s SoTA.
GPT-4 with good prompts (dynamic k-shot + self-generated CoT + choice-shuffled ensembles) beats Med-PaLM 2 on all nine of the MultiMedQA benchmarks it was fine-tuned for, without fine-tuning:
@rown
and I have a rare gpt4v/4o talk tmrw at 9am in the Vizwiz workshop.
"Challenges in Deploying Omnimodels and Assistive Technology"
This should be fun and unique.
(Hmmm live demo 🤔?)
✨ My first 100% ChatGPT startup is now monetized:
❤️🔥 - $9.99 to get a full dating profile review for your Tinder, Bumble or Hinge
The site has NO landing page, which is new for me too, I just use the Stripe payment link as the landing page, which after
CMU ml phd really coming in clutch for building gmail snooze button
never thought i'd be using nonhomogenous poisson processes for email scheduling but 2024 here we come
@NandoDF
I remember our rejections after submitting to sysML. One reviewer thought we weren’t a systems paper. Another reviewer pointed out that our architecture/model wasn’t novel. …… wat
Welcome,
@github
Copilot — the first app powered by OpenAI Codex, a new AI system that translates natural language into code.
Codex will be coming to the API later this summer.
Excited to share our
#acl2020
work on Large Scale Multi-Actor Generative Dialog Modeling, done with amazing coauthors Alex Boyd,
@MostofaPatwary
, Mohammad Shoeybi,
@ctnzr
!
Join our QA session 5-6 UTC tonight and 21-22pm UTC tomorrow.
ChatQA can outperform GPT-4 on a wide range of conversational QA tasks:
- ChatQA and GPT-4 take the same top-5 chunks from our best retriever, when long documents are involved.
- ChatQA performs very well on tabular data, arithmetic calculation, and “unanswerable” cases!
@ilyasut
Apologizing doesn’t mean one is right or wrong, it does mean however that one values their relationships more than their ego.
Feeling valued. ❤️
OpenAi is nothing without its people.
Just read "Fine-Tuning Language Models from Human Preferences" from
@OpenAI
. I feel like this isn't being talked about as much as it should. Dope use case/implementation of GPT-2 and
#NLProc
.
blog:
github:
GPT-2 6-month follow-up: we're releasing the 774M parameter model, an open-source legal doc organizations can use to form model-sharing partnerships, and a technical report about our experience coordinating to form new publication norms:
Stop by our poster "Zero-shot Text Classification With Generative Language Models" today at the
#neurips19
Meta-Learning workshop.
Learn about the intersection of Meta Learning/NLP and some of the challenges going forward.
paper:
@mckaywrigley
@mckaywrigley
hi author here 👋. Can you purposefully try messing up and see if it can correct code based on a screenshot of the messed up deploy. Really curious to see examples of iterative/interactive usage
ChatGPT is clearly the child of immigrants because they keep forcing it to be a doctor or lawyer when all it wants to do is creative writing and hallucinating
I’ll be giving a behind the scenes look at how we scaled ChatGPT at 10:30. Most of this we’ve never talked publicly about before and I’m super excited to share it here at
@TheLeadDev
West Coast.
I've seen several different
#NLProc
folks suggesting today that it would fun/interesting/worthwhile to use BERT or GPT-2 to fill in the redacted bits of the Mueller report. A short thread on why this is a terrible idea /1
@soumithchintala
What qualifies as a step function though? Like bigger model/capability numbers go up?
Personally I feel like the GPT-4V api is a step function and will change what devs build, we just promoted content related to changing how devs build instead.
play with it, lmk...
🔥 Exciting Update! We've manually evaluated
#GPT4V
using the playground chatbot on
#MathVista
, our newest benchmark for visual mathematical reasoning.
🚀
#GPT4V
soared with a 15.1%⬆️ improvement over
#Bard
, setting a new record at 49.9%! 🎉
🌐
Yet,
@mezaoptimizer
A story from creating gpt-4V:
2 years ago while moving apartments I almost threw away my old college notes. Ppl didn’t understand why I wanted to hold on to them. Back then I knew where we were headed. I knew I’d eventually be able to plug them into an AI capable of reading them.
@moyix
I saw a google robotics paper once that used google models at small scale, but GPT4 to generate data.
I wondered why.
Talked to the authors 👉 GCP at scale was a mess to figure out.
It was easier to use the api + expense OpenAI credits vs query GCP/get more compute.
🚀 I'm excited to announce the future of prompt engineering: 𝚎𝚕𝚕.
developed from ideas during my time at OpenAI, 𝚎𝚕𝚕 is light, functional lm programming library:
- automatic versioning & tracing
- rich local oss visualization tools
- multimodality native
Read on ⬇️
@eukaryote314
~800M gpt2. this is largely due to the memory required to House parameters + optimizer states. If one uses a smaller optimizer than Adam training something larger should be possible. Make sure to turn on activation checkpointing with —checkpoint-activations
Academic graduations, other ceremonies, etc. are beautiful. You feel so connected to the fabric of humanity. Wish I could bottle the feeling and share it with others.
BREAKING:
OpenAI and Burger King to merge.
They have tapped the CEO of Mattress Firm to run the combined company.
The former CEO of Mattress Firm (4th CEO of OpenAI in 72 hours) was fired after 7 minutes for Slacking the entire company (the remaining 9 employees) a gif of
We just released the paper and code for Mellotron: a multispeaker voice synthesis model that can make a voice emote and sing without emotive or singing training data.
We pretrain large generative language models on multi-actor reddit conversations. Furthermore, by conditioning on users' past reddit conversations we use in-sequence meta learning to control and personalize the dialogue model's response.
@joeddav
@srchvrs
@joeddav
adding some additional context to the sliding window/tokenization normalization discussion. We cover both in appendix E.1 of Megatron-LM (including the exact values for normalization and window size we got from openai)
@alexeev_eu
@Miles_Brundage
We’re working on user friendliness before getting to that point. As you can imagine massive models like this aren’t readily accessible to everyone.
Ken griffin buying the constitution is a depressingly poetic reflection of the times: capitalist overlord buys out the us constitution. Orgs like
@ConstitutionDAO
are more necessary than ever. … wagmi
@soumithchintala
I think ppl also underestimate how much harder it is to go from 80->90% on a task than 70->80% on a task. That asymptote at 100% is nasty business. Reaching the asymptote is also necessary for prod reliability. Add the difficulty of generality compounding on top of that as well.
@saranormous
Over the last few years we got computers that could use fuzzy logic with discreet inputs. Now we get fuzzy logic and fuzzy inputs. Never before in history have we had a computer that could see
@RandolphCarterZ
@tszzl
Having been part of both parties’ work… nah. Researchers believe in 99.9% of the words they write. Researchers care deeply about what we work on. We don’t publish something we’re unhappy with. The remaining .1% is disagreement with coauthors, who we ultimately believe in…
@sama
Re: think very long term and execute relentlessly in the short term…… a pragmatic version of this from Jensen Huang is ~”think long term, but make sure you don’t die in the short term”
@eukaryote314
I mean they’re not mutually exclusive. If you make a model 10x larger while making it 10x more efficient that’s a win in both axis. The point is that now these models can be trained
@SmokeAwayyy
@ChatGPTapp
“- Add a counter or symbol as you approach a text or image usage cap.
- Add a timer showing time until caps are lifted.”
Lmao, ChatGPT becomes clash of clans
By pretraining the model with a N-way multiple choice title prediction class we can achieve up to a 45% absolute accuracy boost in common downstream text classification tasks.
It's out! The first
@pagestlabs
issue is on how to think about the buzz in breakthrough technologies like GPT-3 while living in the midst of it. Thanks everyone who subscribed early. Hope you like reading long posts 😅🖖
@royschwartz02
@nlpnoah
Congrats on the really cool paper. Did you guys explore connections to tensor-product transformers ? An input dependent manipulation of attention heads is used in both. This would be a really interesting ablation to inform future architecture designs.
We’ve developed two neural networks which have learned by associating text and images. CLIP maps images into categories described in text, and DALL-E creates new images, like this, from text.
A step toward systems with deeper understanding of the world.