@KyleSargentAI

I am a researcher @openai. I work on post-training, improving multimodal reasoning + perception capabilities | Prev: CS PhD @Stanford w/ @jiajunwu_cs @drfeifei

ai.stanford.edu/~ksarge
Joined November 2012
  • Tweets261
  • Following1.3K
  • Followers1.7K
28Media
✨✨✨
OpenAI
@OpenAISep 3
This is GPT-6 Astra. Anything you can do on a computer, Astra can do for you. Fast.
1
1
23
2,000
gFD scaling curve on GPIC 🥹 Looks cool!
Keshigeyan Chandrasegaran
@keshigeyanOct 15
Excited to share Halcyon, our latest research on multimodal pretraining! Building strong text-to-image generative models is simpler than we think! Blog: radicalnumerics.ai/blog/halc… Models: huggingface.co/collections/r… Paper: radicalnumerics.ai/assets/ha…

Text-to-image generative modeling without language pretraining — Radical Numerics

Halcyon is a family of early-fusion multimodal autoregressive models pretrained from scratch on image-text pairs, with no internet-scale language pretraining.

radicalnumerics.ai
2
1
14
1,008
The only appropriate response to an Ezra Klein column tbh
5
18
4
458
79,449
PSA to my fellow pixel-enjoyers 🙏
This quoted post is unavailable.
Try viewing anyway · View on X
2
2
21
1,321
Whoops original post got taken down: xitter.cf/OpenAIDevs/status/2104…
OpenAI Developers
@OpenAIDevsSep 27
We’ve fixed a bug that was degrading image understanding in GPT-6 Sol and GPT-6 Luna. You should now see better results on visual tasks in the API and Codex, including computer use.
1
227
Kyle Sargent retweeted
Brian Huang
@brianryhuangAug 4
Dang it’s crazy that all of this literally did happen
Ed Zitron
@edzitron29 Jul 2024
Newsletter: Based on estimates of their burn rate and historic analyses, I hypothesize that OpenAI will collapse in the next 12-24 months unless it raises more funding than in the history of the valley and creates an entirely new form of AI. wheresyoured.at/to-serve-alt…

How Does OpenAI Survive?

Throughout the last year I’ve written in detail about the rot in tech — the spuriousness of charlatans looking to accumulate money and power, the desperation of the most powerful executives to...

wheresyoured.at
139
203
60
4,005
1,869,647
Kyle Sargent retweeted
Liron Shapira
@lironAug 2
One-time pad of my top predictions for the rest of 2026. I will reveal the key on Jan 1: 8efd6168a055d1e45863c237eb3feab2be94f10b7ccdc5bf8bb54305c8b19f7f0eb026731d9ea3f870511f8c18daeb814eaa2c9e276082b204f2a962212fb5bd532b20b0105c9883348558ed2711c7d23ea6b4ee718364a87e16fad4b3c3a
40
14
11
675
86,887
A hot take: the videogen pipeline of the future is an LLM-coded 3D/4D base layer plus a lightweight diffusion renderer. We already know how to build the latter, but the former requires VLMs to be much better verifiers. So if you want to unblock generation, work on perception
Brandon McKinzie
@mckbrandoJul 14
At last, I can share what the Blender video was making: a slop cannon. Here is the final simulation. All made from scratch in Blender by 5.6-Sol.
18
10
2
146
20,555
🇺🇸Happy 4th of July! 🇺🇸
9
785
Excited by this new preprint that trains T2I models on GPIC with lots of compute! The method is cool and all their code is out now. Their dataloader just points Webdataset at the GPIC HuggingFace shards so the T2I data is stable + streamed online -- easy to startup + compare :)
Bin Lin@LinBin46984Jul 1
🚀Stop freezing your tokenizers! Remember how REPA-E unlocked end-to-end training for diffusion? We brought same magic to Discrete AutoRegression! Introducing GEAR⚙️: Guided End-to-end AutoRegression. Up to 10× faster convergence! github.com/Tencent-Hunyuan/G… linb203.github.io/gear
2
22
5,203
Kyle Sargent retweeted
Justin Johnson
@jcjohnssMay 29
GPIC should be the new standard benchmark for generative modeling. Training 1 epoch on GPIC is the same cost as 100 epochs on ImageNet, but is a much better proxy for real-world problems. If you work in generative modeling, try GPIC for your next project!
Keshigeyan Chandrasegaran
@keshigeyanMay 29
1/ Introducing GPIC: a Giant Permissive Image Corpus and benchmark for visual generation! 🚀100M VLM-captioned image-text pairs for training 📊1M image-text pairs for benchmarking 🖼️~28 trillion pixels 🤗Centrally Hosted ✅Fully permissive for research + commercial use Dataset, benchmark and models🧵👇 Co-led with @KyleSargentAI
9
17
99
48,871
Today we released “GPIC: A Giant Permissive Image Corpus for Visual Generation.” It’s a 100M image dataset for visual generation, with text captions and 100% known+permissive licenses, hosted on HuggingFace. I’m excited to get this out! Check it out: gpic.stanford.edu/
3
15
1
78
8,048
more replies
One practical example is epoch count – “state-of-the-art” models on ImageNet-1K train for 300-1700 epochs (Fig. credit: PixGen). But that’s not the way you would do things outside of an academic comparison – you’d just go get more data!
2
6
431
Correction: the right cite for this table is "PixelDiT: Pixel Diffusion Transformers for Image Generation," not PixGen. Thanks :)
2
196
I’m also proud of this section of the paper, which gives best practices for compliance with our eval protocol. Without calling out anyone in particular, let me just say that using auxiliary foundation models to get a better FD-DINOv2 on GPIC without being very up front about the huge advantages of the extra data and model FLOPs is super bad – please don’t do it!
2
1
10
539
Personally: I spent a portion of my PhD working on strong tokenizers, which is sort of at odds with the current ImageNet-1K meta to add regularization whenever possible, so I’m also personally excited to see how this dataset drives tokenization research. Happy pretraining!
7
273
Load more