Загрузка видео...

Не удалось загрузить видео

На главную

Introducing 💫 Gumbel Straight Flow 💫: Distilling Autoregressive Models into One-step Flow Maps 🚀 Can autoregressive (language) models be used as teachers for flow map parallel samplers?

djjruhe's profile picture

David Ruhe

1,571 subscribers

23,741 просмотров • 11 дней назад •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

Комментарии: 6

Фото профиля David Ruhe
David Ruhe11 дней назад

Turns out they can! Sampling from an AR model can be done with the Gumbel-max trick, picking argmax(logits + noise) at each token. We define a flow matching process directly over this noise-token coupling. The resulting paths are straight and non-crossing.

Фото профиля David Ruhe
David Ruhe11 дней назад

Surprisingly, we can then distill them directly into a bidirectional Flow Map student: the AR teacher supervises the velocity along the path, and a consistency loss trains few-step jumps. At inference, GSF maps noise to a full token sequence in parallel in just a few steps.

Фото профиля David Ruhe
David Ruhe11 дней назад

Across OpenWebText, LM1B, TinyStories, and XSum, GSF outperforms several baselines in 1–4 steps, establishing a strong compute-quality Pareto frontier.

Фото профиля David Ruhe
David Ruhe11 дней назад

📄 Paper: Huge thanks to our amazing intern @k01047339268, and to co-authors @ArnaudDoucet1 , @AndrewC_ML , @ValentinDeBort1 , and Thomas Mensink!

Фото профиля Liran Ringel
Liran Ringel10 дней назад

Impressive research taste 👏🏻 Very interesting

Фото профиля Austin DeWolfe
Austin DeWolfe11 дней назад

200 years ago they would have burned you at the stake for this black magic.

Похожие видео

Introducing ImagineArt Flow (Alpha) 🚀 Unify all your AI models, and edit inside one connected workflow. Not another AI Tool. This is a creative system for professionals! Comment ImagineArt Flow for early access Link :
0:30

Introducing ImagineArt Flow (Alpha) 🚀 Unify all your AI models, and edit inside one connected workflow. Not another AI Tool. This is a creative system for professionals! Comment ImagineArt Flow for early access Link :

ImagineArt

414,592 просмотров • 1 год назад

We're moving beyond autoregressive LLMs! Autoregressive LLMs generate text word-by-word, which can be slow and affect quality, while diffusion models refine noise step-by-step, allowing for faster iterations and error correction. Here's Gemini Diffusion running at 857 tokens/s:
0:40

We're moving beyond autoregressive LLMs! Autoregressive LLMs generate text word-by-word, which can be slow and affect quality, while diffusion models refine noise step-by-step, allowing for faster iterations and error correction. Here's Gemini Diffusion running at 857 tokens/s:

Akshay 🚀

34,524 просмотров • 1 год назад

Images, video, audio, actions — generative modeling has converged on one recipe: compress into continuous latents, generate in latent space, decode. Language is the lone exception: still generated token by token, as a long discrete stream. Latent Thought Flows: We compress 256 text tokens into 8 continuous latents, generate them with a one-step flow model, and read them out as text using an autoregressive decoder. Result: a better inference-compute vs. generation-quality Pareto frontier than a tuned autoregressive baseline — thanks to compression and one-step generation. Co-led w/ Zhengyang Geng 🧵
1:02

Images, video, audio, actions — generative modeling has converged on one recipe: compress into continuous latents, generate in latent space, decode. Language is the lone exception: still generated token by token, as a long discrete stream. Latent Thought Flows: We compress 256 text tokens into 8 continuous latents, generate them with a one-step flow model, and read them out as text using an autoregressive decoder. Result: a better inference-compute vs. generation-quality Pareto frontier than a tuned autoregressive baseline — thanks to compression and one-step generation. Co-led w/ Zhengyang Geng 🧵

Mihir Prabhudesai

83,020 просмотров • 2 месяцев назад

What if robots, humans, and video models could all describe action in the same language? Introducing Hydra-0🐍: a generalist world model that represents robot actions as action flow. One interface. Many embodiments, tasks, and environments. 🧵1/7
1:51

What if robots, humans, and video models could all describe action in the same language? Introducing Hydra-0🐍: a generalist world model that represents robot actions as action flow. One interface. Many embodiments, tasks, and environments. 🧵1/7

Hongyu Li

82,218 просмотров • 1 месяц назад

Did you know you can access Google's advanced image and video generation models all in one place? (Spoiler alert, it's in Flow!) Flow can be part of your creative suite for shot lists, character refinement, and animation. It was recently used to help make Google Search's latest campaign spot for AI Mode just ahead of Thanksgiving. For this ad, the team used Flow to land the characters and animate the scenes using Frames to Video. Watch the Thanksgiving spot below:
0:30

Did you know you can access Google's advanced image and video generation models all in one place? (Spoiler alert, it's in Flow!) Flow can be part of your creative suite for shot lists, character refinement, and animation. It was recently used to help make Google Search's latest campaign spot for AI Mode just ahead of Thanksgiving. For this ad, the team used Flow to land the characters and animate the scenes using Frames to Video. Watch the Thanksgiving spot below:

Google Flow

24,178 просмотров • 11 месяцев назад

We just brought flow maps to language modeling for one-step sequence generation 💥 Discrete diffusion is not necessary -- continuous flows over one-hot encodings achieve SoTA performance and ≥8.3× faster generation 🔥 We believe this is a major step forward for discrete generative modeling and language modeling alike. 🚀 Full thread from first author Chanhyuk Lee:
0:23

We just brought flow maps to language modeling for one-step sequence generation 💥 Discrete diffusion is not necessary -- continuous flows over one-hot encodings achieve SoTA performance and ≥8.3× faster generation 🔥 We believe this is a major step forward for discrete generative modeling and language modeling alike. 🚀 Full thread from first author Chanhyuk Lee:

Nicholas Boffi

44,633 просмотров • 7 месяцев назад

Introducing Fireblocks Flow. Flow is stablecoin acceptance infrastructure for PSPs and fintechs. Consumers can pay with any digital asset. Merchants settle the stablecoin they configure. If you process millions of payments across Africa, LatAm, or Southeast Asia, you know accepting stablecoins is messy. Now, it’s easy with one SDK or API. Drop Flow into your checkout, deposit, withdrawal, and agent flows. Flow will be live on Flutterwave, Nuvion.co, and Blipply.
0:59

Introducing Fireblocks Flow. Flow is stablecoin acceptance infrastructure for PSPs and fintechs. Consumers can pay with any digital asset. Merchants settle the stablecoin they configure. If you process millions of payments across Africa, LatAm, or Southeast Asia, you know accepting stablecoins is messy. Now, it’s easy with one SDK or API. Drop Flow into your checkout, deposit, withdrawal, and agent flows. Flow will be live on Flutterwave, Nuvion.co, and Blipply.

Dynamic

15,036 просмотров • 4 месяцев назад

Today's autoregressive models generate one token at a time. Mercury 2 generates tokens in parallel. Over 1,000 tok/sec on standard GPUs, at comparable quality to speed-optimized models. Since launch, the community has been showing what diffusion LLMs can unlock. Thanks to the team at Clyep for the breakdown.
3:49

Today's autoregressive models generate one token at a time. Mercury 2 generates tokens in parallel. Over 1,000 tok/sec on standard GPUs, at comparable quality to speed-optimized models. Since launch, the community has been showing what diffusion LLMs can unlock. Thanks to the team at Clyep for the breakdown.

Inception

21,231 просмотров • 5 месяцев назад

Today we’re introducing flow, micro1’s next-generation data platform for turning human expertise into measurable capability gains. At the core of flow are Realms, micro1’s real-world RL environments where experts establish what strong performance looks like. flow-gen models expand human judgment into new environments, rubrics, variations, and edge cases, while flow-qc models evaluate performance, identify the highest-value failures, and route them back to experts for review. Each cycle produces a stronger training signal and a measurable gain in capability. Those gains compound across frontier models, enterprise agents through Cortex, and robotics, while improving the suite of data generation models recursively. We’re moving beyond producing data to delivering abundant & predictable units of intelligence improvement.
3:36

Today we’re introducing flow, micro1’s next-generation data platform for turning human expertise into measurable capability gains. At the core of flow are Realms, micro1’s real-world RL environments where experts establish what strong performance looks like. flow-gen models expand human judgment into new environments, rubrics, variations, and edge cases, while flow-qc models evaluate performance, identify the highest-value failures, and route them back to experts for review. Each cycle produces a stronger training signal and a measurable gain in capability. Those gains compound across frontier models, enterprise agents through Cortex, and robotics, while improving the suite of data generation models recursively. We’re moving beyond producing data to delivering abundant & predictable units of intelligence improvement.

micro1

72,768 просмотров • 28 дней назад

Halo Infinite Forgers have proven they can create developer-quality maps, such as WAR 🇺🇦 and Paimon with their map called Nemesis. These maps can be added to the upcoming Forge matchmaking playlist, meaning we will have a constant flow of new maps to play in multiplayer!
1:21

Halo Infinite Forgers have proven they can create developer-quality maps, such as WAR 🇺🇦 and Paimon with their map called Nemesis. These maps can be added to the upcoming Forge matchmaking playlist, meaning we will have a constant flow of new maps to play in multiplayer!

Rebs Gaming

308,441 просмотров • 3 лет назад

Introducing Google Flow Agent 🎬 ✅ Google Flow Agent can help you plan and reason through complex creative tasks with your inputs, all while under your control. 🧠 It's built with Gemini models and brings a deep understanding of your project to help with everything from early brainstorming to editing the final product. 💡 The agent can be a sounding board for dialogue between characters in a scene, and make plot recommendations when you need inspiration. 🎞️ Your Google Flow Agent can also create multiple variations of a scene at one time to give you more options and batch edit. 📂 And when you have all your final assets, your Google Flow Agent can organize them into collections, and even neatly rename them.
0:22

Introducing Google Flow Agent 🎬 ✅ Google Flow Agent can help you plan and reason through complex creative tasks with your inputs, all while under your control. 🧠 It's built with Gemini models and brings a deep understanding of your project to help with everything from early brainstorming to editing the final product. 💡 The agent can be a sounding board for dialogue between characters in a scene, and make plot recommendations when you need inspiration. 🎞️ Your Google Flow Agent can also create multiple variations of a scene at one time to give you more options and batch edit. 📂 And when you have all your final assets, your Google Flow Agent can organize them into collections, and even neatly rename them.

Google

154,347 просмотров • 4 месяцев назад

Google isn’t betting on a single AI architecture. Sundar Pichai, CEO of Google: “We’re going to push the diffusion paradigm as hard as possible.” “All of today’s mainline Gemini models are autoregressive. Diffusion is a different paradigm.” “For the same capability, diffusion can be much faster.” “It’s behind the mainline models today, but there will be areas where it’s the right tool.” “We’re pushing multiple directions in parallel, and bringing them together where it makes sense.”
1:02

Google isn’t betting on a single AI architecture. Sundar Pichai, CEO of Google: “We’re going to push the diffusion paradigm as hard as possible.” “All of today’s mainline Gemini models are autoregressive. Diffusion is a different paradigm.” “For the same capability, diffusion can be much faster.” “It’s behind the mainline models today, but there will be areas where it’s the right tool.” “We’re pushing multiple directions in parallel, and bringing them together where it makes sense.”

Forward Future

152,933 просмотров • 9 месяцев назад

New episode of the Information Bottleneck! We talked with Stefano Ermon about why he thinks diffusion LLMs will replace autoregressive ones. Stefano co-invented DDIM, FlashAttention, DPO, and score-based diffusion models. He's a Stanford professor and now runs Inception AI, where they built Mercury II. We go deep but also cover the bigger picture - the startup journey, PhD vs industry, and where AI is heading. A few things that stuck with me: - He thinks of autoregressive models as typewriters and diffusion models as editors. One goes left to right. The other starts messy and refines. - Mercury II (their text difussion model) already beats the fastest autoregressive models on latency-critical stuff as voice agents, code suggestions, anything where you have a tight time budget. And it does it because diffusion generates tokens in parallel instead of one at a time. - We also got into whether AI will actually replace software engineers (his answer: no), PhD vs industry advice, and what it was like going from an ICML best paper to raising money.
0:54

New episode of the Information Bottleneck! We talked with Stefano Ermon about why he thinks diffusion LLMs will replace autoregressive ones. Stefano co-invented DDIM, FlashAttention, DPO, and score-based diffusion models. He's a Stanford professor and now runs Inception AI, where they built Mercury II. We go deep but also cover the bigger picture - the startup journey, PhD vs industry, and where AI is heading. A few things that stuck with me: - He thinks of autoregressive models as typewriters and diffusion models as editors. One goes left to right. The other starts messy and refines. - Mercury II (their text difussion model) already beats the fastest autoregressive models on latency-critical stuff as voice agents, code suggestions, anything where you have a tight time budget. And it does it because diffusion generates tokens in parallel instead of one at a time. - We also got into whether AI will actually replace software engineers (his answer: no), PhD vs industry advice, and what it was like going from an ICML best paper to raising money.

Ravid Shwartz Ziv

22,350 просмотров • 6 месяцев назад

Introducing Sounds, our latest development. This is our first step into the realm of sound design and auditory UX, unlocking higher levels of flow state and productivity by playing the perfect sounds where applicable. You'll want to turn the volume up for this one.
1:17

Introducing Sounds, our latest development. This is our first step into the realm of sound design and auditory UX, unlocking higher levels of flow state and productivity by playing the perfect sounds where applicable. You'll want to turn the volume up for this one.

Devin Desktop

72,130 просмотров • 1 год назад

.Oliver Cameron says one of the biggest applications for world models will be training AI inside simulated environments. "One of the killer applications of world models will be as a learning environment where language models and other types of AI can learn. It gets very much philosophical."
1:33

.Oliver Cameron says one of the biggest applications for world models will be training AI inside simulated environments. "One of the killer applications of world models will be as a learning environment where language models and other types of AI can learn. It gets very much philosophical."

TBPN

40,172 просмотров • 2 месяцев назад

Introducing Hailuo AI Agent: Your next-gen creative flow — FREE for one week! 🔷 Just talk, and watch your ideas come alive on the infinite canvas. 🔷 The Agent selects the best models, plans the workflow, and brings your vision to life. 🔷 Images, videos, audio — all in one place.
0:27

Introducing Hailuo AI Agent: Your next-gen creative flow — FREE for one week! 🔷 Just talk, and watch your ideas come alive on the infinite canvas. 🔷 The Agent selects the best models, plans the workflow, and brings your vision to life. 🔷 Images, videos, audio — all in one place.

Hailuo AI (MiniMax)

617,849 просмотров • 1 год назад

Sculptor: the missing UI for Claude Code 🎨 Imagine running 5 Claudes in parallel, safely in containers, while you stay in flow. Then bring their work straight into your IDE to test/edit together. This is how one developer ships like a team. Try it with Sonnet 4.5!
1:46

Sculptor: the missing UI for Claude Code 🎨 Imagine running 5 Claudes in parallel, safely in containers, while you stay in flow. Then bring their work straight into your IDE to test/edit together. This is how one developer ships like a team. Try it with Sonnet 4.5!

Kanjun 🐙

519,185 просмотров • 1 год назад

🎉 Introducing: wallet․fund() You can now fund any wallet from fiat via a single API call using the CDP SDK! wallet․fund() consolidates the legacy multi-step fiat onramping process into a single programmatic flow for users. With wallet․fund(): fiat in your bank → crypto in your wallet Without wallet․fund(): fiat in your bank → fiat on an exchange → swap for crypto on exchange → transfer to your wallet Watch Lakshay’s demo below to see two ways to fund a wallet: → A quote-based flow, where you preview the transaction before executing → A direct flow, where you skip the quote and fund your wallet immediately
0:59

🎉 Introducing: wallet․fund() You can now fund any wallet from fiat via a single API call using the CDP SDK! wallet․fund() consolidates the legacy multi-step fiat onramping process into a single programmatic flow for users. With wallet․fund(): fiat in your bank → crypto in your wallet Without wallet․fund(): fiat in your bank → fiat on an exchange → swap for crypto on exchange → transfer to your wallet Watch Lakshay’s demo below to see two ways to fund a wallet: → A quote-based flow, where you preview the transaction before executing → A direct flow, where you skip the quote and fund your wallet immediately

Coinbase Developer Platform🛡️

86,432 просмотров • 1 год назад

🤔The more I studied diffusion language models, the more I came to appreciate the simplicity of autoregressive (AR) language models. AR models are trained to agree with what they generate, and their serving stacks are built to preserve that structure. DLMs often do neither: they lack introspective consistency, and high TPF does not necessarily translate into high real-world TPS. We propose Introspective Diffusion Language Model (I-DLM), which unifies introspection and generation in a single pass: 1. 🧑‍🎓I-DLM brings introspective consistency to DLMs with only 5B training tokens, achieving AR-thinking-level quality. 2. 🚀 I-DLM carefully trades compute for higher TPF while converting that advantage into real TPS under high-concurrency serving. 📖Website: ⌨️Code:
1:17

🤔The more I studied diffusion language models, the more I came to appreciate the simplicity of autoregressive (AR) language models. AR models are trained to agree with what they generate, and their serving stacks are built to preserve that structure. DLMs often do neither: they lack introspective consistency, and high TPF does not necessarily translate into high real-world TPS. We propose Introspective Diffusion Language Model (I-DLM), which unifies introspection and generation in a single pass: 1. 🧑‍🎓I-DLM brings introspective consistency to DLMs with only 5B training tokens, achieving AR-thinking-level quality. 2. 🚀 I-DLM carefully trades compute for higher TPF while converting that advantage into real TPS under high-concurrency serving. 📖Website: ⌨️Code:

Chenfeng_X

344,902 просмотров • 6 месяцев назад