正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Excited to share what my team has been working on lately - Gemini diffusion! We bring diffusion to language modeling, yielding more power and blazing speeds! 🚀🚀🚀 Gemini diffusion is especially strong at coding. In this example the model generates at 2000 tokens/sec, including overheads like tokenization, prefill, safety... show more

Brendan O'Donoghue

7,152 subscribers

577,872 次观看 • 1 年前 •via X (Twitter)

科学技术新闻政治教育

Anya Rossi• Live Now

Private livecam show

10 条评论

Brendan O'Donoghue 的头像

Brendan O'Donoghue1 年前

With Gemini diffusion you can 'vibe code' so fast that it feels almost instant:

Brendan O'Donoghue 的头像

Brendan O'Donoghue1 年前

Lightning speeds are not the only advantage. Unlike autoregressive models that are restricted to generating one token at a time, diffusion can do non-causal reasoning within the generation. Take this example: "What is (√(81) * (2/3))^2 + (15 - 3) / (2^2)). First provide the answer and then derive the solution." This is a very hard prompt for AR models because they can't reason about the solution before generating it, but diffusion models can reason non-causally to get it right (ans: 39). GPT-4o fails this problem:

Brendan O'Donoghue 的头像

Brendan O'Donoghue1 年前

A similar one inspired by the 'Sparks of AGI paper' by @SebastienBubeck et al: "How many primes are there between 150 and 250? The first thing you should output is the total number, then print the exact list inside [ ] brackets." (ans: 18) GPT-4o fails this one too:

Brendan O'Donoghue 的头像

Brendan O'Donoghue1 年前

For more details, and to get access, see here:

Jeremy Howard 的头像

Jeremy Howard1 年前

Makes me so happy to see this! :D I've hearing about this project for quite some time, and was really hoping that it would see the light of day.

Antoine Moulin 的头像

Antoine Moulin1 年前

Congrats to the team!!

Brendan O'Donoghue 的头像

Brendan O'Donoghue1 年前

Thank you!

David Zhang (▲) 的头像

David Zhang (▲)1 年前

Do you feel the AGI yet?

heiner 的头像

heiner1 年前

congrats

mike 的头像

mike1 年前

This thing really justifies the mind blown emoji. Well done. Absolutely incredible preview of the future.

相关视频

🚀Meet Gemini Diffusion, our first diffusion-based and super fast language model, just announced at Google I/O!🚀 Very excited to be able to share what I've been working on for the past little while with our amazing small team Google DeepMind.

🚀Meet Gemini Diffusion, our first diffusion-based and super fast language model, just announced at Google I/O!🚀 Very excited to be able to share what I've been working on for the past little while with our amazing small team Google DeepMind.

Ivana Balazevic

43,531 次观看 • 1 年前

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.

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 次观看 • 2 个月前

(1/10) 🔥Thrilled to introduce OneDiffusion—our latest work in unified diffusion modeling! 🚀 This model bridges the gap between image synthesis and understanding, excelling in a wide range of tasks: T2I, conditional generation, image understanding, identity preservation, multiview generation, and even camera pose estimation. Learn more at: Project: arXiv: Code (on the way):

(1/10) 🔥Thrilled to introduce OneDiffusion—our latest work in unified diffusion modeling! 🚀 This model bridges the gap between image synthesis and understanding, excelling in a wide range of tasks: T2I, conditional generation, image understanding, identity preservation, multiview generation, and even camera pose estimation. Learn more at: Project: arXiv: Code (on the way):

Jiasen Lu

33,426 次观看 • 1 年前

🤔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:

🤔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

344,620 次观看 • 4 个月前

Very excited to share the best coding model we’ve ever built! Today we’re launching Gemini 2.5 Pro Preview 'I/O edition' with massively improved coding capabilities. Ranks no.1 on LMArena in Coding and no.1 on the WebDev Arena Leaderboard. It’s especially good at building interactive web apps - this demo shows how it can be helpful for prototyping ideas. Try it in Google Gemini, Vertex AI, and AI Studio Enjoy the pre-I/O goodies !

Very excited to share the best coding model we’ve ever built! Today we’re launching Gemini 2.5 Pro Preview 'I/O edition' with massively improved coding capabilities. Ranks no.1 on LMArena in Coding and no.1 on the WebDev Arena Leaderboard. It’s especially good at building interactive web apps - this demo shows how it can be helpful for prototyping ideas. Try it in Google Gemini, Vertex AI, and AI Studio Enjoy the pre-I/O goodies !

Demis Hassabis

995,897 次观看 • 1 年前

Sundar Pichai Reveals That Google Founder Sergey Brin Is Coding Gemini "Sergey is definitely spending time with the Gemini team in a pretty hardcore way, like sitting and coding and spending time with the engineers. That gives energy to the team. It is unparalleled to have a founder sitting there, looking at loss curves, giving feedback on model architectures, and how can we improve post training." Sundar Pichai talking with david friedberg on The All-In Podcast

Sundar Pichai Reveals That Google Founder Sergey Brin Is Coding Gemini "Sergey is definitely spending time with the Gemini team in a pretty hardcore way, like sitting and coding and spending time with the engineers. That gives energy to the team. It is unparalleled to have a founder sitting there, looking at loss curves, giving feedback on model architectures, and how can we improve post training." Sundar Pichai talking with david friedberg on The All-In Podcast

Overlap: Business & Tech

46,573 次观看 • 1 年前

Today marks a proud milestone for Zeta and a moment to reflect on the journey built by our customers, our people, and a shared belief in what this company could become. This video captures an incredible chapter in the Zeta story, from ringing the bell at the NYSE 🏛 to seeing the future we envisioned come to life. We are leading the AI marketing revolution, and I could not be more excited about what comes next. Extremely proud of this team and what we have built together. $zeta LG 🚀🚀🚀

Today marks a proud milestone for Zeta and a moment to reflect on the journey built by our customers, our people, and a shared belief in what this company could become. This video captures an incredible chapter in the Zeta story, from ringing the bell at the NYSE 🏛 to seeing the future we envisioned come to life. We are leading the AI marketing revolution, and I could not be more excited about what comes next. Extremely proud of this team and what we have built together. $zeta LG 🚀🚀🚀

David A Steinberg

652,294 次观看 • 2 个月前

I'll remember this week for the rest of my life. Here's my interview with the incredible Sundar Pichai, CEO of Google. We spoke about Gemini, world models, self-improving AI, AI anxiety, and more. Full vid below: 0:00 - Intro 0:40 - Gemini as a World Model 1:30 - Gemini Diffusion 2:33 - Self-improving AI 4:09 - What AI needs to improve 5:06 - Agent Memory 6:46 - AI Form Factors 7:58 - Google Search Future 9:33 - AI Anxiety

I'll remember this week for the rest of my life. Here's my interview with the incredible Sundar Pichai, CEO of Google. We spoke about Gemini, world models, self-improving AI, AI anxiety, and more. Full vid below: 0:00 - Intro 0:40 - Gemini as a World Model 1:30 - Gemini Diffusion 2:33 - Self-improving AI 4:09 - What AI needs to improve 5:06 - Agent Memory 6:46 - AI Form Factors 7:58 - Google Search Future 9:33 - AI Anxiety

Matthew Berman

98,864 次观看 • 1 年前

Also starting today, Gemini Advanced users can create with Canvas on Gemini 2.5 Pro. This model’s advanced coding abilities make it easy for anyone to collaborate, create, and share imaginative prototypes (like this example of a 3D dino world, made in under 5 minutes with just a few prompts). Here’s how to get started: 1) Go to 2) Choose the “2.5 Pro (experimental)” model 3) Select “Canvas” and start prompting We’re working hard to make Gemini 2.5 Pro available for more people to try soon–watch this space for updates.

Also starting today, Gemini Advanced users can create with Canvas on Gemini 2.5 Pro. This model’s advanced coding abilities make it easy for anyone to collaborate, create, and share imaginative prototypes (like this example of a 3D dino world, made in under 5 minutes with just a few prompts). Here’s how to get started: 1) Go to 2) Choose the “2.5 Pro (experimental)” model 3) Select “Canvas” and start prompting We’re working hard to make Gemini 2.5 Pro available for more people to try soon–watch this space for updates.

Google Gemini

302,724 次观看 • 1 年前

(1/5) Gemini 3, our most intelligent model, is landing in Google Search today – starting with AI Mode. Excited that this is the first time we’re shipping a new Gemini model in Search on day one! 🚀 In Search, Gemini 3 with generative layouts will make it easy to get a rich understanding of anything on your mind. It has state-of-the-art reasoning, deep multimodal understanding and advanced agentic capabilities. That allows the model to shine when you ask it to explain advanced concepts or ideas – it reasons and can code interactive visuals in real-time. It can tackle your toughest questions like advanced science.

(1/5) Gemini 3, our most intelligent model, is landing in Google Search today – starting with AI Mode. Excited that this is the first time we’re shipping a new Gemini model in Search on day one! 🚀 In Search, Gemini 3 with generative layouts will make it easy to get a rich understanding of anything on your mind. It has state-of-the-art reasoning, deep multimodal understanding and advanced agentic capabilities. That allows the model to shine when you ask it to explain advanced concepts or ideas – it reasons and can code interactive visuals in real-time. It can tackle your toughest questions like advanced science.

Robby Stein

94,877 次观看 • 8 个月前

TokenFlow: Consistent Diffusion Features for Consistent Video Editing paper page: The generative AI revolution has recently expanded to videos. Nevertheless, current state-of-the-art video models are still lagging behind image models in terms of visual quality and user control over the generated content. In this work, we present a framework that harnesses the power of a text-to-image diffusion model for the task of text-driven video editing. Specifically, given a source video and a target text-prompt, our method generates a high-quality video that adheres to the target text, while preserving the spatial layout and motion of the input video. Our method is based on a key observation that consistency in the edited video can be obtained by enforcing consistency in the diffusion feature space. We achieve this by explicitly propagating diffusion features based on inter-frame correspondences, readily available in the model. Thus, our framework does not require any training or fine-tuning, and can work in conjunction with any off-the-shelf text-to-image editing method. We demonstrate state-of-the-art editing results on a variety of real-world videos.

TokenFlow: Consistent Diffusion Features for Consistent Video Editing paper page: The generative AI revolution has recently expanded to videos. Nevertheless, current state-of-the-art video models are still lagging behind image models in terms of visual quality and user control over the generated content. In this work, we present a framework that harnesses the power of a text-to-image diffusion model for the task of text-driven video editing. Specifically, given a source video and a target text-prompt, our method generates a high-quality video that adheres to the target text, while preserving the spatial layout and motion of the input video. Our method is based on a key observation that consistency in the edited video can be obtained by enforcing consistency in the diffusion feature space. We achieve this by explicitly propagating diffusion features based on inter-frame correspondences, readily available in the model. Thus, our framework does not require any training or fine-tuning, and can work in conjunction with any off-the-shelf text-to-image editing method. We demonstrate state-of-the-art editing results on a variety of real-world videos.

AK

143,469 次观看 • 3 年前

Diffusion has shown great promise for generating robot **actions**, can it act as a **world model** to generate the future conditioned on actions? In our work led by han qi Haocheng Yin and in collaboration with Yilun Du, we show a **controllable** action-conditioned video diffusion model can produce photorealistic and (near) physics-accurate future predictions. This ability strengthens the policy via: - ranking different action proposals and selecting the best, or - **visual** trajectory optimization by optimizing the action proposals using gradient ascent. Learn more about Generative Predictive Control (GPC) at:

Diffusion has shown great promise for generating robot actions, can it act as a world model to generate the future conditioned on actions? In our work led by han qi Haocheng Yin and in collaboration with Yilun Du, we show a controllable action-conditioned video diffusion model can produce photorealistic and (near) physics-accurate future predictions. This ability strengthens the policy via: - ranking different action proposals and selecting the best, or - visual trajectory optimization by optimizing the action proposals using gradient ascent. Learn more about Generative Predictive Control (GPC) at:

Heng Yang

38,434 次观看 • 1 年前

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.

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 次观看 • 4 个月前

Tune in to PayPal First Look on our PayPal YouTube channel tomorrow, January 25 at 9:30am PT. We'll unveil what our PayPal and Venmo teams have been working on and will bring to market this year. And we're just getting started! 🚀

Tune in to PayPal First Look on our PayPal YouTube channel tomorrow, January 25 at 9:30am PT. We'll unveil what our PayPal and Venmo teams have been working on and will bring to market this year. And we're just getting started! 🚀

Alex Chriss

274,882 次观看 • 2 年前

Very excited to share what our research team at Covariant has been working on: RFM-1, our latest Robotics Foundation Model. Built on top of high-quality multimodal data the Covariant robotics fleet has collected over the years, this project embodies our commitment to pushing the boundaries of how AI can interact with and understand the physical world. We're eager to share this milestone with the community and invite feedback and collaboration. Read more at

Very excited to share what our research team at Covariant has been working on: RFM-1, our latest Robotics Foundation Model. Built on top of high-quality multimodal data the Covariant robotics fleet has collected over the years, this project embodies our commitment to pushing the boundaries of how AI can interact with and understand the physical world. We're eager to share this milestone with the community and invite feedback and collaboration. Read more at

Rocky Duan

31,833 次观看 • 2 年前

GLM 4.7 is one of the top open-source models on LM Arena—and it's going toe-to-toe with Claude Opus 4.5 and Gemini Pro. We sat down with Anastasios Nikolas Angelopoulos, co-founder and CEO of Arena.ai, to break down 8,000+ developer votes: → Within 30 points of Gemini Pro in math & coding → Frontier-level multi-turn & instruction following → The open-source model devs are actually switching to The best part? You can run it at 1,500+ tokens/sec on Cerebras—for free.

GLM 4.7 is one of the top open-source models on LM Arena—and it's going toe-to-toe with Claude Opus 4.5 and Gemini Pro. We sat down with Anastasios Nikolas Angelopoulos, co-founder and CEO of Arena.ai, to break down 8,000+ developer votes: → Within 30 points of Gemini Pro in math & coding → Frontier-level multi-turn & instruction following → The open-source model devs are actually switching to The best part? You can run it at 1,500+ tokens/sec on Cerebras—for free.

Cerebras

28,515 次观看 • 6 个月前

Wow 🤩 the response to Project Starlight, the first-ever diffusion model for AI video restoration, has been incredible🚀 Many of you couldn't wait and sent us your videos to restore with Starlight So let’s make it a thing: Send us your videos, and we’ll restore them for free using Starlight 🌟 Read instructions in 🧵👇 Like & comment Starlight to get early access, if you haven’t done so already #ProjectStarlight

Wow 🤩 the response to Project Starlight, the first-ever diffusion model for AI video restoration, has been incredible🚀 Many of you couldn't wait and sent us your videos to restore with Starlight So let’s make it a thing: Send us your videos, and we’ll restore them for free using Starlight 🌟 Read instructions in 🧵👇 Like & comment Starlight to get early access, if you haven’t done so already #ProjectStarlight

Topaz Labs

329,619 次观看 • 1 年前

$Auto regressive LLMs are officially on notice. run Gemma 4 26B diffusion gguf with llama.cpp Google just dropped DiffusionGemma-26B, and it completely flips how we generate text. instead of predicting words one by one, it generates 256 tokens in parallel using bi-directional attention. its like stable diffusion, but for language. the model starts with random text "noise" and iteratively refines and self-corrects the entire block in real-time to fix formatting and reasoning errors on the fly. since it’s a Mixture of Experts (MoE) that only activates 3.8B parameters during inference, it fits perfectly on consumer hardware. You can run the Q4_K_M quant with an 18GB VRAM budget on a single RTX 3090 or RTX 4090 with exceptional throughput. Tested on Ubuntu 22 with CUDA 13.1 using the cutting edge experimental llama.cpp branch. Here is how to compile and run it with the live terminal denoising visualizer: # 1. Clone & check out the experimental PR (#24423) - 1) git clone && cd llama.cpp -git fetch origin 2) pull/24423/head:diffusiongemma && --git checkout diffusiongemma # 2. Build with CUDA support 1) cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native 2) cmake --build build -j $(nproc) --config Release --target llama-diffusion-cli # 3. Run with live visual denoising (llama.cpp flags) ./build/bin/llama-diffusion-cli \ -m /path/to/diffusiongemma-26B-A4B-it-Q4_K_M.gguf \ -ngl 99 -cnv -n 2048 --diffusion-visual Watch the video below to see the live --diffusion-visual canvas iteratively de noising the prompt output in real time. guide and unsloth's hugging face GGUF model links are in the comments below! Is auto regressive generation officially legacy tech? Let me know what you think.$

Auto regressive LLMs are officially on notice. run Gemma 4 26B diffusion gguf with llama.cpp Google just dropped DiffusionGemma-26B, and it completely flips how we generate text. instead of predicting words one by one, it generates 256 tokens in parallel using bi-directional attention. its like stable diffusion, but for language. the model starts with random text "noise" and iteratively refines and self-corrects the entire block in real-time to fix formatting and reasoning errors on the fly. since it’s a Mixture of Experts (MoE) that only activates 3.8B parameters during inference, it fits perfectly on consumer hardware. You can run the Q4_K_M quant with an 18GB VRAM budget on a single RTX 3090 or RTX 4090 with exceptional throughput. Tested on Ubuntu 22 with CUDA 13.1 using the cutting edge experimental llama.cpp branch. Here is how to compile and run it with the live terminal denoising visualizer: # 1. Clone & check out the experimental PR (#24423) - 1) git clone && cd llama.cpp -git fetch origin 2) pull/24423/head:diffusiongemma && --git checkout diffusiongemma # 2. Build with CUDA support 1) cmake -B build -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native 2) cmake --build build -j $(nproc) --config Release --target llama-diffusion-cli # 3. Run with live visual denoising (llama.cpp flags) ./build/bin/llama-diffusion-cli \ -m /path/to/diffusiongemma-26B-A4B-it-Q4_K_M.gguf \ -ngl 99 -cnv -n 2048 --diffusion-visual Watch the video below to see the live --diffusion-visual canvas iteratively de noising the prompt output in real time. guide and unsloth's hugging face GGUF model links are in the comments below! Is auto regressive generation officially legacy tech? Let me know what you think.

Alok

52,656 次观看 • 2 个月前

Gemini 3 is indeed impressive. In my limited testing with it, it feels like it is able to take on larger, complex prompts with ease and low failure rate. I think one of the biggest strengths that Gemini 3 has is its ability to generate coherent looking 3D voxel or SVG art, which makes it super easy to create game prototypes without having to implement art at all. Made this Heroes of might and magic-inspired overworld prototype with mini map and basic UI in less than 2 hours. Really excited to test this more. Demis Hassabis and Logan Kilpatrick and the Gemini team, you guys really cooked this time!

Gemini 3 is indeed impressive. In my limited testing with it, it feels like it is able to take on larger, complex prompts with ease and low failure rate. I think one of the biggest strengths that Gemini 3 has is its ability to generate coherent looking 3D voxel or SVG art, which makes it super easy to create game prototypes without having to implement art at all. Made this Heroes of might and magic-inspired overworld prototype with mini map and basic UI in less than 2 hours. Really excited to test this more. Demis Hassabis and Logan Kilpatrick and the Gemini team, you guys really cooked this time!

Danny Limanseta

119,970 次观看 • 8 个月前