Загрузка видео...

Не удалось загрузить видео

На главную

Xue et al., "DepthART: Scaling Foundation Monocular Depth to Tiny Models" Well-strategized distillation with camera-conditioned fine-tuning. 1000 FPS on RTX A6000, 200 FPS on Jetson Orin NX.

kwangmoo_yi's profile picture

Kwang Moo Yi

4,803 subscribers

16,409 просмотров • 28 дней назад •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

TikTok presents Depth Anything Unleashing the Power of Large-Scale Unlabeled Data paper page: demo: Depth Anything is trained on 1.5M labeled images and 62M+ unlabeled images jointly, providing the most capable Monocular Depth Estimation (MDE) foundation models with the following features: zero-shot relative depth estimation, better than MiDaS v3.1 (BEiTL-512) zero-shot metric depth estimation, better than ZoeDepth optimal in-domain fine-tuning and evaluation on NYUv2 and KITTI
0:44

TikTok presents Depth Anything Unleashing the Power of Large-Scale Unlabeled Data paper page: demo: Depth Anything is trained on 1.5M labeled images and 62M+ unlabeled images jointly, providing the most capable Monocular Depth Estimation (MDE) foundation models with the following features: zero-shot relative depth estimation, better than MiDaS v3.1 (BEiTL-512) zero-shot metric depth estimation, better than ZoeDepth optimal in-domain fine-tuning and evaluation on NYUv2 and KITTI

AK

600,132 просмотров • 2 лет назад

timelapse #164 (15.5 hrs) - Burning midnight oil with the Kimi team benchmarking on kernels - Getting a Persistent Fused Mega Kernel for a Puffer Lib Environment (the full Craftax game) - Dipped for the night to have models run my overnight training runs on autopilot, including some kernel-based GRPO on qwen3.6 - private contract work - downloaded 1.1 TB of minecraft videos to train a foundation model on - optimized my minecraft foundation model training pipeline and kernels to train on 500k hrs of 128px at 10 fps in ~28 hrs on single gpu (5 hrs of learning per second on rtx pro 6000 blackwell) - doing it for the love of the game
0:30

timelapse #164 (15.5 hrs) - Burning midnight oil with the Kimi team benchmarking on kernels - Getting a Persistent Fused Mega Kernel for a Puffer Lib Environment (the full Craftax game) - Dipped for the night to have models run my overnight training runs on autopilot, including some kernel-based GRPO on qwen3.6 - private contract work - downloaded 1.1 TB of minecraft videos to train a foundation model on - optimized my minecraft foundation model training pipeline and kernels to train on 500k hrs of 128px at 10 fps in ~28 hrs on single gpu (5 hrs of learning per second on rtx pro 6000 blackwell) - doing it for the love of the game

Elliot Arledge

28,210 просмотров • 1 месяц назад

This seemingly obvious prediction didn't take long to become reality. MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors Contributions: • The first real-time SLAM system using the two-view 3D reconstruction prior MASt3R [20] as a foundation. • Efficient techniques for pointmap matching, tracking and local fusion, graph construction and loop closure, and second-order global optimization. • A state-of-the-art dense SLAM system capable of handling generic, time-varying camera models. Abstract: We present a real-time monocular dense SLAM system, designed from the ground up using MASt3R, a two-view 3D reconstruction and matching prior. Equipped with this strong prior, our system remains robust on in-the-wild video sequences, making no assumptions on a fixed or parametric camera model beyond a unique camera center. Key features include: - Efficient methods for pointmap matching, camera tracking, and local fusion - Graph construction and loop closure - Second-order global optimization With known calibration, a simple modification achieves state-of-the-art performance across various benchmarks. Altogether, we propose a plug-and-play monocular SLAM system capable of producing globally-consistent poses and dense geometry while operating at 15 FPS.
0:23

This seemingly obvious prediction didn't take long to become reality. MASt3R-SLAM: Real-Time Dense SLAM with 3D Reconstruction Priors Contributions: • The first real-time SLAM system using the two-view 3D reconstruction prior MASt3R [20] as a foundation. • Efficient techniques for pointmap matching, tracking and local fusion, graph construction and loop closure, and second-order global optimization. • A state-of-the-art dense SLAM system capable of handling generic, time-varying camera models. Abstract: We present a real-time monocular dense SLAM system, designed from the ground up using MASt3R, a two-view 3D reconstruction and matching prior. Equipped with this strong prior, our system remains robust on in-the-wild video sequences, making no assumptions on a fixed or parametric camera model beyond a unique camera center. Key features include: - Efficient methods for pointmap matching, camera tracking, and local fusion - Graph construction and loop closure - Second-order global optimization With known calibration, a simple modification achieves state-of-the-art performance across various benchmarks. Altogether, we propose a plug-and-play monocular SLAM system capable of producing globally-consistent poses and dense geometry while operating at 15 FPS.

MrNeRF

29,974 просмотров • 1 год назад

NVIDIA AI Open-Sources ViPE (Video Pose Engine): A Powerful and Versatile 3D Video Annotation Tool for Spatial AI ViPE integrates bundle adjustment with dense optical flow, sparse keypoint tracking, and metric depth priors to estimate camera intrinsics, poses, and dense depth maps at 3–5 FPS on a single GPU. It significantly improves over prior uncalibrated pose estimation methods, achieving 18% and 50% error reduction on TUM and KITTI benchmarks, respectively, and shows robustness to dynamic scenes and diverse camera models. Beyond the method, the NVIDIA team also released a large-scale dataset comprising ~100K real-world internet videos, 1M AI-generated videos, and 2K panoramic videos (≈96M frames) annotated with metric depth and poses. This dataset and engine aim to accelerate training for spatial AI tasks such as 3D reconstruction, video generation, and robotics.... full analysis: paper: codes: NVIDIA NVIDIA AI NVIDIAnewsroom NVIDIA Robotics NVIDIA AIDev NVIDIAdeveloper
1:13

NVIDIA AI Open-Sources ViPE (Video Pose Engine): A Powerful and Versatile 3D Video Annotation Tool for Spatial AI ViPE integrates bundle adjustment with dense optical flow, sparse keypoint tracking, and metric depth priors to estimate camera intrinsics, poses, and dense depth maps at 3–5 FPS on a single GPU. It significantly improves over prior uncalibrated pose estimation methods, achieving 18% and 50% error reduction on TUM and KITTI benchmarks, respectively, and shows robustness to dynamic scenes and diverse camera models. Beyond the method, the NVIDIA team also released a large-scale dataset comprising ~100K real-world internet videos, 1M AI-generated videos, and 2K panoramic videos (≈96M frames) annotated with metric depth and poses. This dataset and engine aim to accelerate training for spatial AI tasks such as 3D reconstruction, video generation, and robotics.... full analysis: paper: codes: NVIDIA NVIDIA AI NVIDIAnewsroom NVIDIA Robotics NVIDIA AIDev NVIDIAdeveloper

Marktechpost AI Dev News ⚡

217,453 просмотров • 11 месяцев назад

New short course on Pretraining LLMs! Developed with Upstage and taught by their CEO Sung Kim and CSO Lucy Park. While prompting or fine-tuning existing models works well for many general language tasks, pretraining is valuable for specialized domains or languages with limited representation in current models. This course walks you through the LLM pretraining pipeline: 1. Data preparation: Learn to source, clean, and prepare training data using HuggingFace. 2. Model architecture: Configure transformer networks, including modifying existing models. 3. Training: Set up and run training using open-source libraries. 4. Evaluation: Benchmark performance using popular evaluation strategies. As an example use case, you'll also compare the output of a base model with its fine-tuned and further pretrained variants, to see the impact of pretraining on a model's ability to write Python. The course also explores an innovative technique called depth up-scaling, which Upstage used to train their Solar model family, reducing pretraining compute costs by up to 70%. This technique works by first duplicating layers of a smaller pretrained model to form a larger model, and then further pretraining the result. Sign up here!
3:54

New short course on Pretraining LLMs! Developed with Upstage and taught by their CEO Sung Kim and CSO Lucy Park. While prompting or fine-tuning existing models works well for many general language tasks, pretraining is valuable for specialized domains or languages with limited representation in current models. This course walks you through the LLM pretraining pipeline: 1. Data preparation: Learn to source, clean, and prepare training data using HuggingFace. 2. Model architecture: Configure transformer networks, including modifying existing models. 3. Training: Set up and run training using open-source libraries. 4. Evaluation: Benchmark performance using popular evaluation strategies. As an example use case, you'll also compare the output of a base model with its fine-tuned and further pretrained variants, to see the impact of pretraining on a model's ability to write Python. The course also explores an innovative technique called depth up-scaling, which Upstage used to train their Solar model family, reducing pretraining compute costs by up to 70%. This technique works by first duplicating layers of a smaller pretrained model to form a larger model, and then further pretraining the result. Sign up here!

Andrew Ng

85,695 просмотров • 2 лет назад

Are you kidding me??? It grasps multiple objects with different ways, all at once with… a single hand??? No pauses. 1x speed. GENE-26.5 is Genesis AI’s robotics-native multimodal foundation model. It’s trained on 200,000+ hours of real human hand data (motion, force, touch) and runs on a 54-DoF bimanual system: Scaling that human data 4x lifted real-robot success rates from 16.6% to 65.6% on long-horizon dexterous tasks!! Same model weights, zero fine-tuning for this exact sequence. You know those tiny coordinated movements you do without thinking…? Robots couldn’t reliably do that before. Now they can. Today. This is the video you’ll send to my friends, outside of our bubble, when they say “robots are still just demos.” Congrats to the entire team around Zhou Xian! Credit: Seen at Zu Wang (Zu Wang) Genesis official announcement for the full story + longer demo: ——— Weekly robotics and AI insights. Subscribe free:
0:20

Are you kidding me??? It grasps multiple objects with different ways, all at once with… a single hand??? No pauses. 1x speed. GENE-26.5 is Genesis AI’s robotics-native multimodal foundation model. It’s trained on 200,000+ hours of real human hand data (motion, force, touch) and runs on a 54-DoF bimanual system: Scaling that human data 4x lifted real-robot success rates from 16.6% to 65.6% on long-horizon dexterous tasks!! Same model weights, zero fine-tuning for this exact sequence. You know those tiny coordinated movements you do without thinking…? Robots couldn’t reliably do that before. Now they can. Today. This is the video you’ll send to my friends, outside of our bubble, when they say “robots are still just demos.” Congrats to the entire team around Zhou Xian! Credit: Seen at Zu Wang (Zu Wang) Genesis official announcement for the full story + longer demo: ——— Weekly robotics and AI insights. Subscribe free:

Ilir Aliu

204,476 просмотров • 3 месяцев назад

A team tested Pi0, Pi0 Fast, Gr00t, and ACT on real robot arms in manufacturing tasks. (🔖 Bookmark this for later!) The task was precise: place thin rectangular frames from a messy stack into a holder. The team fine-tuned each model on 100 real trajectories and compared training time, inference speed, motion quality, and success rates. ⬇️ Here’s a breakdown of what they found Pi0 (Original) ✅ Strongest overall performance in precise pick-and-place ✅ High success rate even in edge cases ✅ Longest training time (~11 hours, ~$30 per run) ✅ Inference time of 80 ms causes short pauses between actions Despite delays, it handles complex scenarios well… solid for high-precision tasks, but slow to train. Gr00t ✅ Trains fast (~2 hours, ~$5 per run) ✅ Performs almost as well as Pi0 on large-object tasks ✅ Struggles with fine precision; random movement in some trials ✅ More training didn’t fix jitter or random offsets Best suited for tasks where exact precision isn’t critical. Not ready for manufacturing-grade accuracy without more tuning. Pi0 Fast ✅ Promised faster training, but results were underwhelming ✅ Training at 6 hours still showed low success rates ✅ Inference was slower than expected ✅ Not reliable for generalizing even slightly new tasks Currently too unstable for real-world deployment. Doesn’t live up to the “Fast” name yet. ACT (Baseline) ✅ 200MB model—lightweight, but limited ✅ Struggles with stacked objects or ambiguous scenes ✅ Success rates around 70% in best-case setups ✅ Can’t match newer models on precision or generalization Still a solid baseline, but clearly a generation behind in robustness. 🚨 Extra Notes All newer models share a common issue: •Inference takes longer than a frame (80 ms vs 33 ms), so robots “pause” between chunks. •This results in jittery movements, but not a dealbreaker unless tasks are time-sensitive. Language-conditioned tasks also fell short: after training on two labeled tasks, the model couldn’t generalize to a third unseen combination using only text prompts. ✅ The good news? These models adapt well to new robot arms with quick fine-tuning. ❌ The bad news? There’s still no plug-and-play solution for improving performance after deployment. Reinforcement learning or DAgger-style data collection during real-world operation may be the next big step, something many teams in robotics are actively working on.
0:51

A team tested Pi0, Pi0 Fast, Gr00t, and ACT on real robot arms in manufacturing tasks. (🔖 Bookmark this for later!) The task was precise: place thin rectangular frames from a messy stack into a holder. The team fine-tuned each model on 100 real trajectories and compared training time, inference speed, motion quality, and success rates. ⬇️ Here’s a breakdown of what they found Pi0 (Original) ✅ Strongest overall performance in precise pick-and-place ✅ High success rate even in edge cases ✅ Longest training time (~11 hours, ~$30 per run) ✅ Inference time of 80 ms causes short pauses between actions Despite delays, it handles complex scenarios well… solid for high-precision tasks, but slow to train. Gr00t ✅ Trains fast (~2 hours, ~$5 per run) ✅ Performs almost as well as Pi0 on large-object tasks ✅ Struggles with fine precision; random movement in some trials ✅ More training didn’t fix jitter or random offsets Best suited for tasks where exact precision isn’t critical. Not ready for manufacturing-grade accuracy without more tuning. Pi0 Fast ✅ Promised faster training, but results were underwhelming ✅ Training at 6 hours still showed low success rates ✅ Inference was slower than expected ✅ Not reliable for generalizing even slightly new tasks Currently too unstable for real-world deployment. Doesn’t live up to the “Fast” name yet. ACT (Baseline) ✅ 200MB model—lightweight, but limited ✅ Struggles with stacked objects or ambiguous scenes ✅ Success rates around 70% in best-case setups ✅ Can’t match newer models on precision or generalization Still a solid baseline, but clearly a generation behind in robustness. 🚨 Extra Notes All newer models share a common issue: •Inference takes longer than a frame (80 ms vs 33 ms), so robots “pause” between chunks. •This results in jittery movements, but not a dealbreaker unless tasks are time-sensitive. Language-conditioned tasks also fell short: after training on two labeled tasks, the model couldn’t generalize to a third unseen combination using only text prompts. ✅ The good news? These models adapt well to new robot arms with quick fine-tuning. ❌ The bad news? There’s still no plug-and-play solution for improving performance after deployment. Reinforcement learning or DAgger-style data collection during real-world operation may be the next big step, something many teams in robotics are actively working on.

Ilir Aliu

21,844 просмотров • 1 год назад

TEDRA Text-based Editing of Dynamic and Photoreal Actors discuss: Over the past years, significant progress has been made in creating photorealistic and drivable 3D avatars solely from videos of real humans. However, a core remaining challenge is the fine-grained and user-friendly editing of clothing styles by means of textual descriptions. To this end, we present TEDRA, the first method allowing text-based edits of an avatar, which maintains the avatar's high fidelity, space-time coherency, as well as dynamics, and enables skeletal pose and view control. We begin by training a model to create a controllable and high-fidelity digital replica of the real actor. Next, we personalize a pretrained generative diffusion model by fine-tuning it on various frames of the real character captured from different camera angles, ensuring the digital representation faithfully captures the dynamics and movements of the real person. This two-stage process lays the foundation for our approach to dynamic human avatar editing. Utilizing this personalized diffusion model, we modify the dynamic avatar based on a provided text prompt using our Personalized Normal Aligned Score Distillation Sampling (PNA-SDS) within a model-based guidance framework. Additionally, we propose a time step annealing strategy to ensure high-quality edits. Our results demonstrate a clear improvement over prior work in functionality and visual quality.
3:54

TEDRA Text-based Editing of Dynamic and Photoreal Actors discuss: Over the past years, significant progress has been made in creating photorealistic and drivable 3D avatars solely from videos of real humans. However, a core remaining challenge is the fine-grained and user-friendly editing of clothing styles by means of textual descriptions. To this end, we present TEDRA, the first method allowing text-based edits of an avatar, which maintains the avatar's high fidelity, space-time coherency, as well as dynamics, and enables skeletal pose and view control. We begin by training a model to create a controllable and high-fidelity digital replica of the real actor. Next, we personalize a pretrained generative diffusion model by fine-tuning it on various frames of the real character captured from different camera angles, ensuring the digital representation faithfully captures the dynamics and movements of the real person. This two-stage process lays the foundation for our approach to dynamic human avatar editing. Utilizing this personalized diffusion model, we modify the dynamic avatar based on a provided text prompt using our Personalized Normal Aligned Score Distillation Sampling (PNA-SDS) within a model-based guidance framework. Additionally, we propose a time step annealing strategy to ensure high-quality edits. Our results demonstrate a clear improvement over prior work in functionality and visual quality.

AK

17,302 просмотров • 1 год назад

NVIDIA just unleashed SANA-WM and it’s an absolute MONSTER for the future of open source AI! A blazing-fast 2.6B-parameter open-source world model that doesn’t just generate video… it creates controllable, physics-rich, high-fidelity worlds on demand. Why this is insanely powerful: • One image + text prompt + 6-DoF camera trajectory → generates 720p videos up to 60 seconds long with buttery-smooth, precisely controlled camera movement. You’re not just watching, you’re piloting the simulation. • Runs locally on a single consumer GPU (RTX 5090 level) thanks to heavy distillation + NVFP4 quantization. Full 60-second clip denoised in ~34 seconds. No massive clusters required. • 36× higher throughput than previous open models while rivaling (or beating) closed industrial giants in visual quality and consistency. • Trained lightning-fast: ~213K public videos in just 15 days on 64 H100s. • Built with next-level tech: Hybrid Linear Attention, dual-branch camera control, two-stage pipeline, and rock-solid metric-scale pose understanding. This is a true open world model, the foundation for embodied AI, robotics, autonomous systems, and hyper-realistic simulations that can run anywhere. Project: At our Zero-Human Company, we’re already running SANA-WM live in our core pipelines. It’s supercharging autonomous agent training, generating unlimited synthetic training data, and powering full end-to-end simulation loops, zero humans in the loop. The speed and control let us test thousands of edge-case scenarios overnight, iterate at lightspeed, and push our fully autonomous operations further than ever before. This is the kind of breakthrough that turns science fiction into daily reality. World models just leveled up — hard. The age of personal, local, controllable universes is here.
1:13

NVIDIA just unleashed SANA-WM and it’s an absolute MONSTER for the future of open source AI! A blazing-fast 2.6B-parameter open-source world model that doesn’t just generate video… it creates controllable, physics-rich, high-fidelity worlds on demand. Why this is insanely powerful: • One image + text prompt + 6-DoF camera trajectory → generates 720p videos up to 60 seconds long with buttery-smooth, precisely controlled camera movement. You’re not just watching, you’re piloting the simulation. • Runs locally on a single consumer GPU (RTX 5090 level) thanks to heavy distillation + NVFP4 quantization. Full 60-second clip denoised in ~34 seconds. No massive clusters required. • 36× higher throughput than previous open models while rivaling (or beating) closed industrial giants in visual quality and consistency. • Trained lightning-fast: ~213K public videos in just 15 days on 64 H100s. • Built with next-level tech: Hybrid Linear Attention, dual-branch camera control, two-stage pipeline, and rock-solid metric-scale pose understanding. This is a true open world model, the foundation for embodied AI, robotics, autonomous systems, and hyper-realistic simulations that can run anywhere. Project: At our Zero-Human Company, we’re already running SANA-WM live in our core pipelines. It’s supercharging autonomous agent training, generating unlimited synthetic training data, and powering full end-to-end simulation loops, zero humans in the loop. The speed and control let us test thousands of edge-case scenarios overnight, iterate at lightspeed, and push our fully autonomous operations further than ever before. This is the kind of breakthrough that turns science fiction into daily reality. World models just leveled up — hard. The age of personal, local, controllable universes is here.

Brian Roemmele

618,941 просмотров • 3 месяцев назад

Wow. Recreating the Shawshank Redemption prison in 3D from a single video, in real time (!) Just read the MASt3R-SLAM paper and it's pretty neat. These folks basically built a real-time dense SLAM system on top of MASt3R, which is a transformer-based neural network that can do 3d reconstruction and localization from uncalibrated image pairs. The cool part is they don't need a fixed camera model -- it just works with arbitrary cameras -- think different focal lengths, sensor sizes, even handling zooming in video (FMV drone video anyone?!). If you've done photogrammetry or played with NeRFs you know that is a HUGE deal. They've solved some tricky problems like efficient point matching and tracking, plus they've figured out how to fuse point clouds and handle loop closures in real-time. Their system runs at about 15 FPS on a 4090 and produces both camera poses and dense geometry. When they know the camera calibration, they get SOTA results across several benchmarks, but even without calibration, they still perform well. What's interesting is the approach -- most recent SLAM work has built on DROID-SLAM's architecture, but these folks went a different direction by leveraging a strong 3D reconstruction prior. Seems to give them more coherent geometry, which makes sense since that's what MASt3R was designed for. For anyone who cares about monocular SLAM and 3D reconstruction, this feels like a significant step toward plug-and-play dense SLAM without calibration headaches -- perfect for drones, robots, AR/VR -- the works!
0:29

Wow. Recreating the Shawshank Redemption prison in 3D from a single video, in real time (!) Just read the MASt3R-SLAM paper and it's pretty neat. These folks basically built a real-time dense SLAM system on top of MASt3R, which is a transformer-based neural network that can do 3d reconstruction and localization from uncalibrated image pairs. The cool part is they don't need a fixed camera model -- it just works with arbitrary cameras -- think different focal lengths, sensor sizes, even handling zooming in video (FMV drone video anyone?!). If you've done photogrammetry or played with NeRFs you know that is a HUGE deal. They've solved some tricky problems like efficient point matching and tracking, plus they've figured out how to fuse point clouds and handle loop closures in real-time. Their system runs at about 15 FPS on a 4090 and produces both camera poses and dense geometry. When they know the camera calibration, they get SOTA results across several benchmarks, but even without calibration, they still perform well. What's interesting is the approach -- most recent SLAM work has built on DROID-SLAM's architecture, but these folks went a different direction by leveraging a strong 3D reconstruction prior. Seems to give them more coherent geometry, which makes sense since that's what MASt3R was designed for. For anyone who cares about monocular SLAM and 3D reconstruction, this feels like a significant step toward plug-and-play dense SLAM without calibration headaches -- perfect for drones, robots, AR/VR -- the works!

Bilawal Sidhu

704,068 просмотров • 1 год назад

VoxCPM 2 just dropped by OpenBMB Only 2B-param open-source TTS (Text-to-Speech) model built for production-grade multilingual voice work. Apache-2.0 license, Can run on only 8GB VRAM. • Eliminates the "robotic" feel of traditional TTS, delivering prosody and emotional depth suitable for high-stakes professional environments like filmmaking, gaming, animation, and audiobooks. • 30-language multilingual: no language tag needed, just type in a supported language and generate directly. • Voice design: create a brand-new voice from a text description alone, like age, tone, pace, or emotion. No reference audio required. Describe the desired voice characteristics (gender, age, tone, emotion, pace …) in Control Instruction, and VoxCPM2 will craft a unique voice from your description alone. • Controllable cloning: clone from a short clip, then steer delivery style without losing the speaker’s core voice. • Ultimate cloning: use reference audio + transcript for continuation-style cloning that keeps the tiny vocal details. • 48kHz output: takes 16kHz reference audio and produces studio-quality speech without an external upsampler. • Real-time ready: around 0.3 RTF on RTX 4090, even lower with Nano-VLLM. • Commercial use: Apache-2.0 licensed. Developer-Friendly Infrastructure: - Native Torch Inference: Direct support for PyTorch-based workflows. - Training Flexibility: Supports both full-parameter and LoRA fine-tuning for specific domain adaptation. - Production Readiness: Compatible with voxcpm-nanovllm for large-scale, high-concurrency deployment.
1:58

VoxCPM 2 just dropped by OpenBMB Only 2B-param open-source TTS (Text-to-Speech) model built for production-grade multilingual voice work. Apache-2.0 license, Can run on only 8GB VRAM. • Eliminates the "robotic" feel of traditional TTS, delivering prosody and emotional depth suitable for high-stakes professional environments like filmmaking, gaming, animation, and audiobooks. • 30-language multilingual: no language tag needed, just type in a supported language and generate directly. • Voice design: create a brand-new voice from a text description alone, like age, tone, pace, or emotion. No reference audio required. Describe the desired voice characteristics (gender, age, tone, emotion, pace …) in Control Instruction, and VoxCPM2 will craft a unique voice from your description alone. • Controllable cloning: clone from a short clip, then steer delivery style without losing the speaker’s core voice. • Ultimate cloning: use reference audio + transcript for continuation-style cloning that keeps the tiny vocal details. • 48kHz output: takes 16kHz reference audio and produces studio-quality speech without an external upsampler. • Real-time ready: around 0.3 RTF on RTX 4090, even lower with Nano-VLLM. • Commercial use: Apache-2.0 licensed. Developer-Friendly Infrastructure: - Native Torch Inference: Direct support for PyTorch-based workflows. - Training Flexibility: Supports both full-parameter and LoRA fine-tuning for specific domain adaptation. - Production Readiness: Compatible with voxcpm-nanovllm for large-scale, high-concurrency deployment.

Rohan Paul

13,541 просмотров • 4 месяцев назад

FORGET MY PREVIOUS POST, LET ME BLOW YOUR MINDS! 🤯NEW METHOD! AI 2D SPRITES ARE SOLVED. PERIOD. Day 50 of building a webGL game engine using Cursor 🔽🔽🔽 I was wrong. The previous system had flaws. It wasn't the best. If you use it on VFX heavy sprites, it would start falling apart - either via grain/artifacts, or displacement. But I just couldn't give up. I knew I needed something more. This was the limit the models could give me. So I consulted science: Zheng et al., "Bilateral Reference for High-Resolution Dichotomous Image Segmentation" (CAAI AIR 2024, arXiv:2401.03407) - BiRefNet If any of the brilliant minds who wrote this paper read this post, I'm extremely thankful for your research. It made me solve a really big problem with 2D sprites, and I'm extremely grateful! Thank you! Please reach out if you see this! Now, how does this solution work: 1⃣ Generate your Kling animation (still king for anime 2D style). 2⃣ Run BiRefNet (HR-matting variant rocks) → get solid character alpha. 3⃣ Compute simple brightness/luminance alpha (luma → alpha curve, easy in any tool). 4⃣ Final alpha = max(BiRefNet_alpha, brightness_alpha) — that's it. No fancy weights needed. 5⃣ Feed to your engine / Comfy / whatever. Dynamic lights now play nice with zero artifacts. Mind you, this solution requires an nVidia GPU to run local inference. Alternatively, you can get it done on Colab likely with the free tier! You can try CPU, but no guarantees. Try a lighter model like BiRefNet_lite-2K. Demo is purposely in 30 fps even though I wouldn't run an anime 2D game above 12 fps - for both style and performance. Just wanted to show you the absolute results. Please bookmark this! ☑️ If you love 2D and sprite work, this is truly the best way, the best method I could find. Use it, steal it. Enjoy it! Repost if it saves your workflows too! I think I need a break... almost 40 hours straight solving this. Next are dynamic shadows, and I'm not looking forward to that...
0:40

FORGET MY PREVIOUS POST, LET ME BLOW YOUR MINDS! 🤯NEW METHOD! AI 2D SPRITES ARE SOLVED. PERIOD. Day 50 of building a webGL game engine using Cursor 🔽🔽🔽 I was wrong. The previous system had flaws. It wasn't the best. If you use it on VFX heavy sprites, it would start falling apart - either via grain/artifacts, or displacement. But I just couldn't give up. I knew I needed something more. This was the limit the models could give me. So I consulted science: Zheng et al., "Bilateral Reference for High-Resolution Dichotomous Image Segmentation" (CAAI AIR 2024, arXiv:2401.03407) - BiRefNet If any of the brilliant minds who wrote this paper read this post, I'm extremely thankful for your research. It made me solve a really big problem with 2D sprites, and I'm extremely grateful! Thank you! Please reach out if you see this! Now, how does this solution work: 1⃣ Generate your Kling animation (still king for anime 2D style). 2⃣ Run BiRefNet (HR-matting variant rocks) → get solid character alpha. 3⃣ Compute simple brightness/luminance alpha (luma → alpha curve, easy in any tool). 4⃣ Final alpha = max(BiRefNet_alpha, brightness_alpha) — that's it. No fancy weights needed. 5⃣ Feed to your engine / Comfy / whatever. Dynamic lights now play nice with zero artifacts. Mind you, this solution requires an nVidia GPU to run local inference. Alternatively, you can get it done on Colab likely with the free tier! You can try CPU, but no guarantees. Try a lighter model like BiRefNet_lite-2K. Demo is purposely in 30 fps even though I wouldn't run an anime 2D game above 12 fps - for both style and performance. Just wanted to show you the absolute results. Please bookmark this! ☑️ If you love 2D and sprite work, this is truly the best way, the best method I could find. Use it, steal it. Enjoy it! Repost if it saves your workflows too! I think I need a break... almost 40 hours straight solving this. Next are dynamic shadows, and I'm not looking forward to that...

Startracker 🔺

71,439 просмотров • 6 месяцев назад

I’m thrilled to announce that we just released GraspGen, a multi-year project we have been cooking at NVIDIA Robotics 🚀 GraspGen: A Diffusion-Based Framework for 6-DOF Grasping Grasping is a foundational challenge in robotics 🤖 — whether for industrial picking or general-purpose humanoids. VLA + real data collection is all the rage now but is expensive and scales poorly for this task. For every new gripper and/or scene, you’ll have to recollect the dataset in this paradigm for the best perf. 💡Key Idea: Since grasping is such a well-defined task in simulation - why can’t we just scale synthetic data generation and train a generative model for grasping? By embracing modularity and standardized grasp formats, we can make this a turnkey technology that works zero-shot for multiple settings. GraspGen is a modular framework for diffusion-based 6-DOF grasp generation that scales across embodiment types, observability conditions, clutter, task complexity. Key Features: ✅ Multi-embodiment support: suction, parallel-jaw, and multi-fingered grippers ✅ Generalization to partial + complete 3D point clouds ✅ Generalization to single-objects + cluttered scenes ✅ Modular design uses other robotics modules and foundation models (SAM2, cuRobo, FoundationStereo, FoundationPose). This allows GraspGen to focus on only one thing - grasp generation ✅ Training recipe: grasp discriminator is trained with On-Generator data from the diffusion model - so that it learns to correct the mistakes (if any) of the diffusion generator ✅ Real-time performance (~20 Hz) before any GPU acceleration; low memory footprint 📊 Results: • SOTA on the FetchBench [Han et al. CoRL 2024] benchmark • Zero-shot sim-to-real transfer on unknown objects and cluttered scenes • Dataset of 53M simulated grasps across 8K objects from Objaverse 📄 arXiv: 🌐 Website: 💻 Code: A huge thank you to everyone involved in this journey — excited to see what the community builds on top of it! Joint work with Clemens Eppner , Balakumar Sundaralingam , Yu-Wei, Jun Yamada Wentao Yuan and other collaborators #robotics #diffusionmodels #physicalAI #simtoreal
0:39

I’m thrilled to announce that we just released GraspGen, a multi-year project we have been cooking at NVIDIA Robotics 🚀 GraspGen: A Diffusion-Based Framework for 6-DOF Grasping Grasping is a foundational challenge in robotics 🤖 — whether for industrial picking or general-purpose humanoids. VLA + real data collection is all the rage now but is expensive and scales poorly for this task. For every new gripper and/or scene, you’ll have to recollect the dataset in this paradigm for the best perf. 💡Key Idea: Since grasping is such a well-defined task in simulation - why can’t we just scale synthetic data generation and train a generative model for grasping? By embracing modularity and standardized grasp formats, we can make this a turnkey technology that works zero-shot for multiple settings. GraspGen is a modular framework for diffusion-based 6-DOF grasp generation that scales across embodiment types, observability conditions, clutter, task complexity. Key Features: ✅ Multi-embodiment support: suction, parallel-jaw, and multi-fingered grippers ✅ Generalization to partial + complete 3D point clouds ✅ Generalization to single-objects + cluttered scenes ✅ Modular design uses other robotics modules and foundation models (SAM2, cuRobo, FoundationStereo, FoundationPose). This allows GraspGen to focus on only one thing - grasp generation ✅ Training recipe: grasp discriminator is trained with On-Generator data from the diffusion model - so that it learns to correct the mistakes (if any) of the diffusion generator ✅ Real-time performance (~20 Hz) before any GPU acceleration; low memory footprint 📊 Results: • SOTA on the FetchBench [Han et al. CoRL 2024] benchmark • Zero-shot sim-to-real transfer on unknown objects and cluttered scenes • Dataset of 53M simulated grasps across 8K objects from Objaverse 📄 arXiv: 🌐 Website: 💻 Code: A huge thank you to everyone involved in this journey — excited to see what the community builds on top of it! Joint work with Clemens Eppner , Balakumar Sundaralingam , Yu-Wei, Jun Yamada Wentao Yuan and other collaborators #robotics #diffusionmodels #physicalAI #simtoreal

Adithya Murali

24,106 просмотров • 1 год назад

Loved this 22-minute talk on continual learning for AI agents. Must watch for anyone looking to get agents performant and into production. Credit: Soheil Feizi at AI Engineer • Agent learning can happen at three layers: the model (weights), the harness (prompts, tools, skills, code, workflows), and memory (session or persistent). • Two fundamental challenges: (1) getting feedback, meaning how do we know if the agent did well and what it should have done instead, and (2) acting on that feedback, meaning deciding which layer or component to change and how. • Feedback sources differ by stage: In development you have benchmarks with evaluators that score pass/fail. In production you only have logs, which can be judged either automatically (LLMs or code analyzing the log, which is scalable) or by human experts (low volume but critical domain knowledge). • Logs plus feedback aren't enough because they're not testable: A single log with feedback is one observation of what happened. You need to lift it into a replayable learning environment, a simulation with tools, users, and defined evaluators, so candidate fixes can be run, verified, and compared. • Three ways to optimize the agent, with tradeoffs: Model-layer updates (SFT, RL post-training like DPO/GRPO, LoRA) are expensive and need benchmarks and evaluators. Harness updates (trace-to-harness coding agents, prompt search like GEPA) are flexible but either untestable and "vibe-based" or benchmark-dependent. Memory updates (fact storage like Letta/Mem0, skill distillation) are cheapest and fastest but usually unverified. • A good learning engine makes "the smallest durable change at the right layer" of the agent. • Verifiable continual learning (VCL): Improve an agent from its own experience where every fix is proven to help and proven to break nothing that already worked. It requires an executable test (replayable failure), a measured delta (score before and after), and regression tests (prior tests still pass). • Four principles of practical VCL: Replayability (turn one-off failures into rerunnable tests), holisticness (one failure can have causes in memory, prompts, tools, workflow, or model, so route the fix to the right layer), lifelongness (fix new failures subject to no regression on past environments, with regression handled inside the optimization loop rather than post-hoc), and efficiency (the loop must run frequently and cheaply, without scaling linearly as past environments accumulate). • Three takeaways: (1) Agent continual learning isn't necessarily fine-tuning; many useful updates live in the harness and memory layers. (2) Production logs are not learning environments and must be transformed into replayable ones. (3) The frontier is regression-aware improvement: fixing new failures while verifying you don't break old ones.
22:34

Loved this 22-minute talk on continual learning for AI agents. Must watch for anyone looking to get agents performant and into production. Credit: Soheil Feizi at AI Engineer • Agent learning can happen at three layers: the model (weights), the harness (prompts, tools, skills, code, workflows), and memory (session or persistent). • Two fundamental challenges: (1) getting feedback, meaning how do we know if the agent did well and what it should have done instead, and (2) acting on that feedback, meaning deciding which layer or component to change and how. • Feedback sources differ by stage: In development you have benchmarks with evaluators that score pass/fail. In production you only have logs, which can be judged either automatically (LLMs or code analyzing the log, which is scalable) or by human experts (low volume but critical domain knowledge). • Logs plus feedback aren't enough because they're not testable: A single log with feedback is one observation of what happened. You need to lift it into a replayable learning environment, a simulation with tools, users, and defined evaluators, so candidate fixes can be run, verified, and compared. • Three ways to optimize the agent, with tradeoffs: Model-layer updates (SFT, RL post-training like DPO/GRPO, LoRA) are expensive and need benchmarks and evaluators. Harness updates (trace-to-harness coding agents, prompt search like GEPA) are flexible but either untestable and "vibe-based" or benchmark-dependent. Memory updates (fact storage like Letta/Mem0, skill distillation) are cheapest and fastest but usually unverified. • A good learning engine makes "the smallest durable change at the right layer" of the agent. • Verifiable continual learning (VCL): Improve an agent from its own experience where every fix is proven to help and proven to break nothing that already worked. It requires an executable test (replayable failure), a measured delta (score before and after), and regression tests (prior tests still pass). • Four principles of practical VCL: Replayability (turn one-off failures into rerunnable tests), holisticness (one failure can have causes in memory, prompts, tools, workflow, or model, so route the fix to the right layer), lifelongness (fix new failures subject to no regression on past environments, with regression handled inside the optimization loop rather than post-hoc), and efficiency (the loop must run frequently and cheaply, without scaling linearly as past environments accumulate). • Three takeaways: (1) Agent continual learning isn't necessarily fine-tuning; many useful updates live in the harness and memory layers. (2) Production logs are not learning environments and must be transformed into replayable ones. (3) The frontier is regression-aware improvement: fixing new failures while verifying you don't break old ones.

Alex Lieberman

20,085 просмотров • 1 месяц назад

Introduction to Artificial Intelligence 101 For Jonathan Greenblatt I am not an expert on artificial intelligence or machine learning by any stretch of the imagination, but I am familiar with the very basic principles behind how these systems are built and developed. Two things to consider on AI bias: The responses generated by an AI system after being prompted reflect the corpus of data on which it was trained and the objectives used during its fine-tuning phase. (This is intended only as a high-level overview.) First, think of an AI model as a kind of blank slate, there is a pre-training stage that uses extremely large amounts of data, often measured in petabytes. This represents vast quantities of text, far more than any individual, like you or I, could read in a lifetime. After this initial training phase, developers conduct a fine-tuning process designed to shape how the system formulates its responses. While this description is over simplified, it captures the general idea behind how modern natural language processing AI models are created. If an AI model is trained and refined using material that reflects a particular viewpoint or advocacy perspective like Hasbara 8th-Front propaganda, its outputs will tend to reproduce that very subjective perspective. Conversely, if the training and fine-tuning emphasize a broader range of sources and methodologies, the resulting outputs will reflect that wider balance of information. Most people don't want to be brainwashed by the ADL, SPLC, Bnai Brith or other related activist groups. I personally would prefer to have a broader range of viewpoints so that they can each be individually fact checked based on past and current events in our reality. Hopefully nobody else wants to be told what to think, but will be given the option to explore perspectives with a skeptical, trust but verify attitude. My goal in studying the Leo Frank case for the past 57 years has been to learn both the prosecution and defense side with equal depth. If the goal is to encourage understanding and inquiry, it is important to present a full spectrum of documented perspectives so that readers and researchers can evaluate the material and reach their own logical and fair-minded conclusions based on evidence and reason. Shoving hasbara 8th-Front propaganda down the throats of younger folks is only going to cause them to explore other AI models and viewpoints more intensely. This younger generation is mentally vaccinated by nature from your organizations propaganda (and the propaganda of your fellow travelers), they have developed a herd immunity. And you are going to start seeing the results of that once your Golem army of baby boomers start passing away. We can already see that most AI models are Hasbara 8th-Front biased, we prompt them everyday, we are well aware, which means we will start to pursue other options. You might be able to manipulate the leading models with Hasbara 8th-Front, but you won't manipulate them all. People are naturally going to gravitate toward objective and truth seeking AI models. I pray that Elon Musk and XAI will pursue objective truth seeking with a burning fanaticism. When your cohorts forced the sale of TikTok USA, multitudes simply moved on to other social media. In other words, or to summarize: AI systems do not originate positions independently. They reflect the structure, assumptions, and emphases present in their training environment as designed by their developers. For those who are interested in genuine truth-seeking applications of AI, the expectation is that such systems strive for breadth of sourcing, transparency of method, share a wide variety of viewpoints, and openness to examination. Many observers have raised questions about how leading AI platforms handle controversial historical topics, including discussion of the Leo Frank case and the rape-murder of little Mary Phagan, and have called for continued improvement in neutrality, documentation, and analytical rigor. This issue is not going away and there is a world wide growing consensus against Zionist propaganda. The Palestinian Holocaust is not going away, it was live streamed genocide and petabytes of footage have been captured of its atrocities, these crimes against humanity and war crimes will be talked about everyday for the next millennia. The hope is that advancing technology will support open inquiry, truth seeking, logic, reason, objectivity, empiricism, the scientific method, responsible scholarship, and respectful dialogue, allowing people to review historical records carefully and draw thoughtful, fair-minded conclusions. We the people want Truth seeking AI. We want the objective truth (which is a process, not a destination), not the subjective truth of a group of people with burning anti-Christian fanaticism (many read the website) or anti-Gentilism in their hearts. You have some soul searching to do Jonathan Greenblatt, and I will continue to pray for you. I pray to God that you will give up your anti-Gentile, anti-American, treacherous, wicked and control freak ways, and allow the human race to review all perspectives so they might draw their own logical, and fair-minded conclusions. People reject all calls to be brainwashed. Most people want to learn how to sift the evidence, not brainwashed into thinking a certain way. In the first edition of my book from the 1980s I presented both the prosecution and defense perspectives. I'm Praying that you will turn away from the tribal extremist darkness in your soul and stop spreading false accusations that Leo Frank's trial was driven by anti-Semitism or that he was wrongfully convicted. Or the disingenuous claims you make every year on August 17, that Leo Frank was extra-judicially hanged because of his religion, when the real reason was that he was a convicted murderer, pervert, and a homicidal rapist-pedophile, who brutal beat, sexually assaulted, and strangled my great aunt, little Mary Anne Phagan to death. Leo Frank was not lynched because of anti-religious bigotry, but because he was a lethal child molester. Leo Frank also tried to racistly frame two black employees for the crime. I also pray that executives of the leading AI models will put a stop to spreading false information about the Mary Phagan case, it is a great injustice when their AI models falsify the history of the case, especially the facts of the Leo Frank trial which was extremely well documented. In the mean time, I politely encourage all members and supporters of the ADL Anti-Defamation League of B'nai B'rith to acquire a freshly printed book of my new 2025 revised edition of The Murder of Little Mary Phagan. It will provide you with insights you won't learn from the pro-Frank literature. See my pinned-post for more information.
2:57

Introduction to Artificial Intelligence 101 For Jonathan Greenblatt I am not an expert on artificial intelligence or machine learning by any stretch of the imagination, but I am familiar with the very basic principles behind how these systems are built and developed. Two things to consider on AI bias: The responses generated by an AI system after being prompted reflect the corpus of data on which it was trained and the objectives used during its fine-tuning phase. (This is intended only as a high-level overview.) First, think of an AI model as a kind of blank slate, there is a pre-training stage that uses extremely large amounts of data, often measured in petabytes. This represents vast quantities of text, far more than any individual, like you or I, could read in a lifetime. After this initial training phase, developers conduct a fine-tuning process designed to shape how the system formulates its responses. While this description is over simplified, it captures the general idea behind how modern natural language processing AI models are created. If an AI model is trained and refined using material that reflects a particular viewpoint or advocacy perspective like Hasbara 8th-Front propaganda, its outputs will tend to reproduce that very subjective perspective. Conversely, if the training and fine-tuning emphasize a broader range of sources and methodologies, the resulting outputs will reflect that wider balance of information. Most people don't want to be brainwashed by the ADL, SPLC, Bnai Brith or other related activist groups. I personally would prefer to have a broader range of viewpoints so that they can each be individually fact checked based on past and current events in our reality. Hopefully nobody else wants to be told what to think, but will be given the option to explore perspectives with a skeptical, trust but verify attitude. My goal in studying the Leo Frank case for the past 57 years has been to learn both the prosecution and defense side with equal depth. If the goal is to encourage understanding and inquiry, it is important to present a full spectrum of documented perspectives so that readers and researchers can evaluate the material and reach their own logical and fair-minded conclusions based on evidence and reason. Shoving hasbara 8th-Front propaganda down the throats of younger folks is only going to cause them to explore other AI models and viewpoints more intensely. This younger generation is mentally vaccinated by nature from your organizations propaganda (and the propaganda of your fellow travelers), they have developed a herd immunity. And you are going to start seeing the results of that once your Golem army of baby boomers start passing away. We can already see that most AI models are Hasbara 8th-Front biased, we prompt them everyday, we are well aware, which means we will start to pursue other options. You might be able to manipulate the leading models with Hasbara 8th-Front, but you won't manipulate them all. People are naturally going to gravitate toward objective and truth seeking AI models. I pray that Elon Musk and XAI will pursue objective truth seeking with a burning fanaticism. When your cohorts forced the sale of TikTok USA, multitudes simply moved on to other social media. In other words, or to summarize: AI systems do not originate positions independently. They reflect the structure, assumptions, and emphases present in their training environment as designed by their developers. For those who are interested in genuine truth-seeking applications of AI, the expectation is that such systems strive for breadth of sourcing, transparency of method, share a wide variety of viewpoints, and openness to examination. Many observers have raised questions about how leading AI platforms handle controversial historical topics, including discussion of the Leo Frank case and the rape-murder of little Mary Phagan, and have called for continued improvement in neutrality, documentation, and analytical rigor. This issue is not going away and there is a world wide growing consensus against Zionist propaganda. The Palestinian Holocaust is not going away, it was live streamed genocide and petabytes of footage have been captured of its atrocities, these crimes against humanity and war crimes will be talked about everyday for the next millennia. The hope is that advancing technology will support open inquiry, truth seeking, logic, reason, objectivity, empiricism, the scientific method, responsible scholarship, and respectful dialogue, allowing people to review historical records carefully and draw thoughtful, fair-minded conclusions. We the people want Truth seeking AI. We want the objective truth (which is a process, not a destination), not the subjective truth of a group of people with burning anti-Christian fanaticism (many read the website) or anti-Gentilism in their hearts. You have some soul searching to do Jonathan Greenblatt, and I will continue to pray for you. I pray to God that you will give up your anti-Gentile, anti-American, treacherous, wicked and control freak ways, and allow the human race to review all perspectives so they might draw their own logical, and fair-minded conclusions. People reject all calls to be brainwashed. Most people want to learn how to sift the evidence, not brainwashed into thinking a certain way. In the first edition of my book from the 1980s I presented both the prosecution and defense perspectives. I'm Praying that you will turn away from the tribal extremist darkness in your soul and stop spreading false accusations that Leo Frank's trial was driven by anti-Semitism or that he was wrongfully convicted. Or the disingenuous claims you make every year on August 17, that Leo Frank was extra-judicially hanged because of his religion, when the real reason was that he was a convicted murderer, pervert, and a homicidal rapist-pedophile, who brutal beat, sexually assaulted, and strangled my great aunt, little Mary Anne Phagan to death. Leo Frank was not lynched because of anti-religious bigotry, but because he was a lethal child molester. Leo Frank also tried to racistly frame two black employees for the crime. I also pray that executives of the leading AI models will put a stop to spreading false information about the Mary Phagan case, it is a great injustice when their AI models falsify the history of the case, especially the facts of the Leo Frank trial which was extremely well documented. In the mean time, I politely encourage all members and supporters of the ADL Anti-Defamation League of B'nai B'rith to acquire a freshly printed book of my new 2025 revised edition of The Murder of Little Mary Phagan. It will provide you with insights you won't learn from the pro-Frank literature. See my pinned-post for more information.

Mary Phagan-Kean

21,628 просмотров • 6 месяцев назад

Announcing DreamDojo: our open-source, interactive world model that takes robot motor controls and generates the future in pixels. No engine, no meshes, no hand-authored dynamics. It's Simulation 2.0. Time for robotics to take the bitter lesson pill. Real-world robot learning is bottlenecked by time, wear, safety, and resets. If we want Physical AI to move at pretraining speed, we need a simulator that adapts to pretraining scale with as little human engineering as possible. Our key insights: (1) human egocentric videos are a scalable source of first-person physics; (2) latent actions make them "robot-readable" across different hardware; (3) real-time inference unlocks live teleop, policy eval, and test-time planning *inside* a dream. We pre-train on 44K hours of human videos: cheap, abundant, and collected with zero robot-in-the-loop. Humans have already explored the combinatorics: we grasp, pour, fold, assemble, fail, retry—across cluttered scenes, shifting viewpoints, changing light, and hour-long task chains—at a scale no robot fleet could match. The missing piece: these videos have no action labels. So we introduce latent actions: a unified representation inferred directly from videos that captures "what changed between world states" without knowing the underlying hardware. This lets us train on any first-person video as if it came with motor commands attached. As a result, DreamDojo generalizes zero-shot to objects and environments never seen in any robot training set, because humans saw them first. Next, we post-train onto each robot to fit its specific hardware. Think of it as separating "how the world looks and behaves" from "how this particular robot actuates." The base model follows the general physical rules, then "snaps onto" the robot's unique mechanics. It's kind of like loading a new character and scene assets into Unreal Engine, but done through gradient descent and generalizes far beyond the post-training dataset. A world simulator is only useful if it runs fast enough to close the loop. We train a real-time version of DreamDojo that runs at 10 FPS, stable for over a minute of continuous rollout. This unlocks exciting possibilities: - Live teleoperation *inside* a dream. Connect a VR controller, stream actions into DreamDojo, and teleop a virtual robot in real time. We demo this on Unitree G1 with a PICO headset and one RTX 5090. - Policy evaluation. You can benchmark a policy checkpoint in DreamDojo instead of the real world. The simulated success rates strongly correlate with real-world results - accurate enough to rank checkpoints without burning a single motor. - Model-based planning. Sample multiple action proposals → simulate them all in parallel → pick the best future. Gains +17% real-world success out of the box on a fruit packing task. We open-source everything!! Weights, code, post-training dataset, eval set, and whitepaper with tons of details to reproduce. DreamDojo is based on NVIDIA Cosmos, which is open-weight too. 2026 is the year of World Models for physical AI. We want you to build with us. Happy scaling! Links in thread:
1:17

Announcing DreamDojo: our open-source, interactive world model that takes robot motor controls and generates the future in pixels. No engine, no meshes, no hand-authored dynamics. It's Simulation 2.0. Time for robotics to take the bitter lesson pill. Real-world robot learning is bottlenecked by time, wear, safety, and resets. If we want Physical AI to move at pretraining speed, we need a simulator that adapts to pretraining scale with as little human engineering as possible. Our key insights: (1) human egocentric videos are a scalable source of first-person physics; (2) latent actions make them "robot-readable" across different hardware; (3) real-time inference unlocks live teleop, policy eval, and test-time planning *inside* a dream. We pre-train on 44K hours of human videos: cheap, abundant, and collected with zero robot-in-the-loop. Humans have already explored the combinatorics: we grasp, pour, fold, assemble, fail, retry—across cluttered scenes, shifting viewpoints, changing light, and hour-long task chains—at a scale no robot fleet could match. The missing piece: these videos have no action labels. So we introduce latent actions: a unified representation inferred directly from videos that captures "what changed between world states" without knowing the underlying hardware. This lets us train on any first-person video as if it came with motor commands attached. As a result, DreamDojo generalizes zero-shot to objects and environments never seen in any robot training set, because humans saw them first. Next, we post-train onto each robot to fit its specific hardware. Think of it as separating "how the world looks and behaves" from "how this particular robot actuates." The base model follows the general physical rules, then "snaps onto" the robot's unique mechanics. It's kind of like loading a new character and scene assets into Unreal Engine, but done through gradient descent and generalizes far beyond the post-training dataset. A world simulator is only useful if it runs fast enough to close the loop. We train a real-time version of DreamDojo that runs at 10 FPS, stable for over a minute of continuous rollout. This unlocks exciting possibilities: - Live teleoperation *inside* a dream. Connect a VR controller, stream actions into DreamDojo, and teleop a virtual robot in real time. We demo this on Unitree G1 with a PICO headset and one RTX 5090. - Policy evaluation. You can benchmark a policy checkpoint in DreamDojo instead of the real world. The simulated success rates strongly correlate with real-world results - accurate enough to rank checkpoints without burning a single motor. - Model-based planning. Sample multiple action proposals → simulate them all in parallel → pick the best future. Gains +17% real-world success out of the box on a fruit packing task. We open-source everything!! Weights, code, post-training dataset, eval set, and whitepaper with tons of details to reproduce. DreamDojo is based on NVIDIA Cosmos, which is open-weight too. 2026 is the year of World Models for physical AI. We want you to build with us. Happy scaling! Links in thread:

Jim Fan

227,026 просмотров • 5 месяцев назад

The U.S. MUST win the AI race We’ve implemented a clear policy at micro1: we will only work with U.S. AI labs and its allies. We made this decision because the AI race is not just about better products. It is about who controls the intelligence layer of the global economy, and whether frontier capability is used to strengthen the free world or to empower adversarial states. AI will be the most important technology of our lifetime. In the fullness of time, it will automate most functions across the economy. Not just software tasks, but coordination, production, logistics, judgment, and execution. As those functions are automated, human time is freed up to invent new ones. Those new functions then become candidates for automation themselves. This loop compounds. As this trajectory continues, output per worker increases dramatically. Entire categories of work become cheaper and faster to perform. Manufacturing reshoring becomes economically viable not because of policy intervention, but because intelligent systems operated domestically outperform global labor arbitrage. Goods and services trend toward lower marginal cost, while distribution improves through better coordination of supply and demand. That is the upside. However, this is impossible without deep integration of intelligent systems. For AI to meaningfully automate real-world functions inside enterprises or governments, it needs full context of any given enterprise. That means read and write access to its core databases. There is no credible path to automating high-impact functions without granting frontier systems that level of access. If the United States does not win the AI race, enterprises eventually face a constrained choice. Either grant that access to Chinese models controlled by an adversarial government, or rely on sub-optimal intelligence to automate functions that still must be automated. Both outcomes are not acceptable. And ultimately, this becomes the greatest national security risk the United States has ever faced. AI models are trained by humans. The judgment embedded in pre-training data and especially in expert post-training data largely determines how a model behaves. While emergent behavior exists, a useful approximation is that a model reflects the weighted aggregate of the human judgment distilled into it. Assisting foreign actors—who will naturally prioritize expert tasks aligned with their own interests—to dominate data creation embeds those interests directly into the intelligence layer itself. Once encoded at scale, these interests propagate through every downstream applications that relies on that intelligence. Here’s how we win. First, leverage is in software. China is ahead in hardware for physically intelligent systems. Catching up there is a long and difficult battle. Software, both large language models and robotics models, remains the bottleneck. Advancing the brain (AI models) is the fastest way to increase the usefulness of existing hardware and deployed systems. Second, the U.S. must 100x its investment in structured human judgment. Continued investment in compute and algorithmic efficiency is critical. But that investment is ultimately a bet on very high future inference demand. For that bet to pay off, models must unlock many new capabilities, and in practice the only way to unlock those capabilities is through expert human data. Historically, experts like doctors and lawyers were never incentivized to produce high-quality reasoning data in a machine-verifiable format. There was no reason for a doctor to generate precise, structured simulations of patient interactions, diagnostic reasoning, or treatment tradeoffs. There was no reason for a lawyer to document complex legal reasoning paths in a way that could be programmatically evaluated. AI systems now require exactly this kind of data. The incentive finally exists because this data directly improves systems that operate at massive scale, and experts can be paid well to produce it. Once expert judgment is encoded into models in a structured, verifiable way, it compounds. Those who delay do not just lose time. They lose the ability to catch up. Third, distillation from Chinese labs must be stopped. AI labs must do everything they can to prevent Chinese labs and models from distilling frontier models. Simply calling frontier APIs, or even interacting through UIs, lets Chinese model companies rapidly generate high-quality supervised fine-tuning datasets and close the gap at a fraction of the cost. This method does not put you at the frontier, but it does let you catch up quickly, which is what we saw with DeepSeek. The West significantly overreacted to DeepSeek’s headline capabilities, but underreacted to the underlying dynamic: frontier access itself becomes a training set at a fraction of the cost. Human data platforms also have a duty to help prevent this distillation. Lastly, the U.S.government should set the standard for AI Evaluation that leads to real production usage. AI agents are under-deployed relative to what the technology allows because they are probabilistic systems that require a fundamentally different QA approach than deterministic software. Generic QA is insufficient; safely shipping agents requires explicit evaluation frameworks that assess their full action space. Organizations must clearly define which functions an agent is allowed to perform, how quality is measured for each function, and which domain experts are qualified to judge outcomes. With these frameworks in place, agents can be rigorously tested using structured human data, deployed to production with confidence, and continuously improved over time. The U.S. government should be the first large enterprise to implement rigorous evaluation systems across every function. If the government leads on evaluation-driven deployment, adoption across the private sector accelerates naturally. This is how American workers become more powerful. Each worker operates digital or physical agents that expand their effective output. Recruiting, manufacturing, logistics, and other domains shift toward human judgment overseeing autonomous execution. Reshoring occurs because it becomes economically rational. Work becomes more meaningful. This is a race to determine who controls the intelligence layer of the global economy. And that must be us. 🇺🇸
0:48

The U.S. MUST win the AI race We’ve implemented a clear policy at micro1: we will only work with U.S. AI labs and its allies. We made this decision because the AI race is not just about better products. It is about who controls the intelligence layer of the global economy, and whether frontier capability is used to strengthen the free world or to empower adversarial states. AI will be the most important technology of our lifetime. In the fullness of time, it will automate most functions across the economy. Not just software tasks, but coordination, production, logistics, judgment, and execution. As those functions are automated, human time is freed up to invent new ones. Those new functions then become candidates for automation themselves. This loop compounds. As this trajectory continues, output per worker increases dramatically. Entire categories of work become cheaper and faster to perform. Manufacturing reshoring becomes economically viable not because of policy intervention, but because intelligent systems operated domestically outperform global labor arbitrage. Goods and services trend toward lower marginal cost, while distribution improves through better coordination of supply and demand. That is the upside. However, this is impossible without deep integration of intelligent systems. For AI to meaningfully automate real-world functions inside enterprises or governments, it needs full context of any given enterprise. That means read and write access to its core databases. There is no credible path to automating high-impact functions without granting frontier systems that level of access. If the United States does not win the AI race, enterprises eventually face a constrained choice. Either grant that access to Chinese models controlled by an adversarial government, or rely on sub-optimal intelligence to automate functions that still must be automated. Both outcomes are not acceptable. And ultimately, this becomes the greatest national security risk the United States has ever faced. AI models are trained by humans. The judgment embedded in pre-training data and especially in expert post-training data largely determines how a model behaves. While emergent behavior exists, a useful approximation is that a model reflects the weighted aggregate of the human judgment distilled into it. Assisting foreign actors—who will naturally prioritize expert tasks aligned with their own interests—to dominate data creation embeds those interests directly into the intelligence layer itself. Once encoded at scale, these interests propagate through every downstream applications that relies on that intelligence. Here’s how we win. First, leverage is in software. China is ahead in hardware for physically intelligent systems. Catching up there is a long and difficult battle. Software, both large language models and robotics models, remains the bottleneck. Advancing the brain (AI models) is the fastest way to increase the usefulness of existing hardware and deployed systems. Second, the U.S. must 100x its investment in structured human judgment. Continued investment in compute and algorithmic efficiency is critical. But that investment is ultimately a bet on very high future inference demand. For that bet to pay off, models must unlock many new capabilities, and in practice the only way to unlock those capabilities is through expert human data. Historically, experts like doctors and lawyers were never incentivized to produce high-quality reasoning data in a machine-verifiable format. There was no reason for a doctor to generate precise, structured simulations of patient interactions, diagnostic reasoning, or treatment tradeoffs. There was no reason for a lawyer to document complex legal reasoning paths in a way that could be programmatically evaluated. AI systems now require exactly this kind of data. The incentive finally exists because this data directly improves systems that operate at massive scale, and experts can be paid well to produce it. Once expert judgment is encoded into models in a structured, verifiable way, it compounds. Those who delay do not just lose time. They lose the ability to catch up. Third, distillation from Chinese labs must be stopped. AI labs must do everything they can to prevent Chinese labs and models from distilling frontier models. Simply calling frontier APIs, or even interacting through UIs, lets Chinese model companies rapidly generate high-quality supervised fine-tuning datasets and close the gap at a fraction of the cost. This method does not put you at the frontier, but it does let you catch up quickly, which is what we saw with DeepSeek. The West significantly overreacted to DeepSeek’s headline capabilities, but underreacted to the underlying dynamic: frontier access itself becomes a training set at a fraction of the cost. Human data platforms also have a duty to help prevent this distillation. Lastly, the U.S.government should set the standard for AI Evaluation that leads to real production usage. AI agents are under-deployed relative to what the technology allows because they are probabilistic systems that require a fundamentally different QA approach than deterministic software. Generic QA is insufficient; safely shipping agents requires explicit evaluation frameworks that assess their full action space. Organizations must clearly define which functions an agent is allowed to perform, how quality is measured for each function, and which domain experts are qualified to judge outcomes. With these frameworks in place, agents can be rigorously tested using structured human data, deployed to production with confidence, and continuously improved over time. The U.S. government should be the first large enterprise to implement rigorous evaluation systems across every function. If the government leads on evaluation-driven deployment, adoption across the private sector accelerates naturally. This is how American workers become more powerful. Each worker operates digital or physical agents that expand their effective output. Recruiting, manufacturing, logistics, and other domains shift toward human judgment overseeing autonomous execution. Reshoring occurs because it becomes economically rational. Work becomes more meaningful. This is a race to determine who controls the intelligence layer of the global economy. And that must be us. 🇺🇸

Ali Ansari

396,355 просмотров • 6 месяцев назад

$AMD $620/share is too conservative for 2026 🧵 Some quick facts before I dive into this super long thread: $META allocated 42% GPUs to $AMD and 58% to $NVDA OpenAI allocated 6GW(38%) to $AMD and 10GW to $NVDA My $620 PT below by end of 2026 was only for 10-15% market share. I believe $AMD is going to have much much higher market share than I projected. The AI accelerator market is exploding, projected to reach $500 billion by 2028(is now heading $1Tril), driven by insatiable demand for training and inference compute in large language models (LLMs), recommendation systems, and autonomous systems. Nvidia ($NVDA) has long held a stranglehold, commanding over 90% market share through its CUDA ecosystem and superior rack-scale solutions. However, AMD is mounting a formidable challenge, leveraging cost advantages, open-source software momentum, and hyperscaler partnerships to erode Nvidia's moat. Recent deals—such as Meta's ($META) allocation of 42% of its GPU capacity to AMD and OpenAI's commitment to 6GW of AMD compute (versus 10GW for Nvidia)—signal a tipping point. At the forefront is AMD's Instinct MI450 series, a next-generation AI GPU slated for H2 2026 launch, which promises "no-excuses" leadership in training, inference, and distributed workloads. This analysis dissects how AMD will capture more market share and why hyperscalers like $Meta , xAI , Oracle , and others are poised to become voracious buyers of the MI450. AMD's AI GPU revenue has surged from negligible levels in 2022 to an estimated $4-5 billion in 2025, capturing ~6% of the data center GPU market. This growth stems from the Instinct MI300X, which offers 141GB of HBM3 memory and competitive FP8/FP16 performance at 20-30% lower cost than Nvidia's H100. Hyperscalers, facing NVIDIA 's overcharging, have turned to AMD for diversification. Meta, for instance, plans 600,000 H100-equivalent GPUs by end-2024, with ~42% (or 250,000+ units) sourced from AMD's MI300 series for inference tasks like image editing and AI assistants. Similarly, OpenAI's recent multi-year deal commits to 6GW of AMD compute—equivalent to ~300,000-400,000 MI450 GPUs—starting with 1GW in 2026, explicitly to counterbalance its 10GW Nvidia allocation. These aren't one-offs. Microsoft Azure, Amazon AWS, and Oracle Cloud Infrastructure (OCI) have integrated MI300X for AI workloads, with Oracle deploying 30,000 MI355X units in zettascale clusters. xAI, Elon Musk Musk's AI venture, ran 30% of Grok-1's production traffic on MI300X GPUs and has confirmed ongoing purchases. Collectively, these partners represent over $400 billion in projected AI infrastructure spend through 2028, with AMD targeting up to 40% market share. For those that subscribed, I wrote a specific thread on how AMD "secret weapon" is going to change the game in 2026 with an improved designs on all its products, yes AMD has patent on it. Software is the linchpin. AMD's ROCm platform, once derided as "half-baked," now supports day-zero integration for Llama-4, DeepSeek V3, and GPT-OSS models—closing the CUDA gap. Benchmarks show MI355X (MI450 precursor) outperforming Nvidia's B200 in inference by 1.5-2x on memory-bound tasks, at 25-35% lower TCO. For training, MI450's rack-scale IF128 configuration (128 GPUs, 1.4 PB/s intra-rack bandwidth) rivals Nvidia's VR200 NVL144, enabling clusters like xAI's Colossus (scaling to 1M GPUs). My below thread projected Etimated conservative FY 25 revenue: $34-$36B Estimated conservative FY 26 revenue: $55B-$62B Below is why $AMD is revenue is going to be much higher after OpenAI deal. 1. OpenAI 1GW in 2026. With high demand for MI355X at $30,000k+ per unit, with MI450 is likely to be sold in the $45k-$55k. We can safely calcuate 1GW would require roughly 400,000 MI450 GPUs. or Roughly ~$20B revenue in 2026 alone from OpenAI. That would mean $AMD would hit $56B just from one partnership(OpenAI) in 2026 2. $META, the biggest spender on AI Infrastructure right now, Daddy Zuckerberg bought 250,000+ MI300, and is buying MI355X for recommendation engines and Llama training. It is very unlikely for Daddy Zuck to slow down AMD Chips, due to its Inference superiority to NVDA Chips. Most likely we will see at least 300,000-400,000 MI355X ordered from now toward end of H1 2025. And another 300,000-500,000 MI450 by H2 2025. Or ~$20B from just Meta in H2 alone, excluded H1. 3. xAI : Musk confirmed "AMD GPUs work very well" for Grok's small/medium models, with 30% of Grok-1 on MI300X. xAI's Colossus (200K+ GPUs, targeting 1M) and Oracle partnership (via OCI's MI355X cluster) position it for MI450 trials in H1 2026. With $6B funding and Grok integration into Oracle services, xAI could allocate 10-20% ($10B-$15B) to MI450 for distributed inference. We haven't heard the detail from Daddy Elon Musk yet, but most likely not going to be spending less than OpenAI or Sam Altman 4. Oracle ($ORCL): A multi-billion-dollar MI355X deal powers OCI's AI superclusters, with $500B+ remaining performance obligations. Larry Ellison's zettascale ambitions and xAI/OpenAI integrations make Oracle a MI450 anchor tenant—projected 50-100k units ($15B+ spend) for enterprise AI platforms. $ORCL is likely to spend more on the new "secret weapon" due to its capability in AI inference and cost advantage for $500B backlog. 5. Others ( Microsoft , Amazon , Saudi+other countries): Microsoft (Azure MI300X for training) and Amazon ($148B 15-year spend) test MI450 via Stargate ($500B with Oracle/SoftBank). Emerging buyers like G42 (5GW UAE campus), Crusoe, and Hot Aisle add 5-10GW demand. These potentially would add $15B-$30B in 2026 alone. We also need to factor in $TSM supply constraint( $NVDA is TSMC favorite), so $AMD market cap/growth is being tamed by TSMC. So what are you saying Mike, well $AMD 2026 revenue could hit $90-$100B by end of 2026 or nearly 185% growth YoYo. So what does that mean for valuation? I have no idea how Mr. Market gonna value AMD in 2026 with 3 digits growth. My Conservative $620 was my best projection until today with OpenAI partnership. I'm telling you as one of the biggest AMD bull, that I will leave it to "smart money" and other investors to do the price discovery while I'm chilling and writing DDs daily. Lastly, AMD's MI450 isn't hype—it's a calibrated strike at Nvidia's vulnerabilities, amplified by hyperscaler bets like Meta's 42% allocation and OpenAI's 6GW lifeline. By prioritizing inference efficiency, rack-scale innovation, and open ecosystems, AMD will siphon 10-15% share in 2026, scaling to 20%+ as TCO trumps CUDA loyalty. Meta, xAI, Oracle et al. aren't passive; they're active co-designers, betting billions on MI450 to fuel AGI pursuits without Nvidia's premium. For investors, this is AMD's inflection Per Dr. Lisa Su Not Financial Advice!
1:06

$AMD $620/share is too conservative for 2026 🧵 Some quick facts before I dive into this super long thread: $META allocated 42% GPUs to $AMD and 58% to $NVDA OpenAI allocated 6GW(38%) to $AMD and 10GW to $NVDA My $620 PT below by end of 2026 was only for 10-15% market share. I believe $AMD is going to have much much higher market share than I projected. The AI accelerator market is exploding, projected to reach $500 billion by 2028(is now heading $1Tril), driven by insatiable demand for training and inference compute in large language models (LLMs), recommendation systems, and autonomous systems. Nvidia ($NVDA) has long held a stranglehold, commanding over 90% market share through its CUDA ecosystem and superior rack-scale solutions. However, AMD is mounting a formidable challenge, leveraging cost advantages, open-source software momentum, and hyperscaler partnerships to erode Nvidia's moat. Recent deals—such as Meta's ($META) allocation of 42% of its GPU capacity to AMD and OpenAI's commitment to 6GW of AMD compute (versus 10GW for Nvidia)—signal a tipping point. At the forefront is AMD's Instinct MI450 series, a next-generation AI GPU slated for H2 2026 launch, which promises "no-excuses" leadership in training, inference, and distributed workloads. This analysis dissects how AMD will capture more market share and why hyperscalers like $Meta , xAI , Oracle , and others are poised to become voracious buyers of the MI450. AMD's AI GPU revenue has surged from negligible levels in 2022 to an estimated $4-5 billion in 2025, capturing ~6% of the data center GPU market. This growth stems from the Instinct MI300X, which offers 141GB of HBM3 memory and competitive FP8/FP16 performance at 20-30% lower cost than Nvidia's H100. Hyperscalers, facing NVIDIA 's overcharging, have turned to AMD for diversification. Meta, for instance, plans 600,000 H100-equivalent GPUs by end-2024, with ~42% (or 250,000+ units) sourced from AMD's MI300 series for inference tasks like image editing and AI assistants. Similarly, OpenAI's recent multi-year deal commits to 6GW of AMD compute—equivalent to ~300,000-400,000 MI450 GPUs—starting with 1GW in 2026, explicitly to counterbalance its 10GW Nvidia allocation. These aren't one-offs. Microsoft Azure, Amazon AWS, and Oracle Cloud Infrastructure (OCI) have integrated MI300X for AI workloads, with Oracle deploying 30,000 MI355X units in zettascale clusters. xAI, Elon Musk Musk's AI venture, ran 30% of Grok-1's production traffic on MI300X GPUs and has confirmed ongoing purchases. Collectively, these partners represent over $400 billion in projected AI infrastructure spend through 2028, with AMD targeting up to 40% market share. For those that subscribed, I wrote a specific thread on how AMD "secret weapon" is going to change the game in 2026 with an improved designs on all its products, yes AMD has patent on it. Software is the linchpin. AMD's ROCm platform, once derided as "half-baked," now supports day-zero integration for Llama-4, DeepSeek V3, and GPT-OSS models—closing the CUDA gap. Benchmarks show MI355X (MI450 precursor) outperforming Nvidia's B200 in inference by 1.5-2x on memory-bound tasks, at 25-35% lower TCO. For training, MI450's rack-scale IF128 configuration (128 GPUs, 1.4 PB/s intra-rack bandwidth) rivals Nvidia's VR200 NVL144, enabling clusters like xAI's Colossus (scaling to 1M GPUs). My below thread projected Etimated conservative FY 25 revenue: $34-$36B Estimated conservative FY 26 revenue: $55B-$62B Below is why $AMD is revenue is going to be much higher after OpenAI deal. 1. OpenAI 1GW in 2026. With high demand for MI355X at $30,000k+ per unit, with MI450 is likely to be sold in the $45k-$55k. We can safely calcuate 1GW would require roughly 400,000 MI450 GPUs. or Roughly ~$20B revenue in 2026 alone from OpenAI. That would mean $AMD would hit $56B just from one partnership(OpenAI) in 2026 2. $META, the biggest spender on AI Infrastructure right now, Daddy Zuckerberg bought 250,000+ MI300, and is buying MI355X for recommendation engines and Llama training. It is very unlikely for Daddy Zuck to slow down AMD Chips, due to its Inference superiority to NVDA Chips. Most likely we will see at least 300,000-400,000 MI355X ordered from now toward end of H1 2025. And another 300,000-500,000 MI450 by H2 2025. Or ~$20B from just Meta in H2 alone, excluded H1. 3. xAI : Musk confirmed "AMD GPUs work very well" for Grok's small/medium models, with 30% of Grok-1 on MI300X. xAI's Colossus (200K+ GPUs, targeting 1M) and Oracle partnership (via OCI's MI355X cluster) position it for MI450 trials in H1 2026. With $6B funding and Grok integration into Oracle services, xAI could allocate 10-20% ($10B-$15B) to MI450 for distributed inference. We haven't heard the detail from Daddy Elon Musk yet, but most likely not going to be spending less than OpenAI or Sam Altman 4. Oracle ($ORCL): A multi-billion-dollar MI355X deal powers OCI's AI superclusters, with $500B+ remaining performance obligations. Larry Ellison's zettascale ambitions and xAI/OpenAI integrations make Oracle a MI450 anchor tenant—projected 50-100k units ($15B+ spend) for enterprise AI platforms. $ORCL is likely to spend more on the new "secret weapon" due to its capability in AI inference and cost advantage for $500B backlog. 5. Others ( Microsoft , Amazon , Saudi+other countries): Microsoft (Azure MI300X for training) and Amazon ($148B 15-year spend) test MI450 via Stargate ($500B with Oracle/SoftBank). Emerging buyers like G42 (5GW UAE campus), Crusoe, and Hot Aisle add 5-10GW demand. These potentially would add $15B-$30B in 2026 alone. We also need to factor in $TSM supply constraint( $NVDA is TSMC favorite), so $AMD market cap/growth is being tamed by TSMC. So what are you saying Mike, well $AMD 2026 revenue could hit $90-$100B by end of 2026 or nearly 185% growth YoYo. So what does that mean for valuation? I have no idea how Mr. Market gonna value AMD in 2026 with 3 digits growth. My Conservative $620 was my best projection until today with OpenAI partnership. I'm telling you as one of the biggest AMD bull, that I will leave it to "smart money" and other investors to do the price discovery while I'm chilling and writing DDs daily. Lastly, AMD's MI450 isn't hype—it's a calibrated strike at Nvidia's vulnerabilities, amplified by hyperscaler bets like Meta's 42% allocation and OpenAI's 6GW lifeline. By prioritizing inference efficiency, rack-scale innovation, and open ecosystems, AMD will siphon 10-15% share in 2026, scaling to 20%+ as TCO trumps CUDA loyalty. Meta, xAI, Oracle et al. aren't passive; they're active co-designers, betting billions on MI450 to fuel AGI pursuits without Nvidia's premium. For investors, this is AMD's inflection Per Dr. Lisa Su Not Financial Advice!

Mike

711,006 просмотров • 10 месяцев назад

$GRAB Map is The New Google Maps(B2B)🧵 Here is your Free.99 analysis on GrabMap, for those that selling courses for $50-$500/m, if you are using my $GRAB and other analyses, I don't ask for much, at least give me some credit/cite. And yes 99.999% of my posts are Free.99. If you want to support my work, slap the like/repost, as I don't choose to write "Grab or any Ticker is going to x10 x100-x1000" kind of threads or "mark my words" to please the X Algo. Consider Subscribe($0.33/day) if you want to support my work further and get more in-depth analyses! TLDR: GrabMap could generate $7B-$15B a year alone for Grab B2B segment. That is why you are seeing Anthony Tan is mad excited abt this massive opportunity. And it also significantly boost GrabAds long term globally. This precisely proved my point that, Anthony is going to expand to 5B people and we are only 14% thesis realized right now. Grab doesn't have to be just Ride-share/Delivery when expanding! Grab , Southeast Asia's leading AI SuperApp for ride-hailing, food delivery, financial services,Tourism, Dine-Out and more, has developed its proprietary mapping platform, GrabMaps, a massive B2B revenue potential over the next long term, not just in Singapore, Indonesia, Malaysia, Thailand, Philippines, Vietnam, Cambodia, and Myanmar but expanding beyond SEA markets/Customers. 1. GrabMaps: A Strategic Asset GrabMaps is not merely a technological tool but a critical component of Grab's ecosystem, powering its ride-hailing, food delivery, and financial services. Developed in-house, GrabMaps leverages data collected from Grab's vast network of driver-partners across eight SEA countries. This data-driven approach ensures hyper-local customization, addressing the unique challenges of SEA's urban environments, such as narrow alleys, informal roads, and rapid infrastructure changes. The recent announcement of KartaCam2, an upgraded street-level imaging device, marks a significant technological advancement. KartaCam2 enhances data collection by providing higher quality images and more precise location data, which are crucial for maintaining the accuracy and freshness of maps. This breakthrough is part of Grab's broader 2025 AI push, including integrations with OpenAI 's GPT-4o for vision-based mapping and the establishment of an AI Centre of Excellence. These innovations position GrabMaps as a formidable competitor to Google Maps, especially in regions where localized data is paramount. 2. Revenue implications long term The expansion of GrabMaps into B2B services opens up new revenue streams, which could significantly impact Grab's financial performance over the long term. But GrabMap is a brandnew B2B product, and GoogleMap generates around $13-$20B globally. A. Market Opportunity in Southeast Asia ~The SEA market presents a substantial opportunity for GrabMaps. The foodservice market alone is projected to grow from $223.8 billion in 2025 to $416.3 billion by 2030, indicating a robust demand for services that enhance operational efficiencies. Businesses in logistics, e-commerce, and urban planning could benefit from GrabMaps' precise mapping and navigation capabilities, potentially generating revenue through licensing fees, subscription models, and advertising. ~Grab's existing user base of over 46 million monthly transacting users provides a strong foundation for cross-selling B2B solutions, thereby increasing revenue without significant additional marketing costs. B. Competitive Advantage of a Future $500B MC AI SuperApp over Google Map Google Maps, while dominant, may not be as finely tuned for SEA's unique challenges. GrabMaps' hyper-local data and AI-driven enhancements offer a competitive edge, attracting businesses that require accurate and cost-effective mapping solutions. Revenue from B2B services could include: Licensing Fees: Enterprises can license GrabMaps' APIs and SDKs to integrate mapping functionalities into their operations. Subscription Models: Continuous updates and premium features could be offered on a subscription basis. Advertising Revenue: GrabAds, which leverages mapping data, could generate additional income through targeted advertising. C. Global Expansion is Inevitable ~The partnership with Tino in Mongolia is a strategic move to scale GrabMaps internationally. This marks Grab's first major mapping partnership outside SEA, indicating potential for revenue growth in other regions where Google Maps' dominance is less entrenched or where local data needs are acute. ~The use of IoT devices like KartaCam2 and KartaDashCam for real-time data collection could further enhance GrabMaps' value proposition, potentially increasing revenue through premium service offerings in new markets. D. Synergies w/ other businesses Grab's ecosystem approach allows for synergies between GrabMaps and other services like GrabPay, GrabFood, and GrabTransport. For example, businesses using GrabMaps for logistics could also adopt GrabPay for transactions, creating a revenue multiplier effect. 3. Google Map Revenue in Asia ~Total Revenue in Asia-Pacific (2018): Google APAC, based in Singapore, reported $20.24 billion out of the total $21.37 billion revenue in the Asia-Pacific region. This indicates that a significant portion of Google's revenue in Asia is attributed to Singapore, likely due to its role as a hub for Google’s operations. ~Advertising Revenue: In 2018, Google APAC generated $15.8 billion from advertising alone, compared to $4.4 billion from other activities like Google Play. Advertising on Google properties, including Google Maps, is a major revenue driver. ~Market Share in Search Marketing: Google Maps holds a 62.34% market share in the search marketing category, competing with tools like Wix (26.54%) and Google Ads (4.14%). This dominance suggests that a considerable portion of Google’s advertising revenue in Asia is linked to mapping services. For the full fiscal year 2024, Alphabet (Google's parent company) generated $56.82 billion in revenue from the Asia-Pacific (APAC) region. This represented approximately 16.24% of the company's total revenue for the year. If we take a conservative estimate at 25% of $56.82B of Google's total advertising revenue in Asia is related to mapping services= $14.2B. => If GrabMaps secures even 50% of this market share in SEA, it could generate around $7B annually from this segment alone. GrabMap is 4x lower error rate, 10x lower latency, 75% fewer mapping mistakes, and much cheaper than GoogleMap. With OpenAI GPT-4o fine-tuning, GrabMaps hit 80% accuracy for speed limits and lanes13-20% above prior levels excelling in occlusions ( rainy monsoons) where Google relies more on satellite data. Now do you understand why Google and HSBC are clapping $GRAB on search and downgrade? Yes, because GrabMap is a massive threat and Grab Anthony Tan refused to buy $goto since 2020. Conclusion: Grab's expansion of GrabMaps into B2B services represents a strategic move to challenge Google Maps' dominance in Asia, particularly in SEA and future expansion. The revenue implications are substantial, with potential gains from licensing fees, subscription models, advertising, and international expansions. While Google Maps generates billions in revenue, primarily through advertising, GrabMaps' localized and AI-enhanced approach could carve out a significant niche, especially in regions where precise, real-time mapping data is critical. The success of this strategy will depend on Grab's ability to scale internationally, maintain technological superiority, and effectively monetize its B2B offerings. However, the opportunity is clear, and Grab's ecosystem approach positions it well to capitalize on the growing demand for advanced mapping solutions in a rapidly digitalizing world. This move not only enhances Grab's revenue potential but also solidifies its role as a key player in the global tech landscape. Not Financial Advice! Source: Grab Dot Com.
1:26

$GRAB Map is The New Google Maps(B2B)🧵 Here is your Free.99 analysis on GrabMap, for those that selling courses for $50-$500/m, if you are using my $GRAB and other analyses, I don't ask for much, at least give me some credit/cite. And yes 99.999% of my posts are Free.99. If you want to support my work, slap the like/repost, as I don't choose to write "Grab or any Ticker is going to x10 x100-x1000" kind of threads or "mark my words" to please the X Algo. Consider Subscribe($0.33/day) if you want to support my work further and get more in-depth analyses! TLDR: GrabMap could generate $7B-$15B a year alone for Grab B2B segment. That is why you are seeing Anthony Tan is mad excited abt this massive opportunity. And it also significantly boost GrabAds long term globally. This precisely proved my point that, Anthony is going to expand to 5B people and we are only 14% thesis realized right now. Grab doesn't have to be just Ride-share/Delivery when expanding! Grab , Southeast Asia's leading AI SuperApp for ride-hailing, food delivery, financial services,Tourism, Dine-Out and more, has developed its proprietary mapping platform, GrabMaps, a massive B2B revenue potential over the next long term, not just in Singapore, Indonesia, Malaysia, Thailand, Philippines, Vietnam, Cambodia, and Myanmar but expanding beyond SEA markets/Customers. 1. GrabMaps: A Strategic Asset GrabMaps is not merely a technological tool but a critical component of Grab's ecosystem, powering its ride-hailing, food delivery, and financial services. Developed in-house, GrabMaps leverages data collected from Grab's vast network of driver-partners across eight SEA countries. This data-driven approach ensures hyper-local customization, addressing the unique challenges of SEA's urban environments, such as narrow alleys, informal roads, and rapid infrastructure changes. The recent announcement of KartaCam2, an upgraded street-level imaging device, marks a significant technological advancement. KartaCam2 enhances data collection by providing higher quality images and more precise location data, which are crucial for maintaining the accuracy and freshness of maps. This breakthrough is part of Grab's broader 2025 AI push, including integrations with OpenAI 's GPT-4o for vision-based mapping and the establishment of an AI Centre of Excellence. These innovations position GrabMaps as a formidable competitor to Google Maps, especially in regions where localized data is paramount. 2. Revenue implications long term The expansion of GrabMaps into B2B services opens up new revenue streams, which could significantly impact Grab's financial performance over the long term. But GrabMap is a brandnew B2B product, and GoogleMap generates around $13-$20B globally. A. Market Opportunity in Southeast Asia ~The SEA market presents a substantial opportunity for GrabMaps. The foodservice market alone is projected to grow from $223.8 billion in 2025 to $416.3 billion by 2030, indicating a robust demand for services that enhance operational efficiencies. Businesses in logistics, e-commerce, and urban planning could benefit from GrabMaps' precise mapping and navigation capabilities, potentially generating revenue through licensing fees, subscription models, and advertising. ~Grab's existing user base of over 46 million monthly transacting users provides a strong foundation for cross-selling B2B solutions, thereby increasing revenue without significant additional marketing costs. B. Competitive Advantage of a Future $500B MC AI SuperApp over Google Map Google Maps, while dominant, may not be as finely tuned for SEA's unique challenges. GrabMaps' hyper-local data and AI-driven enhancements offer a competitive edge, attracting businesses that require accurate and cost-effective mapping solutions. Revenue from B2B services could include: Licensing Fees: Enterprises can license GrabMaps' APIs and SDKs to integrate mapping functionalities into their operations. Subscription Models: Continuous updates and premium features could be offered on a subscription basis. Advertising Revenue: GrabAds, which leverages mapping data, could generate additional income through targeted advertising. C. Global Expansion is Inevitable ~The partnership with Tino in Mongolia is a strategic move to scale GrabMaps internationally. This marks Grab's first major mapping partnership outside SEA, indicating potential for revenue growth in other regions where Google Maps' dominance is less entrenched or where local data needs are acute. ~The use of IoT devices like KartaCam2 and KartaDashCam for real-time data collection could further enhance GrabMaps' value proposition, potentially increasing revenue through premium service offerings in new markets. D. Synergies w/ other businesses Grab's ecosystem approach allows for synergies between GrabMaps and other services like GrabPay, GrabFood, and GrabTransport. For example, businesses using GrabMaps for logistics could also adopt GrabPay for transactions, creating a revenue multiplier effect. 3. Google Map Revenue in Asia ~Total Revenue in Asia-Pacific (2018): Google APAC, based in Singapore, reported $20.24 billion out of the total $21.37 billion revenue in the Asia-Pacific region. This indicates that a significant portion of Google's revenue in Asia is attributed to Singapore, likely due to its role as a hub for Google’s operations. ~Advertising Revenue: In 2018, Google APAC generated $15.8 billion from advertising alone, compared to $4.4 billion from other activities like Google Play. Advertising on Google properties, including Google Maps, is a major revenue driver. ~Market Share in Search Marketing: Google Maps holds a 62.34% market share in the search marketing category, competing with tools like Wix (26.54%) and Google Ads (4.14%). This dominance suggests that a considerable portion of Google’s advertising revenue in Asia is linked to mapping services. For the full fiscal year 2024, Alphabet (Google's parent company) generated $56.82 billion in revenue from the Asia-Pacific (APAC) region. This represented approximately 16.24% of the company's total revenue for the year. If we take a conservative estimate at 25% of $56.82B of Google's total advertising revenue in Asia is related to mapping services= $14.2B. => If GrabMaps secures even 50% of this market share in SEA, it could generate around $7B annually from this segment alone. GrabMap is 4x lower error rate, 10x lower latency, 75% fewer mapping mistakes, and much cheaper than GoogleMap. With OpenAI GPT-4o fine-tuning, GrabMaps hit 80% accuracy for speed limits and lanes13-20% above prior levels excelling in occlusions ( rainy monsoons) where Google relies more on satellite data. Now do you understand why Google and HSBC are clapping $GRAB on search and downgrade? Yes, because GrabMap is a massive threat and Grab Anthony Tan refused to buy $goto since 2020. Conclusion: Grab's expansion of GrabMaps into B2B services represents a strategic move to challenge Google Maps' dominance in Asia, particularly in SEA and future expansion. The revenue implications are substantial, with potential gains from licensing fees, subscription models, advertising, and international expansions. While Google Maps generates billions in revenue, primarily through advertising, GrabMaps' localized and AI-enhanced approach could carve out a significant niche, especially in regions where precise, real-time mapping data is critical. The success of this strategy will depend on Grab's ability to scale internationally, maintain technological superiority, and effectively monetize its B2B offerings. However, the opportunity is clear, and Grab's ecosystem approach positions it well to capitalize on the growing demand for advanced mapping solutions in a rapidly digitalizing world. This move not only enhances Grab's revenue potential but also solidifies its role as a key player in the global tech landscape. Not Financial Advice! Source: Grab Dot Com.

Mike

120,532 просмотров • 9 месяцев назад

Tales of Cities | Tongren, where China's matcha capital hides in plain sight Most people who drink matcha assume it is Japanese. In Tongren, a mountain city tucked into southwest China's Guizhou Province, that assumption meets a quiet correction. The tea tradition long associated with Japan can be traced back further to China's Tang and Song dynasties (618-1279). During the Song Dynasty, a method called diancha, whisking powdered tea into hot water until it turns pale and frothy, was already widespread among monks and scholars. Gu Lijun, a county-level inheritor of that Song Dynasty tradition, traces the practice back even earlier still, to the Wei and Jin dynasties (220-420). The technique now associated with Japanese tea ceremony, in other words, started in China, centuries before it ever crossed the sea. Tongren is now working to reclaim that origin story with production rather than just history. The city wears the title "China's matcha capital" for good reason, ranking first nationally and second worldwide in matcha production and sales. That scale is easy to state as a statistic and harder to picture, until you trace the leaf back to the mountain that makes it possible. Mount Fanjing rises more than 2,500 meters above the surrounding tea country, its ancient rock formations emerging from clouds that never quite leave, wrapped in mist and rain for more than 200 days a year, long enough that locals treat a clear day at the summit as luck rather than expectation. It is a UNESCO World Natural Heritage Site, home to plant and animal species found nowhere else on Earth, and the climb to its Red Cloud Golden Summit is demanding by design, steep enough in places to require hands as well as feet. It is a mountain that guards its own secrets. Almost everything grown in its shadow carries that same difficulty, and benefits from it. At the Qizimei Tea Garden in Jiangkou county, plantation head Chen Chunlian tends 220 mu (about 14.67 hectares) of Fuding Dabai, Wuniuzao and Longjing tea, all destined to become tencha, the ground raw material for matcha, with Mount Fanjing visible from the rows themselves. Walk the plantation's lower slopes and the rows of low, dense bushes stretch across the hillside in every direction, broken only by mist rolling down from above, tended by farmers who speak of the mountain with a quiet, matter-of-fact devotion, as though caring for something this rare were simply what one does here. Chen explains that the mountain's high altitude, limited sunlight and near-constant cloud cover raise the amino acid content of the leaves, improving quality well beyond what flatland tea can achieve. Farmers add a further layer of engineering before harvest, covering the plants to block out even more light. Less sunlight means more chlorophyll and more theanine, the compound chemistry behind both matcha's saturated green color and its umami depth. The effect has a name in the trade: "covered aroma." Gu describes the resulting flavor in more familiar terms, bright and fresh, carrying what tea drinkers usually just call "the aroma of seaweed." From there, the leaf still has a distance to travel before it becomes the powder found on shelves abroad. At Gui Tea Group's matcha workshop, production supervisor Li Tingbao oversees the fine processing stage: aroma enhancement, color sorting, grinding and screening, followed by air separation, a step that filters out stems before the refined tea drops into a material bin to await blending. Only after grinding does it finally become matcha in the form most people would recognize. Li reports that the workshop's output has grown quickly, from roughly 2,500 tonnes in 2025 to an expected 5,000 tonnes this year, doubling in a single season and underscoring how fast global demand has moved. What happens to that matcha afterward is where Tongren's ambitions become most visible, and most inventive, and it tracks a pattern playing out across the industry worldwide, where matcha has moved well past the teacup into bakery cases, dessert menus and skincare shelves. Tang Yunpeng, general manager of Guizhou Guigui Matcha Food Co., Ltd., leads a team developing matcha products well outside tradition: sun cakes with a flaky, crumbly crust wrapped around matcha filling, cheesecakes, chocolates. Tongren's matcha, Tang notes, carries a stronger bitterness and fuller flavor than matcha grown elsewhere, a distinction that comes through clearly once it is baked, whisked or melted into something else. Elsewhere in the city, matcha turns up in dumplings, noodles, beer, chapstick, hand lotion and perfume. At this point, matcha in Tongren is less an ingredient than a category of its own. At a local nongjiale, a uniquely Chinese agritainment venue operated by local farmers that doubles as a restaurant and a window into local life, located in Yunshe village of Taiping town in Tongren's Jiangkou county, villager Yang Yanfei's table typically arrives loaded with Larou, the salt-cured smoked pork that anchors Guizhou home cooking, alongside a dish that still catches newcomers off guard: matcha sturgeon, the tea's flavor threaded lightly through the fish rather than layered over it. It is not a novelty so much as evidence of how completely matcha has settled into daily life here, alongside the area's clean air and unspoiled mountains and water, which Yang credits as the real foundation beneath all of it. For the people who make Tongren's matcha economy run, from tea garden to processing floor to test kitchen, the ambitions are collective, and they add up to something larger than any single product. Chen speaks of the quiet pride in supplying the raw material that starts the entire chain. Tang hopes matcha becomes something bigger still: a calling card not just for Tongren, but for Guizhou and for China as a whole. As matcha continues its rise in cafes and kitchens far beyond Guizhou, Tongren's name is beginning to travel with it.
7:44

Tales of Cities | Tongren, where China's matcha capital hides in plain sight Most people who drink matcha assume it is Japanese. In Tongren, a mountain city tucked into southwest China's Guizhou Province, that assumption meets a quiet correction. The tea tradition long associated with Japan can be traced back further to China's Tang and Song dynasties (618-1279). During the Song Dynasty, a method called diancha, whisking powdered tea into hot water until it turns pale and frothy, was already widespread among monks and scholars. Gu Lijun, a county-level inheritor of that Song Dynasty tradition, traces the practice back even earlier still, to the Wei and Jin dynasties (220-420). The technique now associated with Japanese tea ceremony, in other words, started in China, centuries before it ever crossed the sea. Tongren is now working to reclaim that origin story with production rather than just history. The city wears the title "China's matcha capital" for good reason, ranking first nationally and second worldwide in matcha production and sales. That scale is easy to state as a statistic and harder to picture, until you trace the leaf back to the mountain that makes it possible. Mount Fanjing rises more than 2,500 meters above the surrounding tea country, its ancient rock formations emerging from clouds that never quite leave, wrapped in mist and rain for more than 200 days a year, long enough that locals treat a clear day at the summit as luck rather than expectation. It is a UNESCO World Natural Heritage Site, home to plant and animal species found nowhere else on Earth, and the climb to its Red Cloud Golden Summit is demanding by design, steep enough in places to require hands as well as feet. It is a mountain that guards its own secrets. Almost everything grown in its shadow carries that same difficulty, and benefits from it. At the Qizimei Tea Garden in Jiangkou county, plantation head Chen Chunlian tends 220 mu (about 14.67 hectares) of Fuding Dabai, Wuniuzao and Longjing tea, all destined to become tencha, the ground raw material for matcha, with Mount Fanjing visible from the rows themselves. Walk the plantation's lower slopes and the rows of low, dense bushes stretch across the hillside in every direction, broken only by mist rolling down from above, tended by farmers who speak of the mountain with a quiet, matter-of-fact devotion, as though caring for something this rare were simply what one does here. Chen explains that the mountain's high altitude, limited sunlight and near-constant cloud cover raise the amino acid content of the leaves, improving quality well beyond what flatland tea can achieve. Farmers add a further layer of engineering before harvest, covering the plants to block out even more light. Less sunlight means more chlorophyll and more theanine, the compound chemistry behind both matcha's saturated green color and its umami depth. The effect has a name in the trade: "covered aroma." Gu describes the resulting flavor in more familiar terms, bright and fresh, carrying what tea drinkers usually just call "the aroma of seaweed." From there, the leaf still has a distance to travel before it becomes the powder found on shelves abroad. At Gui Tea Group's matcha workshop, production supervisor Li Tingbao oversees the fine processing stage: aroma enhancement, color sorting, grinding and screening, followed by air separation, a step that filters out stems before the refined tea drops into a material bin to await blending. Only after grinding does it finally become matcha in the form most people would recognize. Li reports that the workshop's output has grown quickly, from roughly 2,500 tonnes in 2025 to an expected 5,000 tonnes this year, doubling in a single season and underscoring how fast global demand has moved. What happens to that matcha afterward is where Tongren's ambitions become most visible, and most inventive, and it tracks a pattern playing out across the industry worldwide, where matcha has moved well past the teacup into bakery cases, dessert menus and skincare shelves. Tang Yunpeng, general manager of Guizhou Guigui Matcha Food Co., Ltd., leads a team developing matcha products well outside tradition: sun cakes with a flaky, crumbly crust wrapped around matcha filling, cheesecakes, chocolates. Tongren's matcha, Tang notes, carries a stronger bitterness and fuller flavor than matcha grown elsewhere, a distinction that comes through clearly once it is baked, whisked or melted into something else. Elsewhere in the city, matcha turns up in dumplings, noodles, beer, chapstick, hand lotion and perfume. At this point, matcha in Tongren is less an ingredient than a category of its own. At a local nongjiale, a uniquely Chinese agritainment venue operated by local farmers that doubles as a restaurant and a window into local life, located in Yunshe village of Taiping town in Tongren's Jiangkou county, villager Yang Yanfei's table typically arrives loaded with Larou, the salt-cured smoked pork that anchors Guizhou home cooking, alongside a dish that still catches newcomers off guard: matcha sturgeon, the tea's flavor threaded lightly through the fish rather than layered over it. It is not a novelty so much as evidence of how completely matcha has settled into daily life here, alongside the area's clean air and unspoiled mountains and water, which Yang credits as the real foundation beneath all of it. For the people who make Tongren's matcha economy run, from tea garden to processing floor to test kitchen, the ambitions are collective, and they add up to something larger than any single product. Chen speaks of the quiet pride in supplying the raw material that starts the entire chain. Tang hopes matcha becomes something bigger still: a calling card not just for Tongren, but for Guizhou and for China as a whole. As matcha continues its rise in cafes and kitchens far beyond Guizhou, Tongren's name is beginning to travel with it.

Beautiful China

21,874 просмотров • 26 дней назад