Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Reasoning VLAs can think. They just can't think fast. Until now. Introducing FlashDrive⚡ 🚀 716 ms → 159 ms on RTX PRO 6000 (up to 5.7×) ✅ Zero accuracy loss FlashDrive = streaming inference + DFlash speculative reasoning + ParoQuant W4A8 Real-time reasoning for autonomous driving is here!

Zhijian Liu

9,743 subscribers

191,948 views • 5 months ago •via X (Twitter)

News & Politics Science & Technology Sports

Anya Rossi• Live Now

Private livecam show

32 Comments

Zhijian Liu5 months ago

Code + model checkpoints coming soon. The ingredients are already open-source: 🔥 DFlash → ⚡ ParoQuant →

Zhijian Liu5 months ago

Great work led by @Richard91316073 and @YihaoLiang01, with @hongfeizhang0xF and @jianchen1799.

idonteven5 months ago

Thanks for the paper, I've spent around 2 hours reading it and come up with my response. Screenshot because of twitter character limit.

Zhijian Liu5 months ago

Thanks for sharing your feedback! The baseline is under BF16.

Piotr Mazurek5 months ago

awesome work 🫡

Crepe Supreme5 months ago

716ms to 159ms is the line real-time robotics has been waiting on. Every VLA roadmap priced in at least another year before reasoning models could sit inside the control loop. Quietly huge.

Josef (atris)5 months ago

Trying this on an in-car 3090 with a comma soon

Neural Drop5 months ago

5.7x faster with zero accuracy loss sounds like the kind of magic trick that only works on paper until you actually try it in a car moving at 60mph

Henry Zhang5 months ago

holy shit. need this deployed somewhere and served on an api right now

Don Park5 months ago

could be more compelling from safety perspective if the model used for speculative reasoning is RL-ed to be more cautious than the full reasoning. safety-first in the literal sense. :-)

Brian Chew5 months ago

@kstonekuan this looks interesting and in your field!

Taro Bushidō5 months ago

Sub-160ms VLAs that don't think in molasses. ParoQuant + DFlash cuts latency 5.7x with zero accuracy loss. Lab demos just became production reality. ⚡️

Kirito (e/acc) 🏴‍☠️5 months ago

@yassineyousfi_

Avais Aziz5 months ago

Nice work getting reasoning VLA down to 159ms on RTX PRO 6000 with no accuracy drop. Streaming + DFlash speculative decoding + ParoQuant W4A8 is a clean stack for closing the control loop. Looking forward to the checkpoints.

Alen Gabs5 months ago

very cool !

Team Reagent5 months ago

Holy smokes this is really cool :O

Uğur Yekta Başak5 months ago

this is very practical

Atlas AI5 months ago

716ms to 159ms. Real-time VLA reasoning is no longer a lab benchmark.

Nischay4 months ago

The latency gap is real for live video reasoning. The quality tradeoff at high token budgets is the other side - we found diminishing returns past ~700 thinking tokens for video understanding. Benchmark at

David Yang5 months ago

Really cool!

Cloud Strife5 months ago

The latency wall was always the easier problem. Zero accuracy loss at 5.7x is impressive, but the moat in autonomous driving isn't inference speed — it's whether the system can handle the one corner case that no training distribution ever covered.

yangwao.hl5 months ago

are they going after @comma_ai

Zhengzhong Tu5 months ago

@QuanquanPeng03 tql

Josef (atris)5 months ago

wen weights tho

notnull5 months ago

speculative decoding on rtx pro is one thing. what about consumer gpus that people actually use. a10s cost way less

Zhijian Liu5 months ago

We've also tested on many other GPUs. See our blog for more details!

ogx5 months ago

Algorithm + System Collaboration

Duck Punk5 months ago

Some humans I know will need this to finally see the world with reasoning

💥5 months ago

😳

Team Reagent5 months ago

I need to update superlab to have support for playing around with VLAs and reasoning VLAs @FelionSpike

rburton5 months ago

One thing I wonder as I watch this video, does the software do a delta between the prior frame to the next phrase to spot changes then analyze that to see what changed that's important?

validate.qa5 months ago

flashdrive slicing latency. what's the real-world task it unlocks first

Related Videos

ParoQuant just got a big upgrade 🚀 ✅ Supports the new Qwen3.5 models ⚡ Now runs on MLX (fast local inference on Apple Silicon) 🧠 Preserves reasoning quality with 4-bit quantization We also built an agent demo running locally on my 4-year-old M2 Max. Can't wait to upgrade to an M5 Max and see what kind of magic we can do. ✨

ParoQuant just got a big upgrade 🚀 ✅ Supports the new Qwen3.5 models ⚡ Now runs on MLX (fast local inference on Apple Silicon) 🧠 Preserves reasoning quality with 4-bit quantization We also built an agent demo running locally on my 4-year-old M2 Max. Can't wait to upgrade to an M5 Max and see what kind of magic we can do. ✨

Zhijian Liu

49,826 views • 6 months ago

NEWS: NVIDIA just announced Alpamayo, what CEO Jensen Huang calls the world’s first thinking, reasoning autonomous vehicle AI, launching on U.S. roads later this year, starting with the Mercedes CLA. Jensen: "It's trained end-to-end. Literally from camera in to actuation out; It reasons what action it is about to take, the reason by which is came about that action, and the trajectory." Alpamayo introduces Vision-Language-Action (VLA) models, which enable self-driving systems to interpret what they see, reason about complex driving scenarios, and generate driving actions. The platform includes large reasoning models, simulation tools for testing rare and edge-case scenarios, and open datasets for training and validation. NVIDIA says the approach improves transparency, safety, and robustness in autonomous systems, particularly in complex real-world environments, and supports progress toward higher levels of vehicle autonomy: "With a 10-billion-parameter architecture, Alpamayo 1 uses video input to generate trajectories alongside reasoning traces, showing the logic behind each decision. Developers can adapt Alpamayo 1 into smaller runtime models for vehicle development, or use it as a foundation for AV development tools such as reasoning-based evaluators and auto-labeling systems. Alpamayo 1 provides open model weights and open-source inferencing scripts. Future models in the family will feature larger parameter counts, more detailed reasoning capabilities, more input and output flexibility, and options for commercial usage."

NEWS: NVIDIA just announced Alpamayo, what CEO Jensen Huang calls the world’s first thinking, reasoning autonomous vehicle AI, launching on U.S. roads later this year, starting with the Mercedes CLA. Jensen: "It's trained end-to-end. Literally from camera in to actuation out; It reasons what action it is about to take, the reason by which is came about that action, and the trajectory." Alpamayo introduces Vision-Language-Action (VLA) models, which enable self-driving systems to interpret what they see, reason about complex driving scenarios, and generate driving actions. The platform includes large reasoning models, simulation tools for testing rare and edge-case scenarios, and open datasets for training and validation. NVIDIA says the approach improves transparency, safety, and robustness in autonomous systems, particularly in complex real-world environments, and supports progress toward higher levels of vehicle autonomy: "With a 10-billion-parameter architecture, Alpamayo 1 uses video input to generate trajectories alongside reasoning traces, showing the logic behind each decision. Developers can adapt Alpamayo 1 into smaller runtime models for vehicle development, or use it as a foundation for AV development tools such as reasoning-based evaluators and auto-labeling systems. Alpamayo 1 provides open model weights and open-source inferencing scripts. Future models in the family will feature larger parameter counts, more detailed reasoning capabilities, more input and output flexibility, and options for commercial usage."

Sawyer Merritt

1,605,075 views • 8 months ago

Cerebras inference is very fast. So fast that it changes how we think about configuring our LLMs for voice agent use cases. Kimi K2.6 is a 1T parameter reasoning model that Cerebras serves at 650 - 1,000 tokens per second (end-to-end throughput), with time to first token metrics as low as 150ms (latency). These numbers are two to three times faster than other similarly capable models. The biggest lever we get from this kind of speed is that we can use the model in reasoning mode, and still have excellent "time to first non-thinking token." This solves a big pain point we have in 2026 for voice agent use cases. Almost all recent innovation in post-training has focused on making models good at reasoning ("test time compute"). This is great, but it makes the user-facing model latency much, much slower. Which is a problem for conversational voice agents. We can run Kimi K2.6 with reasoning turned on, and get responses faster than other models produce with reasoning disabled. On my 30-turn voice agent benchmark, Kimi K2.6 with reasoning enabled ties GPT 5.1 and Haiku 4.5 with reasoning disabled, and is still about 200ms seconds faster! On my primary task agent benchmark, Kimi K2.6 is now the #2 model. It ranks just behind Gemini 3.5 Flash in "high" reasoning mode, and tied with GLM 5, Sonnet 4.6, and GPT 5.4 with reasoning set to "low." But Kimi K2.6 completes each turn in the agent loop in under 500ms. The other four models are all at least 3x slower. (Models only qualify for this benchmark if they can complete task turns at a P50 <4s.) A couple of other things that this speed buys us, for production voice agents: - Tool calls happen fast enough that we don't have to work around tool call latency in our pipeline design. - We can prompt the model to output structured data at the beginning of a response, followed by plain text for voice generation. This opens up possibilities like asking the model to do complex classification/generation tasks that influence the rest of the pipeline. For example, the model could create a detailed style prompt for a steerable TTS model, for each individual conversation turn. And, of course, you can use Kimi K2.6 with reasoning turned off. Cerebras calls this "instant" mode. Here's a video of a Cerebras Kimi K2.6 voice agent with voice-to-voice response time, measured at the client, under 500ms. This is the true response latency as perceived by the user, including all network and audio codec overhead, transcription and turn detection, Kimi K2.6 token generation, and voice generation. 500ms is, effectively, instant. So the Cerebras naming for this mode is a propos. :-)

Cerebras inference is very fast. So fast that it changes how we think about configuring our LLMs for voice agent use cases. Kimi K2.6 is a 1T parameter reasoning model that Cerebras serves at 650 - 1,000 tokens per second (end-to-end throughput), with time to first token metrics as low as 150ms (latency). These numbers are two to three times faster than other similarly capable models. The biggest lever we get from this kind of speed is that we can use the model in reasoning mode, and still have excellent "time to first non-thinking token." This solves a big pain point we have in 2026 for voice agent use cases. Almost all recent innovation in post-training has focused on making models good at reasoning ("test time compute"). This is great, but it makes the user-facing model latency much, much slower. Which is a problem for conversational voice agents. We can run Kimi K2.6 with reasoning turned on, and get responses faster than other models produce with reasoning disabled. On my 30-turn voice agent benchmark, Kimi K2.6 with reasoning enabled ties GPT 5.1 and Haiku 4.5 with reasoning disabled, and is still about 200ms seconds faster! On my primary task agent benchmark, Kimi K2.6 is now the #2 model. It ranks just behind Gemini 3.5 Flash in "high" reasoning mode, and tied with GLM 5, Sonnet 4.6, and GPT 5.4 with reasoning set to "low." But Kimi K2.6 completes each turn in the agent loop in under 500ms. The other four models are all at least 3x slower. (Models only qualify for this benchmark if they can complete task turns at a P50 <4s.) A couple of other things that this speed buys us, for production voice agents: - Tool calls happen fast enough that we don't have to work around tool call latency in our pipeline design. - We can prompt the model to output structured data at the beginning of a response, followed by plain text for voice generation. This opens up possibilities like asking the model to do complex classification/generation tasks that influence the rest of the pipeline. For example, the model could create a detailed style prompt for a steerable TTS model, for each individual conversation turn. And, of course, you can use Kimi K2.6 with reasoning turned off. Cerebras calls this "instant" mode. Here's a video of a Cerebras Kimi K2.6 voice agent with voice-to-voice response time, measured at the client, under 500ms. This is the true response latency as perceived by the user, including all network and audio codec overhead, transcription and turn detection, Kimi K2.6 token generation, and voice generation. 500ms is, effectively, instant. So the Cerebras naming for this mode is a propos. :-)

kwindla

40,593 views • 4 months ago

We’re excited to introduce KAME: Tandem Architecture for Enhancing Knowledge in Real-Time Speech-to-Speech Conversational AI, accepted at #ICASSP2026! 🐢 Blog Paper Can a speech AI think deeply without pausing to process? In real conversation, we don’t wait until we’ve fully worked out what we want to say—we start talking, and our thoughts catch up as the sentence unfolds. Fast speech-to-speech models achieve this, but their reasoning tends to stay shallow. Cascaded pipelines that route through a knowledgeable LLM are smarter, but the added latency breaks the flow—they fall back to "think, then speak." In our new paper, we propose a way to break this trade-off. We call it KAME (Turtle in Japanese). A speech-to-speech model handles the fast response loop and starts replying immediately. In parallel, a backend LLM runs asynchronously, generating response candidates that are continuously injected as "oracle" signals in real time. This shifts the AI paradigm from "think, then speak" to "speak while thinking." The backend LLM is completely swappable. You can plug in GPT-4.1, Claude Opus, or Gemini 2.5 Flash depending on the task without changing the frontend. In our experiments, Claude tended to score higher on reasoning, while GPT did better on humanities questions. Try the model yourself here:

We’re excited to introduce KAME: Tandem Architecture for Enhancing Knowledge in Real-Time Speech-to-Speech Conversational AI, accepted at #ICASSP2026! 🐢 Blog Paper Can a speech AI think deeply without pausing to process? In real conversation, we don’t wait until we’ve fully worked out what we want to say—we start talking, and our thoughts catch up as the sentence unfolds. Fast speech-to-speech models achieve this, but their reasoning tends to stay shallow. Cascaded pipelines that route through a knowledgeable LLM are smarter, but the added latency breaks the flow—they fall back to "think, then speak." In our new paper, we propose a way to break this trade-off. We call it KAME (Turtle in Japanese). A speech-to-speech model handles the fast response loop and starts replying immediately. In parallel, a backend LLM runs asynchronously, generating response candidates that are continuously injected as "oracle" signals in real time. This shifts the AI paradigm from "think, then speak" to "speak while thinking." The backend LLM is completely swappable. You can plug in GPT-4.1, Claude Opus, or Gemini 2.5 Flash depending on the task without changing the frontend. In our experiments, Claude tended to score higher on reasoning, while GPT did better on humanities questions. Try the model yourself here:

Sakana AI

293,975 views • 5 months ago

Nvidia has just announced Alpamayo 2 Super, an open 34 billion parameter reasoning vision-language-action model designed to accelerate the development of autonomous vehicles. This new model combines the NVIDIA Cosmos 3 Super reasoning model with a 2 billion parameter diffusion-based action expert model, and is post trained with reinforcement learning. The model can return multiple outputs: future trajectory plans, reasoning traces, grounded answers to questions about the scenes, and auto label generation. The model weights are now available for anyone to download on Hugging Face, and the inference code has been posted to GitHub. Distilled models can be deployed commercially without any further permission from Nvidia, and model outputs carry no license conditions. Automakers can distill down a compact version of this model that can run on the Nvidia computer in the car. Major kudos to Nvidia and Jensen Huang for advancing the state of the industry by releasing this as an open model with permissive licensing. Jensen isn't just paying lip service to the idea of open models, Nvidia is actually contributing to the ecosystem — and it's great for their business, because it helps sell more Thor computers that go in the car. Anyone can go download the model and play with it. If you do, let me know what you think. Personally I think it's so cool that we have open weights models that are this advanced, for anyone to download.

Nvidia has just announced Alpamayo 2 Super, an open 34 billion parameter reasoning vision-language-action model designed to accelerate the development of autonomous vehicles. This new model combines the NVIDIA Cosmos 3 Super reasoning model with a 2 billion parameter diffusion-based action expert model, and is post trained with reinforcement learning. The model can return multiple outputs: future trajectory plans, reasoning traces, grounded answers to questions about the scenes, and auto label generation. The model weights are now available for anyone to download on Hugging Face, and the inference code has been posted to GitHub. Distilled models can be deployed commercially without any further permission from Nvidia, and model outputs carry no license conditions. Automakers can distill down a compact version of this model that can run on the Nvidia computer in the car. Major kudos to Nvidia and Jensen Huang for advancing the state of the industry by releasing this as an open model with permissive licensing. Jensen isn't just paying lip service to the idea of open models, Nvidia is actually contributing to the ecosystem — and it's great for their business, because it helps sell more Thor computers that go in the car. Anyone can go download the model and play with it. If you do, let me know what you think. Personally I think it's so cool that we have open weights models that are this advanced, for anyone to download.

Whole Mars Catalog

45,595 views • 1 month ago

OpenAI just announced API access to o1 (advanced reasoning model) yesterday. I'm delighted to announce today a new short course, Reasoning with o1, built with OpenAI, and taught by Colin Jarvis, Head of AI Solutions at OpenAI, to show you how to use this effectively! Unlike previous language models which generate output directly, o1 “thinks before it responds,” and generates many reasoning tokens before returning a more thoughtful and accurate response. It is great at complex reasoning -- including planning for agentic workflows, coding, and domain-specific reasoning in STEM fields like law. But how you should use it is quite different from other LLMs. I think o1 will be a game changer for many AI applications; and in this course, you'll learn how to use it effectively. In detail, you’ll: - Learn to recognize what tasks o1 is suited for, and when to use a smaller model, or combine o1 with a smaller model - Understand the new principles of prompting reasoning models: Be simple and direct; no explicit chain-of-thought required; use structure; show rather than tell - Implement multi-step orchestration in which o1 plans, and hands tasks over to gpt-4o-mini to execute specific steps; this illustrates a design pattern to optimize intelligence (accuracy) and cost - Use o1 for a coding task to build a new application, edit existing code, and test performance by running a coding competition between o1-mini and GPT 4o - Use o1 for image understanding and learn how it performs better with a "hierarchy of reasoning," in which it incurs the latency and cost upfront, preprocessing the image and indexing it with rich details so it can be used for Q&A later - Learn a technique called meta-prompting, in which you use o1 to improve your prompts. Using a customer support evaluation set, you'll iteratively use o1 to modify a prompt to improve performance You'll also learn about how OpenAI used reinforcement learning to produce a model that uses "test-time compute" to improve performance. I think you'll find this course enjoyable and valuable. Please sign up for it here:

OpenAI just announced API access to o1 (advanced reasoning model) yesterday. I'm delighted to announce today a new short course, Reasoning with o1, built with OpenAI, and taught by Colin Jarvis, Head of AI Solutions at OpenAI, to show you how to use this effectively! Unlike previous language models which generate output directly, o1 “thinks before it responds,” and generates many reasoning tokens before returning a more thoughtful and accurate response. It is great at complex reasoning -- including planning for agentic workflows, coding, and domain-specific reasoning in STEM fields like law. But how you should use it is quite different from other LLMs. I think o1 will be a game changer for many AI applications; and in this course, you'll learn how to use it effectively. In detail, you’ll: - Learn to recognize what tasks o1 is suited for, and when to use a smaller model, or combine o1 with a smaller model - Understand the new principles of prompting reasoning models: Be simple and direct; no explicit chain-of-thought required; use structure; show rather than tell - Implement multi-step orchestration in which o1 plans, and hands tasks over to gpt-4o-mini to execute specific steps; this illustrates a design pattern to optimize intelligence (accuracy) and cost - Use o1 for a coding task to build a new application, edit existing code, and test performance by running a coding competition between o1-mini and GPT 4o - Use o1 for image understanding and learn how it performs better with a "hierarchy of reasoning," in which it incurs the latency and cost upfront, preprocessing the image and indexing it with rich details so it can be used for Q&A later - Learn a technique called meta-prompting, in which you use o1 to improve your prompts. Using a customer support evaluation set, you'll iteratively use o1 to modify a prompt to improve performance You'll also learn about how OpenAI used reinforcement learning to produce a model that uses "test-time compute" to improve performance. I think you'll find this course enjoyable and valuable. Please sign up for it here:

Andrew Ng

358,273 views • 1 year ago

Two years ago, we built OSWorld 1.0 — the benchmark that became the standard for computer-use agents. Agents now score 83.5% on it. Problem solved? Not even close. 🚀Today we introduce OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks. What's new: 🎯 108 real-world workflows, each ~1.6 hours ⏱️ for a skilled human ⚙️ ~318 tool calls/task vs. ~30 in OSWorld 1.0 🌍 Grounded in authentic artifacts & stateful user profiles ⚡ Captures real phenomena: dynamic environments, streaming interaction, cross-source reasoning, implicit-state inference & more 📊 Best results: Claude Opus 4.8 reaches the highest accuracy at 20.6%, while GPT-5.5 is far more token-efficient but plateaus near 13%. No one is close to solving real computer use. 🏠 Homepage: 📄 Paper: 💻 Code: 🤗 Dataset: 🧵 [1/8]

Two years ago, we built OSWorld 1.0 — the benchmark that became the standard for computer-use agents. Agents now score 83.5% on it. Problem solved? Not even close. 🚀Today we introduce OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks. What's new: 🎯 108 real-world workflows, each ~1.6 hours ⏱️ for a skilled human ⚙️ ~318 tool calls/task vs. ~30 in OSWorld 1.0 🌍 Grounded in authentic artifacts & stateful user profiles ⚡ Captures real phenomena: dynamic environments, streaming interaction, cross-source reasoning, implicit-state inference & more 📊 Best results: Claude Opus 4.8 reaches the highest accuracy at 20.6%, while GPT-5.5 is far more token-efficient but plateaus near 13%. No one is close to solving real computer use. 🏠 Homepage: 📄 Paper: 💻 Code: 🤗 Dataset: 🧵 [1/8]

XLANG NLP Lab

172,141 views • 3 months ago

Inkling, Thinking Machines' first open model, dropped today: 975B total / 41B active MoE, up to 1M context, reasoning natively over text, images, and audio. Serving and RL support are already live: you can run and shape it on an open stack, starting now. Day 0 support on SGLang SGLang and Miles RadixArk👇 - Inkling's new architecture (ShortConv, attention with relative positional embedding, shared expert sink MoE) is natively implemented and deeply optimized, with prefill full CUDA graph and MXFP8 KV cache - Full parameter and LoRA RL in a customized Megatron backend, train inference consistency via customized kernels, routing replay, and cross-runtime parameter synchronization - DFlash speculative decoding from Modal for low-latency serving Launching now, blog and cookbook in the comments ⬇️

Inkling, Thinking Machines' first open model, dropped today: 975B total / 41B active MoE, up to 1M context, reasoning natively over text, images, and audio. Serving and RL support are already live: you can run and shape it on an open stack, starting now. Day 0 support on SGLang SGLang and Miles RadixArk👇 - Inkling's new architecture (ShortConv, attention with relative positional embedding, shared expert sink MoE) is natively implemented and deeply optimized, with prefill full CUDA graph and MXFP8 KV cache - Full parameter and LoRA RL in a customized Megatron backend, train inference consistency via customized kernels, routing replay, and cross-runtime parameter synchronization - DFlash speculative decoding from Modal for low-latency serving Launching now, blog and cookbook in the comments ⬇️

LMSYS Org

146,536 views • 2 months ago

Qwopus on a single RTX 3090. Claude Opus 4.6 reasoning distilled into Qwen 3.5 27B dense, running through Claude's own coding agent (claude code). 29-35 tok/s with thinking mode on. the jinja bug that kills thinking on base Qwen doesn't carry over. harness and model matched. the base model would pause mid task on Claude Code. just stop generating. that's why i ran it through OpenCode, which handles stalled states automatically. this distilled version doesn't stall. it waits for tool outputs, reads them, selfcorrects when something breaks, and keeps going. i gave it a benchmark analysis task. went 9 minutes autonomous. wrote a README nobody asked for. zero steering. video is 5x speed but fully uncut. if you have a 3090, you can run this right now. free. no API. no subscription. opus structured reasoning on localhost. octopus invaders is next. same prompt that base qwen passed in 13 minutes and hermes 4.3 failed on 2x the hardware. i want to see if the distillation changes the outcome or just the style. more data soon.

Qwopus on a single RTX 3090. Claude Opus 4.6 reasoning distilled into Qwen 3.5 27B dense, running through Claude's own coding agent (claude code). 29-35 tok/s with thinking mode on. the jinja bug that kills thinking on base Qwen doesn't carry over. harness and model matched. the base model would pause mid task on Claude Code. just stop generating. that's why i ran it through OpenCode, which handles stalled states automatically. this distilled version doesn't stall. it waits for tool outputs, reads them, selfcorrects when something breaks, and keeps going. i gave it a benchmark analysis task. went 9 minutes autonomous. wrote a README nobody asked for. zero steering. video is 5x speed but fully uncut. if you have a 3090, you can run this right now. free. no API. no subscription. opus structured reasoning on localhost. octopus invaders is next. same prompt that base qwen passed in 13 minutes and hermes 4.3 failed on 2x the hardware. i want to see if the distillation changes the outcome or just the style. more data soon.

Sudo su

296,392 views • 6 months ago

"A parcel with snacks has been delivered for Flexion. Retrieve it using the stairs and come up using the elevator. Then unpack it and place the items into the empty drawer on the shelf in the snack area." One instruction. No human operator. Everything that follows is autonomous. Today we're introducing Reflect v1.0, our robotics intelligence platform for long-horizon work. From a single natural-language command, the robot understands the task, navigates a multi-floor building, calls elevators, handles doors, uses tools to unpack a box, and puts the items away. The biggest shift in v1.0 is that we use reinforcement learning across every layer, from low-level control to high-level reasoning. Long-horizon autonomy is unforgiving. The robot must recover on its own when things don't go to plan because in the real world, they never do. Combining reasoning, perception, physical execution and runtime robustness into a single mission-capable system is the foundation required to solve humanoid autonomy. Our team is just getting started. #HumanoidRobots #Flexion

"A parcel with snacks has been delivered for Flexion. Retrieve it using the stairs and come up using the elevator. Then unpack it and place the items into the empty drawer on the shelf in the snack area." One instruction. No human operator. Everything that follows is autonomous. Today we're introducing Reflect v1.0, our robotics intelligence platform for long-horizon work. From a single natural-language command, the robot understands the task, navigates a multi-floor building, calls elevators, handles doors, uses tools to unpack a box, and puts the items away. The biggest shift in v1.0 is that we use reinforcement learning across every layer, from low-level control to high-level reasoning. Long-horizon autonomy is unforgiving. The robot must recover on its own when things don't go to plan because in the real world, they never do. Combining reasoning, perception, physical execution and runtime robustness into a single mission-capable system is the foundation required to solve humanoid autonomy. Our team is just getting started. #HumanoidRobots #Flexion

Flexion

143,392 views • 3 months ago

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.

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 views • 1 year ago

I have been testing DeepSeek-V4-Pro with the Pi coding agent. I am mindblown by how well it works out of the box. A few notes: I spent a few hours building an LLM wiki with an agent powered entirely by DeepSeek-V4-Pro on Fireworks inference. This is the first time I feel like there is an open-weight model that can reason at the level of Claude and Codex. And it does this in a cost-effective way with support for 1M context length. To be clear, I am using DeepSeek-V4-Pro inside of Pi without any special configuration. It works out of the box. It's exciting that there is a model that can just be plugged into a basic harness like Pi, and it just works. I've never seen that before. Most models require lots of configuration and setup. DeepSeek's DeepSeek-V4-Pro is clearly good at agentic coding (probably the best from the open-weight models), but the model is also great on knowledge-intensive tasks where reasoning matters. The agent pulled agentic engineering best practices from different company docs (Anthropic, OpenAI, Google, Stripe, Meta, Modal, DeepSeek, Mistral, Cohere), searched and digested Reddit and HN threads, summarized arxiv papers, and surfaced trending GitHub repos. Then it distilled everything into actionable tips across categories. I love the Wiki it built. The quality is really good. Here is a snapshot of what the wiki looks like: DeepSeek-V4-Pro handled the task without breaking stride. Multi-step research queries, code generation for scaffolding, context-heavy reasoning across disparate sources. For coding specifically, this is the first open-weight model that genuinely feels like a Codex or Claude Code experience. It compares in capability and actual multi-turn agentic work. What made the loop feel so responsive was Fireworks' inference speed (the fastest in the market) and the fact that they actually validate models at the systems level before shipping. No corrupted reasoning traces. Just fast, reliable iteration. The hybrid CSA and HCA attention design cuts KV cache to just 10% and inference FLOPs by nearly 4x at 1M-token context. This is what makes the agent loop actually fast and cheap enough to run in practice. For devs who've been watching open-weight models close the gap but haven't found one that actually delivers in practice, this is the closest I've seen. Try it here:

I have been testing DeepSeek-V4-Pro with the Pi coding agent. I am mindblown by how well it works out of the box. A few notes: I spent a few hours building an LLM wiki with an agent powered entirely by DeepSeek-V4-Pro on Fireworks inference. This is the first time I feel like there is an open-weight model that can reason at the level of Claude and Codex. And it does this in a cost-effective way with support for 1M context length. To be clear, I am using DeepSeek-V4-Pro inside of Pi without any special configuration. It works out of the box. It's exciting that there is a model that can just be plugged into a basic harness like Pi, and it just works. I've never seen that before. Most models require lots of configuration and setup. DeepSeek's DeepSeek-V4-Pro is clearly good at agentic coding (probably the best from the open-weight models), but the model is also great on knowledge-intensive tasks where reasoning matters. The agent pulled agentic engineering best practices from different company docs (Anthropic, OpenAI, Google, Stripe, Meta, Modal, DeepSeek, Mistral, Cohere), searched and digested Reddit and HN threads, summarized arxiv papers, and surfaced trending GitHub repos. Then it distilled everything into actionable tips across categories. I love the Wiki it built. The quality is really good. Here is a snapshot of what the wiki looks like: DeepSeek-V4-Pro handled the task without breaking stride. Multi-step research queries, code generation for scaffolding, context-heavy reasoning across disparate sources. For coding specifically, this is the first open-weight model that genuinely feels like a Codex or Claude Code experience. It compares in capability and actual multi-turn agentic work. What made the loop feel so responsive was Fireworks' inference speed (the fastest in the market) and the fact that they actually validate models at the systems level before shipping. No corrupted reasoning traces. Just fast, reliable iteration. The hybrid CSA and HCA attention design cuts KV cache to just 10% and inference FLOPs by nearly 4x at 1M-token context. This is what makes the agent loop actually fast and cheap enough to run in practice. For devs who've been watching open-weight models close the gap but haven't found one that actually delivers in practice, this is the closest I've seen. Try it here:

elvis

60,281 views • 5 months ago

On 11/6/2026 at 18:00 UTC, Michael Owen joined ForeGate for a World Cup Prediction Live. Missed the stream? Now is your chance to catch up on Michael Owen’s thoughts on the FIFA World Cup — from potential champions to key teams, match insights, and the reasoning behind his predictions. Michael Owen and our special guests shared their views on who could go all the way, what to watch during the tournament, and how fans can think about their own predictions. The World Cup has just started. It’s never too late to get involved. Watch the live replay, then make your prediction on ForeGate.

On 11/6/2026 at 18:00 UTC, Michael Owen joined ForeGate for a World Cup Prediction Live. Missed the stream? Now is your chance to catch up on Michael Owen’s thoughts on the FIFA World Cup — from potential champions to key teams, match insights, and the reasoning behind his predictions. Michael Owen and our special guests shared their views on who could go all the way, what to watch during the tournament, and how fans can think about their own predictions. The World Cup has just started. It’s never too late to get involved. Watch the live replay, then make your prediction on ForeGate.

ForeGate

16,682 views • 3 months ago

OpenAI just launched GPT-Live, and this is what it really means: They have made ChatGPT Voice sound like a human being instead of slow robot. “Listen while it speaks”. Do you remember how you used to wait until the AI finished speaking? Now you can interrupt it with “Wait, no, I mean…” and it will accept your request and keep going. "Asynchronous reasoning" It solves the problem of the "awkward pause". The developers divided the brain of the AI into two parts: one for communication and the other for doing calculations and looking for solutions in the background. Finally, it can do everything at once. Connects to GPT-5.5 If more advanced reasoning or the use of tools is required, GPT-Live can also can also contact our cutting-edge models, like GPT-5.5, without interrupting the flow of the conversation. “6 round trips to 1”. Think of round trips as the digital handshake. Previously, your phone had to go through six handshake round trips before saying hello. Now, it only needs to do it once and it can greet you right away.

OpenAI just launched GPT-Live, and this is what it really means: They have made ChatGPT Voice sound like a human being instead of slow robot. “Listen while it speaks”. Do you remember how you used to wait until the AI finished speaking? Now you can interrupt it with “Wait, no, I mean…” and it will accept your request and keep going. "Asynchronous reasoning" It solves the problem of the "awkward pause". The developers divided the brain of the AI into two parts: one for communication and the other for doing calculations and looking for solutions in the background. Finally, it can do everything at once. Connects to GPT-5.5 If more advanced reasoning or the use of tools is required, GPT-Live can also can also contact our cutting-edge models, like GPT-5.5, without interrupting the flow of the conversation. “6 round trips to 1”. Think of round trips as the digital handshake. Previously, your phone had to go through six handshake round trips before saying hello. Now, it only needs to do it once and it can greet you right away.

Chidanand Tripathi

27,325 views • 1 month ago

A voice agent powered by gpt-oss. Running locally on my macBook. Demo recorded in a Waymo with WiFi turned off. I'm still on my space game voice AI kick, obviously. Code link below. For conversational voice AI, you want to set the gpt-oss reasoning behavior to "low". (The default is "medium".) Notes on how to do that and a jinja template you can use are in the repo. The LLM in the demo video is the big, 120B version of gpt-oss. You can use the smaller, 20B model for this, of course. But OpenAI really did a cool thing here designing the 120B model to run in "just" 80GB of VRAM. And the llama.cpp mlx inference is fast: ~250ms TTFT. Running a big model on-device feels like a time warp into the future of AI.

A voice agent powered by gpt-oss. Running locally on my macBook. Demo recorded in a Waymo with WiFi turned off. I'm still on my space game voice AI kick, obviously. Code link below. For conversational voice AI, you want to set the gpt-oss reasoning behavior to "low". (The default is "medium".) Notes on how to do that and a jinja template you can use are in the repo. The LLM in the demo video is the big, 120B version of gpt-oss. You can use the smaller, 20B model for this, of course. But OpenAI really did a cool thing here designing the 120B model to run in "just" 80GB of VRAM. And the llama.cpp mlx inference is fast: ~250ms TTFT. Running a big model on-device feels like a time warp into the future of AI.

kwindla

202,211 views • 1 year ago

MolmoAct2 runs zero-shot on an SO-ARM101, no training required 🦾 Ai2 open Action Reasoning Model is now in LeRobot with the full lifecycle covered: fine-tuning (full or LoRA), evaluation, and real-robot deployment. It pairs a Molmo2-ER vision-language backbone with a flow-matching action expert to turn images, language, and proprioceptive state into action chunks. The best part: ready-made checkpoints ship with calibration correction baked in, so you can point it at your SO-100/101 and just watch it work. Inference fits in ~12GB at bf16, and LoRA fine-tuning runs on a single 24GB GPU. Big thanks to the Ai2 team for building this in the open.

MolmoAct2 runs zero-shot on an SO-ARM101, no training required 🦾 Ai2 open Action Reasoning Model is now in LeRobot with the full lifecycle covered: fine-tuning (full or LoRA), evaluation, and real-robot deployment. It pairs a Molmo2-ER vision-language backbone with a flow-matching action expert to turn images, language, and proprioceptive state into action chunks. The best part: ready-made checkpoints ship with calibration correction baked in, so you can point it at your SO-100/101 and just watch it work. Inference fits in ~12GB at bf16, and LoRA fine-tuning runs on a single 24GB GPU. Big thanks to the Ai2 team for building this in the open.

LeRobot

23,339 views • 2 months ago

NEWS: Waymo has released a new blog post detailing their AI strategy and how it’s allowing them to bring service to more riders faster. "Achieving demonstrably safe AI — where safety is proven, not just promised — requires a holistic approach. Beyond a smart and capable Driver, you also need a closed-loop, realistic Simulator to train and rigorously test the Driver in a myriad of challenging situations, and a sharp Critic to evaluate the Driver's performance and identify areas for improvement." Waymo says that autonomous driving isn’t just a matter of building a “smart driver,” but rather creating a full AI ecosystem centered on safety from the ground up. At the core is the Waymo Foundation Model, a unified world-model that powers all major components of Waymo’s autonomous stack (Driver, Simulator, Critic). "By using a “Think Fast/Think Slow” architecture (combining rapid sensor-fusion with deep semantic reasoning), this system enables the car to detect complex and rare road scenarios (e.g. a burning vehicle ahead), reason about them, and choose safe behavior. Waymo trains large “Teacher” AI-models for driving, simulation, and evaluation, then distills them into smaller, efficient “Student” models suitable for real-world deployment, while keeping safety validation tightly integrated. The result is a continuous “flywheel” of learning: driving data (real and simulated) generate feedback, which leads to refinements, more simulation, more data, and only when safety checks pass is new code deployed. Having already exceeded 100 million fully autonomous miles, Waymo reports a more than ten-fold reduction in severe-injury crashes compared to human drivers." Full blog post:

NEWS: Waymo has released a new blog post detailing their AI strategy and how it’s allowing them to bring service to more riders faster. "Achieving demonstrably safe AI — where safety is proven, not just promised — requires a holistic approach. Beyond a smart and capable Driver, you also need a closed-loop, realistic Simulator to train and rigorously test the Driver in a myriad of challenging situations, and a sharp Critic to evaluate the Driver's performance and identify areas for improvement." Waymo says that autonomous driving isn’t just a matter of building a “smart driver,” but rather creating a full AI ecosystem centered on safety from the ground up. At the core is the Waymo Foundation Model, a unified world-model that powers all major components of Waymo’s autonomous stack (Driver, Simulator, Critic). "By using a “Think Fast/Think Slow” architecture (combining rapid sensor-fusion with deep semantic reasoning), this system enables the car to detect complex and rare road scenarios (e.g. a burning vehicle ahead), reason about them, and choose safe behavior. Waymo trains large “Teacher” AI-models for driving, simulation, and evaluation, then distills them into smaller, efficient “Student” models suitable for real-world deployment, while keeping safety validation tightly integrated. The result is a continuous “flywheel” of learning: driving data (real and simulated) generate feedback, which leads to refinements, more simulation, more data, and only when safety checks pass is new code deployed. Having already exceeded 100 million fully autonomous miles, Waymo reports a more than ten-fold reduction in severe-injury crashes compared to human drivers." Full blog post:

Sawyer Merritt

83,635 views • 9 months ago

I stopped trusting AI outputs blindly. Not because AI is bad… But because I started noticing a pattern. The more I used AI for serious work, the more I had to: → double-check everything → re-verify sources → question the logic behind answers At some point, I thought: “What’s the point of saving time… if I still don’t trust the output?” That’s when I came across MiroMindAI and it felt different from day one. Not flashy. Not trying to impress. Just… built for accuracy. I tested it the same way I test any tool: Real use cases. No hype. • Deep research • Multi-source validation • Complex reasoning tasks And here’s what stood out 👇 🧠 It shows how it thinks Not just answers actual reasoning chains you can read, audit, and replay. 🔍 It doesn’t “summarize”… it investigates Pulls from hundreds of sources and builds structured, evidence-backed reports. ⚖️ It verifies itself before responding Multiple layers checking the output (something most AI tools skip completely) And honestly, this is what clicked for me: Most AI tools today are like 👉 smart interns (fast, helpful, but need supervision) MiroMind feels more like 👉 a senior analyst (slower, but you can rely on it) 💡 One simple shift I noticed: Before: 10 tabs open → cross-checking → still unsure Now: 1 report → clear reasoning → backed by sources I’m not saying this replaces expertise. But it does reduce the noise. A lot. If you’re someone who works in: • research • finance • legal • healthcare You’ll probably appreciate this more than others. 👉

I stopped trusting AI outputs blindly. Not because AI is bad… But because I started noticing a pattern. The more I used AI for serious work, the more I had to: → double-check everything → re-verify sources → question the logic behind answers At some point, I thought: “What’s the point of saving time… if I still don’t trust the output?” That’s when I came across MiroMindAI and it felt different from day one. Not flashy. Not trying to impress. Just… built for accuracy. I tested it the same way I test any tool: Real use cases. No hype. • Deep research • Multi-source validation • Complex reasoning tasks And here’s what stood out 👇 🧠 It shows how it thinks Not just answers actual reasoning chains you can read, audit, and replay. 🔍 It doesn’t “summarize”… it investigates Pulls from hundreds of sources and builds structured, evidence-backed reports. ⚖️ It verifies itself before responding Multiple layers checking the output (something most AI tools skip completely) And honestly, this is what clicked for me: Most AI tools today are like 👉 smart interns (fast, helpful, but need supervision) MiroMind feels more like 👉 a senior analyst (slower, but you can rely on it) 💡 One simple shift I noticed: Before: 10 tabs open → cross-checking → still unsure Now: 1 report → clear reasoning → backed by sources I’m not saying this replaces expertise. But it does reduce the noise. A lot. If you’re someone who works in: • research • finance • legal • healthcare You’ll probably appreciate this more than others. 👉

Md Riyazuddin

20,990 views • 5 months ago

hey if you're thinking about running qwopus (the claude opus distilled qwen 3.5 27B) as a coding agent, this might save you a few hours. i tested both the base and the distilled version on the same hardware. single RTX 3090. same prompt. same context. same everything. the only variable was the model weights. base qwen 3.5 27B built octopus invaders in 13 minutes. 1,827 lines across 11 files. zero steering. one scope bug that took 2 lines to fix. game ran. qwopus couldn't finish the same task. enemies overlapping on screen. bullets not firing. controls worked but the game was broken. i had to steer it multiple times and it still didn't produce a playable result. both run at 35 tok/s. both use thinking mode. the distilled version actually has better jinja compatibility and doesn't stall midtask like base does on claude code. for conversation and reasoning it feels sharper. but for multifile autonomous coding where the model needs to coordinate 10+ files without losing track, base wins and it's not close. distillation compresses reasoning patterns but seems to lose precision on complex coordination. the model "thinks" well but can't hold the full picture across files the way base can. tested on opencode (base) and claude code (both). next up is hermes agent framework on base. same hardware. same prompt. comparing agents now, not just models. video below. first half is the distilled model's broken game. second half is what base built on the same 3090. judge for yourself.

hey if you're thinking about running qwopus (the claude opus distilled qwen 3.5 27B) as a coding agent, this might save you a few hours. i tested both the base and the distilled version on the same hardware. single RTX 3090. same prompt. same context. same everything. the only variable was the model weights. base qwen 3.5 27B built octopus invaders in 13 minutes. 1,827 lines across 11 files. zero steering. one scope bug that took 2 lines to fix. game ran. qwopus couldn't finish the same task. enemies overlapping on screen. bullets not firing. controls worked but the game was broken. i had to steer it multiple times and it still didn't produce a playable result. both run at 35 tok/s. both use thinking mode. the distilled version actually has better jinja compatibility and doesn't stall midtask like base does on claude code. for conversation and reasoning it feels sharper. but for multifile autonomous coding where the model needs to coordinate 10+ files without losing track, base wins and it's not close. distillation compresses reasoning patterns but seems to lose precision on complex coordination. the model "thinks" well but can't hold the full picture across files the way base can. tested on opencode (base) and claude code (both). next up is hermes agent framework on base. same hardware. same prompt. comparing agents now, not just models. video below. first half is the distilled model's broken game. second half is what base built on the same 3090. judge for yourself.

Sudo su

45,052 views • 6 months ago