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.

After fixing a few bugs with inference code, I finally have a working pi0 set up! Fine-tuned overnight on single task data, it learns rough controls and no longer outputs unsafe actions. But mastering the task and form factor will require much more compute. Stay tuned!

Ville🤖

2,645 subscribers

33,242 views • 1 year ago •via X (Twitter)

Science & Technology Education

Anya Rossi• Live Now

Private livecam show

11 Comments

Ville 🤖1 year ago

vibe checks - still struggling with instruction following like most VLAs. (Instruction comes from text, the speech is just for the demo and the robot won’t listen to that)

opensourceCM1 year ago

What’s the cost of mistakes in your contracts? If you work with contracts day-to-day, it’s time to automate. Track every detail, streamline workflows ... ✨ Make managing contracts as easy as a few clicks. Visit our new website & book your demo today!

Stone Tao1 year ago

success rate / how many fine tuning demos?

Ville 🤖1 year ago

0% I need to fine-tune much longer to do justice for this model 😅

Vaishak V Kumar1 year ago

What changed?

Ville 🤖1 year ago

fixed several bugs with the infra and re-trained the model - not 100% sure what parts mattered and what didn't robotics is hard

David Bar1 year ago

Cool

SuperMario1 year ago

Great work!

Eugene Mironov1 year ago

Looks cool, looking forward to try it too

Ville 🤖1 year ago

I'm looking forward to trying your RL code once ready!

Pierre-Louis Biojout1 year ago

How much episodes in your dataset? Given the size of the model, without some little tricks you need a decent amount of datapoints

Related Videos

Please retweet this and spread the word! After many hard weeks of finishing and working, we finally have ✨Vs Kyu 1.1✨! It fixed a lot of things and added a few more elements! Stay tuned! FNF News Funkin News CommunityGame

Please retweet this and spread the word! After many hard weeks of finishing and working, we finally have ✨Vs Kyu 1.1✨! It fixed a lot of things and added a few more elements! Stay tuned! FNF News Funkin News CommunityGame

Kyu

149,720 views • 3 years ago

$QVAC SDK will support in 0.9.0 (gonna be release in ~10 days) LoRA fine-tuning directly on-device, letting developers customize LLMs with their own data without sending anything to the cloud. You just load a base model, point it at your training dataset, and get a lightweight LoRA adapter back — all running locally. The fine-tuned model can then be used for inference immediately, with no extra setup. Why it matters: LoRA (Low-Rank Adaptation) fine-tuning lets you specialize a general-purpose language model for your specific use case — like matching a brand's tone, mastering domain terminology, or following a particular output format — using a fraction of the compute a full fine-tune would require. QVAC handles the entire workflow locally: dataset preparation, training with configurable hyperparameters, checkpoint saving, and seamless inference with the resulting adapter. Your data never leaves the device. The developer experience: Fine-tuning with QVAC is as simple as calling "sdk.finetune()" with your dataset and a few hyperparameters. Training runs entirely on your local hardware, produces a compact LoRA adapter file, and supports pause/resume so you can stop a job and pick it back up without losing progress. The result plugs straight into QVAC's inference pipeline — no model conversion, no deployment step, just immediate local completions with your fine-tuned model.$

QVAC SDK will support in 0.9.0 (gonna be release in ~10 days) LoRA fine-tuning directly on-device, letting developers customize LLMs with their own data without sending anything to the cloud. You just load a base model, point it at your training dataset, and get a lightweight LoRA adapter back — all running locally. The fine-tuned model can then be used for inference immediately, with no extra setup. Why it matters: LoRA (Low-Rank Adaptation) fine-tuning lets you specialize a general-purpose language model for your specific use case — like matching a brand's tone, mastering domain terminology, or following a particular output format — using a fraction of the compute a full fine-tune would require. QVAC handles the entire workflow locally: dataset preparation, training with configurable hyperparameters, checkpoint saving, and seamless inference with the resulting adapter. Your data never leaves the device. The developer experience: Fine-tuning with QVAC is as simple as calling "sdk.finetune()" with your dataset and a few hyperparameters. Training runs entirely on your local hardware, produces a compact LoRA adapter file, and supports pause/resume so you can stop a job and pick it back up without losing progress. The result plugs straight into QVAC's inference pipeline — no model conversion, no deployment step, just immediate local completions with your fine-tuned model.

Paolo Ardoino 🤖

42,413 views • 3 months ago

OpenLora is the designated deployment engine for specialized models in the OpenLedger ecosystem. By leveraging just-in-time adapter switching, OpenLora enables the efficient serving of thousands of fine-tuned LoRA adapters on a single GPU, drastically reducing deployment costs. Unlike generic models, OpenLora-powered specialized models require fewer input tokens and produce more optimized, precise outputs - minimizing both input and output token size while delivering task-specific performance at scale. Faster, Smarter, and Specialized.

OpenLora is the designated deployment engine for specialized models in the OpenLedger ecosystem. By leveraging just-in-time adapter switching, OpenLora enables the efficient serving of thousands of fine-tuned LoRA adapters on a single GPU, drastically reducing deployment costs. Unlike generic models, OpenLora-powered specialized models require fewer input tokens and produce more optimized, precise outputs - minimizing both input and output token size while delivering task-specific performance at scale. Faster, Smarter, and Specialized.

OpenLedger

76,396 views • 1 year ago

The problem with humanoid teleoperation is that it is expensive and difficult to scale Enter NVIDIA's EgoScale: - A VLA model pretrained on thousands hours of egocentric human videos. - Mid-trained via 50 hours of human + 4 hours of robot "play" data for human-robot alignment. - Fine-tuned with very few examples of task-specific robot teleoperation (100 or fewer per task). - Successfully transfers across 5-finger (Sharpa) and 3-finger (Unitree G1) robot hands. - Performance scales predictably as data increases.

The problem with humanoid teleoperation is that it is expensive and difficult to scale Enter NVIDIA's EgoScale: - A VLA model pretrained on thousands hours of egocentric human videos. - Mid-trained via 50 hours of human + 4 hours of robot "play" data for human-robot alignment. - Fine-tuned with very few examples of task-specific robot teleoperation (100 or fewer per task). - Successfully transfers across 5-finger (Sharpa) and 3-finger (Unitree G1) robot hands. - Performance scales predictably as data increases.

The Humanoid Hub

44,441 views • 5 months ago

Short update on the T Unit progress as some people have been waiting. Basically, most if not all the graphics are complete, however, there are some palette corruption bugs, not sure what it is about, but will take a look. There's some flickering, maybe due to performance related issues, but will investigate. Sound partially working on the ADPCM, still have a bit to do on the DCS sound board though, been working for a few months on it, it is pretty large. Stay Tuned! #CoinOpCollection #Midway

Short update on the T Unit progress as some people have been waiting. Basically, most if not all the graphics are complete, however, there are some palette corruption bugs, not sure what it is about, but will take a look. There's some flickering, maybe due to performance related issues, but will investigate. Sound partially working on the ADPCM, still have a bit to do on the DCS sound board though, been working for a few months on it, it is pretty large. Stay Tuned! #CoinOpCollection #Midway

Pramod Somashekar

18,042 views • 5 months ago

NVIDIA has published a paper on DREAMGEN – a powerful 4-step pipeline for generating synthetic data for humanoids that enables task and environment generalization. - Step 1: Fine-tune a video generation model using a small number of human teleoperation videos - Step 2: Prompt the fine-tuned model to turn a single real image into new AI-imagined videos - Step 3: Automatically label actions in the generated videos - Step 4: Train a robot AI model with the labeled synthetic dataset This enabled humanoid robots to perform 22 novel behaviors – such as pouring, opening/closing articulated objects, and manipulating a variety of tools. The original teleoperation dataset only included pick-and-place tasks. This takes task extensibility to another level without requiring human teleoperation for every single task. The pipeline will be made open-source soon. Project page:

NVIDIA has published a paper on DREAMGEN – a powerful 4-step pipeline for generating synthetic data for humanoids that enables task and environment generalization. - Step 1: Fine-tune a video generation model using a small number of human teleoperation videos - Step 2: Prompt the fine-tuned model to turn a single real image into new AI-imagined videos - Step 3: Automatically label actions in the generated videos - Step 4: Train a robot AI model with the labeled synthetic dataset This enabled humanoid robots to perform 22 novel behaviors – such as pouring, opening/closing articulated objects, and manipulating a variety of tools. The original teleoperation dataset only included pick-and-place tasks. This takes task extensibility to another level without requiring human teleoperation for every single task. The pipeline will be made open-source soon. Project page:

The Humanoid Hub

12,074 views • 1 year ago

While working on new update I figured out a completely new way of writing text on prop in much cheaper way and which allows more cool stuff :) Stay tuned, soon new update will be released. Just need to tweak some things when I have time.

While working on new update I figured out a completely new way of writing text on prop in much cheaper way and which allows more cool stuff :) Stay tuned, soon new update will be released. Just need to tweak some things when I have time.

AsicsoN

13,697 views • 2 months ago

In just one week, Binh Pham and I trained a full-body Unitree G1. Here's a recap: 1. Secured a Unitree G1 humanoid through a LinkedIn post 2. Deployed TWIST2 full-body teleoperation pipelines 3. Adapted TWIST2 for Zed stereo camera & collected full-body teleoperation samples (carried by Binh Pham ) 4. Adapted & fine-tuned NVIDIA Gr00T N1.5 VLA on the TWIST2 public datasets, which I fine-tuned on an 8xNVIDIA H100 Cluster. We picked Gr00T N1.5 as it was trained with Unitree G1 embodiment data. 5. Adapted the TWIST2 codebase to stream in the actions from Gr00T via ZMQ using a co-located NVIDIA H100 for ~200ms inference latency 6. Tested the model in sim, then deployed to the real-world Unitree G1. We streamed a training sample observation to the VLA (as we didn't want to break robot in case real observations were OOD) We were the first team in the world to deploy the full TWIST2 data collection pipeline to the unitree g1 :) Much more work ahead though, which I'll work on as a side-project over the next months: 1. Exploring the various types of 'world models': video backbones, dynamics models, v-jepa-2 models. I believe these will generalize better & train much more data-efficiently than VLM backbones 2. Speeding up inference - I believe low-latency robotics inference will be a big challenge. There are many works in video diffusion which I'd like to test (e.g. SageAttention, SparseAttention, Drifting Models). Perhaps also writing custom CUDA kernels. 3. Economics of inference scaling :) What will be the compute demands as we scale inference up to millions of humanoids? Will it run on edge or on distributed 'co-located' inference clusters? These are questions I'd like to answer. Adapted TWIST2 codebase: Adapted Gr00T-N1.5 codebase: The ETH Robotics Club are doing a cool GTC Golden ticket competition with NVIDIA , so this is my submission :) The DGX Spark compute will get me a long way with initial prototyping & especially working on inference optimization for next-gen Blackwell GPUs #NVIDIAGTC #GOLDENTICKET #ETHRC

In just one week, Binh Pham and I trained a full-body Unitree G1. Here's a recap: 1. Secured a Unitree G1 humanoid through a LinkedIn post 2. Deployed TWIST2 full-body teleoperation pipelines 3. Adapted TWIST2 for Zed stereo camera & collected full-body teleoperation samples (carried by Binh Pham ) 4. Adapted & fine-tuned NVIDIA Gr00T N1.5 VLA on the TWIST2 public datasets, which I fine-tuned on an 8xNVIDIA H100 Cluster. We picked Gr00T N1.5 as it was trained with Unitree G1 embodiment data. 5. Adapted the TWIST2 codebase to stream in the actions from Gr00T via ZMQ using a co-located NVIDIA H100 for ~200ms inference latency 6. Tested the model in sim, then deployed to the real-world Unitree G1. We streamed a training sample observation to the VLA (as we didn't want to break robot in case real observations were OOD) We were the first team in the world to deploy the full TWIST2 data collection pipeline to the unitree g1 :) Much more work ahead though, which I'll work on as a side-project over the next months: 1. Exploring the various types of 'world models': video backbones, dynamics models, v-jepa-2 models. I believe these will generalize better & train much more data-efficiently than VLM backbones 2. Speeding up inference - I believe low-latency robotics inference will be a big challenge. There are many works in video diffusion which I'd like to test (e.g. SageAttention, SparseAttention, Drifting Models). Perhaps also writing custom CUDA kernels. 3. Economics of inference scaling :) What will be the compute demands as we scale inference up to millions of humanoids? Will it run on edge or on distributed 'co-located' inference clusters? These are questions I'd like to answer. Adapted TWIST2 codebase: Adapted Gr00T-N1.5 codebase: The ETH Robotics Club are doing a cool GTC Golden ticket competition with NVIDIA , so this is my submission :) The DGX Spark compute will get me a long way with initial prototyping & especially working on inference optimization for next-gen Blackwell GPUs #NVIDIAGTC #GOLDENTICKET #ETHRC

Arnie Ramesh

14,815 views • 5 months ago

It’s been an incredibly exciting few weeks for all of us involved with #Alvara. We’re delighted to finally be up and running but the real work has just started. All eyes are firmly on the testnet launch in a few weeks. An exciting few months lay ahead. Stay tuned. $ALVA

It’s been an incredibly exciting few weeks for all of us involved with #Alvara. We’re delighted to finally be up and running but the real work has just started. All eyes are firmly on the testnet launch in a few weeks. An exciting few months lay ahead. Stay tuned. $ALVA

Alvara

31,911 views • 2 years ago

I predicted both #Overwatch and #RegularShow in #Fortnite 3 years ago! Now I hope #MichaelJackson can get a skin too, what do you think? I made a new Fortnite game mode in UEFN and will publish it in the next few days, stay tuned, I'm sure you'll have so much fun!

I predicted both #Overwatch and #RegularShow in #Fortnite 3 years ago! Now I hope #MichaelJackson can get a skin too, what do you think? I made a new Fortnite game mode in UEFN and will publish it in the next few days, stay tuned, I'm sure you'll have so much fun!

João Filipe Santiago

137,028 views • 2 months ago

Vibe coding is cool. But have you tried vibe-tuning? Here’s Claude Code using the Lightning Rod Labs SDK to turn messy Fed PDFs into a small, specialized economic forecaster. The agent: - parses raw historical PDFs - builds a training set - fine-tunes a smaller model - evaluates it against GPT-5 Result: a fine-tuned model that beats GPT-5 and runs anywhere, on a single GPU. Try it in your coding agent of choice.

Vibe coding is cool. But have you tried vibe-tuning? Here’s Claude Code using the Lightning Rod Labs SDK to turn messy Fed PDFs into a small, specialized economic forecaster. The agent: - parses raw historical PDFs - builds a training set - fine-tunes a smaller model - evaluates it against GPT-5 Result: a fine-tuned model that beats GPT-5 and runs anywhere, on a single GPU. Try it in your coding agent of choice.

Ben Turtel

423,542 views • 2 months ago

We’re excited to announce our partnership with Use Morpheus as they develop a specialized AI model on OpenLedger. Leveraging our decentralized infrastructure, Morpheus will create a fine-tuned Web3 LLM focused on secure code generation and autonomous workflows. This collaboration will push the boundaries of intelligent contract creation, enabling decentralized applications to scale smarter, faster, and more securely. The next wave of AI-powered smart contracts is here, stay tuned for more updates! Hear adamzhao.eth , founder of Use Morpheus, break it down in the video.

We’re excited to announce our partnership with Use Morpheus as they develop a specialized AI model on OpenLedger. Leveraging our decentralized infrastructure, Morpheus will create a fine-tuned Web3 LLM focused on secure code generation and autonomous workflows. This collaboration will push the boundaries of intelligent contract creation, enabling decentralized applications to scale smarter, faster, and more securely. The next wave of AI-powered smart contracts is here, stay tuned for more updates! Hear adamzhao.eth , founder of Use Morpheus, break it down in the video.

OpenLedger

17,229 views • 1 year ago

VIDEO | Delhi: Anurag Singh, Lieutenant Colonel, Officer Commanding, Yamuna Task Force, says, "The Indian Army and the Government of India's Ministry of Defence decided to form the Yamuna Task Force here...The Indian Army's involvement in this brings with it people's expectations and hopes, and we will always live up to them." On being asked about the possibility of taking a holy dip, he says, "People will start taking a dip in the Yamuna from next year itself. But let me make it clear, this effort will require everyone's participation." (Full video available on PTI Videos - #Delhi

VIDEO | Delhi: Anurag Singh, Lieutenant Colonel, Officer Commanding, Yamuna Task Force, says, "The Indian Army and the Government of India's Ministry of Defence decided to form the Yamuna Task Force here...The Indian Army's involvement in this brings with it people's expectations and hopes, and we will always live up to them." On being asked about the possibility of taking a holy dip, he says, "People will start taking a dip in the Yamuna from next year itself. But let me make it clear, this effort will require everyone's participation." (Full video available on PTI Videos - #Delhi

Press Trust of India

88,262 views • 1 month ago

I was working on videos for UE 5.4 and wanted to give a sneak peak on Layered Control Rigs. You no longer need to bake to edit anim clips! Lots of fun uses with this. Stay tuned for the video and the example content! #UE5 #gamedev #ControlRig #animation #rigTip #Sequencer

I was working on videos for UE 5.4 and wanted to give a sneak peak on Layered Control Rigs. You no longer need to bake to edit anim clips! Lots of fun uses with this. Stay tuned for the video and the example content! #UE5 #gamedev #ControlRig #animation #rigTip #Sequencer

Greg Richardson

32,474 views • 2 years ago

PONDPHUWIN NIVEA SKINGLOW #NIVEASkinGlowXPondPhuwin #ปอนด์ภูวินทร์ R: It seems in the future u will collab/working together (with NIVEA) too Phu: Yes. Stay tuned to find out what event/activity will be like R: I heard just now the MC hv spill about it about the activity Pond: Just wait and see Phu: Might hv to stay tuned. But hv for sure, in a while. Might not anytime soon

PONDPHUWIN NIVEA SKINGLOW #NIVEASkinGlowXPondPhuwin #ปอนด์ภูวินทร์ R: It seems in the future u will collab/working together (with NIVEA) too Phu: Yes. Stay tuned to find out what event/activity will be like R: I heard just now the MC hv spill about it about the activity Pond: Just wait and see Phu: Might hv to stay tuned. But hv for sure, in a while. Might not anytime soon

Az91saifar :)

14,170 views • 2 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

⚡ Speed of flash. Just 2 days after launch, DFlash is already running in SGLang (SGLang). With serving-engine support, we can now unlock speedup with higher concurrency, and we’ve quickly worked on a new demo based on it. We'll be cooking up more and better draft models over the next few weeks.🔥 Stay tuned!

⚡ Speed of flash. Just 2 days after launch, DFlash is already running in SGLang (SGLang). With serving-engine support, we can now unlock speedup with higher concurrency, and we’ve quickly worked on a new demo based on it. We'll be cooking up more and better draft models over the next few weeks.🔥 Stay tuned!

Zhijian Liu

21,233 views • 6 months ago

i watched gemma 4 12b build something genuinely impressive today, and then loop itself to death right in front of me. the full run is in the video, sped up but completely uncut, watch it to the end and you will catch the exact moment it stops building and starts looping right in the middle of the work. the task was clean, build a single file gravity simulator, n-body physics, orbits, collisions, running locally on one 3090 through an agent. and for ten minutes it was a joy to watch. it reached for a symplectic integrator on its own, the correct one, the kind that keeps orbits stable instead of spiralling out. real gravity with softening, proper orbital velocities, momentum conserved on collision. the physics was right. the thing actually worked. then on the very last step, writing a few tests to prove its own code, it fell into a loop. not a crash, a loop. it started repeating itself and would not stop. ten more minutes, thirty four thousand tokens into a single answer, the same fragments over and over, until i killed it myself. so it's not that gemma can't code. it did the hard part beautifully. it cannot finish. it cannot hold a long task together without unravelling, and finishing is the entire job in agentic work. here's the part that stings. i run this exact task, same harness, same card, on the chinese open models, qwen especially, and i never see this. they build it, they test it, they stop. every single time. google has the raw capability, you can see it sitting right there in the code, and then the model loops itself to death on a task a 27b from alibaba finishes clean. open weights, apache 2.0, so much to love on paper. i just need it to know when to stop talking.

i watched gemma 4 12b build something genuinely impressive today, and then loop itself to death right in front of me. the full run is in the video, sped up but completely uncut, watch it to the end and you will catch the exact moment it stops building and starts looping right in the middle of the work. the task was clean, build a single file gravity simulator, n-body physics, orbits, collisions, running locally on one 3090 through an agent. and for ten minutes it was a joy to watch. it reached for a symplectic integrator on its own, the correct one, the kind that keeps orbits stable instead of spiralling out. real gravity with softening, proper orbital velocities, momentum conserved on collision. the physics was right. the thing actually worked. then on the very last step, writing a few tests to prove its own code, it fell into a loop. not a crash, a loop. it started repeating itself and would not stop. ten more minutes, thirty four thousand tokens into a single answer, the same fragments over and over, until i killed it myself. so it's not that gemma can't code. it did the hard part beautifully. it cannot finish. it cannot hold a long task together without unravelling, and finishing is the entire job in agentic work. here's the part that stings. i run this exact task, same harness, same card, on the chinese open models, qwen especially, and i never see this. they build it, they test it, they stop. every single time. google has the raw capability, you can see it sitting right there in the code, and then the model loops itself to death on a task a 27b from alibaba finishes clean. open weights, apache 2.0, so much to love on paper. i just need it to know when to stop talking.

Sudo su

39,574 views • 1 month ago