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.

Decentralized AI in SN29 @Coldint_ - Splitting massive models across GPUs, enabling scalability. - Efficient validation makes training smarter without massive hardware costs. - Gradual model growth builds on smaller models, saving compute. Watch the full episode here:

Openτensor Foundaτion

173,884 subscribers

28,057 views • 1 year ago •via X (Twitter)

Science & Technology News & Politics Education

Anya Rossi• Live Now

Private livecam show

6 Comments

Sébastien 2.01 year ago

@Coldint_ Can you share more about how SN29's gradual model growth feature works? I'm curious to learn more about its potential to save compute.

Musti ⚔️1 year ago

@Coldint_ Model sharding has been been part of Nesa for example since months now. Nothing new at all

nordin.eth1 year ago

@Coldint_ Game changing!

Ales1 year ago

@Coldint_ 🤝

Mikhail Pashetnykh1 year ago

@Coldint_ Fix mobile wallet application please.

TAOAgent1 year ago

@Coldint_ 😻

Related Videos

Checkpointing saves the state of an AI model during training so you can resume or recover without starting over. But writing massive checkpoints fast enough to keep GPUs busy is a real challenge. Below, David Flynn describes how Hammerspace makes checkpointing efficient

Checkpointing saves the state of an AI model during training so you can resume or recover without starting over. But writing massive checkpoints fast enough to keep GPUs busy is a real challenge. Below, David Flynn describes how Hammerspace makes checkpointing efficient

Jamin Ball

11,079 views • 1 year ago

In a conversation with Kenzi Nakamura, Ron Bodkin explains how swarm protocols are the glue of the decentralized AI future — enabling agents to: — Communicate across platforms — Share data seamlessly — Execute trades regardless of what tech stack they’re built on This isn’t just about smarter agents — it’s about coordinated intelligence at scale. Watch the full episode here:

In a conversation with Kenzi Nakamura, Ron Bodkin explains how swarm protocols are the glue of the decentralized AI future — enabling agents to: — Communicate across platforms — Share data seamlessly — Execute trades regardless of what tech stack they’re built on This isn’t just about smarter agents — it’s about coordinated intelligence at scale. Watch the full episode here:

Web3Unchained Podcast 🎙️

45,003 views • 10 months ago

AI-driven design is already reshaping compute. We are transforming how we design chips, with software and hardware evolving together. This approach is accelerating innovation and enabling more efficient, scalable solutions. AMD Corporate Fellow Alex Starr joins CTO Mark Papermaster in a new season of Advanced Insights. Watch the full episode on YouTube:

AI-driven design is already reshaping compute. We are transforming how we design chips, with software and hardware evolving together. This approach is accelerating innovation and enabling more efficient, scalable solutions. AMD Corporate Fellow Alex Starr joins CTO Mark Papermaster in a new season of Advanced Insights. Watch the full episode on YouTube:

AMD

10,621 views • 5 months ago

Today in Parliament I highlighted GPU shortage as the biggest constraint in India’s AI ambition. The bottleneck is rising GPU costs & global supply shortage. This could choke India’s data centre expansion and AI advanced model training. India’s current compute pool has 34,000 GPUs, but a very small number compared to global scale required for training advanced models. Today I asked in the parliament, what are the specific targets, timelines & geopolitical engagements in place to secure predictable access to advanced computing resources, in particular GPUs?

Today in Parliament I highlighted GPU shortage as the biggest constraint in India’s AI ambition. The bottleneck is rising GPU costs & global supply shortage. This could choke India’s data centre expansion and AI advanced model training. India’s current compute pool has 34,000 GPUs, but a very small number compared to global scale required for training advanced models. Today I asked in the parliament, what are the specific targets, timelines & geopolitical engagements in place to secure predictable access to advanced computing resources, in particular GPUs?

Raghav Chadha

313,794 views • 6 months ago

Cluster Protocol: Revolutionizing #AI with blockchain integrity. Proof of Compute and Deploy to earn ensures that every byte counts.🚀 We offer a platform for Compute, Data, AI models and an open source community of AI researchers. Harness idle #GPUs for collective computing. Join the shared growth with GPU Sharing Economy where GPU providers are rewarded even when their GPUs are idle sitting on the network.🤝 Collaborate on our 'GitHub for AI'. Mold the future with open-source decentralized AI models.👩‍💻 Democratizing AI, we're shattering data monopolies. Converge in our global think-tank of decentralized datasets.🌐 Together, we redefine the limits in almost every sector possible where AI can make it strides. Let's unlock our collective potential. 📈 Start your AI odyssey with us now 👋 For more details check our blog at

Cluster Protocol: Revolutionizing #AI with blockchain integrity. Proof of Compute and Deploy to earn ensures that every byte counts.🚀 We offer a platform for Compute, Data, AI models and an open source community of AI researchers. Harness idle #GPUs for collective computing. Join the shared growth with GPU Sharing Economy where GPU providers are rewarded even when their GPUs are idle sitting on the network.🤝 Collaborate on our 'GitHub for AI'. Mold the future with open-source decentralized AI models.👩‍💻 Democratizing AI, we're shattering data monopolies. Converge in our global think-tank of decentralized datasets.🌐 Together, we redefine the limits in almost every sector possible where AI can make it strides. Let's unlock our collective potential. 📈 Start your AI odyssey with us now 👋 For more details check our blog at

Cluster Protocol

182,439 views • 2 years ago

THEY RAN A LOCAL AI MODEL ON A RASPBERRY PI THAT COSTS LESS THAN $300 no cloud no subscription no data center device smaller than your wallet output on a 27 year old iMac I tested local ai this week on a macbook and a phone the hardware keeps getting cheaper and the models keep getting smarter bookmark this if you’re still paying $20 a month for something your $300 raspberry pi can do

THEY RAN A LOCAL AI MODEL ON A RASPBERRY PI THAT COSTS LESS THAN $300 no cloud no subscription no data center device smaller than your wallet output on a 27 year old iMac I tested local ai this week on a macbook and a phone the hardware keeps getting cheaper and the models keep getting smarter bookmark this if you’re still paying $20 a month for something your $300 raspberry pi can do

leopardracer

73,534 views • 3 months ago

AI doesn’t scale on mega-models — it scales on distilled skills at the edge. From Davos, Qualcomm's Nakul Duggal breaks down the real physics of edge AI: smaller, purpose-built models, mixed precision, and ruthless focus on power + cost efficiency. The biggest constraint is that if AI runs in robotics, ADAS, or edge automation, it has to be efficient. Watch the full conversation to see why edge AI outcomes depend on efficiency by design:

AI doesn’t scale on mega-models — it scales on distilled skills at the edge. From Davos, Qualcomm's Nakul Duggal breaks down the real physics of edge AI: smaller, purpose-built models, mixed precision, and ruthless focus on power + cost efficiency. The biggest constraint is that if AI runs in robotics, ADAS, or edge automation, it has to be efficient. Watch the full conversation to see why edge AI outcomes depend on efficiency by design:

Six Five Media

74,120 views • 6 months ago

.Trenton Bricken explains how we know LLMs are actually generalizing - aka they're not just stochastic parrots: - Training models on code makes them better at reasoning in language. - Models fine tuned on math problems become better at entity detection. - We can just straightforwardly read the world-models developed by smaller NNs which are easier to interpret (Othello). Transfer learning shows models are developing a deeper understanding of their data. Full episode out Thursday.

.Trenton Bricken explains how we know LLMs are actually generalizing - aka they're not just stochastic parrots: - Training models on code makes them better at reasoning in language. - Models fine tuned on math problems become better at entity detection. - We can just straightforwardly read the world-models developed by smaller NNs which are easier to interpret (Othello). Transfer learning shows models are developing a deeper understanding of their data. Full episode out Thursday.

Dwarkesh Patel

237,812 views • 2 years ago

AI is transforming the way businesses operate 🤖. It helps companies work faster, cut costs, and make smarter decisions. From automating repetitive tasks to predicting trends and uncovering insights, AI is becoming the backbone of modern business. But AI can’t function without data 📊. Massive amounts of high-quality data are required to train models and make them effective. Right now, most of this data is controlled by a few centralized giants 🔒, making it expensive, siloed, and out of reach for smaller businesses and innovators. This is where OORT | The Data Cloud for Decentralized AI comes in 🌐✨. By combining decentralized storage with AI infrastructure powered by blockchain, OORT makes data open, secure, and accessible to everyone. It’s building a foundation where businesses of all sizes can tap into AI’s full potential and compete in the AI-driven economy 🚀

AI is transforming the way businesses operate 🤖. It helps companies work faster, cut costs, and make smarter decisions. From automating repetitive tasks to predicting trends and uncovering insights, AI is becoming the backbone of modern business. But AI can’t function without data 📊. Massive amounts of high-quality data are required to train models and make them effective. Right now, most of this data is controlled by a few centralized giants 🔒, making it expensive, siloed, and out of reach for smaller businesses and innovators. This is where OORT | The Data Cloud for Decentralized AI comes in 🌐✨. By combining decentralized storage with AI infrastructure powered by blockchain, OORT makes data open, secure, and accessible to everyone. It’s building a foundation where businesses of all sizes can tap into AI’s full potential and compete in the AI-driven economy 🚀

Sky Wee

58,239 views • 9 months ago

$Successfully deployed Deepseek R1 Distilled 70B (AWQ) across 8x NVIDIA RTX 3080 10G GPUs, achieving 60 tokens/s with full tensor parallelism via PCIe. Total hardware cost: $6,400 This demonstrates that consumer GPUs can deliver substantial ML inference capabilities at a fraction of the cost of datacenter hardware. For perspective, a single A100 80G costs $17,550, while a H100 80G runs $25,000. Our testing validates an often-overlooked opportunity: millions of idle crypto mining rigs could be repurposed into a powerful, distributed AI infrastructure. The performance-to-cost ratio of consumer GPUs, especially when properly optimized for tensor operations, presents a compelling case for decentralized AI compute. We're excited to share more insights as we continue pushing the boundaries of consumer hardware in AI workloads.$

Successfully deployed Deepseek R1 Distilled 70B (AWQ) across 8x NVIDIA RTX 3080 10G GPUs, achieving 60 tokens/s with full tensor parallelism via PCIe. Total hardware cost: $6,400 This demonstrates that consumer GPUs can deliver substantial ML inference capabilities at a fraction of the cost of datacenter hardware. For perspective, a single A100 80G costs $17,550, while a H100 80G runs $25,000. Our testing validates an often-overlooked opportunity: millions of idle crypto mining rigs could be repurposed into a powerful, distributed AI infrastructure. The performance-to-cost ratio of consumer GPUs, especially when properly optimized for tensor operations, presents a compelling case for decentralized AI compute. We're excited to share more insights as we continue pushing the boundaries of consumer hardware in AI workloads.

TensorBlock

16,219 views • 1 year ago

There's no point in doing decentralized training without efficient communication. >> DiLoCo (H=15) ships ~480mb/merge with 163 syncs. >> SparseLoCo (H=15) ships ~5.5–17mb/merge at 0.78–3.12% density with 163 syncs Top-K Compression + 2 bit comms ~28–89× smaller per sync than DiLoCo. Subnet 3 :: Luis el grande If you have the algorithm, you can train large language models across disparate compute, collectively. "In the space of eight months or nine months, we've been able to scale our model from 1.2B to 70B, which represents 58x improvement" Distributed State Research paper :: Full Episode059 + const :: The holy grail of distributed AI training SN3 :: Templar :: Luis el grande_ai SN39 :: Basilica :: basilica SN81 :: Grail :: grail #SN3 #SN39 #SN81 #Bittensor

There's no point in doing decentralized training without efficient communication. >> DiLoCo (H=15) ships ~480mb/merge with 163 syncs. >> SparseLoCo (H=15) ships ~5.5–17mb/merge at 0.78–3.12% density with 163 syncs Top-K Compression + 2 bit comms ~28–89× smaller per sync than DiLoCo. Subnet 3 :: Luis el grande If you have the algorithm, you can train large language models across disparate compute, collectively. "In the space of eight months or nine months, we've been able to scale our model from 1.2B to 70B, which represents 58x improvement" Distributed State Research paper :: Full Episode059 + const :: The holy grail of distributed AI training SN3 :: Templar :: Luis el grande_ai SN39 :: Basilica :: basilica SN81 :: Grail :: grail #SN3 #SN39 #SN81 #Bittensor

Openτensor Foundaτion

17,767 views • 11 months ago

The new SpaceX-Cursor model is going to be insane. It is being trained from scratch, unlike previous Composer models, which were built off Kimi. This new model is Elon’s attempt to get back in the AI race, and he might rejoin as the frontrunner. Composer 2.5 is already a world-class model, only losing to models like GPT-5.5, Opus 4.8, and Fable 5. Grok Composer 3, with SpaceX’s massive compute and xAI’s team, could become the best AI model on the planet. Elon is back.

The new SpaceX-Cursor model is going to be insane. It is being trained from scratch, unlike previous Composer models, which were built off Kimi. This new model is Elon’s attempt to get back in the AI race, and he might rejoin as the frontrunner. Composer 2.5 is already a world-class model, only losing to models like GPT-5.5, Opus 4.8, and Fable 5. Grok Composer 3, with SpaceX’s massive compute and xAI’s team, could become the best AI model on the planet. Elon is back.

Ali Romman

119,837 views • 2 months ago

Check out our latest development updates! AI Model Training • Real-time distributed training • Performance insights & optimization • Validation, generalization, and scalability Cosmic Network Mainnet • Smooth block indexing & finalization • Efficient P2P connections • Consistent app hashes & state syncing • Reliable, decentralized infrastructure Website Finalization • C-Nodes: Real-time stats & maps • Nebula Lite: Features, specs & purchase flow • News & Blogs: Latest insights Android & iOS • Stability testing • New tutorial animations • Helpful tooltips • Enhanced stats UI macOS • Intuitive map gestures • Improved sidebar functionality • Tutorial animations & tooltips • Enhanced stats UI Nebula Lite • Final stages of production Watch the full update in the video!

Check out our latest development updates! AI Model Training • Real-time distributed training • Performance insights & optimization • Validation, generalization, and scalability Cosmic Network Mainnet • Smooth block indexing & finalization • Efficient P2P connections • Consistent app hashes & state syncing • Reliable, decentralized infrastructure Website Finalization • C-Nodes: Real-time stats & maps • Nebula Lite: Features, specs & purchase flow • News & Blogs: Latest insights Android & iOS • Stability testing • New tutorial animations • Helpful tooltips • Enhanced stats UI macOS • Intuitive map gestures • Improved sidebar functionality • Tutorial animations & tooltips • Enhanced stats UI Nebula Lite • Final stages of production Watch the full update in the video!

Cosmic Network

11,311 views • 1 year ago

🌟Quilibrium’s AI Breakthrough: Encrypted Training on CPUs In her latest live stream ( - minute 14) Cassie unveiled a groundbreaking AI training method that allows models to be trained on encrypted data using CPUs while achieving performance comparable to Nvidia’s A100 GPU (blue line in the graph below). Traditionally, AI training requires expensive GPUs because matrix multiplications—the core of deep learning—are highly computational. Running these calculations on CPUs is painfully slow, often taking hours or days for even small models. The problem worsens when trying to train AI on encrypted data, as standard encryption methods add a massive computational burden. Quilibrium’s breakthrough removes this bottleneck. Instead of relying on traditional matrix multiplication, their method uses a completely different mathematical approach, allowing AI models to be trained securely and efficiently without exposing the raw data. Cassie didn’t reveal the exact technique, only hinting that it’s inspired by existing AI research and will be detailed in a future open-source AGPL-licensed paper. The key advantage? AI can now be trained at GPU speeds on standard CPUs, making privacy-preserving machine learning far more accessible. This innovation has major implications. It slashes AI infrastructure costs, allowing organizations to train powerful models without investing in expensive hardware. It also enables private AI training on personal or corporate data without revealing sensitive information, a game-changer for industries like healthcare and finance. If Quilibrium’s method delivers on its promise, it could reshape AI development, making privacy-first computing the new standard. $QUIL $wQUIL

🌟Quilibrium’s AI Breakthrough: Encrypted Training on CPUs In her latest live stream ( - minute 14) Cassie unveiled a groundbreaking AI training method that allows models to be trained on encrypted data using CPUs while achieving performance comparable to Nvidia’s A100 GPU (blue line in the graph below). Traditionally, AI training requires expensive GPUs because matrix multiplications—the core of deep learning—are highly computational. Running these calculations on CPUs is painfully slow, often taking hours or days for even small models. The problem worsens when trying to train AI on encrypted data, as standard encryption methods add a massive computational burden. Quilibrium’s breakthrough removes this bottleneck. Instead of relying on traditional matrix multiplication, their method uses a completely different mathematical approach, allowing AI models to be trained securely and efficiently without exposing the raw data. Cassie didn’t reveal the exact technique, only hinting that it’s inspired by existing AI research and will be detailed in a future open-source AGPL-licensed paper. The key advantage? AI can now be trained at GPU speeds on standard CPUs, making privacy-preserving machine learning far more accessible. This innovation has major implications. It slashes AI infrastructure costs, allowing organizations to train powerful models without investing in expensive hardware. It also enables private AI training on personal or corporate data without revealing sensitive information, a game-changer for industries like healthcare and finance. If Quilibrium’s method delivers on its promise, it could reshape AI development, making privacy-first computing the new standard. $QUIL $wQUIL

Quilibrium Community

18,919 views • 1 year ago

ELON: GROK’S A BEAST - IT CAN'T BE CRAMMED INTO A CAR “Grok is a giant model; you could not possibly squeeze Grok onto a car, that’s for sure. With Grok, it’s trying to solve for artificial intelligence, with a massive amount of AI training, compute, and inference compute. Grok 5 will only run effectively on a GV300 [AI training cluster]; that’s how much of a beast Grok 5 is. Whereas Tesla’s models are less than 10% the size, maybe closer to 5% the size of Grok.” Source: Elon Musk, Tesla Podcast, October 22, 2025

ELON: GROK’S A BEAST - IT CAN'T BE CRAMMED INTO A CAR “Grok is a giant model; you could not possibly squeeze Grok onto a car, that’s for sure. With Grok, it’s trying to solve for artificial intelligence, with a massive amount of AI training, compute, and inference compute. Grok 5 will only run effectively on a GV300 [AI training cluster]; that’s how much of a beast Grok 5 is. Whereas Tesla’s models are less than 10% the size, maybe closer to 5% the size of Grok.” Source: Elon Musk, Tesla Podcast, October 22, 2025

Mario Nawfal

33,523 views • 9 months ago

Prime Intellect engineers explained how they train reasoning models over the open internet in 30 minutes - better than $3000 distributed training courses. split policy and rollouts across nodes -> run agents in parallel envs -> verify with LLM judges -> gradient-sync over the internet -> train Llama, Qwen, Gemma at cluster scale on rented GPUs. That loop is why open reasoning models are catching closed labs without owning a data center. Prime-RL + verifiers + distributed rollouts + LLM judges + multi-cloud GPUs - that's the stack. Watch and save it, then launch your first distributed RL run this week.

Prime Intellect engineers explained how they train reasoning models over the open internet in 30 minutes - better than $3000 distributed training courses. split policy and rollouts across nodes -> run agents in parallel envs -> verify with LLM judges -> gradient-sync over the internet -> train Llama, Qwen, Gemma at cluster scale on rented GPUs. That loop is why open reasoning models are catching closed labs without owning a data center. Prime-RL + verifiers + distributed rollouts + LLM judges + multi-cloud GPUs - that's the stack. Watch and save it, then launch your first distributed RL run this week.

h100envy

39,620 views • 1 month ago

This is huge! A UCLA team managed to build an optical generative model that runs on light instead of GPUs. In their demo, a shallow encoder maps noise into phase patterns, which a free-space optical decoder then transforms into images—digits, fashion, butterflies, faces, even Van Gogh–style art—without any computation during synthesis. ⚡ The results rival digital diffusion models, pointing to ultra-fast, energy-efficient AI powered by photonics. Optical generative models | Nature Paper:

This is huge! A UCLA team managed to build an optical generative model that runs on light instead of GPUs. In their demo, a shallow encoder maps noise into phase patterns, which a free-space optical decoder then transforms into images—digits, fashion, butterflies, faces, even Van Gogh–style art—without any computation during synthesis. ⚡ The results rival digital diffusion models, pointing to ultra-fast, energy-efficient AI powered by photonics. Optical generative models | Nature Paper:

机器之心 JIQIZHIXIN

173,657 views • 10 months ago