Loading video...

Video Failed to Load

Go Home

🚀 SANA-Video 2.0 is here! A full-stack optimized video model designed for efficiency — while still delivering high quality. We combine a hybrid architecture closely related to the recent Kimi K3 design, adopt Self-Flow from FLUX 3, and further accelerate it with our own Sol-Engine. Key technical ingredients: 🧠...

xieenze_jr's profile picture

Enze Xie

1,936 subscribers

23,585 views • 4 days ago •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

After 8+ years on the Tesla Autopilot team and 3 years at Intel, I started Apex Compute to design a new architecture for efficient AI inference. For the past 9 months, we’ve been building our custom inference accelerator. Today we’re releasing Unified Engine v1. Last June we raised our seed round with Maxitech , DeepFin Research, Soma Capital and an incredible group of angel investors. In less than 9 months, we completed our RTL architecture and brought our first pre-silicon prototype to life on FPGA. Our architecture combines systolic array and vector processing in a single compute engine with multiple architectural optimizations, achieving very high FLOPs utilization. A single engine is super lean and it uses less than 90K LUTs and 1 MB Block RAM. It may also be one of the smallest logic-footprint compute engines developed so far. Our Unified Engine v1 supports: -matrix-matrix multiplication (~95% FLOPs utilization) -softmax (~90% FLOPs utilization) -broadcast and element-wise operations -RMSNorm / LayerNorm -block quantization/dequantization (fp4, int4) -multi-engine synchronization and many other operations. We even implemented memory-efficient attention similar to FlashAttention, reaching ~90% FLOP utilization. Full benchmarks and the software stack are available on our GitHub: We have basic compiler written in Python and it supports PyTorch tensors directly to easily test and transfer tensors between the accelerator and host using bf16, fp4 and int4 formats. Our FPGA prototype can already run LLM inference and outperform NVIDIA Jetson Orin Nano, even on a mid-tier FPGA setup (6.4x lower memory bandwidth, 18% slower clock speed at 4.5 Watts). Check the side-by-side comparison video below. Our GitHub includes low-level operator implementations, examples for tiled matrix multiplication, operation chaining, tensor parallelism, attention kernel and a full Gemma 3 1B model implementation. Many more models(Vision Transformers and VLA) are coming soon. Our accelerator IP is AXI-ready for deployment on any AMD(Xilinx) FPGA platform today. Even better, our two-engine prototype runs on an entry-level AMD(Xilinx) FPGA as a PCIe accelerator card. You can purchase it here for $50 to experiment our pre-silicon prototype on your desktop PC or Raspberry Pi 5. We will be releasing hardware bitstream updates as the architecture gets new features. More to come soon! We are expanding our team and looking for compiler engineers and floating-point hardware design engineers. If you're interested, please send me a DM.
1:17

After 8+ years on the Tesla Autopilot team and 3 years at Intel, I started Apex Compute to design a new architecture for efficient AI inference. For the past 9 months, we’ve been building our custom inference accelerator. Today we’re releasing Unified Engine v1. Last June we raised our seed round with Maxitech , DeepFin Research, Soma Capital and an incredible group of angel investors. In less than 9 months, we completed our RTL architecture and brought our first pre-silicon prototype to life on FPGA. Our architecture combines systolic array and vector processing in a single compute engine with multiple architectural optimizations, achieving very high FLOPs utilization. A single engine is super lean and it uses less than 90K LUTs and 1 MB Block RAM. It may also be one of the smallest logic-footprint compute engines developed so far. Our Unified Engine v1 supports: -matrix-matrix multiplication (~95% FLOPs utilization) -softmax (~90% FLOPs utilization) -broadcast and element-wise operations -RMSNorm / LayerNorm -block quantization/dequantization (fp4, int4) -multi-engine synchronization and many other operations. We even implemented memory-efficient attention similar to FlashAttention, reaching ~90% FLOP utilization. Full benchmarks and the software stack are available on our GitHub: We have basic compiler written in Python and it supports PyTorch tensors directly to easily test and transfer tensors between the accelerator and host using bf16, fp4 and int4 formats. Our FPGA prototype can already run LLM inference and outperform NVIDIA Jetson Orin Nano, even on a mid-tier FPGA setup (6.4x lower memory bandwidth, 18% slower clock speed at 4.5 Watts). Check the side-by-side comparison video below. Our GitHub includes low-level operator implementations, examples for tiled matrix multiplication, operation chaining, tensor parallelism, attention kernel and a full Gemma 3 1B model implementation. Many more models(Vision Transformers and VLA) are coming soon. Our accelerator IP is AXI-ready for deployment on any AMD(Xilinx) FPGA platform today. Even better, our two-engine prototype runs on an entry-level AMD(Xilinx) FPGA as a PCIe accelerator card. You can purchase it here for $50 to experiment our pre-silicon prototype on your desktop PC or Raspberry Pi 5. We will be releasing hardware bitstream updates as the architecture gets new features. More to come soon! We are expanding our team and looking for compiler engineers and floating-point hardware design engineers. If you're interested, please send me a DM.

Hasan

37,603 views • 4 months ago

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

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

Brian Roemmele

618,431 views • 2 months ago

Most video-action robot models are a content-creation video generator with an action module attached. LingBot-VA 2.0 from Robbyant, a video-action foundation model, throws that starting point out and trains the whole stack natively for control. And it runs closed-loop at a peak 225 Hz. It's so important because A robot cannot move responsively when its controller pauses to imagine the next few frames. LingBot-VA 2.0 predicts during execution, then corrects using each real observation. And it carries only about 13B video parameters while activating roughly 1.9B per token. Bigger robot models usually mean slower reactions, creating a direct conflict between intelligence and control. LingBot-VA 2.0 is trained from scratch for robot control rather than adapted from a video generator built for content creation. Robbyant, an embodied AI company under Ant Group, built it to learn how scenes change under actions, predict what should happen next, and turn those predictions into real-time robot movements. Most video-action systems inherit a tokenizer and video backbone trained mainly to reproduce visual appearance. LingBot-VA 2.0 rebuilds both parts around physical control. Its semantic visual-action tokenizer maps observations toward features from a frozen vision foundation model and learns compact latent actions from frame-to-frame changes using self-supervised inverse and forward dynamics. Unlabeled web video can therefore carry action-relevant training signals without robot action labels. The policy is causal from the start, so every prediction can use only past observations. Its sparse Mixture-of-Experts video backbone has about 13B total parameters, while about 1.9B are active per token, keeping the compute lower during each step. A high-level vision-language planner breaks long tasks into smaller instructions, while the low-level video-action policy handles continuous movement. Foresight Reasoning predicts future visual states while the robot is already acting, then replaces imagined states with every new real observation. Combined with few-step distillation and systems acceleration, the paper reports a peak asynchronous execution frequency of 225 Hz. The model adapts from 10–15 demonstrations, transfers across robot embodiments, and handles some new tasks zero-shot. In the paper’s own evaluations, it reaches 93.6 average on RoboTwin 2.0 and reports stronger real-world results than LingBot-VA and π0.5 across the tested tasks. 🧵 1.
1:17

Most video-action robot models are a content-creation video generator with an action module attached. LingBot-VA 2.0 from Robbyant, a video-action foundation model, throws that starting point out and trains the whole stack natively for control. And it runs closed-loop at a peak 225 Hz. It's so important because A robot cannot move responsively when its controller pauses to imagine the next few frames. LingBot-VA 2.0 predicts during execution, then corrects using each real observation. And it carries only about 13B video parameters while activating roughly 1.9B per token. Bigger robot models usually mean slower reactions, creating a direct conflict between intelligence and control. LingBot-VA 2.0 is trained from scratch for robot control rather than adapted from a video generator built for content creation. Robbyant, an embodied AI company under Ant Group, built it to learn how scenes change under actions, predict what should happen next, and turn those predictions into real-time robot movements. Most video-action systems inherit a tokenizer and video backbone trained mainly to reproduce visual appearance. LingBot-VA 2.0 rebuilds both parts around physical control. Its semantic visual-action tokenizer maps observations toward features from a frozen vision foundation model and learns compact latent actions from frame-to-frame changes using self-supervised inverse and forward dynamics. Unlabeled web video can therefore carry action-relevant training signals without robot action labels. The policy is causal from the start, so every prediction can use only past observations. Its sparse Mixture-of-Experts video backbone has about 13B total parameters, while about 1.9B are active per token, keeping the compute lower during each step. A high-level vision-language planner breaks long tasks into smaller instructions, while the low-level video-action policy handles continuous movement. Foresight Reasoning predicts future visual states while the robot is already acting, then replaces imagined states with every new real observation. Combined with few-step distillation and systems acceleration, the paper reports a peak asynchronous execution frequency of 225 Hz. The model adapts from 10–15 demonstrations, transfers across robot embodiments, and handles some new tasks zero-shot. In the paper’s own evaluations, it reaches 93.6 average on RoboTwin 2.0 and reports stronger real-world results than LingBot-VA and π0.5 across the tested tasks. 🧵 1.

Rohan Paul

10,996 views • 14 days ago

Want to create an avatar from a single image? FlexAvatar is a transformer model that creates full 360°, high-quality, and expressive 3D head avatar from just a single portrait image in minutes. Real-time Demo: FlexAvatar's lightweight architecture allows both animation and rendering in real-time, enabling interactive user experiences. To create a new 3D head avatar, only one image is required, e.g., from a webcam. The final avatar is ready after 2 minutes. Architecture: Under the hood, FlexAvatar adopts a transformer-based encoder-decoder design. The encoder maps the input image onto a latent avatar space, while the decoder produces 3D Gaussian attribute maps by incorporating the animation signal via cross-attention. The model learns all facial animations directly from the data without relying on pre-built 3D face models. This equips the avatars with realistic facial expressions. The internal avatar latent space can be conveniently used to integrate additional observations of a person via fitting. This enables use-cases where more than one image of a person is available, e.g., from a phone scan of the person. We train jointly on 2D monocular videos and multi-view data. However, in monocular videos, the animation signal leaks the target viewpoint, causing the model to produce incomplete 3D heads. We call this phenomenon entanglement of driving signal and target viewpoint. To prevent entanglement, we introduce bias sinks. These are learnable tokens that indicate whether a training sample stems from a monocular or a multi-view dataset. During training, the model learns to produce incomplete 3D heads only when the monocular token is present. During inference, FlexAvatar then always uses the multi-view token for which the model has learned to produce complete 3D heads. This simple design allows to combine the generalizability from monocular data with the quality of multi-view data. FlexAvatar summary: - Input: Single-image, phone scan, or monocular video - Output: Full 360° head avatar - Expressive animations - Real-time rendering and animation - Generalization to any portrait - Create a new avatar in 2 minutes - Use bias sinks to combine 2D and 3D data 🏠 🌍 🎥 Great work by Tobias Kirschstein and Simon Giebenhain!
0:30

Want to create an avatar from a single image? FlexAvatar is a transformer model that creates full 360°, high-quality, and expressive 3D head avatar from just a single portrait image in minutes. Real-time Demo: FlexAvatar's lightweight architecture allows both animation and rendering in real-time, enabling interactive user experiences. To create a new 3D head avatar, only one image is required, e.g., from a webcam. The final avatar is ready after 2 minutes. Architecture: Under the hood, FlexAvatar adopts a transformer-based encoder-decoder design. The encoder maps the input image onto a latent avatar space, while the decoder produces 3D Gaussian attribute maps by incorporating the animation signal via cross-attention. The model learns all facial animations directly from the data without relying on pre-built 3D face models. This equips the avatars with realistic facial expressions. The internal avatar latent space can be conveniently used to integrate additional observations of a person via fitting. This enables use-cases where more than one image of a person is available, e.g., from a phone scan of the person. We train jointly on 2D monocular videos and multi-view data. However, in monocular videos, the animation signal leaks the target viewpoint, causing the model to produce incomplete 3D heads. We call this phenomenon entanglement of driving signal and target viewpoint. To prevent entanglement, we introduce bias sinks. These are learnable tokens that indicate whether a training sample stems from a monocular or a multi-view dataset. During training, the model learns to produce incomplete 3D heads only when the monocular token is present. During inference, FlexAvatar then always uses the multi-view token for which the model has learned to produce complete 3D heads. This simple design allows to combine the generalizability from monocular data with the quality of multi-view data. FlexAvatar summary: - Input: Single-image, phone scan, or monocular video - Output: Full 360° head avatar - Expressive animations - Real-time rendering and animation - Generalization to any portrait - Create a new avatar in 2 minutes - Use bias sinks to combine 2D and 3D data 🏠 🌍 🎥 Great work by Tobias Kirschstein and Simon Giebenhain!

Matthias Niessner

95,991 views • 7 months ago

PrismML Releases Bonsai 27B: 1-bit and Ternary Builds of Qwen3.6-27B Hitting 89.5% of FP16 at 3.9GB. No new pretrain. No higher-precision escape hatches. No multi-GPU rig. Here's how it works. 👇 1: Codes, not floats Every weight becomes a code, with one shared FP16 scale per group of 128. Ternary is {−1, 0, +1}, binary is {−1, +1}. Sharing the scale across 128 weights keeps its cost at 16/128 = 0.125 bits. → Ternary: log2(3) + 16/128 ≈ 1.71 bits/weight → 5.9GB → Binary: 1 + 16/128 = 1.125 bits/weight → 3.9GB 2: Post-training, not from scratch No BitNet-style low-bit pretrain. It starts from off-the-shelf Qwen3.6-27B, architecture unchanged. The representation runs end to end across embeddings, attention projections, MLP projections, and the LM head. → 9.4× (ternary) and 14.2× (binary) vs the 54GB FP16 baseline 3: Labels are not bit-widths Conventional low-bit builds are mixed-precision by construction. The advertised name describes the most-compressed tensors, not the model. → Q4_K_XL, labeled "4-bit," is really 5.2 bits/weight at 17.6GB → IQ2_XXS, labeled "2-bit," is really 2.8 bits/weight at 9.4GB 4: Fitting a phone is two budgets iOS caps a single app near half of RAM, so a 12GB iPhone exposes ~6GB. The KV cache grows on top. Hybrid attention at ~75% linear means only 16 of 64 layers cache. → 4-bit KV: 4.3GB at 262K context, down from 17.2GB → 11.0 tok/s on iPhone 17 Pro Max 5: The numbers (15 benchmarks, thinking mode) → Ternary: 80.49 avg at 5.9GB — 94.6% of FP16 → 1-bit: 76.11 avg at 3.9GB — 89.5% of FP16 → IQ2_XXS falls to 57.5 on AIME26 while still scoring 88.93 on MMLU-Redux The key takeaway: 27B-class reasoning without the 54GB checkpoint — group-wise ternary and binary codes, an end-to-end low-bit language stack, 4-bit KV, on one phone. Full analysis: Repo: Model weight: Technical details: PrismML
0:37

PrismML Releases Bonsai 27B: 1-bit and Ternary Builds of Qwen3.6-27B Hitting 89.5% of FP16 at 3.9GB. No new pretrain. No higher-precision escape hatches. No multi-GPU rig. Here's how it works. 👇 1: Codes, not floats Every weight becomes a code, with one shared FP16 scale per group of 128. Ternary is {−1, 0, +1}, binary is {−1, +1}. Sharing the scale across 128 weights keeps its cost at 16/128 = 0.125 bits. → Ternary: log2(3) + 16/128 ≈ 1.71 bits/weight → 5.9GB → Binary: 1 + 16/128 = 1.125 bits/weight → 3.9GB 2: Post-training, not from scratch No BitNet-style low-bit pretrain. It starts from off-the-shelf Qwen3.6-27B, architecture unchanged. The representation runs end to end across embeddings, attention projections, MLP projections, and the LM head. → 9.4× (ternary) and 14.2× (binary) vs the 54GB FP16 baseline 3: Labels are not bit-widths Conventional low-bit builds are mixed-precision by construction. The advertised name describes the most-compressed tensors, not the model. → Q4_K_XL, labeled "4-bit," is really 5.2 bits/weight at 17.6GB → IQ2_XXS, labeled "2-bit," is really 2.8 bits/weight at 9.4GB 4: Fitting a phone is two budgets iOS caps a single app near half of RAM, so a 12GB iPhone exposes ~6GB. The KV cache grows on top. Hybrid attention at ~75% linear means only 16 of 64 layers cache. → 4-bit KV: 4.3GB at 262K context, down from 17.2GB → 11.0 tok/s on iPhone 17 Pro Max 5: The numbers (15 benchmarks, thinking mode) → Ternary: 80.49 avg at 5.9GB — 94.6% of FP16 → 1-bit: 76.11 avg at 3.9GB — 89.5% of FP16 → IQ2_XXS falls to 57.5 on AIME26 while still scoring 88.93 on MMLU-Redux The key takeaway: 27B-class reasoning without the 54GB checkpoint — group-wise ternary and binary codes, an end-to-end low-bit language stack, 4-bit KV, on one phone. Full analysis: Repo: Model weight: Technical details: PrismML

Marktechpost AI

31,860 views • 13 days ago

The Kimi AI Model And What If This Carnegie Mellon, PhD student was incentivized to open source AI in the US? The answer is Kimi would have been an open source US model. Yang Zhilin Moonshot’s founder and maker of the Kimi series turned his deep research expertise into one of the world’s most capable open AI systems. After earning his PhD at Carnegie Mellon under leading researchers and interning at Google Brain and Meta, he returned to China and co-founded Moonshot AI in March 2023 with Tsinghua classmates Zhou Xinyu and Wu Yuxin. Why? The US VC world and tax structure did not favor Zhilin’s proposal to open source the AI models as a strategy. So he left. He named the company after his favorite Pink Floyd album, reflecting his ambitious vision for scalable. Yang assembled a core technical team of inventors behind breakthroughs like Transformer-XL and RoPE. Together they focused on turning massive compute into efficient intelligence through innovative architectures. The journey began with the Kimi chatbot in October 2023, rapidly scaling context from 200,000 to millions of characters. This evolved into the Kimi series: K1.5 matched top reasoning models, K2 introduced a 1-trillion-parameter Mixture-of-Experts design trained on 15.5 trillion tokens and released openly, and K2 Thinking added advanced agentic capabilities. Kimi K3 represents the pinnacle, this 2.8-trillion-parameter model uses a sparse MoE architecture with 896 experts (only 16 active per token), new Kimi Delta Attention and attention residuals for efficiency, and a 1-million-token context window. It delivers frontier performance in long-horizon coding, reasoning, and multimodal tasks at competitive cost, with weights set for open release. Yang’s approach emphasizes openness, efficiency, and continuous self-improvement — enabling solo developers and teams to achieve what once required massive resources. By sharing technical insights in public talks, he has accelerated global progress toward more accessible, powerful AI. Imagine if we held open source higher than the fear theater games of Anthropic? We are chasing out some of the best minds. This is how you lose…
17:29

The Kimi AI Model And What If This Carnegie Mellon, PhD student was incentivized to open source AI in the US? The answer is Kimi would have been an open source US model. Yang Zhilin Moonshot’s founder and maker of the Kimi series turned his deep research expertise into one of the world’s most capable open AI systems. After earning his PhD at Carnegie Mellon under leading researchers and interning at Google Brain and Meta, he returned to China and co-founded Moonshot AI in March 2023 with Tsinghua classmates Zhou Xinyu and Wu Yuxin. Why? The US VC world and tax structure did not favor Zhilin’s proposal to open source the AI models as a strategy. So he left. He named the company after his favorite Pink Floyd album, reflecting his ambitious vision for scalable. Yang assembled a core technical team of inventors behind breakthroughs like Transformer-XL and RoPE. Together they focused on turning massive compute into efficient intelligence through innovative architectures. The journey began with the Kimi chatbot in October 2023, rapidly scaling context from 200,000 to millions of characters. This evolved into the Kimi series: K1.5 matched top reasoning models, K2 introduced a 1-trillion-parameter Mixture-of-Experts design trained on 15.5 trillion tokens and released openly, and K2 Thinking added advanced agentic capabilities. Kimi K3 represents the pinnacle, this 2.8-trillion-parameter model uses a sparse MoE architecture with 896 experts (only 16 active per token), new Kimi Delta Attention and attention residuals for efficiency, and a 1-million-token context window. It delivers frontier performance in long-horizon coding, reasoning, and multimodal tasks at competitive cost, with weights set for open release. Yang’s approach emphasizes openness, efficiency, and continuous self-improvement — enabling solo developers and teams to achieve what once required massive resources. By sharing technical insights in public talks, he has accelerated global progress toward more accessible, powerful AI. Imagine if we held open source higher than the fear theater games of Anthropic? We are chasing out some of the best minds. This is how you lose…

Brian Roemmele

37,185 views • 7 days ago

Today we're announcing #GAIA1: a 9B parameter world model, trained on 4,700 hours of driving data, able to simulate complex and diverse driving scenes from video, text and action inputs. This model is 480x larger than the preview we shared earlier this year and the results are incredible. These videos are entirely synthetically generated by Wayve's generative AI, GAIA-1. But there is more here than just generating videos, GAIA is an entire world model. A world model allows us to simulate the future, conditioned on video, text and action inputs, which can be leveraged for making informed decisions when driving. Why is this game-changing for autonomous driving? 1. Safety. One limitation with AI systems like today's Large Language Models is that they are autoregressive, next-word prediction algorithms, but aren't necessarily aware of the implications of their decisions. A world model allows us to give our AI the capability to be aware of its decisions, by simulating the future, which is important for self-driving safety. 2. Synthetic training data. I believe synthetic training data is the future for AI, because it is safer, cheaper, and infinitely scalable. GAIA-1 unlocks unprecedented realism and diversity of synthetic data for self-driving. 3. Long-tail robustness. One of the biggest challenges for self-driving is long-tail robustness: dealing with the enormous magnitude of edge cases we see on the road. An advantage of generative AI is its incredible ability to recombine experiences in new ways. This is exciting for self-driving as it means we can learn from two edge case scenarios, and combine them to become a corner case. For example, we can experience driving in fog, and experience of jay-walking pedestrians, and GAIA can learn from these experiences to understand how to generate a fog+jay walking scenario. Check out many more videos in our blog or further technical details in our paper: Or come chat with our team who are at the International Conference on Computer Vision (#ICCV2023) this week in Paris in Booth 32 Jamie Shotton
0:41

Today we're announcing #GAIA1: a 9B parameter world model, trained on 4,700 hours of driving data, able to simulate complex and diverse driving scenes from video, text and action inputs. This model is 480x larger than the preview we shared earlier this year and the results are incredible. These videos are entirely synthetically generated by Wayve's generative AI, GAIA-1. But there is more here than just generating videos, GAIA is an entire world model. A world model allows us to simulate the future, conditioned on video, text and action inputs, which can be leveraged for making informed decisions when driving. Why is this game-changing for autonomous driving? 1. Safety. One limitation with AI systems like today's Large Language Models is that they are autoregressive, next-word prediction algorithms, but aren't necessarily aware of the implications of their decisions. A world model allows us to give our AI the capability to be aware of its decisions, by simulating the future, which is important for self-driving safety. 2. Synthetic training data. I believe synthetic training data is the future for AI, because it is safer, cheaper, and infinitely scalable. GAIA-1 unlocks unprecedented realism and diversity of synthetic data for self-driving. 3. Long-tail robustness. One of the biggest challenges for self-driving is long-tail robustness: dealing with the enormous magnitude of edge cases we see on the road. An advantage of generative AI is its incredible ability to recombine experiences in new ways. This is exciting for self-driving as it means we can learn from two edge case scenarios, and combine them to become a corner case. For example, we can experience driving in fog, and experience of jay-walking pedestrians, and GAIA can learn from these experiences to understand how to generate a fog+jay walking scenario. Check out many more videos in our blog or further technical details in our paper: Or come chat with our team who are at the International Conference on Computer Vision (#ICCV2023) this week in Paris in Booth 32 Jamie Shotton

Alex Kendall

631,856 views • 2 years ago

kimi k3 vs gpt 5.6 sol vs fable 5 vs grok 4.5 Kimi.ai just dropped kimi k3 – a 2.8t param native multimodal model, the first open 3t-class release. key facts: • 1m token context. stable latentmoe activating 16 of 896 experts, built on kimi delta attention (kda) and attention residuals • quantization-aware training from the sft stage onward – mxfp4 weights, mxfp8 activations. moonshot claims ~2.5x scaling efficiency over k2 • max thinking effort by default. low- and high-effort modes are "coming in updates" – there is no way to turn the thinking down today, and you feel it in every run • pricing: $0.30/mtok cache-hit input, $3.00/mtok cache-miss, $15.00/mtok output. claims >90% cache hit rate on coding workloads • benchmarks: swe marathon 42.0 (1st – fable 5: 35.0, sol: 39.0, opus 4.8: 40.0), terminal bench 2.1 88.3, browsecomp 91.2 (1st), program bench 77.8 (1st), gpqa-diamond 93.5. loses frontierswe 81.2 vs fable's 86.6, and deepswe 67.5 vs sol's 73.0 our test – 3 prompts, single-file html, Three.js, fully procedural, no assets: 1. photorealistic european roulette wheel – 37 pockets in the real sequence, mahogany clearcoat bowl, chrome turret, diamond deflectors, flick-to-spin, ball that spirals inward and settles on a mathematically real number 2. las vegas slot machine – 3 reels behind transmissive glass, drag the chrome lever to play, mechanical odometer counters modelled in 3d, coin physics on win 3. full pinball table – 6.5° tilted playfield, flipper impulse physics, spline ramps, drop targets, 6 bumpers, mechanical score reels in the backbox we ran the test on AI/ML API platform results: - cost #1 grok 4.5 – $0.30 #2 kimi k3 – $0.71 #3 gpt 5.6 sol – $2.05 #4 fable 5 – $7.69 - tokens #1 grok 4.5 – 34,241 #2 gpt 5.6 sol – 51,748 #3 fable 5 – 144,126 #4 kimi k3 – 157,999 - lines of code #1 gpt 5.6 sol – 3,054 #2 grok 4.5 – 3,047 #3 kimi k3 – 2,255 #4 fable 5 – 1,950 - generation time #1 grok 4.5 – 5.1 min #2 gpt 5.6 sol – 22.0 min #3 fable 5 – 31.5 min #4 kimi k3 – 75.6 min observations: • kimi k3 is cheap and it is slow. 75.6 minutes across three prompts against grok's 5.1. it is 2.4x grok's price and 15x grok's wall clock. the roulette took 15 min, the slot 18, the pinball 42 • it failed 2 of 3. only the roulette works. the slot machine has reel cutouts on both faces of the cabinet and the symbols face backwards – you can only read your spin by walking around to the rear of the machine. the pinball table stands vertically on its edge with the legs floating detached beside it. • 81% of kimi's output tokens are reasoning, not code. grok: 22%. you are not paying for a bigger answer, you are paying for a longer argument with itself • price per 100 shipped lines – grok $0.010, kimi $0.031, sol $0.067, fable $0.394. a 39x spread for the same three files kimi k3's code quality: upsides: • the roulette is genuinely good – procedural wood grain with real specular breakup, correct european sequence (0-32-15-19-4...), chrome turret, diamond deflectors, clean console • the pinball artwork is the best in the test – a synthwave "nova strike / deep space" field with six individually coloured neon bumper rings, a retro sun on a grid horizon, a nova burst, and a scoring legend printed on the apron. no other model printed the rules on the machine. it is a beautiful texture on a broken object • physics reasoning is real – it derived a 480hz substep for the collider, worked out ball settle conditions and termination guarantees, and checked every ramp exit vector by hand before writing any of it • it is the only model that saw the importmap trap coming. sol shipped a blank white page twice because three.js addons import the bare specifier 'three' and die without an import map downsides: • it dodged that trap on the slot by loading three.js r128 through classic script tags – a 2021 build with no working transmission. its slot glass rendered fully opaque and buried all three reels behind a white pane. the code asks for transmission: 0.93, ior: 1.5 – correct, and silently ignored by a renderer that predates the feature • after 42 minutes and 212k characters of reasoning, the pinball cabinet is not assembled. the table stands vertically on its edge like a wardrobe – the prompt asked for 6.5° from horizontal, it delivered 90°. the legs float detached in the void beside it. head-on it photographs beautifully; orbit ten degrees and it is a painted slab with four chrome rods hovering nearby • the playfield z-fights with the glass – hard black banding across the whole field as soon as you pull the camera back a note on the pinball, in fairness to kimi: nobody passed it. every model shipped broken ball physics and controls you cannot trust. it is the hardest prompt we have run and the whole field failed it, each in its own way kimi k3 reasons better than anything else here and it shows exactly where reasoning pays – physics constants, sequences, edge cases, traps the others walked into follow thehype. for 24/7 ai news, analysis and breakdowns
0:36

kimi k3 vs gpt 5.6 sol vs fable 5 vs grok 4.5 Kimi.ai just dropped kimi k3 – a 2.8t param native multimodal model, the first open 3t-class release. key facts: • 1m token context. stable latentmoe activating 16 of 896 experts, built on kimi delta attention (kda) and attention residuals • quantization-aware training from the sft stage onward – mxfp4 weights, mxfp8 activations. moonshot claims ~2.5x scaling efficiency over k2 • max thinking effort by default. low- and high-effort modes are "coming in updates" – there is no way to turn the thinking down today, and you feel it in every run • pricing: $0.30/mtok cache-hit input, $3.00/mtok cache-miss, $15.00/mtok output. claims >90% cache hit rate on coding workloads • benchmarks: swe marathon 42.0 (1st – fable 5: 35.0, sol: 39.0, opus 4.8: 40.0), terminal bench 2.1 88.3, browsecomp 91.2 (1st), program bench 77.8 (1st), gpqa-diamond 93.5. loses frontierswe 81.2 vs fable's 86.6, and deepswe 67.5 vs sol's 73.0 our test – 3 prompts, single-file html, Three.js, fully procedural, no assets: 1. photorealistic european roulette wheel – 37 pockets in the real sequence, mahogany clearcoat bowl, chrome turret, diamond deflectors, flick-to-spin, ball that spirals inward and settles on a mathematically real number 2. las vegas slot machine – 3 reels behind transmissive glass, drag the chrome lever to play, mechanical odometer counters modelled in 3d, coin physics on win 3. full pinball table – 6.5° tilted playfield, flipper impulse physics, spline ramps, drop targets, 6 bumpers, mechanical score reels in the backbox we ran the test on AI/ML API platform results: - cost #1 grok 4.5 – $0.30 #2 kimi k3 – $0.71 #3 gpt 5.6 sol – $2.05 #4 fable 5 – $7.69 - tokens #1 grok 4.5 – 34,241 #2 gpt 5.6 sol – 51,748 #3 fable 5 – 144,126 #4 kimi k3 – 157,999 - lines of code #1 gpt 5.6 sol – 3,054 #2 grok 4.5 – 3,047 #3 kimi k3 – 2,255 #4 fable 5 – 1,950 - generation time #1 grok 4.5 – 5.1 min #2 gpt 5.6 sol – 22.0 min #3 fable 5 – 31.5 min #4 kimi k3 – 75.6 min observations: • kimi k3 is cheap and it is slow. 75.6 minutes across three prompts against grok's 5.1. it is 2.4x grok's price and 15x grok's wall clock. the roulette took 15 min, the slot 18, the pinball 42 • it failed 2 of 3. only the roulette works. the slot machine has reel cutouts on both faces of the cabinet and the symbols face backwards – you can only read your spin by walking around to the rear of the machine. the pinball table stands vertically on its edge with the legs floating detached beside it. • 81% of kimi's output tokens are reasoning, not code. grok: 22%. you are not paying for a bigger answer, you are paying for a longer argument with itself • price per 100 shipped lines – grok $0.010, kimi $0.031, sol $0.067, fable $0.394. a 39x spread for the same three files kimi k3's code quality: upsides: • the roulette is genuinely good – procedural wood grain with real specular breakup, correct european sequence (0-32-15-19-4...), chrome turret, diamond deflectors, clean console • the pinball artwork is the best in the test – a synthwave "nova strike / deep space" field with six individually coloured neon bumper rings, a retro sun on a grid horizon, a nova burst, and a scoring legend printed on the apron. no other model printed the rules on the machine. it is a beautiful texture on a broken object • physics reasoning is real – it derived a 480hz substep for the collider, worked out ball settle conditions and termination guarantees, and checked every ramp exit vector by hand before writing any of it • it is the only model that saw the importmap trap coming. sol shipped a blank white page twice because three.js addons import the bare specifier 'three' and die without an import map downsides: • it dodged that trap on the slot by loading three.js r128 through classic script tags – a 2021 build with no working transmission. its slot glass rendered fully opaque and buried all three reels behind a white pane. the code asks for transmission: 0.93, ior: 1.5 – correct, and silently ignored by a renderer that predates the feature • after 42 minutes and 212k characters of reasoning, the pinball cabinet is not assembled. the table stands vertically on its edge like a wardrobe – the prompt asked for 6.5° from horizontal, it delivered 90°. the legs float detached in the void beside it. head-on it photographs beautifully; orbit ten degrees and it is a painted slab with four chrome rods hovering nearby • the playfield z-fights with the glass – hard black banding across the whole field as soon as you pull the camera back a note on the pinball, in fairness to kimi: nobody passed it. every model shipped broken ball physics and controls you cannot trust. it is the hardest prompt we have run and the whole field failed it, each in its own way kimi k3 reasons better than anything else here and it shows exactly where reasoning pays – physics constants, sequences, edge cases, traps the others walked into follow thehype. for 24/7 ai news, analysis and breakdowns

thehype.

2,153,239 views • 11 days ago

*** Mega Parodius Sega Megadrive Update *** Lately I've been having some fun seeing how far we can push the Parodius game engine . This is not arcade accurate - just for testing, decreased bullet timers and upped the bullet count to see what would happen re cpu usage when things are made a lot busier ! Maybe something like this could be insane mode in the options etc. Since last update we are a lot more optimised under the hood . The C based sprite engine, particularly visibility checking was speedup , overall 25 % faster. Then re-wrote the sprite engine completely in 68k assembly , 30% faster again. It took about 3 days and 1500 lines of assembly , thankfully the gains were worth it. All these gains will be back ported to S.O.T.A also as its using the same sprite engine. Usage is around 19% of cpu per frame with a full sprite load - 33 % cpu left in the busiest frame currently . Parodius is using a write the sprite list every frame type engine , so priority and meta-sprite objects can be handled with ease albiet its still a bit slower than static allocation sprite engines such as the one in lufthoheit but this is a bit easier to code for in the long run. All collision checks are been done , we are using a spatial grid system to get the collision checks done faster than a brute force approach . To get to maximum sprite count destruction is disabled in the video. We hit 80 sprites onscreen in this sequence , the sprite counter is one the Left side. We use about 16 sprites in the top hud and water line , about 21 for player attacks / missiles / shots / options . Up to 35 bullets + enemies . No lazer usage here as the Lazers use zero sprites thanks to the raster tricks , this video is all about the sprites ! Improved the water line when the Catboss is active , as he is Sprites + Forground we are doing some tile rotation tricks eg bit scrolling to give the impression of parrallax at that point . Vector Orbitex has updated the stage 1 music again and has made things even higher quality , hes mixed pcm and fm channels together to create higher quality orchestral hits for example . Pyron has started converting assets from stage 2/3 , some amazing work there . Hes well ahead of me at present which is a good thing. Still lots of incomplete animations missing logic etc , its slow going with RL getting in the way haha and 2 other projects !! Still its fun !! #SGDK #SegaMegadrive #SegaGenesis #Parodius
0:41

*** Mega Parodius Sega Megadrive Update *** Lately I've been having some fun seeing how far we can push the Parodius game engine . This is not arcade accurate - just for testing, decreased bullet timers and upped the bullet count to see what would happen re cpu usage when things are made a lot busier ! Maybe something like this could be insane mode in the options etc. Since last update we are a lot more optimised under the hood . The C based sprite engine, particularly visibility checking was speedup , overall 25 % faster. Then re-wrote the sprite engine completely in 68k assembly , 30% faster again. It took about 3 days and 1500 lines of assembly , thankfully the gains were worth it. All these gains will be back ported to S.O.T.A also as its using the same sprite engine. Usage is around 19% of cpu per frame with a full sprite load - 33 % cpu left in the busiest frame currently . Parodius is using a write the sprite list every frame type engine , so priority and meta-sprite objects can be handled with ease albiet its still a bit slower than static allocation sprite engines such as the one in lufthoheit but this is a bit easier to code for in the long run. All collision checks are been done , we are using a spatial grid system to get the collision checks done faster than a brute force approach . To get to maximum sprite count destruction is disabled in the video. We hit 80 sprites onscreen in this sequence , the sprite counter is one the Left side. We use about 16 sprites in the top hud and water line , about 21 for player attacks / missiles / shots / options . Up to 35 bullets + enemies . No lazer usage here as the Lazers use zero sprites thanks to the raster tricks , this video is all about the sprites ! Improved the water line when the Catboss is active , as he is Sprites + Forground we are doing some tile rotation tricks eg bit scrolling to give the impression of parrallax at that point . Vector Orbitex has updated the stage 1 music again and has made things even higher quality , hes mixed pcm and fm channels together to create higher quality orchestral hits for example . Pyron has started converting assets from stage 2/3 , some amazing work there . Hes well ahead of me at present which is a good thing. Still lots of incomplete animations missing logic etc , its slow going with RL getting in the way haha and 2 other projects !! Still its fun !! #SGDK #SegaMegadrive #SegaGenesis #Parodius

Shannon Birt

16,084 views • 7 months ago

Dear ICP community, the Internet Computer has now been running strong for 5 years 👏👏👏 Here is a celebratory preview of ICP "cloud engines," the sovereign frontier cloud technology the network shall soon provide from Main points: — Cloud engines enable anyone to spin up their own sovereign frontier cloud. The technology involves an extraordinary inventive step, in which cloud is created from a mathematically secure network of nodes. The nodes run as part of the Internet Computer network ( but are selected and configured by the cloud engine's owner. — The frontier cloud provided by engines is strongly focused on enabling AI agents to build and update online applications and services for us. The world is changing fast, and nearly all new online apps and services are already being built with the help of AI, and thus cloud engines target the future of cloud. — Software hosted on cloud engines is tamperproof, which means that it is immune to infrastructure hacks, because it runs inside a mathematically secure network protocol, rather than on computers directly. This means that AI agents, and those building with them, don't need to have a security team in the loop, or to trust someone else's security team. This is crucial, because in the future, non technical people will demand the freedom to build with full automation — where they just need to issue instructions to AI about what to build, and don't need to worry about anything or anyone else. Of course, apps and services running on engines are also vastly safer from the new breed of hacker being enabled by frontier AI. (The cloud engines themselves are also "tamperproof." Even if a hacker gains physical access to some portion of a cloud engine's nodes, and can make arbitrary changes, the computations and data of the hosted apps and services cannot be corrupted or interrupted so long as the network's fault bounds aren't exceeded. The recent hack of Vercel, a major cloud platform, which gave hackers access to the apps it hosted, provides additional perspective on the importance of this advantage.) — Software hosted on cloud engines is guaranteed to run, so long as a sufficient number of the engine's nodes are running. This means that AI can build applications and services without the need to have a human systems admin team constantly tinkering with the underlying platform to keep it running, which is again crucial, because in the future, non technical people will expect the freedom to use AI to build without the support of others. — New frontier programming language technology, in the form of the Motoko language developed by Caffeine Labs, leverages seminal "orthogonal persistence" technology that unifies program logic and data to deliver further unlocks for AI (Motoko is the first computer language being developed that targets agents that are writing software rather than humans engineers per se). Nowadays, AI can build and update production apps at a prodigious rate, even at the speed of conversation. But it can also make mistakes, and there's a risk that an update it creates might be "lossy" in the sense it causes some transformed data to be lost. Again, in this new world, it's both undesirable and impractical for everyone to have to have a systems admin team on-hand to detect lossy updates and roll them back, but Motoko provides a solution: it can detect new software updates are lossy before they are applied, reducing potentially catastrophic errors by AI to harmless coding retries. — Software hosted on cloud engines is "serverless" but unlike traditional serverless software, directly it directly incorporates data through "orthogonal persistence." Another key purpose is simplify backend software logic and fuel the modeling power of AI by increasing abstraction (sorry for the technical language!!!). Put simply, this enables AI to produce more sophisticated backends, faster, and at dramatically lower costs, as measured by the number AI API tokens consumed during coding. (Tip for the technical: orthogonal persistence is a new paradigm where "the program is the database," and data lives inside program variables, which is possible because it's as if hosted software runs forever in persistent memory). — An expanding database of skills at shall make it possible to develop and directly deploy apps and services to your cloud engines directly from Claude Code, Perplexity, Codex and other AI platforms. Further, your account on can be connected, so that new apps and updates created through conversation automatically appear hosted from your cloud engine. In the future, R&D is going to be very seamless. You converse with AI, and your secure and unstoppable apps or services are created or updated. Cloud engines are designed to directly support this "self-writing cloud" future where we can work hands-free. — Tech sovereignty is becoming a huge issue worldwide, with governments and corporations seeking to create sovereign tech stacks owing to geopolitical tensions. Increasingly, people are realizing that tech provided by foreign nations can come with hidden backdoors and kills switches, from the base platform, right up through hosted apps and services. ICP technology is open source, and those building on ICP using AI own their own source code. When you have the source code, you can verify that there are no backdoors, and when you own the source code thanks to AI, you can update it at will, freeing you from vendor lock-in. But cloud engines take sovereignty much further... — You create a cloud engine by selecting the nodes that will be combined. You can choose the class of nodes used, and their number, but more importantly, you can choose who operates the nodes, and where they are located. Almost any configuration is possible, because the Internet Computer scales the security privileges afforded to hosted software within the network according to configuration (software hosted on cloud engines can directly interoperate with software on other engines and traditional subnets, but base restrictions are applied according to security rules). A cloud engine can be created within a region such as Europe, to comply with regs such as GDPR, or completely within a sovereign state like Switzerland or Pakistan. But cloud engines go further still... — Sovereignty is also about freedom from vendor lock-in. Cloud engines are essentially ICP (Internet Computer Protocol) network configurations, and this means the underlying compute nodes they combine can be swapped out without interrupting their hosted apps and services. This is a big deal. In addition, cloud engines now support nodes that are instances running on Big Tech's clouds, in addition to nodes that are dedicated specialized hardware, as per the Gen I and Gen II nodes that dominate the Internet Computer today. For example, it is possible to have an engine running across different AWS data centers, say, and then reconfigure the engine to run across a mixture of AWS, Google, Azure and Hetzner for even more resilience, without the users of hosted apps and services noticing a thing. That's true freedom. — Sovereign AI is becoming increasingly important too, and cloud engines allow special "AI nodes" to be added to them, so that hosted software can perform inference on hardware provisioned by the owner from a location the owner has selected. Even though the AI nodes are only accessible within the cloud engine, they can still benefit from the forthcoming Internet Intelligence Gateway (IG), which will make it possible to validate inference performed on key frontier open weights LLMs, even when the inference is performed on completely independent AI clouds. When the results of inference are received, this technology can verify that neither the prompt+context (input) nor the inference result (output) have been modified, and that the results were produced by the precise LLM expected. This ensures that AI clouds don't cheat by running inference on cheaper models than are being paid for, and bad actors aren't modifying the inputs or outputs to surreptitiously insert advertising into results, say, or change facts, or insert malware when code is being generated. What's super cool about this technology is the cost of the verification is scalable. A very valuable additional security can be achieved with only 1-2% of extra cost. — Scaling apps and services when they hit capacity limits is another thorny problem that cloud engines help the world address. Engines make scaling possible without rewriting or reconfiguring software. The query workload capacity of hosted software can be horizontally scaled simply by adding new nodes to an engine, and nodes can also be added in geographical proximity to demand. Meanwhile, update workload capacity can first be scaled-up by swapping an engine's nodes out for the next class up, and then when no larger class of node is available, horizontally scaled-out by "splitting" the engine into two, which doubles available capacity. (Technical tip: horizontally scaling update capacity by splitting engines requires multi-canister architectures). — For those who have been following how Caffeine builds apps that can efficiently store large numbers of files, I should mention that apps built on cloud engines will also support the new ICP Blob Storage cloud network (since cloud engines currently have up to about 3 TB of memory, which apps storing large amounts of files can easily exceed). We are also working on allowing blob storage nodes to be added to cloud engines, to enable sovereign mass blob storage within an engine, similarly to how AI nodes can be added currently. — Lastly, but certainly not least, I should mention that cloud engines are multi-blockchain capable, and ready for digital assets, thanks to the clever math at their core. For example, an e-commerce service built on a cloud engine can securely accept and custody stablecoin payments, or a multi-chain DEX could be hosted. Further, engines can support software autonomy (software orchestrated and controlled by other autonomous software, in a decentralized way) and can themselves be orchestrated by SNS technology, and thus run autonomously too. Today, though, the focus is on *mainstream* cloud. This year, the cloud industry will generate approximately one trillion dollars in revenue. That number is already huge, but is expected to grow to two trillion dollars by 2030. After years of continuous development, which have seen more than $500m spent on R&D, the Internet Computer network is now tacking directly toward this mainstream cloud market with cloud engine technology. In their first version, cloud engines are not meant to be a cloud panacea. For example, currently they are not ideal for working with big data. You should use something like DataBricks for that. Cloud engines are carefully targeted at enabling AI to produce traditional online applications and services, including SaaS, in a safer and more productive way, which represents a new market segment with tremendous potential. Of course, DFINITY will continue to work relentlessly to push forward ICP's capabilities, so expect further developments. It's worth mentioning that this cloud segment isn't just about creating new apps and services using AI, it's also about replacing legacy systems and apps built on super expensive SaaS services. Caffeine Labs is working to produce technology (Caffeine Snorkel) that can study an enterprise's legacy systems and app built on SaaS, create replacement systems and apps, and migrate the data, while supporting key stakeholders through the process over email and chat, with full automation. Thus the legacy systems and SaaS markets shall also be addressed by cloud engines. Zooming out, and reasoning in a more metaphysical way, we believe, as we always have, that there is room for a new kind of cloud created by mathematical networks, that provides seminal advances in the fields of security and resilience, as well as true sovereignty and freedom from lock-in. That this same technology, with the help of additional technologies like orthogonal persistence and Motoko, enables AI to build for us without the need for so much oversight, and to create more backend sophistication while consuming fewer AI API tokens, enables ICP to bring game-changing advances to the world. Cloud engines will work synergistically with the Intelligence Gateway, which will enable apps and services running on engines to seamlessly leverage AI, wherever that AI is running, while providing verifiability at extremely low cost for open weights frontier models. We believe that cloud engines represent an inflection point in the storied history of the Internet Computer project, and I'm very proud to be sharing the details with you on the network's fifth birthday 💪 I'll be back with more news soon!!
19:06

Dear ICP community, the Internet Computer has now been running strong for 5 years 👏👏👏 Here is a celebratory preview of ICP "cloud engines," the sovereign frontier cloud technology the network shall soon provide from Main points: — Cloud engines enable anyone to spin up their own sovereign frontier cloud. The technology involves an extraordinary inventive step, in which cloud is created from a mathematically secure network of nodes. The nodes run as part of the Internet Computer network ( but are selected and configured by the cloud engine's owner. — The frontier cloud provided by engines is strongly focused on enabling AI agents to build and update online applications and services for us. The world is changing fast, and nearly all new online apps and services are already being built with the help of AI, and thus cloud engines target the future of cloud. — Software hosted on cloud engines is tamperproof, which means that it is immune to infrastructure hacks, because it runs inside a mathematically secure network protocol, rather than on computers directly. This means that AI agents, and those building with them, don't need to have a security team in the loop, or to trust someone else's security team. This is crucial, because in the future, non technical people will demand the freedom to build with full automation — where they just need to issue instructions to AI about what to build, and don't need to worry about anything or anyone else. Of course, apps and services running on engines are also vastly safer from the new breed of hacker being enabled by frontier AI. (The cloud engines themselves are also "tamperproof." Even if a hacker gains physical access to some portion of a cloud engine's nodes, and can make arbitrary changes, the computations and data of the hosted apps and services cannot be corrupted or interrupted so long as the network's fault bounds aren't exceeded. The recent hack of Vercel, a major cloud platform, which gave hackers access to the apps it hosted, provides additional perspective on the importance of this advantage.) — Software hosted on cloud engines is guaranteed to run, so long as a sufficient number of the engine's nodes are running. This means that AI can build applications and services without the need to have a human systems admin team constantly tinkering with the underlying platform to keep it running, which is again crucial, because in the future, non technical people will expect the freedom to use AI to build without the support of others. — New frontier programming language technology, in the form of the Motoko language developed by Caffeine Labs, leverages seminal "orthogonal persistence" technology that unifies program logic and data to deliver further unlocks for AI (Motoko is the first computer language being developed that targets agents that are writing software rather than humans engineers per se). Nowadays, AI can build and update production apps at a prodigious rate, even at the speed of conversation. But it can also make mistakes, and there's a risk that an update it creates might be "lossy" in the sense it causes some transformed data to be lost. Again, in this new world, it's both undesirable and impractical for everyone to have to have a systems admin team on-hand to detect lossy updates and roll them back, but Motoko provides a solution: it can detect new software updates are lossy before they are applied, reducing potentially catastrophic errors by AI to harmless coding retries. — Software hosted on cloud engines is "serverless" but unlike traditional serverless software, directly it directly incorporates data through "orthogonal persistence." Another key purpose is simplify backend software logic and fuel the modeling power of AI by increasing abstraction (sorry for the technical language!!!). Put simply, this enables AI to produce more sophisticated backends, faster, and at dramatically lower costs, as measured by the number AI API tokens consumed during coding. (Tip for the technical: orthogonal persistence is a new paradigm where "the program is the database," and data lives inside program variables, which is possible because it's as if hosted software runs forever in persistent memory). — An expanding database of skills at shall make it possible to develop and directly deploy apps and services to your cloud engines directly from Claude Code, Perplexity, Codex and other AI platforms. Further, your account on can be connected, so that new apps and updates created through conversation automatically appear hosted from your cloud engine. In the future, R&D is going to be very seamless. You converse with AI, and your secure and unstoppable apps or services are created or updated. Cloud engines are designed to directly support this "self-writing cloud" future where we can work hands-free. — Tech sovereignty is becoming a huge issue worldwide, with governments and corporations seeking to create sovereign tech stacks owing to geopolitical tensions. Increasingly, people are realizing that tech provided by foreign nations can come with hidden backdoors and kills switches, from the base platform, right up through hosted apps and services. ICP technology is open source, and those building on ICP using AI own their own source code. When you have the source code, you can verify that there are no backdoors, and when you own the source code thanks to AI, you can update it at will, freeing you from vendor lock-in. But cloud engines take sovereignty much further... — You create a cloud engine by selecting the nodes that will be combined. You can choose the class of nodes used, and their number, but more importantly, you can choose who operates the nodes, and where they are located. Almost any configuration is possible, because the Internet Computer scales the security privileges afforded to hosted software within the network according to configuration (software hosted on cloud engines can directly interoperate with software on other engines and traditional subnets, but base restrictions are applied according to security rules). A cloud engine can be created within a region such as Europe, to comply with regs such as GDPR, or completely within a sovereign state like Switzerland or Pakistan. But cloud engines go further still... — Sovereignty is also about freedom from vendor lock-in. Cloud engines are essentially ICP (Internet Computer Protocol) network configurations, and this means the underlying compute nodes they combine can be swapped out without interrupting their hosted apps and services. This is a big deal. In addition, cloud engines now support nodes that are instances running on Big Tech's clouds, in addition to nodes that are dedicated specialized hardware, as per the Gen I and Gen II nodes that dominate the Internet Computer today. For example, it is possible to have an engine running across different AWS data centers, say, and then reconfigure the engine to run across a mixture of AWS, Google, Azure and Hetzner for even more resilience, without the users of hosted apps and services noticing a thing. That's true freedom. — Sovereign AI is becoming increasingly important too, and cloud engines allow special "AI nodes" to be added to them, so that hosted software can perform inference on hardware provisioned by the owner from a location the owner has selected. Even though the AI nodes are only accessible within the cloud engine, they can still benefit from the forthcoming Internet Intelligence Gateway (IG), which will make it possible to validate inference performed on key frontier open weights LLMs, even when the inference is performed on completely independent AI clouds. When the results of inference are received, this technology can verify that neither the prompt+context (input) nor the inference result (output) have been modified, and that the results were produced by the precise LLM expected. This ensures that AI clouds don't cheat by running inference on cheaper models than are being paid for, and bad actors aren't modifying the inputs or outputs to surreptitiously insert advertising into results, say, or change facts, or insert malware when code is being generated. What's super cool about this technology is the cost of the verification is scalable. A very valuable additional security can be achieved with only 1-2% of extra cost. — Scaling apps and services when they hit capacity limits is another thorny problem that cloud engines help the world address. Engines make scaling possible without rewriting or reconfiguring software. The query workload capacity of hosted software can be horizontally scaled simply by adding new nodes to an engine, and nodes can also be added in geographical proximity to demand. Meanwhile, update workload capacity can first be scaled-up by swapping an engine's nodes out for the next class up, and then when no larger class of node is available, horizontally scaled-out by "splitting" the engine into two, which doubles available capacity. (Technical tip: horizontally scaling update capacity by splitting engines requires multi-canister architectures). — For those who have been following how Caffeine builds apps that can efficiently store large numbers of files, I should mention that apps built on cloud engines will also support the new ICP Blob Storage cloud network (since cloud engines currently have up to about 3 TB of memory, which apps storing large amounts of files can easily exceed). We are also working on allowing blob storage nodes to be added to cloud engines, to enable sovereign mass blob storage within an engine, similarly to how AI nodes can be added currently. — Lastly, but certainly not least, I should mention that cloud engines are multi-blockchain capable, and ready for digital assets, thanks to the clever math at their core. For example, an e-commerce service built on a cloud engine can securely accept and custody stablecoin payments, or a multi-chain DEX could be hosted. Further, engines can support software autonomy (software orchestrated and controlled by other autonomous software, in a decentralized way) and can themselves be orchestrated by SNS technology, and thus run autonomously too. Today, though, the focus is on *mainstream* cloud. This year, the cloud industry will generate approximately one trillion dollars in revenue. That number is already huge, but is expected to grow to two trillion dollars by 2030. After years of continuous development, which have seen more than $500m spent on R&D, the Internet Computer network is now tacking directly toward this mainstream cloud market with cloud engine technology. In their first version, cloud engines are not meant to be a cloud panacea. For example, currently they are not ideal for working with big data. You should use something like DataBricks for that. Cloud engines are carefully targeted at enabling AI to produce traditional online applications and services, including SaaS, in a safer and more productive way, which represents a new market segment with tremendous potential. Of course, DFINITY will continue to work relentlessly to push forward ICP's capabilities, so expect further developments. It's worth mentioning that this cloud segment isn't just about creating new apps and services using AI, it's also about replacing legacy systems and apps built on super expensive SaaS services. Caffeine Labs is working to produce technology (Caffeine Snorkel) that can study an enterprise's legacy systems and app built on SaaS, create replacement systems and apps, and migrate the data, while supporting key stakeholders through the process over email and chat, with full automation. Thus the legacy systems and SaaS markets shall also be addressed by cloud engines. Zooming out, and reasoning in a more metaphysical way, we believe, as we always have, that there is room for a new kind of cloud created by mathematical networks, that provides seminal advances in the fields of security and resilience, as well as true sovereignty and freedom from lock-in. That this same technology, with the help of additional technologies like orthogonal persistence and Motoko, enables AI to build for us without the need for so much oversight, and to create more backend sophistication while consuming fewer AI API tokens, enables ICP to bring game-changing advances to the world. Cloud engines will work synergistically with the Intelligence Gateway, which will enable apps and services running on engines to seamlessly leverage AI, wherever that AI is running, while providing verifiability at extremely low cost for open weights frontier models. We believe that cloud engines represent an inflection point in the storied history of the Internet Computer project, and I'm very proud to be sharing the details with you on the network's fifth birthday 💪 I'll be back with more news soon!!

dom | icp

268,830 views • 2 months ago

This is one-shot assembly: you show examples of what to build, and the robot just does it. (see original post: To share more on how this works, the robot is controlled in real time by a neural network that takes in video pixels and outputs 100Hz actions. The video below is part of the raw input passed directly into the model. I also like this view (at 1x speed) because it shows more of the (I think very cool) subtle moments of dexterity near the fingertips 👌 One-shot assembly seemed like a dream even just a year ago — it's not easy. It requires both the high-level reasoning of "what to build" (recognizing the geometry of the structures presented by the human), and the low-level visuomotor control of "how to build it" (purposefully re-orienting individual pieces and nudging them together in place). While possible to manually engineer a complex system for this (e.g. w/ hierarchical control, or explicit state representations), we were curious if our own Foundation model could do it all end-to-end with just some post-training data. Surprisingly, it just worked. Nothing about the recipe is substantially different than any other demo we’ve run in the past, and we’re excited about its implications on model capabilities: • On contextual reasoning, these models can (i) attend to task-related pixels in the peripheral view of the video inputs, and (ii) retain this knowledge in-context while ignoring irrelevant background. This is useful for generalizing to a wide range of real workflows: e.g. paying attention to what’s coming down the conveyor line, or glancing at the instructions displayed on a nearby monitor. • On dexterity, these models can produce contact-rich "commonsense" behaviors that can be difficult to pre-program or write language instructions for e.g. rolling a brick slightly to align its studs against the bottom of another, re-grasping to get a better grip or to move out of the way before a forceful press, or gently pushing the corners of a brick against the mat to rotate it in hand and stand it up vertically (i.e. extrinsic dexterity). These aspects work together to form a capability that resembles fast adaptation — a hallmark of intelligence, relevant for real use cases. This has also expanded my own perspective on what's possible with robot learning, using a recipe that's repeatable for many more skills. This milestone stands on top of the solid technical foundations we’ve built here at Generalist: hardcore controls & hardware, all in-house built models, and a data engine that "just works." We're a small group of hyper-focused engineers, and hands-down the highest talent-density team I’ve ever worked with. We're accelerating and scaling aggressively towards unlocking next-generation robot intelligence. Building Legos is just one example, and it's clear to me that we're headed towards a future where robots can do just about anything we want them to. Its coming, and we're going to make it happen.
1:01

This is one-shot assembly: you show examples of what to build, and the robot just does it. (see original post: To share more on how this works, the robot is controlled in real time by a neural network that takes in video pixels and outputs 100Hz actions. The video below is part of the raw input passed directly into the model. I also like this view (at 1x speed) because it shows more of the (I think very cool) subtle moments of dexterity near the fingertips 👌 One-shot assembly seemed like a dream even just a year ago — it's not easy. It requires both the high-level reasoning of "what to build" (recognizing the geometry of the structures presented by the human), and the low-level visuomotor control of "how to build it" (purposefully re-orienting individual pieces and nudging them together in place). While possible to manually engineer a complex system for this (e.g. w/ hierarchical control, or explicit state representations), we were curious if our own Foundation model could do it all end-to-end with just some post-training data. Surprisingly, it just worked. Nothing about the recipe is substantially different than any other demo we’ve run in the past, and we’re excited about its implications on model capabilities: • On contextual reasoning, these models can (i) attend to task-related pixels in the peripheral view of the video inputs, and (ii) retain this knowledge in-context while ignoring irrelevant background. This is useful for generalizing to a wide range of real workflows: e.g. paying attention to what’s coming down the conveyor line, or glancing at the instructions displayed on a nearby monitor. • On dexterity, these models can produce contact-rich "commonsense" behaviors that can be difficult to pre-program or write language instructions for e.g. rolling a brick slightly to align its studs against the bottom of another, re-grasping to get a better grip or to move out of the way before a forceful press, or gently pushing the corners of a brick against the mat to rotate it in hand and stand it up vertically (i.e. extrinsic dexterity). These aspects work together to form a capability that resembles fast adaptation — a hallmark of intelligence, relevant for real use cases. This has also expanded my own perspective on what's possible with robot learning, using a recipe that's repeatable for many more skills. This milestone stands on top of the solid technical foundations we’ve built here at Generalist: hardcore controls & hardware, all in-house built models, and a data engine that "just works." We're a small group of hyper-focused engineers, and hands-down the highest talent-density team I’ve ever worked with. We're accelerating and scaling aggressively towards unlocking next-generation robot intelligence. Building Legos is just one example, and it's clear to me that we're headed towards a future where robots can do just about anything we want them to. Its coming, and we're going to make it happen.

Andy Zeng

49,317 views • 10 months ago

The video shows the trials of the VT-1-1, a turretless tank with 2 x 105 mm guns, firing on the move at the Putlos training ground in 1976. The casemate (turretless) tank, designed for combat while moving, was introduced in Germany in the mid-1970s as a twin-gun casemate tank (the Soviet term for a "turretless tank"). For practical firing tests in the "target pass" mode, two prototypes were built in the mid-1970s. The first prototype, VT 1-1, was armed with two 105 mm guns, while the second, VT 1-2, was equipped with two 120 mm smoothbore guns. Additionally, the VT 1-2 featured a functioning autoloader behind one of the guns, with a firing rate of 10 rounds per minute. The vehicles were developed as part of the KPz 3 or Leopard 3 project. In both prototypes, the main guns were semi-fixed (with aiming and stabilization only in elevation). The chassis solutions were derived from the KPz 70 (MBT 70) program, with the running gear shortened by one road wheel (five road wheels per side). The vehicles weighed 36.8 and 43.5 tons, respectively, with a chassis rotation speed of 60 degrees per second. To achieve high maneuverability on rough terrain, 12-cylinder diesel engines from the MB 873 series with enhanced power were used, equipped with four turbochargers: 2000 hp for VT 1-1 and 2200 hp for VT 1-2. This provided an impressive power-to-weight ratio of 54 and 50 hp/ton, respectively, with acceleration to 55 km/h in 11 seconds, though only in a temporary "turbo" mode, as the standard power was 1600 hp. The accuracy of firing with two guns was unmatched by single-gun tanks, as confirmed by the tests. However, due to the novel combat approach, this revolutionary tank concept was rejected by the customer after trials in favor of the conventionally designed Leopard 2. In essence, although the project was developed as the Leopard 3, a tank for the future, it was, in reality, a parallel project and a potential competitor to the Leopard 2. Achieving the firepower, protection, and mobility of the Leopard 3 within reasonable weight limits was impossible with a conventional layout. At the same time, a significant tactical drawback of the casemate concept (including twin-gun casemate vehicles) is the linkage between the direction of fire and the direction of movement, which in many cases could complicate unit and formation control (according to German experts in the 1970s). Moreover, the VT 1-1 and VT 1-2 can hardly be considered balanced vehicles—they could have been simpler. The vehicles' mobility was exaggerated, with the main engine, transmission, auxiliary engine, batteries, and other systems occupying two-thirds of the vehicle's length. Pros and Cons of VT 1-1 and VT 1-2: Considering the key challenges, the twin-gun casemate concept can be evaluated as follows: Pros: Compact design due to a small internal volume. Two guns provide high firepower and hit probability. Cons: The weight advantage of the casemate design is largely offset by the integration of a second gun. Fire control alone results in high complexity, leading to increased maintenance costs and overall expenses. In 1975/76, five Gefechtsfeldversuchträger (GVT, combat test platforms) were developed and built for further mobility and concept trials. These were smaller than the VT 1-1 and VT 1-2, weighing 30 tons. The GVT 01-05 were equipped only with mock-up guns and laser firing simulators (TALLISSI) and telemetry systems, built using chassis components from the Leopard 1. The GVTs were used at the IABG facility in Lichtenau and the tank training school in Munster to test the twin-gun turretless tank concept in realistic exercise conditions, which is why five vehicles were built. However, the Leopard 2 was already in production, and tank crews showed little enthusiasm for this unconventional vehicle requiring a new approach.
0:20

The video shows the trials of the VT-1-1, a turretless tank with 2 x 105 mm guns, firing on the move at the Putlos training ground in 1976. The casemate (turretless) tank, designed for combat while moving, was introduced in Germany in the mid-1970s as a twin-gun casemate tank (the Soviet term for a "turretless tank"). For practical firing tests in the "target pass" mode, two prototypes were built in the mid-1970s. The first prototype, VT 1-1, was armed with two 105 mm guns, while the second, VT 1-2, was equipped with two 120 mm smoothbore guns. Additionally, the VT 1-2 featured a functioning autoloader behind one of the guns, with a firing rate of 10 rounds per minute. The vehicles were developed as part of the KPz 3 or Leopard 3 project. In both prototypes, the main guns were semi-fixed (with aiming and stabilization only in elevation). The chassis solutions were derived from the KPz 70 (MBT 70) program, with the running gear shortened by one road wheel (five road wheels per side). The vehicles weighed 36.8 and 43.5 tons, respectively, with a chassis rotation speed of 60 degrees per second. To achieve high maneuverability on rough terrain, 12-cylinder diesel engines from the MB 873 series with enhanced power were used, equipped with four turbochargers: 2000 hp for VT 1-1 and 2200 hp for VT 1-2. This provided an impressive power-to-weight ratio of 54 and 50 hp/ton, respectively, with acceleration to 55 km/h in 11 seconds, though only in a temporary "turbo" mode, as the standard power was 1600 hp. The accuracy of firing with two guns was unmatched by single-gun tanks, as confirmed by the tests. However, due to the novel combat approach, this revolutionary tank concept was rejected by the customer after trials in favor of the conventionally designed Leopard 2. In essence, although the project was developed as the Leopard 3, a tank for the future, it was, in reality, a parallel project and a potential competitor to the Leopard 2. Achieving the firepower, protection, and mobility of the Leopard 3 within reasonable weight limits was impossible with a conventional layout. At the same time, a significant tactical drawback of the casemate concept (including twin-gun casemate vehicles) is the linkage between the direction of fire and the direction of movement, which in many cases could complicate unit and formation control (according to German experts in the 1970s). Moreover, the VT 1-1 and VT 1-2 can hardly be considered balanced vehicles—they could have been simpler. The vehicles' mobility was exaggerated, with the main engine, transmission, auxiliary engine, batteries, and other systems occupying two-thirds of the vehicle's length. Pros and Cons of VT 1-1 and VT 1-2: Considering the key challenges, the twin-gun casemate concept can be evaluated as follows: Pros: Compact design due to a small internal volume. Two guns provide high firepower and hit probability. Cons: The weight advantage of the casemate design is largely offset by the integration of a second gun. Fire control alone results in high complexity, leading to increased maintenance costs and overall expenses. In 1975/76, five Gefechtsfeldversuchträger (GVT, combat test platforms) were developed and built for further mobility and concept trials. These were smaller than the VT 1-1 and VT 1-2, weighing 30 tons. The GVT 01-05 were equipped only with mock-up guns and laser firing simulators (TALLISSI) and telemetry systems, built using chassis components from the Leopard 1. The GVTs were used at the IABG facility in Lichtenau and the tank training school in Munster to test the twin-gun turretless tank concept in realistic exercise conditions, which is why five vehicles were built. However, the Leopard 2 was already in production, and tank crews showed little enthusiasm for this unconventional vehicle requiring a new approach.

Andrei_bt

81,636 views • 1 year ago

Google just proved that bigger isn't always better. Their 308M parameter model is outperforming models 2x its size. Google just released 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝗚𝗲𝗺𝗺𝗮, and it's proving that lightweight embedding models can punch way above their weight class. At just 308M parameters (578MB), it's the new state-of-the-art for models under 500M parameters across MTEB multilingual, English, and code benchmarks. But the really impressive part is that it ranks 8th overall on MTEB(Multilingual, v2) - that's 𝟭𝟳 𝗽𝗹𝗮𝗰𝗲𝘀 above the second-best sub-500M model, and it's delivering performance 𝗰𝗼𝗺𝗽𝗮𝗿𝗮𝗯𝗹𝗲 𝘁𝗼 𝗺𝗼𝗱𝗲𝗹𝘀 𝗻𝗲𝗮𝗿𝗹𝘆 𝗱𝗼𝘂𝗯𝗹𝗲 𝗶𝘁𝘀 𝘀𝗶𝘇𝗲. There are three key parts of their training recipe that sets it apart: 𝟭. 𝗘𝗻𝗰𝗼𝗱𝗲𝗿-𝗗𝗲𝗰𝗼𝗱𝗲𝗿 𝗜𝗻𝗶𝘁𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 Instead of starting from a decoder-only Gemma 3 model, they first adapted it to encoder-decoder, then used just the encoder. By basing EmbeddingGemma off an LLM that already has world and language understanding, it gives it a stronger starting point. 𝟮. 𝗧𝗵𝗿𝗲𝗲-𝗟𝗼𝘀𝘀 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 They combine three different loss functions, instead of just having one: • Contrastive loss (NCE) with in-batch negatives and hardness weighting • Spread-out regularization to ensure embeddings utilize the full space (for quantization and ANN retrieval) • Embedding matching distillation from Gemini Embedding - not just learning from relevance scores, but directly aligning the embedding space with the teacher model 𝟯. 𝗠𝗼𝗱𝗲𝗹 𝗦𝗼𝘂𝗽𝗶𝗻𝗴 Rather than just averaging checkpoints from the same training run, they use optimization techniques to find multiple specialized training mixtures. Each mixture creates an "expert" model in different domains, and averaging all their parameters creates a final model that's actually better than individual models. Extras: • Matryoshka embeddings supporting 768, 512, 256, and 128 dimensions • Quantization-aware training - maintains quality even at int4 precision • 100+ languages from Gemma 3 pretraining • Exceptional performance on low-resource languages (check their XTREME-UP results) Is it the absolute best embedding model? No - Gemini Embedding still leads overall. But that's not really the point. EmbeddingGemma proves you can achieve state-of-the-art performance in a small package that's actually deployable on-device, in low-latency applications, and in resource-constrained environments. This makes good embeddings accessible for use cases that I'm seeing more and more: offline applications, privacy-sensitive deployments, and high-throughput scenarios where inference cost actually matters. Full paper: Shoutout to the EmbeddingGemma team at Google DeepMind for this awesome open source work 💙 and to Daniel Williams for helping me with this video! 🫶
5:14

Google just proved that bigger isn't always better. Their 308M parameter model is outperforming models 2x its size. Google just released 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝗚𝗲𝗺𝗺𝗮, and it's proving that lightweight embedding models can punch way above their weight class. At just 308M parameters (578MB), it's the new state-of-the-art for models under 500M parameters across MTEB multilingual, English, and code benchmarks. But the really impressive part is that it ranks 8th overall on MTEB(Multilingual, v2) - that's 𝟭𝟳 𝗽𝗹𝗮𝗰𝗲𝘀 above the second-best sub-500M model, and it's delivering performance 𝗰𝗼𝗺𝗽𝗮𝗿𝗮𝗯𝗹𝗲 𝘁𝗼 𝗺𝗼𝗱𝗲𝗹𝘀 𝗻𝗲𝗮𝗿𝗹𝘆 𝗱𝗼𝘂𝗯𝗹𝗲 𝗶𝘁𝘀 𝘀𝗶𝘇𝗲. There are three key parts of their training recipe that sets it apart: 𝟭. 𝗘𝗻𝗰𝗼𝗱𝗲𝗿-𝗗𝗲𝗰𝗼𝗱𝗲𝗿 𝗜𝗻𝗶𝘁𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 Instead of starting from a decoder-only Gemma 3 model, they first adapted it to encoder-decoder, then used just the encoder. By basing EmbeddingGemma off an LLM that already has world and language understanding, it gives it a stronger starting point. 𝟮. 𝗧𝗵𝗿𝗲𝗲-𝗟𝗼𝘀𝘀 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 They combine three different loss functions, instead of just having one: • Contrastive loss (NCE) with in-batch negatives and hardness weighting • Spread-out regularization to ensure embeddings utilize the full space (for quantization and ANN retrieval) • Embedding matching distillation from Gemini Embedding - not just learning from relevance scores, but directly aligning the embedding space with the teacher model 𝟯. 𝗠𝗼𝗱𝗲𝗹 𝗦𝗼𝘂𝗽𝗶𝗻𝗴 Rather than just averaging checkpoints from the same training run, they use optimization techniques to find multiple specialized training mixtures. Each mixture creates an "expert" model in different domains, and averaging all their parameters creates a final model that's actually better than individual models. Extras: • Matryoshka embeddings supporting 768, 512, 256, and 128 dimensions • Quantization-aware training - maintains quality even at int4 precision • 100+ languages from Gemma 3 pretraining • Exceptional performance on low-resource languages (check their XTREME-UP results) Is it the absolute best embedding model? No - Gemini Embedding still leads overall. But that's not really the point. EmbeddingGemma proves you can achieve state-of-the-art performance in a small package that's actually deployable on-device, in low-latency applications, and in resource-constrained environments. This makes good embeddings accessible for use cases that I'm seeing more and more: offline applications, privacy-sensitive deployments, and high-throughput scenarios where inference cost actually matters. Full paper: Shoutout to the EmbeddingGemma team at Google DeepMind for this awesome open source work 💙 and to Daniel Williams for helping me with this video! 🫶

Victoria Slocum

21,592 views • 8 months ago

#Jasmy and #Janction are entering a new phase of expansion. The team aims to make its platform even more accessible, particularly by simplifying the development environment for application creators. A key focus is the introduction of an English interface, clearer documentation, and enhanced technical support. At the same time, #Jasmy is intensifying its international efforts, with particular attention to Southeast Asia. The year 2025 promises stronger communications and several major announcements ahead. Janction, on its part, has undergone a major transformation. It is no longer just a side project but now a true decentralized physical infrastructure built on a simple idea: everyone should be able to own and benefit from their own assets, including their data and computing power. Today, artificial intelligence, image generation, video rendering, and large-scale data processing all heavily rely on a single resource: the GPU. Originally designed for gaming, GPUs have become essential for any task that requires massive parallel processing. Unlike CPUs, GPUs can execute thousands of operations simultaneously, making them ideal for machine learning and AI models. However, this exponential demand has led to a global shortage. GPU prices are skyrocketing, lead times stretch up to a year, and the market is dominated by a few major players. In Japan and across Asia, the situation is especially strained. This is where Janction steps in with a disruptive approach. The idea is to allow any user to share an unused GPU, whether it’s in a gaming PC, a company server, a university lab, or even a cybercafé. In return, the owner gets paid. And to make this process smooth, simple, and secure, Janction relies on Docker technology. To visualize this, imagine a box containing everything needed to run an application, the code, libraries, and required files. Thanks to Docker, this box can be sent and run on any computer without conflicts or manual setup. This allows Janction to distribute AI or processing tasks across its network seamlessly. Each user receives a container, runs it via their GPU, and is paid automatically through smart contracts deployed on the network. The system is based on a fixed-rate subleasing model. Even if the GPU isn’t used 24/7, the owner still earns income. This is an ideal solution for schools, creative studios, researchers, or startups that have available resources but variable needs. Today, over 4,500 GPU nodes are already active in Japan, Hong Kong, and Singapore. The network offers fast block times and 99.9% reliability. The goal is ambitious: reach 100,000 nodes. To achieve this, Janction is targeting six main markets: AI startups, 3D and video studios, streaming platforms, research centers, game developers, and of course, owners of underutilized GPUs. At the same time, an Ethereum-based JANCTION token is in preparation. It will be used to reserve GPU power, participate in the ecosystem, and unlock additional rewards, including JASMY tokens. This dual-incentive system is designed to encourage the large-scale acquisition, sharing, and use of GPU power. The tokens will be tradable, storable, or reinvestable into hardware to further strengthen the network. #Janction’s strategy is clear, first, establish strong liquidity on recognized exchanges, then open access to a broad investor base, especially in #Japan, South Korea, and the United States.
3:23

#Jasmy and #Janction are entering a new phase of expansion. The team aims to make its platform even more accessible, particularly by simplifying the development environment for application creators. A key focus is the introduction of an English interface, clearer documentation, and enhanced technical support. At the same time, #Jasmy is intensifying its international efforts, with particular attention to Southeast Asia. The year 2025 promises stronger communications and several major announcements ahead. Janction, on its part, has undergone a major transformation. It is no longer just a side project but now a true decentralized physical infrastructure built on a simple idea: everyone should be able to own and benefit from their own assets, including their data and computing power. Today, artificial intelligence, image generation, video rendering, and large-scale data processing all heavily rely on a single resource: the GPU. Originally designed for gaming, GPUs have become essential for any task that requires massive parallel processing. Unlike CPUs, GPUs can execute thousands of operations simultaneously, making them ideal for machine learning and AI models. However, this exponential demand has led to a global shortage. GPU prices are skyrocketing, lead times stretch up to a year, and the market is dominated by a few major players. In Japan and across Asia, the situation is especially strained. This is where Janction steps in with a disruptive approach. The idea is to allow any user to share an unused GPU, whether it’s in a gaming PC, a company server, a university lab, or even a cybercafé. In return, the owner gets paid. And to make this process smooth, simple, and secure, Janction relies on Docker technology. To visualize this, imagine a box containing everything needed to run an application, the code, libraries, and required files. Thanks to Docker, this box can be sent and run on any computer without conflicts or manual setup. This allows Janction to distribute AI or processing tasks across its network seamlessly. Each user receives a container, runs it via their GPU, and is paid automatically through smart contracts deployed on the network. The system is based on a fixed-rate subleasing model. Even if the GPU isn’t used 24/7, the owner still earns income. This is an ideal solution for schools, creative studios, researchers, or startups that have available resources but variable needs. Today, over 4,500 GPU nodes are already active in Japan, Hong Kong, and Singapore. The network offers fast block times and 99.9% reliability. The goal is ambitious: reach 100,000 nodes. To achieve this, Janction is targeting six main markets: AI startups, 3D and video studios, streaming platforms, research centers, game developers, and of course, owners of underutilized GPUs. At the same time, an Ethereum-based JANCTION token is in preparation. It will be used to reserve GPU power, participate in the ecosystem, and unlock additional rewards, including JASMY tokens. This dual-incentive system is designed to encourage the large-scale acquisition, sharing, and use of GPU power. The tokens will be tradable, storable, or reinvestable into hardware to further strengthen the network. #Janction’s strategy is clear, first, establish strong liquidity on recognized exchanges, then open access to a broad investor base, especially in #Japan, South Korea, and the United States.

NeoXtrix

44,354 views • 1 year ago

If an AI can control 1,000 robots to perform 1 million skills in 1 billion different simulations, then it may "just work" in our real world, which is simply another point in the vast space of possible realities. This is the fundamental principle behind why simulation works so effectively for robotics. Real-world teleoperation data scales linearly with human time (< 24 hrs/robot/day). Sim data scales exponentially with compute. There are 3 big trends for simulators in the near future: 1. Massive parallelization on large clusters. Physics equations are "just" matrix math at their core. I hear GPUs are good at matrix math 🔥. One can run 100K copies of simulation on a single GPU. To put this number in perspective: 1 hour of wallclock compute time gives a robot 10 years (!!) of training experience. That's how Neo was able to learn martial arts in a blink of an eye in the Matrix Dojo. 2. Generative graphics pipeline. Traditionally, simulators require a huge amount of manual effort from artists: 3D assets, textures, scene layouts, etc. But every component in the workflow can be automated: text-to-image, text-to-3D mesh, and LLMs that write Universal Scene Description (USD) files as a coding exercise. RoboCasa is one example of a prior work. 3. End2end neural net that acts as simulator itself. This is still bluesky research and quite far from replacing a graphics pipeline, but we are seeing some exciting signs-of-life based on video gen models: Sora, Veo2, CogVideoX, Hunyuan (text-to-video); and action-driven world models: GameNGen, Oasis, Genie-2, etc. Genesis does great on (1) for certain tasks, shows good promises on (2), and could become a data generation tool for reaching (3). Its sim2real capabilities for locomotion are good, but there's still a long way to go for contact-rich, dexterous manipulation. It shows a bold vision and is on the right path to providing a virtual cradle for embodied AI. It is open-source and puts a streamlined user journey at the front and center. I had the privilege to know Zhou Xian and play a small part in his project since a year ago. Xian has been crunching code non-stop on Genesis with a very small group of core devs. He often replied to my messages at 3 am. Zhenjia Xu from our GEAR team helped with sim2real experiments in his spare time. Genesis is truly a grassroot effort with an intense focus on quality engineering. Nothing gives me more joy than seeing the simulation ecosystem bloom. Robotics should be a moonshot initiative owned by all of humanity. Congratulations.
3:26

If an AI can control 1,000 robots to perform 1 million skills in 1 billion different simulations, then it may "just work" in our real world, which is simply another point in the vast space of possible realities. This is the fundamental principle behind why simulation works so effectively for robotics. Real-world teleoperation data scales linearly with human time (< 24 hrs/robot/day). Sim data scales exponentially with compute. There are 3 big trends for simulators in the near future: 1. Massive parallelization on large clusters. Physics equations are "just" matrix math at their core. I hear GPUs are good at matrix math 🔥. One can run 100K copies of simulation on a single GPU. To put this number in perspective: 1 hour of wallclock compute time gives a robot 10 years (!!) of training experience. That's how Neo was able to learn martial arts in a blink of an eye in the Matrix Dojo. 2. Generative graphics pipeline. Traditionally, simulators require a huge amount of manual effort from artists: 3D assets, textures, scene layouts, etc. But every component in the workflow can be automated: text-to-image, text-to-3D mesh, and LLMs that write Universal Scene Description (USD) files as a coding exercise. RoboCasa is one example of a prior work. 3. End2end neural net that acts as simulator itself. This is still bluesky research and quite far from replacing a graphics pipeline, but we are seeing some exciting signs-of-life based on video gen models: Sora, Veo2, CogVideoX, Hunyuan (text-to-video); and action-driven world models: GameNGen, Oasis, Genie-2, etc. Genesis does great on (1) for certain tasks, shows good promises on (2), and could become a data generation tool for reaching (3). Its sim2real capabilities for locomotion are good, but there's still a long way to go for contact-rich, dexterous manipulation. It shows a bold vision and is on the right path to providing a virtual cradle for embodied AI. It is open-source and puts a streamlined user journey at the front and center. I had the privilege to know Zhou Xian and play a small part in his project since a year ago. Xian has been crunching code non-stop on Genesis with a very small group of core devs. He often replied to my messages at 3 am. Zhenjia Xu from our GEAR team helped with sim2real experiments in his spare time. Genesis is truly a grassroot effort with an intense focus on quality engineering. Nothing gives me more joy than seeing the simulation ecosystem bloom. Robotics should be a moonshot initiative owned by all of humanity. Congratulations.

Jim Fan

157,343 views • 1 year ago

After 8 months of building in stealth and testing our infrastructure on 10000+ hours of real-world data and hundreds of unique environments, we're bringing FPV Labs into the open today. FPV Labs started with the following bet - if human data proves to be the underlying factor that determines scaling laws in general-purpose robotics, it will trigger the largest economic transformation in human history, and the underlying infrastructure that captures that data will determine how fast we get there. We will achieve this by building the full-stack infrastructure for capturing, processing, transferring, and evaluating human experience into spatial, temporal, and semantic knowledge for machines. Despite all the research novelty behind ChatGPT, its success can be attributed to one foundational fact - the scaling law of transformers. We believe the same dynamics have made their way into robotics. Recent studies showed task completion rates jumping from 30% to 70% when human demonstration data scaled from 1,000 to 20,000 hours, a log-linear trend that mirrors exactly what we saw in language and vision. Seeing these emergent signs of scaling law curves in robotics, we believe we are entering the era of general-purpose robotics policies, which makes the next few years the most exciting time in the history of this field. But the library of physical interactions required to train general-purpose robot policies does not exist yet. Over the last 8 months, we've seen dozens of companies emerge in this space. We were really happy to see new companies pushing this space forward, but we also saw the same pattern repeat: every egocentric data company was making some tradeoffs between quality, scale, and diversity. We have built FPV labs on the core principle that high-quality data is orders of magnitude more valuable than sheer volume. Case in point, self-driving cars collect thousands of hours of data per day, but only a small fraction of that data is actually useful for training better models. Several studies, like RT-2, have shown that as little as 1% of data improves as much as 25% on task success. The quality and diversity of data matter a lot more than scale, so there is clearly a power law curve in the downstream impact of data. We've spent months obsessing over data quality by building our stack, discarding it, rebuilding it, and iterating until we found a formula that doesn't compromise downstream quality at scale. We believe the downstream impact here is far more profound than most people realize. Workers globally are paid around $60 trillion per year in aggregate, and a lion's share of that compensation goes to physical labor - tasks that require navigating real spaces, manipulating real objects, and negotiating the infinite variability of the physical world. Human-to-robot transfer will be one of the most important infrastructures that will shape our society in the near future, and if it works, the economic impact will dwarf every technology transition that came before it in an exponential manner and lead to the creation of goods and services we can’t imagine today. Our mission is to lay the groundwork for us to transition into this future - the future of abundance. We are deeply grateful to our earliest believers, Paras Chopra and Lossfunk, who played a critical role in shaping our thinking.
0:36

After 8 months of building in stealth and testing our infrastructure on 10000+ hours of real-world data and hundreds of unique environments, we're bringing FPV Labs into the open today. FPV Labs started with the following bet - if human data proves to be the underlying factor that determines scaling laws in general-purpose robotics, it will trigger the largest economic transformation in human history, and the underlying infrastructure that captures that data will determine how fast we get there. We will achieve this by building the full-stack infrastructure for capturing, processing, transferring, and evaluating human experience into spatial, temporal, and semantic knowledge for machines. Despite all the research novelty behind ChatGPT, its success can be attributed to one foundational fact - the scaling law of transformers. We believe the same dynamics have made their way into robotics. Recent studies showed task completion rates jumping from 30% to 70% when human demonstration data scaled from 1,000 to 20,000 hours, a log-linear trend that mirrors exactly what we saw in language and vision. Seeing these emergent signs of scaling law curves in robotics, we believe we are entering the era of general-purpose robotics policies, which makes the next few years the most exciting time in the history of this field. But the library of physical interactions required to train general-purpose robot policies does not exist yet. Over the last 8 months, we've seen dozens of companies emerge in this space. We were really happy to see new companies pushing this space forward, but we also saw the same pattern repeat: every egocentric data company was making some tradeoffs between quality, scale, and diversity. We have built FPV labs on the core principle that high-quality data is orders of magnitude more valuable than sheer volume. Case in point, self-driving cars collect thousands of hours of data per day, but only a small fraction of that data is actually useful for training better models. Several studies, like RT-2, have shown that as little as 1% of data improves as much as 25% on task success. The quality and diversity of data matter a lot more than scale, so there is clearly a power law curve in the downstream impact of data. We've spent months obsessing over data quality by building our stack, discarding it, rebuilding it, and iterating until we found a formula that doesn't compromise downstream quality at scale. We believe the downstream impact here is far more profound than most people realize. Workers globally are paid around $60 trillion per year in aggregate, and a lion's share of that compensation goes to physical labor - tasks that require navigating real spaces, manipulating real objects, and negotiating the infinite variability of the physical world. Human-to-robot transfer will be one of the most important infrastructures that will shape our society in the near future, and if it works, the economic impact will dwarf every technology transition that came before it in an exponential manner and lead to the creation of goods and services we can’t imagine today. Our mission is to lay the groundwork for us to transition into this future - the future of abundance. We are deeply grateful to our earliest believers, Paras Chopra and Lossfunk, who played a critical role in shaping our thinking.

Abhishek Anand

81,573 views • 3 months ago

Chamath Palihapitiya just dropped the number that explains the entire AI infrastructure trade (Save this). A gigawatt of compute now costs $100 billion and when he started his Arizona data center project it was $4 to $5 billion, it has gone up 20x in a single investment cycle. The implication is not just that AI infrastructure is expensive but rather that the capital barrier to owning meaningful compute has become so high that only a handful of entities in the world can actually build it and the companies who got there early are sitting on what may be the most durable pricing power in the history of the technology industry. This is the neocloud trade. The neocloud market, purpose-built GPU cloud providers like CoreWeave, Nebius, and Lambda Labs was worth $35 billion in 2026 and is projected to reach $236 billion by 2031, compounding at 46% annually. For context, that is faster growth than cloud computing itself posted in its first decade. The reason is very simple, hyperscalers like AWS, Azure, and Google are building for everything, storage, databases, enterprise software, networking and their GPU pricing reflects the overhead of that full-stack infrastructure. Neoclouds build for one thing only, AI compute. The result is a 60% to 85% cost advantage on the same Nvidia silicon, bare metal H100s at $0.78 to $2.79 per GPU-hour on a neocloud versus $3.43 to $5.07 per GPU-hour on a hyperscaler. That spread does not close as AI demand scales but rather it widens, because hyperscalers have to amortize legacy infrastructure and margin expectations that neoclouds do not carry. Gartner projects that by 2030, neoclouds will capture 20% of the $267 billion AI cloud market, and Vultr's own analysis says at least 80% of GPU market share by end of 2026 will be held by a small group of scaled neocloud providers. Now zoom into Nebius specifically, because it is the most interesting publicly traded proxy for this trade. Nebius is the infrastructure arm of the former Yandex Russia's equivalent of Google rebuilt from the ground up after Russia's invasion of Ukraine by Arkady Volozh and relisted on Nasdaq in October 2024. The team that built it already knew how to run internet-scale infrastructure at the lowest possible cost, which is exactly the operational DNA a neocloud requires. In Q1 2026, Nebius reported revenue of $399 million and already generating serious cash on a young business with revenue growing nearly eightfold year-over-year. Then in March 2026, Meta signed a five-year infrastructure agreement with Nebius worth up to $27 billion, $12 billion in committed dedicated GPU capacity deployments beginning early 2027, plus up to $15 billion more tied to Meta purchasing Nebius's unsold third-party capacity. The deal will be executed on one of the first large-scale deployments of Nvidia's Vera Rubin platform, the next-generation architecture after Blackwell making Nebius one of a tiny number of operators in the world with confirmed priority access to the most advanced AI hardware available. Following the contract, Nebius guided to $7 to $9 billion in annualized recurring revenue for 2026 representing 540% year-over-year growth. Chamath Palihapitiya point about the $100 billion capital moat is the bear case for new entrants and the bull case for incumbents. No one can afford to build the next CoreWeave or Nebius from scratch at current hardware and power costs. The companies that are already built, already contracted, and already deploying Nvidia's latest silicon have a moat that compounds with every GPU generation cycle because they get allocations first, they deploy fastest, and their customers re-sign rather than wait for a new operator that does not yet exist. Come join Milk Road Pro for our full breakdown, the complete neocloud competitive landscape, how to think about Nebius's valuation versus CoreWeave and AI entire thesis. Link below.
0:20

Chamath Palihapitiya just dropped the number that explains the entire AI infrastructure trade (Save this). A gigawatt of compute now costs $100 billion and when he started his Arizona data center project it was $4 to $5 billion, it has gone up 20x in a single investment cycle. The implication is not just that AI infrastructure is expensive but rather that the capital barrier to owning meaningful compute has become so high that only a handful of entities in the world can actually build it and the companies who got there early are sitting on what may be the most durable pricing power in the history of the technology industry. This is the neocloud trade. The neocloud market, purpose-built GPU cloud providers like CoreWeave, Nebius, and Lambda Labs was worth $35 billion in 2026 and is projected to reach $236 billion by 2031, compounding at 46% annually. For context, that is faster growth than cloud computing itself posted in its first decade. The reason is very simple, hyperscalers like AWS, Azure, and Google are building for everything, storage, databases, enterprise software, networking and their GPU pricing reflects the overhead of that full-stack infrastructure. Neoclouds build for one thing only, AI compute. The result is a 60% to 85% cost advantage on the same Nvidia silicon, bare metal H100s at $0.78 to $2.79 per GPU-hour on a neocloud versus $3.43 to $5.07 per GPU-hour on a hyperscaler. That spread does not close as AI demand scales but rather it widens, because hyperscalers have to amortize legacy infrastructure and margin expectations that neoclouds do not carry. Gartner projects that by 2030, neoclouds will capture 20% of the $267 billion AI cloud market, and Vultr's own analysis says at least 80% of GPU market share by end of 2026 will be held by a small group of scaled neocloud providers. Now zoom into Nebius specifically, because it is the most interesting publicly traded proxy for this trade. Nebius is the infrastructure arm of the former Yandex Russia's equivalent of Google rebuilt from the ground up after Russia's invasion of Ukraine by Arkady Volozh and relisted on Nasdaq in October 2024. The team that built it already knew how to run internet-scale infrastructure at the lowest possible cost, which is exactly the operational DNA a neocloud requires. In Q1 2026, Nebius reported revenue of $399 million and already generating serious cash on a young business with revenue growing nearly eightfold year-over-year. Then in March 2026, Meta signed a five-year infrastructure agreement with Nebius worth up to $27 billion, $12 billion in committed dedicated GPU capacity deployments beginning early 2027, plus up to $15 billion more tied to Meta purchasing Nebius's unsold third-party capacity. The deal will be executed on one of the first large-scale deployments of Nvidia's Vera Rubin platform, the next-generation architecture after Blackwell making Nebius one of a tiny number of operators in the world with confirmed priority access to the most advanced AI hardware available. Following the contract, Nebius guided to $7 to $9 billion in annualized recurring revenue for 2026 representing 540% year-over-year growth. Chamath Palihapitiya point about the $100 billion capital moat is the bear case for new entrants and the bull case for incumbents. No one can afford to build the next CoreWeave or Nebius from scratch at current hardware and power costs. The companies that are already built, already contracted, and already deploying Nvidia's latest silicon have a moat that compounds with every GPU generation cycle because they get allocations first, they deploy fastest, and their customers re-sign rather than wait for a new operator that does not yet exist. Come join Milk Road Pro for our full breakdown, the complete neocloud competitive landscape, how to think about Nebius's valuation versus CoreWeave and AI entire thesis. Link below.

Milk Road AI

138,663 views • 1 month ago

Understanding the BitTorrent Swarm — A Broader Look With Real Data Dynamics BitTorrent isn’t just a file-sharing protocol; it’s one of the most efficient large-scale distribution systems ever designed. At its core lies a simple but powerful principle: when users contribute bandwidth, the entire network accelerates. This is the swarm and its efficiency can be explained through clear data patterns and network behavior. 🔹 The Swarm Model: How Participation Becomes Performance In a traditional client-server setup, bandwidth is fixed. If 10,000 users try to download a 1 GB file from one server with 1 Gbps bandwidth: ➠ Maximum theoretical throughput per user: 0.1 Mbps ➠ Average download time: 2–3 hours ➠ Server overload: very likely BitTorrent rewrites this logic. When 10,000 users join a swarm and each contributes only 50–200 Kbps of upload bandwidth, the network’s total available throughput multiplies thousands of times. This is why, in real swarm studies: ➠ Larger swarms consistently show 30–400% faster download speeds ➠ Popular torrents reach equilibrium within minutes, not hours ➠ Throughput per user remains stable even under heavy demand BitTorrent’s efficiency grows with usage — something centralized systems struggle with. 🔹 Why More Peers = More Speed (Backed by Data Behavior) BitTorrent breaks files into hundreds or thousands of small pieces. Each piece circulates among peers using a strategy called rarest-first ensuring no piece becomes a bottleneck. Here’s what the data shows: 1. Bandwidth multiplication effect If each peer contributes: ➠ 100 peers × 100 Kbps upload = 10 Mbps swarm capacity ➠ 5,000 peers × 150 Kbps upload = 750 Mbps swarm capacity ➠ 20,000 peers × 200 Kbps upload = 4 Gbps swarm capacity This turning point when collective bandwidth surpasses any server is why torrents of large files often download faster than centralized sources. 2. Availability resilience Even if 90% of peers leave, as long as one full copy exists across the swarm’s collective pieces, the file is recoverable without interruption. 3. Load balancing automatically occurs BitTorrent’s choking/unchoking algorithm ensures: ➠ High-bandwidth peers exchange more data ➠ Low-bandwidth peers still participate ➠ No single peer becomes a bottleneck The data flow adapts in real time based on peer performance. 🔹 The Swarm’s Global Impact: Why It Still Matters BitTorrent traffic routinely accounts for: ➠ 10–20% of global internet upload traffic (varies by region) ➠ Multiple petabytes of data exchanged daily ➠ Millions of active swarms at any given time The model works because it scales with demand: ➠ More users → more bandwidth. ➠ More bandwidth → faster delivery. ➠ Faster delivery → stronger swarm health. This “self-reinforcing cycle” is a core reason decentralized systems from Web3 storage to blockchain data sync borrow heavily from BitTorrent’s architecture. 🔹 The Big Picture The BitTorrent swarm illustrates an important truth about decentralized networks: Efficiency doesn’t come from the center it comes from participation. When thousands of people contribute small amounts of bandwidth, the result is a global system capable of speeds that outperform traditional content delivery models. This is not just technology; it’s cooperative acceleration at internet scale. In One Line Files move faster when everyone contributes and BitTorrent proves it with real data. H.E. Justin Sun 👨‍🚀 🌞 BitTorrent #TRONEcoStar #BitTorrent #SwarmNetwork #DataAnalysis #DecentralizedSystems #P2P
1:05

Understanding the BitTorrent Swarm — A Broader Look With Real Data Dynamics BitTorrent isn’t just a file-sharing protocol; it’s one of the most efficient large-scale distribution systems ever designed. At its core lies a simple but powerful principle: when users contribute bandwidth, the entire network accelerates. This is the swarm and its efficiency can be explained through clear data patterns and network behavior. 🔹 The Swarm Model: How Participation Becomes Performance In a traditional client-server setup, bandwidth is fixed. If 10,000 users try to download a 1 GB file from one server with 1 Gbps bandwidth: ➠ Maximum theoretical throughput per user: 0.1 Mbps ➠ Average download time: 2–3 hours ➠ Server overload: very likely BitTorrent rewrites this logic. When 10,000 users join a swarm and each contributes only 50–200 Kbps of upload bandwidth, the network’s total available throughput multiplies thousands of times. This is why, in real swarm studies: ➠ Larger swarms consistently show 30–400% faster download speeds ➠ Popular torrents reach equilibrium within minutes, not hours ➠ Throughput per user remains stable even under heavy demand BitTorrent’s efficiency grows with usage — something centralized systems struggle with. 🔹 Why More Peers = More Speed (Backed by Data Behavior) BitTorrent breaks files into hundreds or thousands of small pieces. Each piece circulates among peers using a strategy called rarest-first ensuring no piece becomes a bottleneck. Here’s what the data shows: 1. Bandwidth multiplication effect If each peer contributes: ➠ 100 peers × 100 Kbps upload = 10 Mbps swarm capacity ➠ 5,000 peers × 150 Kbps upload = 750 Mbps swarm capacity ➠ 20,000 peers × 200 Kbps upload = 4 Gbps swarm capacity This turning point when collective bandwidth surpasses any server is why torrents of large files often download faster than centralized sources. 2. Availability resilience Even if 90% of peers leave, as long as one full copy exists across the swarm’s collective pieces, the file is recoverable without interruption. 3. Load balancing automatically occurs BitTorrent’s choking/unchoking algorithm ensures: ➠ High-bandwidth peers exchange more data ➠ Low-bandwidth peers still participate ➠ No single peer becomes a bottleneck The data flow adapts in real time based on peer performance. 🔹 The Swarm’s Global Impact: Why It Still Matters BitTorrent traffic routinely accounts for: ➠ 10–20% of global internet upload traffic (varies by region) ➠ Multiple petabytes of data exchanged daily ➠ Millions of active swarms at any given time The model works because it scales with demand: ➠ More users → more bandwidth. ➠ More bandwidth → faster delivery. ➠ Faster delivery → stronger swarm health. This “self-reinforcing cycle” is a core reason decentralized systems from Web3 storage to blockchain data sync borrow heavily from BitTorrent’s architecture. 🔹 The Big Picture The BitTorrent swarm illustrates an important truth about decentralized networks: Efficiency doesn’t come from the center it comes from participation. When thousands of people contribute small amounts of bandwidth, the result is a global system capable of speeds that outperform traditional content delivery models. This is not just technology; it’s cooperative acceleration at internet scale. In One Line Files move faster when everyone contributes and BitTorrent proves it with real data. H.E. Justin Sun 👨‍🚀 🌞 BitTorrent #TRONEcoStar #BitTorrent #SwarmNetwork #DataAnalysis #DecentralizedSystems #P2P

catalina ossa

50,297 views • 8 months ago

I pay Claude $20 a month. Most $TAO holders do too. There is a stack you can build in 15 minutes that fixes that completely. It runs on Bittensor. It costs $10. You do not write a single line of code. Here is how every AI chat product actually works under the hood. Three layers. Always three. The model. The brain. GPT, Claude, DeepSeek, Kimi, GLM. The inference layer. The GPU that runs the model when you hit send. The interface. The chat box you actually look at. ChatGPT and Claude bundle all three and hand you the result. You cannot change the model. You cannot change the inference. The interface is non-negotiable. Every prompt you type goes to a server run by a private company whose terms of service can quietly change next month. The anti-ChatGPT move is to pick each layer yourself. This is where $TAO comes in. Chutes is Subnet 64 on Bittensor. It is the inference layer. Open source models like DeepSeek, Kimi, GLM, and Llama get served by a global network of miner-operated GPUs. Validators score the output quality. The best inference wins the emissions. You hit send. A miner somewhere runs your prompt. You get the answer back. The TAO you hold is in part paying for the GPU you just used. The basic stack is one URL. chutes. ai/chat No account. No API key. No setup. Switch models mid-conversation. Web search built in. Image generation. File uploads. Free. The advanced stack is Chutes plus TypingMind. One-time license. No recurring fee. Plugins, agents, custom personas, a prompt library you build over months. Full model switching between Chutes, OpenAI, and Anthropic from the same window. Total cost: $10 a month to Chutes for inference. That $10 buys you $50 in actual usage. But here is the signal most people missed inside this story. Chutes ran a free tier until February. Then they killed it. Then they raised the minimum to $10 in May. Most people saw that as bad news. It is the opposite. Free things on the internet do not last. Real products do. Chutes is becoming a real product. A subnet that generates actual revenue from actual users paying actual money for actual AI inference. That is what $43 million in Q1 network revenue looks like at the individual subnet level. And there is one more thing ChatGPT and Claude cannot offer that Chutes already has. Trusted Execution Environments. Your prompt gets encrypted on your device, shipped to a confidential compute GPU, and the lock only breaks inside the chip. The miner running the model physically cannot read your prompt. ChatGPT cannot promise that. Claude cannot promise that. Bittensor already built it. You are holding a network where the subnets are generating real revenue, shipping real privacy infrastructure, and replacing $20 a month centralised subscriptions with $10 a month decentralised inference. The people who use the product always understand the investment better than the people who only watch the price.
8:27

I pay Claude $20 a month. Most $TAO holders do too. There is a stack you can build in 15 minutes that fixes that completely. It runs on Bittensor. It costs $10. You do not write a single line of code. Here is how every AI chat product actually works under the hood. Three layers. Always three. The model. The brain. GPT, Claude, DeepSeek, Kimi, GLM. The inference layer. The GPU that runs the model when you hit send. The interface. The chat box you actually look at. ChatGPT and Claude bundle all three and hand you the result. You cannot change the model. You cannot change the inference. The interface is non-negotiable. Every prompt you type goes to a server run by a private company whose terms of service can quietly change next month. The anti-ChatGPT move is to pick each layer yourself. This is where $TAO comes in. Chutes is Subnet 64 on Bittensor. It is the inference layer. Open source models like DeepSeek, Kimi, GLM, and Llama get served by a global network of miner-operated GPUs. Validators score the output quality. The best inference wins the emissions. You hit send. A miner somewhere runs your prompt. You get the answer back. The TAO you hold is in part paying for the GPU you just used. The basic stack is one URL. chutes. ai/chat No account. No API key. No setup. Switch models mid-conversation. Web search built in. Image generation. File uploads. Free. The advanced stack is Chutes plus TypingMind. One-time license. No recurring fee. Plugins, agents, custom personas, a prompt library you build over months. Full model switching between Chutes, OpenAI, and Anthropic from the same window. Total cost: $10 a month to Chutes for inference. That $10 buys you $50 in actual usage. But here is the signal most people missed inside this story. Chutes ran a free tier until February. Then they killed it. Then they raised the minimum to $10 in May. Most people saw that as bad news. It is the opposite. Free things on the internet do not last. Real products do. Chutes is becoming a real product. A subnet that generates actual revenue from actual users paying actual money for actual AI inference. That is what $43 million in Q1 network revenue looks like at the individual subnet level. And there is one more thing ChatGPT and Claude cannot offer that Chutes already has. Trusted Execution Environments. Your prompt gets encrypted on your device, shipped to a confidential compute GPU, and the lock only breaks inside the chip. The miner running the model physically cannot read your prompt. ChatGPT cannot promise that. Claude cannot promise that. Bittensor already built it. You are holding a network where the subnets are generating real revenue, shipping real privacy infrastructure, and replacing $20 a month centralised subscriptions with $10 a month decentralised inference. The people who use the product always understand the investment better than the people who only watch the price.

2xnmore

27,019 views • 2 months ago

This is probably the most entertaining way to understand one of AI’s hardest AI debates. Transformer vs Post-Transformer, argued by leading researchers, inside a real physical boxing ring. Both technically deep and genuinely entertaining. I was glued for the entire 1 hour 20 minutes. So many super cool points to learn. 🥊 Transformers - Transformers still own the present because they work at scale. They are simple, trainable, hardware-friendly, and already power the strongest AI systems we use today. - The Transformer is basically a memory machine. It stores information as keys and values, then uses attention to pull back the most useful parts when answering. - The real Transformer advantage is not just “attention.” The bigger advantage is that it fits modern hardware extremely well, so it can process huge batches of tokens fast. - Scaling is still the brutal rule. If you give Transformers more compute, more data, and more parameters, they usually keep getting better. Any Post-Transformer architecture has to scale just as well, or better. - It is not enough to look clever on small tests, because the real question is whether it improves faster than Transformers when scaled up. - A replacement cannot be slightly better. Because the whole AI stack is already built around Transformers, the next architecture may need to be around 10x better to force everyone to switch. - Transformers are powerful, but they may be brute force. A human does not need to read the entire internet many times to become smart, but current LLMs need enormous data and compute. 🥊 Post-Transformer - Post-Transformer people are not saying Transformers are bad. They are saying Transformers may be the best current tool, not the final form of machine intelligence. - The biggest Post-Transformer target is native reasoning and continual learning. Today’s LLM reasoning often feels like text-based step-by-step work added on top, instead of thinking happening naturally inside the model. - Latent reasoning is one possible next step. That means the model reasons inside its own hidden internal space, instead of writing every thought out as words. - Continual learning is still a major weakness. Humans keep learning from experience, but most Transformer-based models are trained, frozen, and then only adapt inside the prompt. - Long context is not the same as real memory. A model can read a huge prompt, but that is different from building a life history, learning from mistakes, and updating beliefs over time. - The future may be hybrid, not a clean replacement. Transformers may stay as 1 building block while newer systems add better memory, better reasoning, and better learning loops. - The most interesting possibility is that Transformers may help discover their own successor. AI agents are already getting better at research and coding, so the next architecture may come from AI-assisted architecture search. ------- - Benchmarks are a problem. Many public benchmarks are easy to game, so they may show leaderboard strength without proving deeper intelligence. - Perplexity is still probably a great metric to evaluate frontier models,, because it tests prediction quality. --- Overall, Transformers continue to dominate, but the frontier is clearly widening. Pathway’s BDH (Dragon Hatchling — brain-inspired reasoning architecture), Sakana AI’s CTMs (Continuous Thought Machines — models that think over time), and Liquid AI’s LFMs (Liquid Foundation Models — efficient multimodal foundation models) - all of these show how the frontier is expanding. --- From “Pathway (pathway[.]com)” Youtube channel (link in comment) Zuzanna Stamirowska
2:35

This is probably the most entertaining way to understand one of AI’s hardest AI debates. Transformer vs Post-Transformer, argued by leading researchers, inside a real physical boxing ring. Both technically deep and genuinely entertaining. I was glued for the entire 1 hour 20 minutes. So many super cool points to learn. 🥊 Transformers - Transformers still own the present because they work at scale. They are simple, trainable, hardware-friendly, and already power the strongest AI systems we use today. - The Transformer is basically a memory machine. It stores information as keys and values, then uses attention to pull back the most useful parts when answering. - The real Transformer advantage is not just “attention.” The bigger advantage is that it fits modern hardware extremely well, so it can process huge batches of tokens fast. - Scaling is still the brutal rule. If you give Transformers more compute, more data, and more parameters, they usually keep getting better. Any Post-Transformer architecture has to scale just as well, or better. - It is not enough to look clever on small tests, because the real question is whether it improves faster than Transformers when scaled up. - A replacement cannot be slightly better. Because the whole AI stack is already built around Transformers, the next architecture may need to be around 10x better to force everyone to switch. - Transformers are powerful, but they may be brute force. A human does not need to read the entire internet many times to become smart, but current LLMs need enormous data and compute. 🥊 Post-Transformer - Post-Transformer people are not saying Transformers are bad. They are saying Transformers may be the best current tool, not the final form of machine intelligence. - The biggest Post-Transformer target is native reasoning and continual learning. Today’s LLM reasoning often feels like text-based step-by-step work added on top, instead of thinking happening naturally inside the model. - Latent reasoning is one possible next step. That means the model reasons inside its own hidden internal space, instead of writing every thought out as words. - Continual learning is still a major weakness. Humans keep learning from experience, but most Transformer-based models are trained, frozen, and then only adapt inside the prompt. - Long context is not the same as real memory. A model can read a huge prompt, but that is different from building a life history, learning from mistakes, and updating beliefs over time. - The future may be hybrid, not a clean replacement. Transformers may stay as 1 building block while newer systems add better memory, better reasoning, and better learning loops. - The most interesting possibility is that Transformers may help discover their own successor. AI agents are already getting better at research and coding, so the next architecture may come from AI-assisted architecture search. ------- - Benchmarks are a problem. Many public benchmarks are easy to game, so they may show leaderboard strength without proving deeper intelligence. - Perplexity is still probably a great metric to evaluate frontier models,, because it tests prediction quality. --- Overall, Transformers continue to dominate, but the frontier is clearly widening. Pathway’s BDH (Dragon Hatchling — brain-inspired reasoning architecture), Sakana AI’s CTMs (Continuous Thought Machines — models that think over time), and Liquid AI’s LFMs (Liquid Foundation Models — efficient multimodal foundation models) - all of these show how the frontier is expanding. --- From “Pathway (pathway[.]com)” Youtube channel (link in comment) Zuzanna Stamirowska

Rohan Paul

89,110 views • 1 month ago