Loading video...

Video Failed to Load

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

we made a speech model 13× smaller by making it repeat itself. 52 layers → 2 shared blocks. 2.04B parameters → 153M. The encoder loops 24 times. The decoder loops 28. Smaller weights, roughly the same compute. Oruk

Nathan Roll

1,603 subscribers

53,634 views • 6 days ago •via X (Twitter)

Education Health & Wellness Science & Technology

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

4/ to achieve maximum memory efficiency, we quantize model weights to ~4-bit, getting the language model under 20GB with room for the kv cache, perception encoder, and drafter alongside it. a dflash drafter proposes blocks of tokens the main model verifies in parallel, so it stays responsive.

4/ to achieve maximum memory efficiency, we quantize model weights to ~4-bit, getting the language model under 20GB with room for the kv cache, perception encoder, and drafter alongside it. a dflash drafter proposes blocks of tokens the main model verifies in parallel, so it stays responsive.

Alexandr Wang

69,532 views • 2 months ago

[VAE] by Hand ✍️ A Variational Auto Encoder (VAE) learns the structure (mean and variance) of hidden features and generates new data from the learned structure. In contrast, GANs only learn to generate new data to fool a discriminator; they may not necessarily know the underlying structure of the data. The International Conference on Learning Representations (ICLR) this year announced its first ever "Test of Time Award" to recognizes the VAE paper, published 10 years ago. This exercise demonstrates how to calculate a VAE by hand. [1] Given: ↳ Three training examples X1, X2, X3 ↳ Copy training examples to the bottom ↳ The purpose is to train the network to reconstruct the training examples. ↳ Since each target is a training example itself, we use the Greek word "auto" which means "self." This crucial step is what makes an autoencoder "auto." [2] Encoder: Layer 1 + ReLU ↳ Multiply inputs with weights and biases ↳ Apply ReLU, crossing out negative values (-1 -> 0) [3] Encoder: Mean and Variance ↳ Multiply features with two sets of weights and biases ↳ 🟩 The first set predicts the means (𝜇) of latent distributions ↳ 🟪 The second set predicts the standard deviation (𝜎) of latent distributions [4] Reparameterization Trick: Random Offset ↳ Sample epsilon ε from the normal distribution with mean = 0 and variance = 1. ↳ The purpose is to randomly pick a offset away from the mean. ↳ Multiply the standard deviation values with epsilon values. ↳ The purpose is to scale the offset by the standard deviation. [5] Reparameterization Trick: Mean + Offset ↳ Add the sampled offset to predicted mean ↳ The result are new parameters or features 🟨 as inputs to the Decoder. [6] Decoder: Layer 1 + ReLU ↳ Multiply input features with weights and biases ↳ Apply ReLU, crossing out negative values. Here, -4 is crossed out. [7] Decoder: Layer 2 ↳ Multiply features with weights and biases ↳ The output is Decoder's attempt to reconstruct the input data X from reparameterized distributions described by 𝜇 and 𝜎. [8]-[10] KL Divergence Loss [8] Loss Gradient: Mean 𝜇 ↳ We want 𝜇 to approach 0. ↳ A lot of math called SGVB simplifies the calculation of loss gradients to simply 𝜇 [9,10] Loss Gradient: Stdev 𝜎 ↳ We want 𝜎 to approach 1. ↳ A lot of math simplifies the calculation to 𝜎 - (1/ 𝜎) [11] Reconstruction Loss ↳ We want the reconstructed data Y (dark 🟧) to be the same as the input data X. ↳ Some math involving Mean Square Error simplifies the calculation to Y - X.

[VAE] by Hand ✍️ A Variational Auto Encoder (VAE) learns the structure (mean and variance) of hidden features and generates new data from the learned structure. In contrast, GANs only learn to generate new data to fool a discriminator; they may not necessarily know the underlying structure of the data. The International Conference on Learning Representations (ICLR) this year announced its first ever "Test of Time Award" to recognizes the VAE paper, published 10 years ago. This exercise demonstrates how to calculate a VAE by hand. [1] Given: ↳ Three training examples X1, X2, X3 ↳ Copy training examples to the bottom ↳ The purpose is to train the network to reconstruct the training examples. ↳ Since each target is a training example itself, we use the Greek word "auto" which means "self." This crucial step is what makes an autoencoder "auto." [2] Encoder: Layer 1 + ReLU ↳ Multiply inputs with weights and biases ↳ Apply ReLU, crossing out negative values (-1 -> 0) [3] Encoder: Mean and Variance ↳ Multiply features with two sets of weights and biases ↳ 🟩 The first set predicts the means (𝜇) of latent distributions ↳ 🟪 The second set predicts the standard deviation (𝜎) of latent distributions [4] Reparameterization Trick: Random Offset ↳ Sample epsilon ε from the normal distribution with mean = 0 and variance = 1. ↳ The purpose is to randomly pick a offset away from the mean. ↳ Multiply the standard deviation values with epsilon values. ↳ The purpose is to scale the offset by the standard deviation. [5] Reparameterization Trick: Mean + Offset ↳ Add the sampled offset to predicted mean ↳ The result are new parameters or features 🟨 as inputs to the Decoder. [6] Decoder: Layer 1 + ReLU ↳ Multiply input features with weights and biases ↳ Apply ReLU, crossing out negative values. Here, -4 is crossed out. [7] Decoder: Layer 2 ↳ Multiply features with weights and biases ↳ The output is Decoder's attempt to reconstruct the input data X from reparameterized distributions described by 𝜇 and 𝜎. [8]-[10] KL Divergence Loss [8] Loss Gradient: Mean 𝜇 ↳ We want 𝜇 to approach 0. ↳ A lot of math called SGVB simplifies the calculation of loss gradients to simply 𝜇 [9,10] Loss Gradient: Stdev 𝜎 ↳ We want 𝜎 to approach 1. ↳ A lot of math simplifies the calculation to 𝜎 - (1/ 𝜎) [11] Reconstruction Loss ↳ We want the reconstructed data Y (dark 🟧) to be the same as the input data X. ↳ Some math involving Mean Square Error simplifies the calculation to Y - X.

Tom Yeh

48,529 views • 2 years ago

Single vs Multi-hand Attention by hand ✍️ Resize matrices yourself 👉 The most important fact about multi-head attention: it has the same parameter count as single-head attention. The difference is purely structural — same total Wqkv weights, partitioned into smaller q–k–v triples. Look at the two diagrams below. Both Wqkv matrices have the same height — same number of weight rows, same number of parameters. What changes is how that single tall block is sliced. • Left. One head. The full Wqkv produces one big QKV: a tall Q (36 rows), a tall K, a tall V. One scoring computation runs over those full-width tensors. • Right. 3 heads. The same-height Wqkv is sliced into 3 smaller q–k–v triples — each 12 rows tall. 3 scoring computations run in parallel, each a thinner version of the left. The compute trade-off — kind of. Same Wqkv weights. Multi-head runs the attention scoring S = Kᵀ × Q once per head, so the dot-product count multiplies by H. • Single-head: seq × seq = 40² = 1600 dot products • Multi-head: seq × seq × H = 40² × 3 = 4800 dot products (3×) But each multi-head dot product is narrower — its inner dimension is head_dim instead of H × head_dim. So when you count actual scalar multiplications, the totals are equal: • Single-head: seq² × (H × head_dim) = 40² × 36 = 57600 • Multi-head: seq² × H × head_dim = 40² × 3 × 12 = 57600 Same FLOPs. Multi-head buys you H independent attention patterns at no extra weight cost and no extra arithmetic cost — it's the same total compute, sliced into H finer-grained heads.

Single vs Multi-hand Attention by hand ✍️ Resize matrices yourself 👉 The most important fact about multi-head attention: it has the same parameter count as single-head attention. The difference is purely structural — same total Wqkv weights, partitioned into smaller q–k–v triples. Look at the two diagrams below. Both Wqkv matrices have the same height — same number of weight rows, same number of parameters. What changes is how that single tall block is sliced. • Left. One head. The full Wqkv produces one big QKV: a tall Q (36 rows), a tall K, a tall V. One scoring computation runs over those full-width tensors. • Right. 3 heads. The same-height Wqkv is sliced into 3 smaller q–k–v triples — each 12 rows tall. 3 scoring computations run in parallel, each a thinner version of the left. The compute trade-off — kind of. Same Wqkv weights. Multi-head runs the attention scoring S = Kᵀ × Q once per head, so the dot-product count multiplies by H. • Single-head: seq × seq = 40² = 1600 dot products • Multi-head: seq × seq × H = 40² × 3 = 4800 dot products (3×) But each multi-head dot product is narrower — its inner dimension is head_dim instead of H × head_dim. So when you count actual scalar multiplications, the totals are equal: • Single-head: seq² × (H × head_dim) = 40² × 36 = 57600 • Multi-head: seq² × H × head_dim = 40² × 3 × 12 = 57600 Same FLOPs. Multi-head buys you H independent attention patterns at no extra weight cost and no extra arithmetic cost — it's the same total compute, sliced into H finer-grained heads.

Tom Yeh

35,911 views • 5 months ago

NEW WORLD MODEL: Yann LeCun's team is back with an efficient model! This project involves Yann LeCun, Lukas Kuhn, Lucas Maes, Quentin Le Lidec, and Randall Balestriero. A couple definitions first: - DINO: self-DIstillation with NO labels. A self-supervised image model (Meta, 2021) where a student network learns to match a teacher (an EMA copy of itself) across two crops of the same image, with no labels and no negatives. - SIGReg: a regularizer that prevents embedding collapse by forcing the embeddings to match an isotropic Gaussian, tested with a normality test (Epps–Pulley) on many random 1-D projections instead of in full dimension. LeVJEPA is a self-supervised video pretraining method, released with open code, weights, and checkpoints. It learns a video representation by pushing the embeddings of global and local crops of the same clip together (an invariance loss), while a regularizer called SIGReg forces the embeddings toward an isotropic Gaussian to provably prevent representation collapse. Unlike V-JEPA and V-JEPA 2 it uses a single shared encoder with a projector and no target network, no predictor and no stop-gradient. It drops 95% of tokens per view, uses block-causal attention (each frame attends only to past frames), and has a single loss weight. It is evaluated purely as a representation learner via frozen probing on ImageNet-1K, Something-Something-v2 and Kinetics-400, not on any robot. What I find interesting, is that V-JEPA and V-JEPA 2 need an EMA target encoder, stop-gradients and a capacity-limited predictor to avoid collapse; LeVJEPA drops all of it for one shared encoder plus projector, preventing collapse instead with the SIGReg regularizer under a provable guarantee and a single hyperparameter. The "P" (predictor) in JEPA is effectively gone. LeVJEPA is also less compute intensive: - 5.6x to 20.8x lower total pretraining compute than V-JEPA 2 - 7.6 points higher on ImageNet-1K at matched FLOPs - trains at batch size 128 within 8GB where V-JEPA 2 saturates at batch size 28 Also worth mentioning: ImageNet-1K accuracy rises monotonically with the token-drop rate, from 33.9% at rho = 0 to 47.6% at rho = 0.95. The aggressive dropping is actually doing regularization work. On the JEPA-versus-DINO debate: - it loses to DINOv2 by 3.1 points on ImageNet-1K (appearance, static) - but wins on Something-Something-v2 by nearly 2x (motion, temporal) - and beats V-JEPA 2 by 1.9 points on ViT-L at 5.6x lower cost. -> optimized for temporal and motion understanding per compute dollar.

NEW WORLD MODEL: Yann LeCun's team is back with an efficient model! This project involves Yann LeCun, Lukas Kuhn, Lucas Maes, Quentin Le Lidec, and Randall Balestriero. A couple definitions first: - DINO: self-DIstillation with NO labels. A self-supervised image model (Meta, 2021) where a student network learns to match a teacher (an EMA copy of itself) across two crops of the same image, with no labels and no negatives. - SIGReg: a regularizer that prevents embedding collapse by forcing the embeddings to match an isotropic Gaussian, tested with a normality test (Epps–Pulley) on many random 1-D projections instead of in full dimension. LeVJEPA is a self-supervised video pretraining method, released with open code, weights, and checkpoints. It learns a video representation by pushing the embeddings of global and local crops of the same clip together (an invariance loss), while a regularizer called SIGReg forces the embeddings toward an isotropic Gaussian to provably prevent representation collapse. Unlike V-JEPA and V-JEPA 2 it uses a single shared encoder with a projector and no target network, no predictor and no stop-gradient. It drops 95% of tokens per view, uses block-causal attention (each frame attends only to past frames), and has a single loss weight. It is evaluated purely as a representation learner via frozen probing on ImageNet-1K, Something-Something-v2 and Kinetics-400, not on any robot. What I find interesting, is that V-JEPA and V-JEPA 2 need an EMA target encoder, stop-gradients and a capacity-limited predictor to avoid collapse; LeVJEPA drops all of it for one shared encoder plus projector, preventing collapse instead with the SIGReg regularizer under a provable guarantee and a single hyperparameter. The "P" (predictor) in JEPA is effectively gone. LeVJEPA is also less compute intensive: - 5.6x to 20.8x lower total pretraining compute than V-JEPA 2 - 7.6 points higher on ImageNet-1K at matched FLOPs - trains at batch size 128 within 8GB where V-JEPA 2 saturates at batch size 28 Also worth mentioning: ImageNet-1K accuracy rises monotonically with the token-drop rate, from 33.9% at rho = 0 to 47.6% at rho = 0.95. The aggressive dropping is actually doing regularization work. On the JEPA-versus-DINO debate: - it loses to DINOv2 by 3.1 points on ImageNet-1K (appearance, static) - but wins on Something-Something-v2 by nearly 2x (motion, temporal) - and beats V-JEPA 2 by 1.9 points on ViT-L at 5.6x lower cost. -> optimized for temporal and motion understanding per compute dollar.

Léo

90,751 views • 1 month ago

Experience a once-in-a-lifetime perspective of a solar eclipse, seen from the window of an airplane.. A solar eclipse occurs when the Moon moves directly between Earth and the Sun, aligning so precisely that it blocks the Sun’s light. Although the Sun is roughly 400 times larger than the Moon, it is also about 400 times farther away, making them appear nearly the same size in the sky. During totality, the Moon completely obscures the Sun, unveiling its outer atmosphere—the corona—and turning daylight into an eerie twilight. Because this exact alignment is rare for any given location, total solar eclipses are truly unforgettable events.

Experience a once-in-a-lifetime perspective of a solar eclipse, seen from the window of an airplane.. A solar eclipse occurs when the Moon moves directly between Earth and the Sun, aligning so precisely that it blocks the Sun’s light. Although the Sun is roughly 400 times larger than the Moon, it is also about 400 times farther away, making them appear nearly the same size in the sky. During totality, the Moon completely obscures the Sun, unveiling its outer atmosphere—the corona—and turning daylight into an eerie twilight. Because this exact alignment is rare for any given location, total solar eclipses are truly unforgettable events.

Historic Vids

103,130 views • 7 months ago

The AI race is getting a new question: how smart can you afford to run all day? Grok 4.7 keeps the same $2 input / $6 output pricing as 4.6, while getting noticeably stronger at coding, engineering and long-running agent work. That means more retries. Longer loops. More users. More actual work done before the bill gets silly. The smartest answer once is impressive. The model you can afford to ask a billion times is hard to beat. Grok SpaceXAI / Writer: Annette, Grok Imagine Designer: Janné

The AI race is getting a new question: how smart can you afford to run all day? Grok 4.7 keeps the same $2 input / $6 output pricing as 4.6, while getting noticeably stronger at coding, engineering and long-running agent work. That means more retries. Longer loops. More users. More actual work done before the bill gets silly. The smartest answer once is impressive. The model you can afford to ask a billion times is hard to beat. Grok SpaceXAI / Writer: Annette, Grok Imagine Designer: Janné

Mario Nawfal

33,746 views • 23 days ago

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Akshay 🚀

245,379 views • 3 months ago

$180,000 VOLVO SEMI TRUCK. EVERY WIRE AND PIPE IN THE CAB AND ENGINE BAY MAPPED IN ADVANCE. BUILT BY TYPING SENTENCES INSTEAD OF PLACING THEM BY HAND. A cab and engine bay this dense needed its wiring, cooling lines, and air and brake lines modeled and checked for clashes before a physical prototype could be built. The design team connected Kimi K3 to their CAD software through MCP, then had it read photos of the cab and engine bay alongside plain English descriptions of each system and build the model directly instead of placing every line by hand. Old process: 2 engineers, 3 weeks, around $19,000. New process: 1 engineer reviewing and refining the AI's output, 5 days, roughly $3,500 combined, engineer's time plus AI compute. Routing clashes caught this way cost a few hundred dollars to fix. The same clash caught after a physical prototype is already built costs 10 to 100 times more. See the article below for a smaller example of the Kimi K3 + Blender MCP workflow.

$180,000 VOLVO SEMI TRUCK. EVERY WIRE AND PIPE IN THE CAB AND ENGINE BAY MAPPED IN ADVANCE. BUILT BY TYPING SENTENCES INSTEAD OF PLACING THEM BY HAND. A cab and engine bay this dense needed its wiring, cooling lines, and air and brake lines modeled and checked for clashes before a physical prototype could be built. The design team connected Kimi K3 to their CAD software through MCP, then had it read photos of the cab and engine bay alongside plain English descriptions of each system and build the model directly instead of placing every line by hand. Old process: 2 engineers, 3 weeks, around $19,000. New process: 1 engineer reviewing and refining the AI's output, 5 days, roughly $3,500 combined, engineer's time plus AI compute. Routing clashes caught this way cost a few hundred dollars to fix. The same clash caught after a physical prototype is already built costs 10 to 100 times more. See the article below for a smaller example of the Kimi K3 + Blender MCP workflow.

Solvaix

65,388 views • 2 months ago

Qwen 3.7-max beats Opus 4.7 and GPT-5.5 We tested three frontier models on a real agentic task: write a Tetris bot that plays the game and trains itself. Each model could read its own code, run benchmarks, and rewrite itself across 10 iterations. Then we compared the final bots head to head. Qwen 3.7-Max: training cost $1.32, bot improvement +56% Claude Opus 4.7: training cost $12.15, bot improvement +28% GPT-5.5: training cost $2.85, bot improvement +7% Qwen won on every dimension - biggest jump, 9× cheaper than Claude, 2× cheaper than GPT. Long agentic loops is where Qwen Max actually delivers.

Qwen 3.7-max beats Opus 4.7 and GPT-5.5 We tested three frontier models on a real agentic task: write a Tetris bot that plays the game and trains itself. Each model could read its own code, run benchmarks, and rewrite itself across 10 iterations. Then we compared the final bots head to head. Qwen 3.7-Max: training cost $1.32, bot improvement +56% Claude Opus 4.7: training cost $12.15, bot improvement +28% GPT-5.5: training cost $2.85, bot improvement +7% Qwen won on every dimension - biggest jump, 9× cheaper than Claude, 2× cheaper than GPT. Long agentic loops is where Qwen Max actually delivers.

atomic.chat

870,448 views • 4 months ago

The human brain is truly a marvel of nature. If you horribly reductive, and boiled it down to a language model, you'd be looking at roughly 100 trillon parameters running as a sparse MoE architecture Only about 1-5% of neurons fire at any given moment, meaning the brain "activates" maybe 1-5 trillion parameters per inference step. For context, the largest AI models we've built probably top out around 5 trillion parameters. The brain is roughly 100x larger. Even its active params at any given moment are larger than almost every model in existence today. Here's what melts my brain (pun intnended) though Your brain does all of this on about 20 watts of power, less than a dim light bulb. Training a frontier AI model consumes enough electricity to power small cities for months. Running inference across data centers pulls megawatts. Your brain runs 24/7 for 80+ years on the equivalent of a phone charger. We haven't come close to matching the brain's scale. And we're not even in the same universe when it comes to efficiency. Evolution spent 500 million yrs optimizing the most energy-efficient intelligence architecture ever known. we're trying to brute force our way there with compute and electricity. Nature is still the best engineer in the room.

The human brain is truly a marvel of nature. If you horribly reductive, and boiled it down to a language model, you'd be looking at roughly 100 trillon parameters running as a sparse MoE architecture Only about 1-5% of neurons fire at any given moment, meaning the brain "activates" maybe 1-5 trillion parameters per inference step. For context, the largest AI models we've built probably top out around 5 trillion parameters. The brain is roughly 100x larger. Even its active params at any given moment are larger than almost every model in existence today. Here's what melts my brain (pun intnended) though Your brain does all of this on about 20 watts of power, less than a dim light bulb. Training a frontier AI model consumes enough electricity to power small cities for months. Running inference across data centers pulls megawatts. Your brain runs 24/7 for 80+ years on the equivalent of a phone charger. We haven't come close to matching the brain's scale. And we're not even in the same universe when it comes to efficiency. Evolution spent 500 million yrs optimizing the most energy-efficient intelligence architecture ever known. we're trying to brute force our way there with compute and electricity. Nature is still the best engineer in the room.

am.will

130,954 views • 6 months ago

In my class, I teach the autoencoder by asking everyone to stand up. 🙆 Stretch your arms out wide. Imagine you are holding a heavy textbook (like Introduction to Algorithms by Prof. Cormen), the whole thing, every page. Now bring your hands slowly together until they almost touch your neck. The bottle "neck." The final exam is tomorrow and you are allowed one cheat sheet: whatever you can scribble on your palm. All nine hundred pages have to survive the squeeze. That is the encoder. Now imagine you sit in the exam. Push your arms back out to where they started. You try to rebuild the textbook from your palm notes. That is the decoder. Of course you cannot get every page back. What you get back is what mattered enough to write down, and the gap between the two is the loss the network is trying to shrink. I call it AI by Arms 🙆. It gets a laugh, and then it gets remembered. Goal: squeeze four numbers down to two, then rebuild the original four from them. 1. Given Let us start with four training examples: X1, X2, X3, X4. 2. Auto (copy to targets) We copy the training examples straight into the targets. That is the whole trick behind the name: "auto" is Greek for "self", and the data is its own label. 3. Encoder, layer 1 Let us multiply the inputs by the weights, add the biases, and apply ReLU. Negative values get crossed out and become zero. 4. Encoder, layer 2 (the bottleneck) We do it again, and now the four dimensions have become two. This layer is called the bottleneck, because everything has to fit through it. 5. Decoder, layer 1 Let us go back the other way: multiply, add, ReLU. This time there are no negatives to cross out. 6. Decoder, layer 2 We multiply once more and get the outputs Y. This is the decoder's attempt to rebuild the four original numbers from the two it was given. 7. MSE loss gradients Let us compare Y with the targets Y'. The gradient is 2 x (Y - Y'): subtract, then double. Those gradients kick off backpropagation, and the weights start to learn. Your entire education is all about encoding and decoding!

In my class, I teach the autoencoder by asking everyone to stand up. 🙆 Stretch your arms out wide. Imagine you are holding a heavy textbook (like Introduction to Algorithms by Prof. Cormen), the whole thing, every page. Now bring your hands slowly together until they almost touch your neck. The bottle "neck." The final exam is tomorrow and you are allowed one cheat sheet: whatever you can scribble on your palm. All nine hundred pages have to survive the squeeze. That is the encoder. Now imagine you sit in the exam. Push your arms back out to where they started. You try to rebuild the textbook from your palm notes. That is the decoder. Of course you cannot get every page back. What you get back is what mattered enough to write down, and the gap between the two is the loss the network is trying to shrink. I call it AI by Arms 🙆. It gets a laugh, and then it gets remembered. Goal: squeeze four numbers down to two, then rebuild the original four from them. 1. Given Let us start with four training examples: X1, X2, X3, X4. 2. Auto (copy to targets) We copy the training examples straight into the targets. That is the whole trick behind the name: "auto" is Greek for "self", and the data is its own label. 3. Encoder, layer 1 Let us multiply the inputs by the weights, add the biases, and apply ReLU. Negative values get crossed out and become zero. 4. Encoder, layer 2 (the bottleneck) We do it again, and now the four dimensions have become two. This layer is called the bottleneck, because everything has to fit through it. 5. Decoder, layer 1 Let us go back the other way: multiply, add, ReLU. This time there are no negatives to cross out. 6. Decoder, layer 2 We multiply once more and get the outputs Y. This is the decoder's attempt to rebuild the four original numbers from the two it was given. 7. MSE loss gradients Let us compare Y with the targets Y'. The gradient is 2 x (Y - Y'): subtract, then double. Those gradients kick off backpropagation, and the weights start to learn. Your entire education is all about encoding and decoding!

Tom Yeh

22,749 views • 2 months ago

Batch Normalization by hand ✍️ ~ 7 steps walkthrough below Batch normalization is common practice for improving training and achieving faster convergence. It sounds simple. But it is often misunderstood. 🤔 Does batch normalization involve trainable parameters, tunable hyper-parameters, or both? 🤔 Is batch normalization applied to inputs, features, weights, biases, or outputs? 🤔 How is batch normalization different from layer normalization? So I drew and calculated one entirely by hand. Goal: normalize a mini-batch of 4 examples to mean 0 and variance 1, then let the network scale it back. = 1. Given = A mini-batch of 4 training examples, each with 3 features. = 2. Linear layer = Let us multiply by the weights and add the biases. Batch norm sits after this, which answers the second question: what gets normalized is features, not inputs, weights or biases. = 3. ReLU = We apply the activation, and -2 becomes 0. Negative values are suppressed before any statistic is taken. = 4. Batch statistics = Let us compute the sum, mean, variance and standard deviation, one row at a time. A row is a feature and the four columns are the four examples, so every number here measures one feature against the rest of the batch. That is the "batch" in batch normalization, and it is exactly what layer normalization does not do. The statistics are rounded to whole numbers, which is what keeps the rest of the page doable in pen. = 5. Shift to mean 0 = We subtract the mean, in green. The four values in each feature now average to zero. = 6. Scale to variance 1 = Let us divide by the standard deviation, in orange. Each feature now has variance one, whatever scale it arrived at. = 7. Scale and shift = We multiply by a linear transformation and pass the result on. The diagonal and the last column are trainable, so having just forced every feature to mean 0 and variance 1, we hand the network the means to undo it. The outputs: Mean of each feature = [2, 1, 2] Std dev of each feature = [1, 1, 2] To the next layer = [2, -2, 2, 0], [-3, 3, 6, -3], [2, 0, 1, 2] The answers: 🤔 Both. The scale and shift are trainable, the statistics are not. Epsilon and the momentum on the running statistics are the hyper-parameters, and one mini-batch by hand needs neither. 🤔 Features, after the linear layer, not inputs, weights or biases. 🤔 Batch norm measures across the batch, one feature at a time. Layer norm measures across the features, one example at a time. 💾 Save this post!

Batch Normalization by hand ✍️ ~ 7 steps walkthrough below Batch normalization is common practice for improving training and achieving faster convergence. It sounds simple. But it is often misunderstood. 🤔 Does batch normalization involve trainable parameters, tunable hyper-parameters, or both? 🤔 Is batch normalization applied to inputs, features, weights, biases, or outputs? 🤔 How is batch normalization different from layer normalization? So I drew and calculated one entirely by hand. Goal: normalize a mini-batch of 4 examples to mean 0 and variance 1, then let the network scale it back. = 1. Given = A mini-batch of 4 training examples, each with 3 features. = 2. Linear layer = Let us multiply by the weights and add the biases. Batch norm sits after this, which answers the second question: what gets normalized is features, not inputs, weights or biases. = 3. ReLU = We apply the activation, and -2 becomes 0. Negative values are suppressed before any statistic is taken. = 4. Batch statistics = Let us compute the sum, mean, variance and standard deviation, one row at a time. A row is a feature and the four columns are the four examples, so every number here measures one feature against the rest of the batch. That is the "batch" in batch normalization, and it is exactly what layer normalization does not do. The statistics are rounded to whole numbers, which is what keeps the rest of the page doable in pen. = 5. Shift to mean 0 = We subtract the mean, in green. The four values in each feature now average to zero. = 6. Scale to variance 1 = Let us divide by the standard deviation, in orange. Each feature now has variance one, whatever scale it arrived at. = 7. Scale and shift = We multiply by a linear transformation and pass the result on. The diagonal and the last column are trainable, so having just forced every feature to mean 0 and variance 1, we hand the network the means to undo it. The outputs: Mean of each feature = [2, 1, 2] Std dev of each feature = [1, 1, 2] To the next layer = [2, -2, 2, 0], [-3, 3, 6, -3], [2, 0, 1, 2] The answers: 🤔 Both. The scale and shift are trainable, the statistics are not. Epsilon and the momentum on the running statistics are the hyper-parameters, and one mini-batch by hand needs neither. 🤔 Features, after the linear layer, not inputs, weights or biases. 🤔 Batch norm measures across the batch, one feature at a time. Layer norm measures across the features, one example at a time. 💾 Save this post!

Tom Yeh

20,848 views • 2 months ago

🇨🇳🇺🇸 China's military may be learning from America's AI, without building it from scratch Why spend billions training a frontier AI model when you can let someone else do the expensive part? Chinese military-linked researchers have repeatedly used outputs from OpenAI and Anthropic models to train smaller domestic AI systems for surveillance, cyber operations, drone targeting, battlefield decision-making, and software analysis. The trick is called distillation: instead of copying the model itself, researchers use its answers, and more importantly, its reasoning, to teach a smaller model. The result? AI that isn't as powerful as the original, but is cheap, runs on local hardware, and can be deployed on drones, military networks, and tactical systems without relying on U.S technology. The AI arms race isn't just about who builds the smartest model anymore; it's about who can squeeze the most military value out of everyone else's. Source: Reuters / Writer: Ian

🇨🇳🇺🇸 China's military may be learning from America's AI, without building it from scratch Why spend billions training a frontier AI model when you can let someone else do the expensive part? Chinese military-linked researchers have repeatedly used outputs from OpenAI and Anthropic models to train smaller domestic AI systems for surveillance, cyber operations, drone targeting, battlefield decision-making, and software analysis. The trick is called distillation: instead of copying the model itself, researchers use its answers, and more importantly, its reasoning, to teach a smaller model. The result? AI that isn't as powerful as the original, but is cheap, runs on local hardware, and can be deployed on drones, military networks, and tactical systems without relying on U.S technology. The AI arms race isn't just about who builds the smartest model anymore; it's about who can squeeze the most military value out of everyone else's. Source: Reuters / Writer: Ian

Mario Nawfal

47,775 views • 2 months ago

$Imagine you have a point dancing in space, following three simple rules. You’d think it would eventually settle down or repeat itself, right? Nature isn't always that predictable. This is the Rössler Attractor. It’s a beautiful piece of "organized chaos." By just tweaking a few non-linear equations, we get this marvelous fractal shape, a curve that weaves and loops, forever exploring new ground without ever retracing its exact same path. Most people try to solve this with standard linear tools, but that's like trying to understand a symphony by looking at a single note. To see the real "magic" the bifurcations and the patterns, you have to embrace the non-linearity. It shows us how the smallest nudge at the start can lead to a completely different, stunningly complex world. It’s a simple set of rules, yet it generates infinite complexity. Isn't that just the kick in the head? (📹Credit; equationsinmotion / IG)$

Imagine you have a point dancing in space, following three simple rules. You’d think it would eventually settle down or repeat itself, right? Nature isn't always that predictable. This is the Rössler Attractor. It’s a beautiful piece of "organized chaos." By just tweaking a few non-linear equations, we get this marvelous fractal shape, a curve that weaves and loops, forever exploring new ground without ever retracing its exact same path. Most people try to solve this with standard linear tools, but that's like trying to understand a symphony by looking at a single note. To see the real "magic" the bifurcations and the patterns, you have to embrace the non-linearity. It shows us how the smallest nudge at the start can lead to a completely different, stunningly complex world. It’s a simple set of rules, yet it generates infinite complexity. Isn't that just the kick in the head? (📹Credit; equationsinmotion / IG)

Philosophy Of Physics

40,168 views • 5 months ago

Code mode lost by 39 seconds. It is supposed to be the faster way for agents to call MCP tools. Fewer round trips, smaller context. We raced it against classic tool calling on the same Linear task and it lost. The reason isn't a code mode problem. Most MCP servers ship input schemas and nothing else, so the agent knows exactly what to send and has to guess what comes back. It guessed wrong three times, and every wrong guess is a full extra turn. So we built a shape cache that remembers what each tool returns and carries it across sessions. Rematch: 27% faster than classic, 47% cheaper, 42% less context. Mark Hetherington on making [code]smith fast:

Code mode lost by 39 seconds. It is supposed to be the faster way for agents to call MCP tools. Fewer round trips, smaller context. We raced it against classic tool calling on the same Linear task and it lost. The reason isn't a code mode problem. Most MCP servers ship input schemas and nothing else, so the agent knows exactly what to send and has to guess what comes back. It guessed wrong three times, and every wrong guess is a full extra turn. So we built a shape cache that remembers what each tool returns and carries it across sessions. Rematch: 27% faster than classic, 47% cheaper, 42% less context. Mark Hetherington on making [code]smith fast:

Blacksmith

13,748 views • 1 month ago

You can run a 27B AI model on a Mac with just 16GB of memory. PrismML’s Bonsai 2 27B is a ternary model built on Qwen3.8-27B. Its reported model footprint is around 7GB, while retaining 98.2% of the FP16 model’s quality according to its developers. That’s a substantial compression of a model that would otherwise require considerably more memory at full precision. Bonsai 2 uses ternary compression, representing weights with three possible values rather than conventional multi-bit numerical representations. This can dramatically reduce the storage needed for model weights. Actual speed improvements depend on whether the inference engine and hardware can take advantage of the compressed format. The reported 7GB footprint makes Bonsai 2 particularly interesting for Macs with 16GB of unified memory, although runtime overhead, context length, and other applications still affect how comfortably it runs. It can do more than chat In one reported demonstration, Bonsai 2 generated a voxel-style Japanese pagoda using Three.js from a single prompt. That involves translating a natural-language request into code that builds a recognizable 3D structure. It’s a useful example of what a compact local model might do beyond answering questions, particularly for developers experimenting with procedural graphics and interactive projects. The model reportedly retains 98.2% of FP16 quality despite its much smaller footprint. However, that number needs context. The evaluation method, benchmark selection, and task-specific results determine how meaningful the comparison is. Preserving 98.2% on one evaluation doesn’t guarantee identical coding ability across every workload. Still, the combination of a 27B architecture and a roughly 7GB model is worth investigating. The direction is clear: better compression could let developers run capable models on hardware they already own, without automatically upgrading to a more expensive machine. Model: Bonsai 2 27B by PrismML.

You can run a 27B AI model on a Mac with just 16GB of memory. PrismML’s Bonsai 2 27B is a ternary model built on Qwen3.8-27B. Its reported model footprint is around 7GB, while retaining 98.2% of the FP16 model’s quality according to its developers. That’s a substantial compression of a model that would otherwise require considerably more memory at full precision. Bonsai 2 uses ternary compression, representing weights with three possible values rather than conventional multi-bit numerical representations. This can dramatically reduce the storage needed for model weights. Actual speed improvements depend on whether the inference engine and hardware can take advantage of the compressed format. The reported 7GB footprint makes Bonsai 2 particularly interesting for Macs with 16GB of unified memory, although runtime overhead, context length, and other applications still affect how comfortably it runs. It can do more than chat In one reported demonstration, Bonsai 2 generated a voxel-style Japanese pagoda using Three.js from a single prompt. That involves translating a natural-language request into code that builds a recognizable 3D structure. It’s a useful example of what a compact local model might do beyond answering questions, particularly for developers experimenting with procedural graphics and interactive projects. The model reportedly retains 98.2% of FP16 quality despite its much smaller footprint. However, that number needs context. The evaluation method, benchmark selection, and task-specific results determine how meaningful the comparison is. Preserving 98.2% on one evaluation doesn’t guarantee identical coding ability across every workload. Still, the combination of a 27B architecture and a roughly 7GB model is worth investigating. The direction is clear: better compression could let developers run capable models on hardware they already own, without automatically upgrading to a more expensive machine. Model: Bonsai 2 27B by PrismML.

FHILY👑

11,797 views • 6 days ago

$Transformer and Mixture of Experts, explained visually! Mixture of Experts (MoE) is a popular architecture that uses different experts to improve Transformer models. Transformer and MoE differ in the decoder block: - Transformer uses a feed-forward network. - MoE uses experts, which are feed-forward networks but smaller compared to those Transformer. During inference, a subset of experts are selected. This makes inference faster in MoE. Also, since the network has multiple decoder layers: - The text passes through different experts across layers. - The chosen experts also differ between tokens. But how does the model decide which experts should be ideal? The router does that. It is a multi-class classifier that produces softmax scores over experts to select the top K experts. The router is trained with the network, and it learns to select the best experts. But it isn't straightforward. There are challenges! Challenge 1) Notice this pattern at the start of training: - Say, the model selects "Expert 2" - This expert gets a bit better - It may get selected again since it's the "best" - It learns more - It gets selected again in the next iteration - It learns more, and so on! This means many experts can go under-trained due to the overselection of a few experts! We solve this in two steps: - Add noise to the feed-forward output of the router so that other experts can get higher logits. - Set all but the top K logits to -infinity. After softmax, these scores become zero. This way, other experts also get the opportunity to train. Challenge 2) Some experts may get exposed to more tokens than others, leading to under-trained experts. We prevent this by limiting the number of tokens an expert can process. If an expert reaches the limit, the token is passed to the next best expert. Overall, MoEs have more parameters to load. But a fraction of them are activated during inference. This leads to faster inference. Mixtral 8x7B and Llama 4 are two popular MoE-based LLMs. Have you used MoEs in production yet? To dive deeper into how MoE inference works in production, we wrote a full article covering token dispatch, grouped expert computation, model-weight memory, multi-GPU communication, expert placement, load imbalance, and performance diagnosis. Read it below.$

Transformer and Mixture of Experts, explained visually! Mixture of Experts (MoE) is a popular architecture that uses different experts to improve Transformer models. Transformer and MoE differ in the decoder block: - Transformer uses a feed-forward network. - MoE uses experts, which are feed-forward networks but smaller compared to those Transformer. During inference, a subset of experts are selected. This makes inference faster in MoE. Also, since the network has multiple decoder layers: - The text passes through different experts across layers. - The chosen experts also differ between tokens. But how does the model decide which experts should be ideal? The router does that. It is a multi-class classifier that produces softmax scores over experts to select the top K experts. The router is trained with the network, and it learns to select the best experts. But it isn't straightforward. There are challenges! Challenge 1) Notice this pattern at the start of training: - Say, the model selects "Expert 2" - This expert gets a bit better - It may get selected again since it's the "best" - It learns more - It gets selected again in the next iteration - It learns more, and so on! This means many experts can go under-trained due to the overselection of a few experts! We solve this in two steps: - Add noise to the feed-forward output of the router so that other experts can get higher logits. - Set all but the top K logits to -infinity. After softmax, these scores become zero. This way, other experts also get the opportunity to train. Challenge 2) Some experts may get exposed to more tokens than others, leading to under-trained experts. We prevent this by limiting the number of tokens an expert can process. If an expert reaches the limit, the token is passed to the next best expert. Overall, MoEs have more parameters to load. But a fraction of them are activated during inference. This leads to faster inference. Mixtral 8x7B and Llama 4 are two popular MoE-based LLMs. Have you used MoEs in production yet? To dive deeper into how MoE inference works in production, we wrote a full article covering token dispatch, grouped expert computation, model-weight memory, multi-GPU communication, expert placement, load imbalance, and performance diagnosis. Read it below.

Daily Dose of Data Science

55,488 views • 29 days ago

The state of indie development has gotten so insane we got this guy who is solo developing a side game while making his main game in the same universe It looks incredible. Inspired by NieR, Gray Raven, and Metal Gear Solid 2

The state of indie development has gotten so insane we got this guy who is solo developing a side game while making his main game in the same universe It looks incredible. Inspired by NieR, Gray Raven, and Metal Gear Solid 2

Jake Lucky

75,474 views • 1 year ago