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.

As someone who builds with LLMs and diffusion models, this 14-minute heat equation video is the closest thing to a "why noise reversal actually works" explainer I've ever seen released for free. Everyone thinks diffusion is a novel AI concept. It's the heat equation from 1822. The forward process... show more

fintex

2,315 subscribers

15,604 views • 1 month ago •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

As someone who's built with modern LLM architectures, this Laplace transform lecture is the closest thing to a "why Mamba works" explainer I've ever seen released for free. Everyone thinks transformers are the only path to modern LLMs. State space models like Mamba and S4 challenge that using math Laplace built in the 1780s. This video shows why. 25 minutes. Bookmark & watch today.

As someone who's built with modern LLM architectures, this Laplace transform lecture is the closest thing to a "why Mamba works" explainer I've ever seen released for free. Everyone thinks transformers are the only path to modern LLMs. State space models like Mamba and S4 challenge that using math Laplace built in the 1780s. This video shows why. 25 minutes. Bookmark & watch today.

fintex

110,467 views • 2 months ago

As someone who ships LLM systems in production, this 22-minute Laplace transform video is the closest thing to a "why Mamba solves what transformers can't" explainer I've ever seen released for free. Everyone thinks scaling transformers is the only path forward. State space models beat them on long context using math Laplace built in 1785. This video shows the trick. Bookmark & watch today. The math is older than the United States.

As someone who ships LLM systems in production, this 22-minute Laplace transform video is the closest thing to a "why Mamba solves what transformers can't" explainer I've ever seen released for free. Everyone thinks scaling transformers is the only path forward. State space models beat them on long context using math Laplace built in 1785. This video shows the trick. Bookmark & watch today. The math is older than the United States.

fintex

26,102 views • 2 months ago

As someone who trains LLMs, this 20-minute gradient descent video is the closest thing to a "why the memorization vs generalization debate matters" explainer I've ever seen released for free. Everyone thinks neural nets learn features. They often just memorize. The video walks through the 2017 paper that shocked the field: shuffle your labels randomly and modern networks will still fit them perfectly. 25 minutes. Bookmark & watch today.

As someone who trains LLMs, this 20-minute gradient descent video is the closest thing to a "why the memorization vs generalization debate matters" explainer I've ever seen released for free. Everyone thinks neural nets learn features. They often just memorize. The video walks through the 2017 paper that shocked the field: shuffle your labels randomly and modern networks will still fit them perfectly. 25 minutes. Bookmark & watch today.

fintex

15,922 views • 2 months ago

As someone who ships LLM systems in production, this Laplace transform video is the closest thing to a "why Mamba beats transformers" explainer I've ever seen released for free. Everyone thinks attention is what makes LLMs powerful. Laplace proved in 1785 that any signal breaks into exponential pieces - locate the poles, expose the frequencies. Mamba and Claude run on this exact trick. 34 minutes. Bookmark & watch today. The math is older than the United States.

As someone who ships LLM systems in production, this Laplace transform video is the closest thing to a "why Mamba beats transformers" explainer I've ever seen released for free. Everyone thinks attention is what makes LLMs powerful. Laplace proved in 1785 that any signal breaks into exponential pieces - locate the poles, expose the frequencies. Mamba and Claude run on this exact trick. 34 minutes. Bookmark & watch today. The math is older than the United States.

Eron

67,007 views • 1 month ago

As someone who ships LLM systems in production, this backpropagation video is the closest thing to a "why GPT actually learns" explainer I've ever seen released for free. Everyone thinks training a neural network is calculus. It's not. It's 175 billion weights getting nudged proportionally to how much each one caused the mistake. PyTorch does this backwards pass on every single training step. 20 minutes. Bookmark & watch today. The algorithm is older than the iPhone.

As someone who ships LLM systems in production, this backpropagation video is the closest thing to a "why GPT actually learns" explainer I've ever seen released for free. Everyone thinks training a neural network is calculus. It's not. It's 175 billion weights getting nudged proportionally to how much each one caused the mistake. PyTorch does this backwards pass on every single training step. 20 minutes. Bookmark & watch today. The algorithm is older than the iPhone.

Eron

20,078 views • 1 month ago

As someone who trains and fine-tunes LLMs, this backpropagation video is the closest thing to a "the calc 1 rule that trains every LLM on earth" explainer I've ever seen released for free. Everyone thinks training GPT-5 requires exotic math. It doesn't. The entire $500B AI industry runs on the chain rule from freshman calculus - the exact rule Marvin Minsky said would never work. Bookmark this 30-min video and watch tonight. Same math your high school teacher taught you, now updating every weight in Claude, GPT-5, and Llama.

As someone who trains and fine-tunes LLMs, this backpropagation video is the closest thing to a "the calc 1 rule that trains every LLM on earth" explainer I've ever seen released for free. Everyone thinks training GPT-5 requires exotic math. It doesn't. The entire $500B AI industry runs on the chain rule from freshman calculus - the exact rule Marvin Minsky said would never work. Bookmark this 30-min video and watch tonight. Same math your high school teacher taught you, now updating every weight in Claude, GPT-5, and Llama.

fintex

50,749 views • 1 month ago

As someone who's been shipping LLMs since the GPT-2 days, this lecture on cross-entropy from a Stanford math grad is the closest thing to an ML PhD qualifying exam I've ever seen released publicly for free. Everyone thinks language models predict the next word. They don't. They compress language. Once you see the math, you can't unsee it. 33 minutes. Bookmark & watch today.

As someone who's been shipping LLMs since the GPT-2 days, this lecture on cross-entropy from a Stanford math grad is the closest thing to an ML PhD qualifying exam I've ever seen released publicly for free. Everyone thinks language models predict the next word. They don't. They compress language. Once you see the math, you can't unsee it. 33 minutes. Bookmark & watch today.

fintex

356,334 views • 2 months ago

As someone who ships LLM systems in production, this double pendulum video is the closest thing to a "why one changed word breaks Claude and GPT-5" explainer I've ever seen released for free. Everyone thinks LLMs are deterministic machines. They're not. Every prompt to GPT-5 is a chaotic double pendulum - a tiny perturbation in the initial state (one comma, one whitespace) sends the entire response into a completely different attractor. Bookmark this 26-min video and watch tonight. Same butterfly-effect chaos that broke Poincaré's math, now the exact reason your API calls to Claude feel unreliable.

As someone who ships LLM systems in production, this double pendulum video is the closest thing to a "why one changed word breaks Claude and GPT-5" explainer I've ever seen released for free. Everyone thinks LLMs are deterministic machines. They're not. Every prompt to GPT-5 is a chaotic double pendulum - a tiny perturbation in the initial state (one comma, one whitespace) sends the entire response into a completely different attractor. Bookmark this 26-min video and watch tonight. Same butterfly-effect chaos that broke Poincaré's math, now the exact reason your API calls to Claude feel unreliable.

fintex

51,856 views • 1 month ago

As someone who trains and fine-tunes LLMs, this double descent video is the closest thing to a "why every ML textbook was wrong about GPT-5" explainer I've ever seen released for free. Everyone thinks bigger models overfit worse. They don't. Past a certain scale the test error dives again - the exact reason OpenAI and Anthropic keep scaling up instead of down. Bookmark & watch this weekend. The 2019 discovery that quietly forced Stanford professors to rewrite their textbooks and reshaped how every frontier LLM gets trained.

As someone who trains and fine-tunes LLMs, this double descent video is the closest thing to a "why every ML textbook was wrong about GPT-5" explainer I've ever seen released for free. Everyone thinks bigger models overfit worse. They don't. Past a certain scale the test error dives again - the exact reason OpenAI and Anthropic keep scaling up instead of down. Bookmark & watch this weekend. The 2019 discovery that quietly forced Stanford professors to rewrite their textbooks and reshaped how every frontier LLM gets trained.

fintex

68,769 views • 1 month ago

As someone who reads every transformer paper that drops, this Fourier series lecture is the closest thing to a Google FNet explainer I've ever seen released for free. Everyone thinks attention is what makes transformers powerful. Google replaced attention with a 200-year-old Fourier transform. It nearly matched BERT and ran 7x faster. 25 minutes. Bookmark & watch today. Then read the article below - I broke down the 5 pieces of math the AI hype skips.

As someone who reads every transformer paper that drops, this Fourier series lecture is the closest thing to a Google FNet explainer I've ever seen released for free. Everyone thinks attention is what makes transformers powerful. Google replaced attention with a 200-year-old Fourier transform. It nearly matched BERT and ran 7x faster. 25 minutes. Bookmark & watch today. Then read the article below - I broke down the 5 pieces of math the AI hype skips.

fintex

101,169 views • 2 months ago

As someone who ships LLM systems in production, this sliding puzzle video is the closest thing to a "why sampling harder never finds the answer" explainer I've ever seen released for free. Everyone thinks a hard problem means a huge search space. It doesn't. Klotski has just 25,955 states, yet random search dies in the same pit every time. That's the exact reason OpenAI and Anthropic pay for guided search instead of more samples. Bookmark & watch this weekend. Same graph traversal, from a wooden toy on a coffee table to every reasoning model burning test-time compute today.

As someone who ships LLM systems in production, this sliding puzzle video is the closest thing to a "why sampling harder never finds the answer" explainer I've ever seen released for free. Everyone thinks a hard problem means a huge search space. It doesn't. Klotski has just 25,955 states, yet random search dies in the same pit every time. That's the exact reason OpenAI and Anthropic pay for guided search instead of more samples. Bookmark & watch this weekend. Same graph traversal, from a wooden toy on a coffee table to every reasoning model burning test-time compute today.

fintex

57,343 views • 1 month ago

As someone who ships LLM systems in production, this hairy ball theorem video is the closest thing to a "why positional encoding in 3D robotics always breaks" explainer I've ever seen released for free. Everyone thinks orientation in 3D space is a solved problem. A theorem from 1912 proves there is always at least one direction where any continuous vector field hits zero — the exact reason every robotics LLM glitches at certain angles. 25 minutes. Bookmark & watch today. The math is older than the airplane.

As someone who ships LLM systems in production, this hairy ball theorem video is the closest thing to a "why positional encoding in 3D robotics always breaks" explainer I've ever seen released for free. Everyone thinks orientation in 3D space is a solved problem. A theorem from 1912 proves there is always at least one direction where any continuous vector field hits zero — the exact reason every robotics LLM glitches at certain angles. 25 minutes. Bookmark & watch today. The math is older than the airplane.

Eron

58,608 views • 1 month ago

As someone who ships LLM systems in production, this billiards video is the closest thing to a "why every stable LLM behavior is just a lucky island in a sea of chaos" explainer I've ever seen released for free. Everyone thinks GPT-5 always behaves the same way. It doesn't. Every consistent Claude answer lives on a tiny island of stability inside an infinite chaotic plane - the exact "winning zone" this video maps by bouncing balls off polygons. Bookmark this 11-min video and watch tonight. Same emergent stability math from 2007 chaos research, now the reason a single prompt tweak can drop your LLM into total gibberish.

As someone who ships LLM systems in production, this billiards video is the closest thing to a "why every stable LLM behavior is just a lucky island in a sea of chaos" explainer I've ever seen released for free. Everyone thinks GPT-5 always behaves the same way. It doesn't. Every consistent Claude answer lives on a tiny island of stability inside an infinite chaotic plane - the exact "winning zone" this video maps by bouncing balls off polygons. Bookmark this 11-min video and watch tonight. Same emergent stability math from 2007 chaos research, now the reason a single prompt tweak can drop your LLM into total gibberish.

fintex

42,313 views • 12 days ago

As someone who's spent 3 years fine-tuning ML models, this lecture on neural networks from a Stanford math grad is the closest thing to a no-bullshit "day one of ML" briefing I've ever seen released publicly for free. Everyone thinks neural nets are magic. They aren't. They're 13,000 dials - a matrix multiplication, a sigmoid squish, done. Once you see the math, "AI" stops feeling mysterious. 18 minutes. Bookmark & watch today.

As someone who's spent 3 years fine-tuning ML models, this lecture on neural networks from a Stanford math grad is the closest thing to a no-bullshit "day one of ML" briefing I've ever seen released publicly for free. Everyone thinks neural nets are magic. They aren't. They're 13,000 dials - a matrix multiplication, a sigmoid squish, done. Once you see the math, "AI" stops feeling mysterious. 18 minutes. Bookmark & watch today.

fintex

332,082 views • 2 months ago

As someone who ships LLM systems in production, this scaling video is the closest thing to a "why 96% of Claude and GPT-5's weights are literally useless" explainer I've ever seen released for free. Everyone thinks trillion-parameter models need every parameter. They don't. A 2019 pruning experiment proved you can delete 96% of a neural net's weights with zero performance loss - meaning most of Claude and GPT-5 is empty scaffolding around a tiny "winning lottery ticket" network doing all the real work. Bookmark this 18-min video and watch tonight. Same lottery ticket math from 2019 MIT research, now the reason every AI lab wastes 90%+ of its Nvidia budget.

As someone who ships LLM systems in production, this scaling video is the closest thing to a "why 96% of Claude and GPT-5's weights are literally useless" explainer I've ever seen released for free. Everyone thinks trillion-parameter models need every parameter. They don't. A 2019 pruning experiment proved you can delete 96% of a neural net's weights with zero performance loss - meaning most of Claude and GPT-5 is empty scaffolding around a tiny "winning lottery ticket" network doing all the real work. Bookmark this 18-min video and watch tonight. Same lottery ticket math from 2019 MIT research, now the reason every AI lab wastes 90%+ of its Nvidia budget.

fintex

787,845 views • 21 days ago

LLM-grounded Diffusion: Enhancing Prompt Understanding of Text-to-Image Diffusion Models with Large Language Models paper page: github: Recent advancements in text-to-image generation with diffusion models have yielded remarkable results synthesizing highly realistic and diverse images. However, these models still encounter difficulties when generating images from prompts that demand spatial or common sense reasoning. We propose to equip diffusion models with enhanced reasoning capabilities by using off-the-shelf pretrained large language models (LLMs) in a novel two-stage generation process. First, we adapt an LLM to be a text-guided layout generator through in-context learning. When provided with an image prompt, an LLM outputs a scene layout in the form of bounding boxes along with corresponding individual descriptions. Second, we steer a diffusion model with a novel controller to generate images conditioned on the layout. Both stages utilize frozen pretrained models without any LLM or diffusion model parameter optimization. We validate the superiority of our design by demonstrating its ability to outperform the base diffusion model in accurately generating images according to prompts that necessitate both language and spatial reasoning. Additionally, our method naturally allows dialog-based scene specification and is able to handle prompts in a language that is not well-supported by the underlying diffusion model.

LLM-grounded Diffusion: Enhancing Prompt Understanding of Text-to-Image Diffusion Models with Large Language Models paper page: github: Recent advancements in text-to-image generation with diffusion models have yielded remarkable results synthesizing highly realistic and diverse images. However, these models still encounter difficulties when generating images from prompts that demand spatial or common sense reasoning. We propose to equip diffusion models with enhanced reasoning capabilities by using off-the-shelf pretrained large language models (LLMs) in a novel two-stage generation process. First, we adapt an LLM to be a text-guided layout generator through in-context learning. When provided with an image prompt, an LLM outputs a scene layout in the form of bounding boxes along with corresponding individual descriptions. Second, we steer a diffusion model with a novel controller to generate images conditioned on the layout. Both stages utilize frozen pretrained models without any LLM or diffusion model parameter optimization. We validate the superiority of our design by demonstrating its ability to outperform the base diffusion model in accurately generating images according to prompts that necessitate both language and spatial reasoning. Additionally, our method naturally allows dialog-based scene specification and is able to handle prompts in a language that is not well-supported by the underlying diffusion model.

AK

83,681 views • 3 years ago

Diffusion is the foundational ML framework behind state-of-the-art AI image and video generation, including Sora, Midjourney and Google Veo. In this episode of Decoded, Ankit Gupta sits down with Francois Chaubard to discuss how diffusion works, walk through a code sample, and explain why everyone training models should understand it. 00:00 Intro 00:33 What is diffusion? 02:50 What are applications of diffusion today? 04:06 Key innovations 07:01 Code examples 19:25 The "squint test" 22:27 Other areas diffusion is widely accessible 24:49 Outro

Diffusion is the foundational ML framework behind state-of-the-art AI image and video generation, including Sora, Midjourney and Google Veo. In this episode of Decoded, Ankit Gupta sits down with Francois Chaubard to discuss how diffusion works, walk through a code sample, and explain why everyone training models should understand it. 00:00 Intro 00:33 What is diffusion? 02:50 What are applications of diffusion today? 04:06 Key innovations 07:01 Code examples 19:25 The "squint test" 22:27 Other areas diffusion is widely accessible 24:49 Outro

Y Combinator

74,604 views • 8 months ago

Fourier wrapped a signal around a circle in 1822. Every Spotify equalizer, every noise filter in your AirPods, and every frequency layer inside ChatGPT runs on what he found. The trick is simple. Spin the signal at different speeds. When the spin matches a hidden frequency, the center of mass shifts. That shift is the answer. Claude and GPT use this same idea to separate relevant tokens from noise. 20 minutes. Bookmark & watch today. The math is older than the telephone.

Fourier wrapped a signal around a circle in 1822. Every Spotify equalizer, every noise filter in your AirPods, and every frequency layer inside ChatGPT runs on what he found. The trick is simple. Spin the signal at different speeds. When the spin matches a hidden frequency, the center of mass shifts. That shift is the answer. Claude and GPT use this same idea to separate relevant tokens from noise. 20 minutes. Bookmark & watch today. The math is older than the telephone.

Eron

57,645 views • 1 month ago

3DTopia-XL High-Quality 3D PBR Asset Generation via Primitive Diffusion demo: model: 3DTopia-XL scales high-quality 3D asset generation using Diffusion Transformer (DiT) built upon an expressive and efficient 3D representation, PrimX. The denoising process takes 5 seconds to generate a 3D PBR asset from text/image input which is ready for the graphics pipeline to use.

3DTopia-XL High-Quality 3D PBR Asset Generation via Primitive Diffusion demo: model: 3DTopia-XL scales high-quality 3D asset generation using Diffusion Transformer (DiT) built upon an expressive and efficient 3D representation, PrimX. The denoising process takes 5 seconds to generate a 3D PBR asset from text/image input which is ready for the graphics pipeline to use.

AK

87,282 views • 2 years ago

This is the craziest model you'll see this week! This model powers Figma. It's also behind Wayfair and a bunch of major e-commerce retailers. This model is a PIXEL-PERFECT image editing model. Literally, "pixel perfect". The model is capable of modifying one image while keeping everything else untouched. It's a model designed, developed, and deployed by the research team at Jasper. This beats OpenAI models, Nano Banana, and every single general image model I've seen so far. It's up to 100x less expensive to run, and up to 10X faster. I recorded a quick video to show you how impressive this is. There's a ton of research behind this model. I'm adding two links below so you can read about how this model works, specifically about "Latent Bridge Matching" and "Flash Diffusion".

This is the craziest model you'll see this week! This model powers Figma. It's also behind Wayfair and a bunch of major e-commerce retailers. This model is a PIXEL-PERFECT image editing model. Literally, "pixel perfect". The model is capable of modifying one image while keeping everything else untouched. It's a model designed, developed, and deployed by the research team at Jasper. This beats OpenAI models, Nano Banana, and every single general image model I've seen so far. It's up to 100x less expensive to run, and up to 10X faster. I recorded a quick video to show you how impressive this is. There's a ton of research behind this model. I'm adding two links below so you can read about how this model works, specifically about "Latent Bridge Matching" and "Flash Diffusion".

Santiago

25,861 views • 11 months ago