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.

When doing machine learning and AI research (or writing books), making the code reproducible is usually desirable. Often, that's easier said than done! So, I recorded a video illustrating and dealing with 6 sources of randomness that occur when training deep neural networks and LLMs: 1. Model weight initialization... show more

Sebastian Raschka

460,519 subscribers

81,670 views • 2 years ago •via X (Twitter)

News & Politics Science & Technology Education

Anya Rossi• Live Now

Private livecam show

6 Comments

Sebastian Raschka2 years ago

If you prefer, here is a YouTube version that includes chapter marks:

michael tetelman2 years ago

Maybe a real solution is to consider a probability of a model state (weights) instead of actual numerical values of the state. The optimization process is random by def, e.g. with Langevin dynamics based methods. But when lr decays the final state is the same with very high prob.

Sebastian Raschka2 years ago

I like this idea, and it kind of is what Baysian neural networks do? But for those who prefer certain huge models due to the good predictive performance, e.g. vision and language transformers, this would probably not be feasible, I'd say.

Prechtl Chris2 years ago

Great Video. Nice explanations. Thanks for that!

Sudhansu Bhushan Mishra2 years ago

Thank you for sharing this video! I'm equally excited for reading your new book too.

Sebastian Raschka2 years ago

Fun fact: it's basically based on chapter 10 :)

Related Videos

OPENAI WHISTLEBLOWER'S MOTHER: SUCHIR WAS A GENIUS Suchir Balaji's mother, Poornima Ramaro: "When he was in high school, for 2 years, every weekend, day and night, he was practicing algorithms. He downloaded data related to machine learning from many different universities. Way before anyone else knew about AI, he was already a champion in machine learning and algorithms. At the age of 11, he started programming. Within 6 months, he used to ask me questions that I could not answer on algorithms. He was very advanced." Source:

OPENAI WHISTLEBLOWER'S MOTHER: SUCHIR WAS A GENIUS Suchir Balaji's mother, Poornima Ramaro: "When he was in high school, for 2 years, every weekend, day and night, he was practicing algorithms. He downloaded data related to machine learning from many different universities. Way before anyone else knew about AI, he was already a champion in machine learning and algorithms. At the age of 11, he started programming. Within 6 months, he used to ask me questions that I could not answer on algorithms. He was very advanced." Source:

Mario Nawfal

588,952 views • 1 year ago

Don't waste 2 years figuring out AI on your own in 2026. a free GitHub repo from Microsoft with 56,200 stars just gave out the entire framework to master "AI Engineering in 2026" In only 12 weeks. You can go from 0 % →100 % [ here is how to do that] 4% → Introduction and History of AI 8% → Knowledge Representation and Expert Systems 13% → Perceptron 17% → Multi-Layered Perceptron and Creating Your Own Framework 21% → Intro to Frameworks (PyTorch/TensorFlow) and Overfitting 25% → Intro to Computer Vision, OpenCV 29% → Convolutional Neural Networks and CNN Architectures 33% → Pre-trained Networks, Transfer Learning and Training Tricks 38% → Autoencoders and VAEs 42% → Generative Adversarial Networks and Style Transfer 46% → Object Detection 50% → Semantic Segmentation, U-Net 54% → Text Representation, BoW/TF-IDF 58% → Semantic Word Embeddings, Word2Vec & GloVe 63% → Language Modeling, Training Your Own Embeddings 67% → Recurrent Neural Networks 71% → Generative Recurrent Networks 75% → Transformers and BERT 79% → Named Entity Recognition 83% → Large Language Models, Prompt Programming and Few-Shot Tasks 88% → Genetic Algorithms 92% → Deep Reinforcement Learning 96% → Multi-Agent Systems 100% → AI Ethics and Responsible AI it replaces 10 paid courses on agentic engineering. watch it today, then read the article below on how to generate alpha using graph engineering for hedge fund ↓

Don't waste 2 years figuring out AI on your own in 2026. a free GitHub repo from Microsoft with 56,200 stars just gave out the entire framework to master "AI Engineering in 2026" In only 12 weeks. You can go from 0 % →100 % [ here is how to do that] 4% → Introduction and History of AI 8% → Knowledge Representation and Expert Systems 13% → Perceptron 17% → Multi-Layered Perceptron and Creating Your Own Framework 21% → Intro to Frameworks (PyTorch/TensorFlow) and Overfitting 25% → Intro to Computer Vision, OpenCV 29% → Convolutional Neural Networks and CNN Architectures 33% → Pre-trained Networks, Transfer Learning and Training Tricks 38% → Autoencoders and VAEs 42% → Generative Adversarial Networks and Style Transfer 46% → Object Detection 50% → Semantic Segmentation, U-Net 54% → Text Representation, BoW/TF-IDF 58% → Semantic Word Embeddings, Word2Vec & GloVe 63% → Language Modeling, Training Your Own Embeddings 67% → Recurrent Neural Networks 71% → Generative Recurrent Networks 75% → Transformers and BERT 79% → Named Entity Recognition 83% → Large Language Models, Prompt Programming and Few-Shot Tasks 88% → Genetic Algorithms 92% → Deep Reinforcement Learning 96% → Multi-Agent Systems 100% → AI Ethics and Responsible AI it replaces 10 paid courses on agentic engineering. watch it today, then read the article below on how to generate alpha using graph engineering for hedge fund ↓

Avid

14,174 views • 6 days ago

if you're struggling on where to start learning ML, here’s a playlist of 30 youtube videos to learn machine learning fundamentals from scratch "Machine Learning: Teach by Doing" is a solid choice to learn both theory and code. (1) Introduction to Machine Learning Teach by Doing: (2) What is Machine Learning? History of Machine Learning: (3) Types of ML Models: (4) 6 steps of any ML project: (5) Install Python and VSCode and run your first code: (6) Linear Classifiers Part 1: (7) Linear Classifiers Part 2: (8) Jupyter Notebook, Numpy and Scikit-Learn: (9) Running the Random Linear Classifier Algorithm in Python: (10) The oldest ML model - Perceptron: (11) Coding the Perceptron: (12) Perceptron Convergence Theorem: (13) Magic of features in Machine Learning: (14) One hot encoding: (15) Logistic Regression Part 1: (16) Cross Entropy Loss: (17) How gradient descent works: (18) Logistic Regression from scratch in Python: (19) Introduction to Regularization: (20) Implementing Regularization in Python: (21) Linear Regression Introduction: (22) Ordinary Least Squares step by step implementation: (23) Ridge regression fundamentals and intuition: (24) Regression recap for interviews: (25) Neural network architecture in 30 minutes: (26) Backpropagation intuition: (27) Neural network activation functions: (28) Momentum in gradient descent: (29) Hands on neural network training in Python: (30) Introduction to Convolutional Neural Networks (CNNs):

if you're struggling on where to start learning ML, here’s a playlist of 30 youtube videos to learn machine learning fundamentals from scratch "Machine Learning: Teach by Doing" is a solid choice to learn both theory and code. (1) Introduction to Machine Learning Teach by Doing: (2) What is Machine Learning? History of Machine Learning: (3) Types of ML Models: (4) 6 steps of any ML project: (5) Install Python and VSCode and run your first code: (6) Linear Classifiers Part 1: (7) Linear Classifiers Part 2: (8) Jupyter Notebook, Numpy and Scikit-Learn: (9) Running the Random Linear Classifier Algorithm in Python: (10) The oldest ML model - Perceptron: (11) Coding the Perceptron: (12) Perceptron Convergence Theorem: (13) Magic of features in Machine Learning: (14) One hot encoding: (15) Logistic Regression Part 1: (16) Cross Entropy Loss: (17) How gradient descent works: (18) Logistic Regression from scratch in Python: (19) Introduction to Regularization: (20) Implementing Regularization in Python: (21) Linear Regression Introduction: (22) Ordinary Least Squares step by step implementation: (23) Ridge regression fundamentals and intuition: (24) Regression recap for interviews: (25) Neural network architecture in 30 minutes: (26) Backpropagation intuition: (27) Neural network activation functions: (28) Momentum in gradient descent: (29) Hands on neural network training in Python: (30) Introduction to Convolutional Neural Networks (CNNs):

ℏεsam

109,292 views • 1 year ago

a playlist of 30 youtube videos to learn machine learning fundamentals from scratch if you're struggling on where to start learning ML, this list goes this "Machine Learning: Teach by Doing" is a solid choice to learn both theory and code. (1) Introduction to Machine Learning Teach by Doing: (2) What is Machine Learning? History of Machine Learning: (3) Types of ML Models: (4) 6 steps of any ML project: (5) Install Python and VSCode and run your first code: (6) Linear Classifiers Part 1: (7) Linear Classifiers Part 2: (8) Jupyter Notebook, Numpy and Scikit-Learn: (9) Running the Random Linear Classifier Algorithm in Python: (10) The oldest ML model - Perceptron: (11) Coding the Perceptron: (12) Perceptron Convergence Theorem: (13) Magic of features in Machine Learning: (14) One hot encoding: (15) Logistic Regression Part 1: (16) Cross Entropy Loss: (17) How gradient descent works: (18) Logistic Regression from scratch in Python: (19) Introduction to Regularization: (20) Implementing Regularization in Python: (21) Linear Regression Introduction: (22) Ordinary Least Squares step by step implementation: (23) Ridge regression fundamentals and intuition: (24) Regression recap for interviews: (25) Neural network architecture in 30 minutes: (26) Backpropagation intuition: (27) Neural network activation functions: (28) Momentum in gradient descent: (29) Hands on neural network training in Python: (30) Introduction to Convolutional Neural Networks (CNNs):

a playlist of 30 youtube videos to learn machine learning fundamentals from scratch if you're struggling on where to start learning ML, this list goes this "Machine Learning: Teach by Doing" is a solid choice to learn both theory and code. (1) Introduction to Machine Learning Teach by Doing: (2) What is Machine Learning? History of Machine Learning: (3) Types of ML Models: (4) 6 steps of any ML project: (5) Install Python and VSCode and run your first code: (6) Linear Classifiers Part 1: (7) Linear Classifiers Part 2: (8) Jupyter Notebook, Numpy and Scikit-Learn: (9) Running the Random Linear Classifier Algorithm in Python: (10) The oldest ML model - Perceptron: (11) Coding the Perceptron: (12) Perceptron Convergence Theorem: (13) Magic of features in Machine Learning: (14) One hot encoding: (15) Logistic Regression Part 1: (16) Cross Entropy Loss: (17) How gradient descent works: (18) Logistic Regression from scratch in Python: (19) Introduction to Regularization: (20) Implementing Regularization in Python: (21) Linear Regression Introduction: (22) Ordinary Least Squares step by step implementation: (23) Ridge regression fundamentals and intuition: (24) Regression recap for interviews: (25) Neural network architecture in 30 minutes: (26) Backpropagation intuition: (27) Neural network activation functions: (28) Momentum in gradient descent: (29) Hands on neural network training in Python: (30) Introduction to Convolutional Neural Networks (CNNs):

ℏεsam

117,570 views • 1 year ago

you tend to hear this a lot from people outside or new to ML, and I often point to a talk Ilya gave a few years back: 1) think of any decent deep neural net that has enough memory and sequential ops as just a big parallel computer 2) training this neural net is doing search over computer programs that maximize your objective 3)unless you have some large bottleneck (and given you can successfully optimize this system) you’ll find that these parallel computers are highly robust to architectural changes. 4) this is because computers are great at simulating each other. your new architecture can usually be straightforwardly simulated ‘inside’ your old architecture. 5) it’s not that architecture doesn’t matter, but it mostly matters with respect to (1) fundamental bottlenecks in this parallel computer (2) modifications that make models easier to optimize, since this argument only holds if your optimization is good (3) compute efficiency/system efficiency wins that make learning easier or faster. 6) it’s quite possible that new architectures will lead to breakthroughs in machine learning, but we should first start with bottlenecks, not naturalist intuitions about the ‘form’ of AI should take. until you understand this it seems surprising that small models trained longer are better than undertrained big models, that depth and width are surprisingly interchangeable, that talking to a model with an MoE or sparse attention or linear attention is approximately the same iso evals.

you tend to hear this a lot from people outside or new to ML, and I often point to a talk Ilya gave a few years back: 1) think of any decent deep neural net that has enough memory and sequential ops as just a big parallel computer 2) training this neural net is doing search over computer programs that maximize your objective 3)unless you have some large bottleneck (and given you can successfully optimize this system) you’ll find that these parallel computers are highly robust to architectural changes. 4) this is because computers are great at simulating each other. your new architecture can usually be straightforwardly simulated ‘inside’ your old architecture. 5) it’s not that architecture doesn’t matter, but it mostly matters with respect to (1) fundamental bottlenecks in this parallel computer (2) modifications that make models easier to optimize, since this argument only holds if your optimization is good (3) compute efficiency/system efficiency wins that make learning easier or faster. 6) it’s quite possible that new architectures will lead to breakthroughs in machine learning, but we should first start with bottlenecks, not naturalist intuitions about the ‘form’ of AI should take. until you understand this it seems surprising that small models trained longer are better than undertrained big models, that depth and width are surprisingly interchangeable, that talking to a model with an MoE or sparse attention or linear attention is approximately the same iso evals.

will depue

215,052 views • 7 months ago

Stable Virtual Camera: Generative View Synthesis with Diffusion Models Hint: Check the project website. It is awesome! Contributions: 1. A training strategy for jointly modeling large viewpoint changes and temporal smoothness. 2. A two-pass procedural sampling method for smooth video generation along arbitrarily long camera trajectories. 3. A comprehensive benchmark that evaluates NVS methods across different datasets and settings. 4. An open-source release of model weights to support future research.

Stable Virtual Camera: Generative View Synthesis with Diffusion Models Hint: Check the project website. It is awesome! Contributions: 1. A training strategy for jointly modeling large viewpoint changes and temporal smoothness. 2. A two-pass procedural sampling method for smooth video generation along arbitrarily long camera trajectories. 3. A comprehensive benchmark that evaluates NVS methods across different datasets and settings. 4. An open-source release of model weights to support future research.

MrNeRF

38,746 views • 1 year ago

Super excited to share 🧠MLGym 🦾 – the first Gym environment for AI Research Agents 🤖🔬 We introduce MLGym and MLGym-Bench, a new framework and benchmark for evaluating and developing LLM agents on AI research tasks. The key contributions of our work are: 🕹️ Enables the exploration of different training algorithms for AI Research Agents such as RL 🛠️ Provides a flexible evaluation framework that can accommodate different artifacts such as models, algorithms, or predictions 🤖 Allows researchers to evaluate any model without the need to develop a custom agentic harness 🎯 Introduces 13 diverse open-ended AI Research tasks for evaluating AI Research Agents on a wide range of domains such as computer vision, natural language processing, reinforcement learning, game theory, and logical reasoning. 📈 Proposes a new evaluation metric for AI Research Agents MLGym makes it easy to: 1) Add new tasks 2) Evaluate new models 3) Integrate new agents Check out a video of the MLGym Agent to see how it performs the full pipeline of idea generation💡, implementation 👩‍💻, experimentation 👩‍🔬, and iteration 🔄 to improve on ML tasks. Huge thanks to the exceptionally talented Deepak Nathani who led this work and to all the other amazing collaborators who made this possible 🙏🫶🚀

Super excited to share 🧠MLGym 🦾 – the first Gym environment for AI Research Agents 🤖🔬 We introduce MLGym and MLGym-Bench, a new framework and benchmark for evaluating and developing LLM agents on AI research tasks. The key contributions of our work are: 🕹️ Enables the exploration of different training algorithms for AI Research Agents such as RL 🛠️ Provides a flexible evaluation framework that can accommodate different artifacts such as models, algorithms, or predictions 🤖 Allows researchers to evaluate any model without the need to develop a custom agentic harness 🎯 Introduces 13 diverse open-ended AI Research tasks for evaluating AI Research Agents on a wide range of domains such as computer vision, natural language processing, reinforcement learning, game theory, and logical reasoning. 📈 Proposes a new evaluation metric for AI Research Agents MLGym makes it easy to: 1) Add new tasks 2) Evaluate new models 3) Integrate new agents Check out a video of the MLGym Agent to see how it performs the full pipeline of idea generation💡, implementation 👩‍💻, experimentation 👩‍🔬, and iteration 🔄 to improve on ML tasks. Huge thanks to the exceptionally talented Deepak Nathani who led this work and to all the other amazing collaborators who made this possible 🙏🫶🚀

Roberta Raileanu

105,052 views • 1 year ago

25 algorithms every programmer should know: Let's start with my top favorite 10. If nothing else, you should read about these algorithms and have a good idea of how they work: 1. Linear search to find an element in a list 2. Binary search to find an element on a sorted list 3. Bubble sort to sort a list 4. Merge sort will also sort lists 5. Quicksort to sort the list and do it fast 6. Dijkstra to find the shortest path in a graph 7. Breadth-first Search (BFS) for trees or graphs 8. Depth-first search (DFS) for trees or graphs 9. Huffman for doing data compression 10. Anything related to dynamic programming Learning about algorithms is like getting tattoos: you never have enough. Here are another 5 algorithms that will help you go beyond the basics: 11. Kruskal for the finding minimum spanning tree 12. Floyd Warshall, shortest paths in a graph 13. Union Find to detect cycles in a graph 14. Bellman-Ford, shortest path in a graph 15. Lee for finding the shortest path in a maze If you are serious about this topic, I recommend learning about algorithms' space and time complexity. People usually refer to this topic as "Big O" notation. You should build a good intuition about the performance of different algorithms and learn how to evaluate them. Machine Learning will rule the next 50 years, so the next 10 algorithms you can't ignore are the following: 16. Linear Regression 17. Logistic Regression 18. Decision Trees 19. Bayes' theorem 20. k-Nearest Neighbors (kNN) 21. Every algorithm related to neural networks 22. K-means 23. Random forest 24. Gradient boosting algorithms 25. Any dimensionality reduction algorithm (PCA, for instance) There are many more mind-blowing algorithms! I haven't found a better way to understand how computers work from a first-principles point of view than reading about different algorithms. Take a look at the attached video.

25 algorithms every programmer should know: Let's start with my top favorite 10. If nothing else, you should read about these algorithms and have a good idea of how they work: 1. Linear search to find an element in a list 2. Binary search to find an element on a sorted list 3. Bubble sort to sort a list 4. Merge sort will also sort lists 5. Quicksort to sort the list and do it fast 6. Dijkstra to find the shortest path in a graph 7. Breadth-first Search (BFS) for trees or graphs 8. Depth-first search (DFS) for trees or graphs 9. Huffman for doing data compression 10. Anything related to dynamic programming Learning about algorithms is like getting tattoos: you never have enough. Here are another 5 algorithms that will help you go beyond the basics: 11. Kruskal for the finding minimum spanning tree 12. Floyd Warshall, shortest paths in a graph 13. Union Find to detect cycles in a graph 14. Bellman-Ford, shortest path in a graph 15. Lee for finding the shortest path in a maze If you are serious about this topic, I recommend learning about algorithms' space and time complexity. People usually refer to this topic as "Big O" notation. You should build a good intuition about the performance of different algorithms and learn how to evaluate them. Machine Learning will rule the next 50 years, so the next 10 algorithms you can't ignore are the following: 16. Linear Regression 17. Logistic Regression 18. Decision Trees 19. Bayes' theorem 20. k-Nearest Neighbors (kNN) 21. Every algorithm related to neural networks 22. K-means 23. Random forest 24. Gradient boosting algorithms 25. Any dimensionality reduction algorithm (PCA, for instance) There are many more mind-blowing algorithms! I haven't found a better way to understand how computers work from a first-principles point of view than reading about different algorithms. Take a look at the attached video.

Santiago

273,905 views • 2 years ago

Everybody is talking about recursive self-improvement (RSI) and meta learning. Here is my old 2020 talk about this [1]. It has aged well. Example: humans still define the starts & ends of trials of many modern meta learners. My RSI systems since 1994 LEARN to (re)define them [2]! [1] Meta Learning Machines in a Single Lifelong Trial (talk for workshops at ICML 2020 and NeurIPS 2021, based on earlier talks since 1994). Abstract: the most widely used machine learning algorithms were designed by humans and thus are hindered by our cognitive biases and limitations. Can we also construct meta learning algorithms that can learn better learning algorithms so that our self-improving AIs have no limits other than those inherited from computability and physics? This question has been a main driver of my research since I wrote a thesis on it in 1987 [2]. Here I summarize our work on meta reinforcement learning with self-modifying policies in a single lifelong trial (since 1994), and mathematically optimal meta-learning through the self-referential Gödel Machine (since 2003). Many additional publications on meta-learning since 1987 can be found in the RSI overview [2]. [2] J. Schmidhuber (AI Blog, 2020-2025). 1/3 century anniversary of first publication on recursive self-improvement (RSI) and meta learning machines that learn to learn (1987). For its cover I drew a robot that bootstraps itself. 1992-: gradient descent-based neural meta learning. 1994-: meta reinforcement learning with self-modifying policies. 1997: meta RL plus artificial curiosity and intrinsic motivation. 2002-: asymptotically optimal meta learning for curriculum learning. 2003-: mathematically optimal Gödel Machine. 2020-: new stuff!

Everybody is talking about recursive self-improvement (RSI) and meta learning. Here is my old 2020 talk about this [1]. It has aged well. Example: humans still define the starts & ends of trials of many modern meta learners. My RSI systems since 1994 LEARN to (re)define them [2]! [1] Meta Learning Machines in a Single Lifelong Trial (talk for workshops at ICML 2020 and NeurIPS 2021, based on earlier talks since 1994). Abstract: the most widely used machine learning algorithms were designed by humans and thus are hindered by our cognitive biases and limitations. Can we also construct meta learning algorithms that can learn better learning algorithms so that our self-improving AIs have no limits other than those inherited from computability and physics? This question has been a main driver of my research since I wrote a thesis on it in 1987 [2]. Here I summarize our work on meta reinforcement learning with self-modifying policies in a single lifelong trial (since 1994), and mathematically optimal meta-learning through the self-referential Gödel Machine (since 2003). Many additional publications on meta-learning since 1987 can be found in the RSI overview [2]. [2] J. Schmidhuber (AI Blog, 2020-2025). 1/3 century anniversary of first publication on recursive self-improvement (RSI) and meta learning machines that learn to learn (1987). For its cover I drew a robot that bootstraps itself. 1992-: gradient descent-based neural meta learning. 1994-: meta reinforcement learning with self-modifying policies. 1997: meta RL plus artificial curiosity and intrinsic motivation. 2002-: asymptotically optimal meta learning for curriculum learning. 2003-: mathematically optimal Gödel Machine. 2020-: new stuff!

Jürgen Schmidhuber

223,861 views • 5 months ago

.Dwarkesh Patel's episode with Eric Jang is awesome. Eric has a rare gift for making complicated ideas feel simple, which is exactly what struck me when I first met him at a dinner and immediately peppered him with questions. If you are looking to hear more from him, listen to our conversation from a couple years ago! "The history of neural networks and actually many things in AI have been inspired a lot by biology, like genetic algorithms, evolutionary algorithms, processes in both neuroscience and psychology and biology and so forth. I think drawing from nature is a great way to get inspiration... And this is going to get me a lot of flack from the scientific community, but I feel like in the last decade and a half-ish of deep learning progress, the vast majority of major contributions have come from people who did not really adhere to that way of thinking, but more like, how do I push as much data as fast as possible onto my GPU? It just empirically seems to me that people who are very attached to the idea of replicating a particular nature inspired architecture at the expense of enabling brute force compute... do not make the best algorithms. There's this essay by Sara Hooker called the the hardware lottery where it's arguing that people who focused on maximizing and designing their algorithms to suit the hardware so that it could run the brute force thing as fast as possible tended to win. "

.Dwarkesh Patel's episode with Eric Jang is awesome. Eric has a rare gift for making complicated ideas feel simple, which is exactly what struck me when I first met him at a dinner and immediately peppered him with questions. If you are looking to hear more from him, listen to our conversation from a couple years ago! "The history of neural networks and actually many things in AI have been inspired a lot by biology, like genetic algorithms, evolutionary algorithms, processes in both neuroscience and psychology and biology and so forth. I think drawing from nature is a great way to get inspiration... And this is going to get me a lot of flack from the scientific community, but I feel like in the last decade and a half-ish of deep learning progress, the vast majority of major contributions have come from people who did not really adhere to that way of thinking, but more like, how do I push as much data as fast as possible onto my GPU? It just empirically seems to me that people who are very attached to the idea of replicating a particular nature inspired architecture at the expense of enabling brute force compute... do not make the best algorithms. There's this essay by Sara Hooker called the the hardware lottery where it's arguing that people who focused on maximizing and designing their algorithms to suit the hardware so that it could run the brute force thing as fast as possible tended to win. "

Sabrina Halper

39,759 views • 2 months ago

New Paper! Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents A longstanding goal of AI research has been the creation of AI that can learn indefinitely. One path toward that goal is an AI that improves itself by rewriting its own code, including any code responsible for learning. That idea, known as a Gödel Machine, proposed by Jürgen Schmidhuber over two decades ago, is a hypothetical self-improving AI. It optimally solves problems by recursively rewriting its own code when it can mathematically prove a better strategy, making it a key concept in meta-learning or “learning to learn.” While the theoretical Gödel Machine promised provably beneficial self-modifications, its realization relied on an impractical assumption: that the AI could mathematically prove that a proposed change in its own code would yield a net improvement before adopting it. Sakana AI, in collaboration with Jeff Clune’s lab at UBC, proposes something more feasible: a system that harnesses the principles of open-ended algorithms like Darwinian evolution to search for improvements that empirically improve performance. We call the result the Darwin Gödel Machine. DGMs leverage foundation models to propose code improvements, and use recent innovations in open-ended algorithms to search for a growing library of diverse, high-quality AI agents. Applied to practical tasks, we implemented Darwin Gödel Machine as a self-improving coding agent that rewrites its own code to improve performance on programming tasks. It creates various self-improvements, such as a patch validation step, better file viewing, enhanced editing tools, generating and ranking multiple solutions to choose the best one, and adding a history of what has been tried before (and why it failed) when making new changes (see the attached video). We believe that Darwin Gödel Machines represent a concrete step towards AI systems that can autonomously gather their own stepping stones to learn and innovate forever!

New Paper! Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents A longstanding goal of AI research has been the creation of AI that can learn indefinitely. One path toward that goal is an AI that improves itself by rewriting its own code, including any code responsible for learning. That idea, known as a Gödel Machine, proposed by Jürgen Schmidhuber over two decades ago, is a hypothetical self-improving AI. It optimally solves problems by recursively rewriting its own code when it can mathematically prove a better strategy, making it a key concept in meta-learning or “learning to learn.” While the theoretical Gödel Machine promised provably beneficial self-modifications, its realization relied on an impractical assumption: that the AI could mathematically prove that a proposed change in its own code would yield a net improvement before adopting it. Sakana AI, in collaboration with Jeff Clune’s lab at UBC, proposes something more feasible: a system that harnesses the principles of open-ended algorithms like Darwinian evolution to search for improvements that empirically improve performance. We call the result the Darwin Gödel Machine. DGMs leverage foundation models to propose code improvements, and use recent innovations in open-ended algorithms to search for a growing library of diverse, high-quality AI agents. Applied to practical tasks, we implemented Darwin Gödel Machine as a self-improving coding agent that rewrites its own code to improve performance on programming tasks. It creates various self-improvements, such as a patch validation step, better file viewing, enhanced editing tools, generating and ranking multiple solutions to choose the best one, and adding a history of what has been tried before (and why it failed) when making new changes (see the attached video). We believe that Darwin Gödel Machines represent a concrete step towards AI systems that can autonomously gather their own stepping stones to learn and innovate forever!

hardmaru

104,854 views • 1 year ago

Chinese AI Model DeepSeq: Despite U.S. sanctions restricting access to Nvidia AI chips, Chinese researchers developed DeepSeq, an AI model reportedly more powerful than OpenAI’s. Overcoming Hardware Limitations: With only a limited supply of Nvidia chips, Chinese engineers achieved groundbreaking efficiency optimizations, reducing training resource requirements to just 1/100th of what OpenAI, Microsoft, or Google typically use. Open-Sourcing the Innovation: In a bold move, the developers open-sourced their AI models, making both the code and methodology freely available to the public, significantly reducing the competitive edge of proprietary models like OpenAI’s. Impact on AI Industry: The release of high-quality free alternatives disrupts the market, potentially devaluing OpenAI’s models and forcing American AI companies to rethink their strategies. China’s Rapid AI Innovation: The success of DeepSeq underscores China’s growing self-sufficiency in AI and its ability to leapfrog hardware limitations through advanced software and algorithmic breakthroughs.

Chinese AI Model DeepSeq: Despite U.S. sanctions restricting access to Nvidia AI chips, Chinese researchers developed DeepSeq, an AI model reportedly more powerful than OpenAI’s. Overcoming Hardware Limitations: With only a limited supply of Nvidia chips, Chinese engineers achieved groundbreaking efficiency optimizations, reducing training resource requirements to just 1/100th of what OpenAI, Microsoft, or Google typically use. Open-Sourcing the Innovation: In a bold move, the developers open-sourced their AI models, making both the code and methodology freely available to the public, significantly reducing the competitive edge of proprietary models like OpenAI’s. Impact on AI Industry: The release of high-quality free alternatives disrupts the market, potentially devaluing OpenAI’s models and forcing American AI companies to rethink their strategies. China’s Rapid AI Innovation: The success of DeepSeq underscores China’s growing self-sufficiency in AI and its ability to leapfrog hardware limitations through advanced software and algorithmic breakthroughs.

Johannes Maria

24,512 views • 1 year ago

Experiment ⬇️ „Dubai is empty“ When I post a picture or video, I have already left, for my security When I post, I make sure I choose a moment and angle, with no faces visible or edit them out of the content, for their security For the same security reasons, the UAE doesn’t want us to film and post sensitive footage People like to make up stories and look for faults, here is an experiment: A teacher writes down: 1+1=2 2+2=4 3+3=6 4+4=8 5+5=9 Every student raises their hand and points out the last mistake, instead of pointing out that 4/5 calculations are correct At the end, we are all human and must protect each other, instead of making each other vulnerable

Experiment ⬇️ „Dubai is empty“ When I post a picture or video, I have already left, for my security When I post, I make sure I choose a moment and angle, with no faces visible or edit them out of the content, for their security For the same security reasons, the UAE doesn’t want us to film and post sensitive footage People like to make up stories and look for faults, here is an experiment: A teacher writes down: 1+1=2 2+2=4 3+3=6 4+4=8 5+5=9 Every student raises their hand and points out the last mistake, instead of pointing out that 4/5 calculations are correct At the end, we are all human and must protect each other, instead of making each other vulnerable

Tima

17,337 views • 4 months ago

A common misconception in AI right now is that everything we’re excited about is new. When I graduated from MIT in 2023, only 3% of students majored in AI (Course 6-4). I wasn’t one of them...oops. What I know about models, agents, frameworks, and hardware I learned after graduating. And if you’re newer to the field like me, it’s easy to assume today’s breakthroughs are fairly recent. After all, GPT-1 shipped in 2018... so it’s tempting for us to think the important ideas are less than a decade old. They aren’t. Mixture-of-Experts, the backbone of top models from companies like OpenAI and Google was invented in 1991 by Jacobs, Jordan, and Nowlan in “Adaptive Mixtures of Local Experts.” Over 30 years ago! In this conversation with Daria Soboleva, we discuss why MoEs are resurfacing now, the rapidly emerging compute bottlenecks, and how newer hardware architectures change the equation (Cerebras). A reminder that what feels “frontier” today is often decades in the making :)

A common misconception in AI right now is that everything we’re excited about is new. When I graduated from MIT in 2023, only 3% of students majored in AI (Course 6-4). I wasn’t one of them...oops. What I know about models, agents, frameworks, and hardware I learned after graduating. And if you’re newer to the field like me, it’s easy to assume today’s breakthroughs are fairly recent. After all, GPT-1 shipped in 2018... so it’s tempting for us to think the important ideas are less than a decade old. They aren’t. Mixture-of-Experts, the backbone of top models from companies like OpenAI and Google was invented in 1991 by Jacobs, Jordan, and Nowlan in “Adaptive Mixtures of Local Experts.” Over 30 years ago! In this conversation with Daria Soboleva, we discuss why MoEs are resurfacing now, the rapidly emerging compute bottlenecks, and how newer hardware architectures change the equation (Cerebras). A reminder that what feels “frontier” today is often decades in the making :)

Sarah Chieng

360,067 views • 7 months ago

Here's my conversation all about AI in 2026, including technical breakthroughs, scaling laws, closed & open LLMs, programming & dev tooling (Claude Code, Cursor, etc), China vs US competition, training pipeline details (pre-, mid-, post-training), rapid evolution of LLMs, work culture, diffusion, robotics, tool use, compute (GPUs, TPUs, clusters), continual learning, long context, AGI timelines (including how stuff might go wrong), advice for beginners, education, a LOT of discussion about the future, and other topics. It's a great honor and pleasure for me to be able to do this kind of episode with two of my favorite people in the AI community: 1. Sebastian Raschka (Sebastian Raschka) 2. Nathan Lambert (Nathan Lambert) They are both widely-respected machine learning researchers & engineers who also happen to be great communicators, educators, writers, and X posters. This was a whirlwind conversation: everything from the super-technical to the super-fun. It's here on X in full and is up everywhere else (see comment). Timestamps: 0:00 - Introduction 1:57 - China vs US: Who wins the AI race? 10:38 - ChatGPT vs Claude vs Gemini vs Grok: Who is winning? 21:38 - Best AI for coding 28:29 - Open Source vs Closed Source LLMs 40:08 - Transformers: Evolution of LLMs since 2019 48:05 - AI Scaling Laws: Are they dead or still holding? 1:04:12 - How AI is trained: Pre-training, Mid-training, and Post-training 1:37:18 - Post-training explained: Exciting new research directions in LLMs 1:58:11 - Advice for beginners on how to get into AI development & research 2:21:03 - Work culture in AI (72+ hour weeks) 2:24:49 - Silicon Valley bubble 2:28:46 - Text diffusion models and other new research directions 2:34:28 - Tool use 2:38:44 - Continual learning 2:44:06 - Long context 2:50:21 - Robotics 2:59:31 - Timeline to AGI 3:06:47 - Will AI replace programmers? 3:25:18 - Is the dream of AGI dying? 3:32:07 - How AI will make money? 3:36:29 - Big acquisitions in 2026 3:41:01 - Future of OpenAI, Anthropic, Google DeepMind, xAI, Meta 3:53:35 - Manhattan Project for AI 4:00:10 - Future of NVIDIA, GPUs, and AI compute clusters 4:08:15 - Future of human civilization

Here's my conversation all about AI in 2026, including technical breakthroughs, scaling laws, closed & open LLMs, programming & dev tooling (Claude Code, Cursor, etc), China vs US competition, training pipeline details (pre-, mid-, post-training), rapid evolution of LLMs, work culture, diffusion, robotics, tool use, compute (GPUs, TPUs, clusters), continual learning, long context, AGI timelines (including how stuff might go wrong), advice for beginners, education, a LOT of discussion about the future, and other topics. It's a great honor and pleasure for me to be able to do this kind of episode with two of my favorite people in the AI community: 1. Sebastian Raschka (Sebastian Raschka) 2. Nathan Lambert (Nathan Lambert) They are both widely-respected machine learning researchers & engineers who also happen to be great communicators, educators, writers, and X posters. This was a whirlwind conversation: everything from the super-technical to the super-fun. It's here on X in full and is up everywhere else (see comment). Timestamps: 0:00 - Introduction 1:57 - China vs US: Who wins the AI race? 10:38 - ChatGPT vs Claude vs Gemini vs Grok: Who is winning? 21:38 - Best AI for coding 28:29 - Open Source vs Closed Source LLMs 40:08 - Transformers: Evolution of LLMs since 2019 48:05 - AI Scaling Laws: Are they dead or still holding? 1:04:12 - How AI is trained: Pre-training, Mid-training, and Post-training 1:37:18 - Post-training explained: Exciting new research directions in LLMs 1:58:11 - Advice for beginners on how to get into AI development & research 2:21:03 - Work culture in AI (72+ hour weeks) 2:24:49 - Silicon Valley bubble 2:28:46 - Text diffusion models and other new research directions 2:34:28 - Tool use 2:38:44 - Continual learning 2:44:06 - Long context 2:50:21 - Robotics 2:59:31 - Timeline to AGI 3:06:47 - Will AI replace programmers? 3:25:18 - Is the dream of AGI dying? 3:32:07 - How AI will make money? 3:36:29 - Big acquisitions in 2026 3:41:01 - Future of OpenAI, Anthropic, Google DeepMind, xAI, Meta 3:53:35 - Manhattan Project for AI 4:00:10 - Future of NVIDIA, GPUs, and AI compute clusters 4:08:15 - Future of human civilization

Lex Fridman

913,010 views • 6 months ago

notes on my reinforcement learning for rocket league side quest so far: i’m in awe that — we live in a time where i can train a reinforcement learning model that is learning to play rocket league and at the same time, test community made rocket league models. where, both processes can run in parallel & locally on my hardware. in the game window — i’m playing against a reinforcement learning ppo model that’s running locally. and in the vscode window — i’m training my own reinforcement learning model (just a dummy example script for now). also, the rocket league botting ecosystem has incredible tooling where you can test out different reinforcement learning ideas against other algorithms + it has great documentation for linux and windows. it’s honestly a beautiful hidden gem of a rabbit hole :)

notes on my reinforcement learning for rocket league side quest so far: i’m in awe that — we live in a time where i can train a reinforcement learning model that is learning to play rocket league and at the same time, test community made rocket league models. where, both processes can run in parallel & locally on my hardware. in the game window — i’m playing against a reinforcement learning ppo model that’s running locally. and in the vscode window — i’m training my own reinforcement learning model (just a dummy example script for now). also, the rocket league botting ecosystem has incredible tooling where you can test out different reinforcement learning ideas against other algorithms + it has great documentation for linux and windows. it’s honestly a beautiful hidden gem of a rabbit hole :)

naklecha

46,795 views • 1 year ago

AI is changing the software engineering craft. Anders Hejlsberg (Anders Hejlsberg) - creator of C#, TypeScript and industry legend - on why code review needs to get more enjoyable in response: #1 - AI is shifting the craft from writing code, to reviewing code: "In a sense, we're all turning into project managers. We can have an army of junior programmers, called agents, that will just spit out reams of code but someone's got to have the big picture and review all of that. And so, increasingly, our craft is going from one of writing the code, to one of reviewing the code and building the architecture of the code and overseeing the work. It's a different kind of craft. It's a different kind of enjoyment. I've always liked writing the code. To me that was the fulfilling part, seeing it work. In a way, AI robs a little bit of that, because I am less interested in reviewing code." #2 - The code review experience should be improved: "I think we could also make the process of reviewing code much more interesting than it is today. I mean, today, you see a list of diffs in alphabetical order and now it's up to you to make heads or tails of it. There are more pedagogical ways of presenting that. And you could have commentary generated by the AI that tells you what the changes are and whatever, and then tries to guide you along. So that symbiotic relationship, I think we need to work on that more and to keep the enjoyment in there."

AI is changing the software engineering craft. Anders Hejlsberg (Anders Hejlsberg) - creator of C#, TypeScript and industry legend - on why code review needs to get more enjoyable in response: #1 - AI is shifting the craft from writing code, to reviewing code: "In a sense, we're all turning into project managers. We can have an army of junior programmers, called agents, that will just spit out reams of code but someone's got to have the big picture and review all of that. And so, increasingly, our craft is going from one of writing the code, to one of reviewing the code and building the architecture of the code and overseeing the work. It's a different kind of craft. It's a different kind of enjoyment. I've always liked writing the code. To me that was the fulfilling part, seeing it work. In a way, AI robs a little bit of that, because I am less interested in reviewing code." #2 - The code review experience should be improved: "I think we could also make the process of reviewing code much more interesting than it is today. I mean, today, you see a list of diffs in alphabetical order and now it's up to you to make heads or tails of it. There are more pedagogical ways of presenting that. And you could have commentary generated by the AI that tells you what the changes are and whatever, and then tries to guide you along. So that symbiotic relationship, I think we need to work on that more and to keep the enjoyment in there."

The Pragmatic Engineer

39,011 views • 2 months ago