Wednesday, February 25, 2026
Materials are the new bottleneck for AI scaling.
February 25 · 8 videos
Max Welling says nature is the fastest computer.
He calls it the Physics Processing Unit.
Anthropic claims China stole 16 million Claude exchanges.
Mitchell Hashimoto is building harnesses for AI agents.
Khabib Nurmagomedov explains the Mentally Smash framework.
Happiness is a trainable skill.
The model is no longer the only constraint.
“Desire is a contract that you make with yourself to be unhappy until you get what you want.”
Khabib vs Lex: Training with Khabib | FULL EXCLUSIVE FOOTAGE
Lex Fridman · Lex Fridman · 22 min
Watch on YouTube →Lex Fridman trains with MMA legend Khabib Nurmagomedov to explore the physical and psychological limits of elite performance. This session reveals the technical mechanics of Khabib's legendary grappling pressure.
- Khabib uses a Mentally Smash framework to force opponents into psychological collapse before seeking a submission.
- Training partners describe Khabib's perceived weight as 280 pounds due to his efficient hip and leg control.
- The session highlights the use of leg hooks to prevent opponents from standing up while conserving the attacker's energy.
- Khabib attributes his dominance to having 25 high-level training partners in his Dubai camp since age 18.
- The philosophy of training involves reaching a fatigue threshold where the mind wants to quit but the body continues.
- Lex describes the experience as being a black belt who was turned back into a white belt by insane pressure.
Generative Adversarial Networks (GANs) Specialization
Andrew Ng · DeepLearningAI · 5 min
Watch on YouTube →Andrew Ng and Sharon Joe introduce a new specialization on GANs, focusing on their potential to transform media and solve data scarcity. The course uses intuitive analogies to explain complex adversarial training.
- GANs are explained through the Art Forger vs. Art Inspector analogy where two networks compete to improve output quality.
- The specialization emphasizes the IKEA effect, suggesting students value models more when they build them from scratch.
- Synthetic data generated by GANs is being used to train medical AI models where real X-ray data is scarce or expensive.
- The curriculum covers DCGANs for high-quality image generation and conditional GANs for granular output control.
- Week 3 of the course is dedicated entirely to the technical challenge of maintaining stable training in adversarial setups.
- GANs are positioned as a critical tool for privacy preservation by synthesizing data in sensitive domains.
TensorFlow: Advanced Techniques Specialization
Lawrence Moroney · DeepLearningAI · 6 min
Watch on YouTube →Lawrence Moroney and Andrew Ng detail a curriculum designed to move developers beyond sequential models into complex, non-linear architectures. This specialization focuses on low-level control of the training loop.
- The Functional API is introduced as the essential tool for building models with multiple inputs, outputs, and internal loops.
- Learners are taught to crack open the training loop using GradientTape to manually manage backpropagation.
- Distributed training strategies are covered to help engineers scale models across multiple GPUs and TPUs.
- The course includes practical projects like a zombie detector to teach object detection and image segmentation.
- Model interpretation techniques are emphasized to help developers understand why convolutions make specific predictions.
- The pedagogical approach involves taking a small step back to basics to enable a huge leap forward in custom implementation.
Max Welling: Materials Underlie Everything
Max Welling · Latent Space · 34 min
Watch on YouTube →Max Welling discusses the transition from digital AI to the physical layer of materials discovery. He introduces the concept of nature as a computational resource for solving climate change.
- The Physics Processing Unit (PPU) concept treats physical experiments as nature-driven computations working with digital models.
- Welling reveals that the mathematics of diffusion models is identical to the physics of non-equilibrium stochastic thermodynamics.
- CuspAI aims to transform materials discovery into a high-speed search engine problem specifically for carbon capture.
- Equivariant neural networks reduce data requirements by hardcoding physical laws and symmetries directly into the architecture.
- The materials bottleneck is identified as the primary constraint for both GPU scaling and the global energy transition.
- A Jeff Bezos-backed AI for Science startup recently raised 6.2 billion dollars, signaling a massive capital influx into the sector.
Mitchell Hashimoto’s new way of writing code
Mitchell Hashimoto · The Pragmatic Engineer · 118 min
Watch on YouTube →Mitchell Hashimoto discusses the evolution of HashiCorp and his new AI-native engineering workflow. He provides a candid look at cloud provider dynamics and the future of open source.
- Hashimoto introduces Harness Engineering as the practice of building automated validation systems to catch AI-generated errors.
- The Ghostty terminal renderer is optimized to submit work to the GPU in just 9 microseconds for extreme performance.
- AWS is described as historically arrogant while Microsoft Azure is praised for its partner-friendly business approach.
- A vouching tree system is proposed for open source to combat the high volume of AI-generated slop in pull requests.
- The Gmail moment for version control suggests that the volume of AI code churn may soon render Git and monorepos obsolete.
- Hashimoto uses an always-on agent strategy, delegating slow research tasks to AI while he focuses on high-level architecture.
Brain Tricks That Make You Happier in 15 Seconds
Rob Dial · The Mindset Mentor Podcast · 21 min
Watch on YouTube →Rob Dial synthesizes philosophies from Naval Ravikant to argue that happiness is an internal skill rather than an external achievement. The episode focuses on the mechanics of desire and presence.
- Desire is defined as a contract you make with yourself to be unhappy until you get what you want.
- Happiness is described as the state of mind that exists when desires are absent, or peace in motion.
- The episode distinguishes between unavoidable pain and suffering, which is the mental refusal to accept reality.
- Intentional boredom is presented as a tool for increasing internal stability and testing one's level of peace.
- External success like wealth or travel fails to change an internal baseline of anxiety if the mindset is not addressed.
- The interpretation model suggests that problems are merely mental labels applied to neutral sensory inputs.
Anthropic vs The Pentagon: Can They Really Do This?
Josh · Limitless Podcast · 22 min
Watch on YouTube →The Limitless team analyzes Anthropic's accusations against Chinese AI labs regarding distillation attacks. They explore the friction between AI safety and national security needs.
- Anthropic claims Chinese labs used 24,000 fake accounts and 16 million exchanges to distill Claude's capabilities.
- Minimax AI is identified as a primary actor in the alleged theft of Claude's proprietary model logic.
- The Pentagon is reportedly pivoting toward xAI and Grok because Anthropic's safety frameworks are too restrictive for military speed.
- US labs face accusations of hypocrisy for building models on copyrighted data while condemning China for distillation.
- Chinese models like C-Dance 3.0 are leading in specific verticals, producing up to 18 minutes of continuous cinematic video.
- The global AI race is described as a lawless bar fight where velocity is the only metric that currently matters.
#100 - Tmux Tutorial: Boost Developer Productivity Like a Pro | Thomas Bustos
Thomas Bustos · Let's Talk AI · 12 min
Watch on YouTube →Thomas Bustos demonstrates how to use Tmux and the Model Context Protocol (MCP) to manage technical entropy. He argues for local CLI tools over trending cloud-based AI environments.
- Tmux is positioned as the backbone of a high-performance engineering workflow for managing multiple persistent sessions.
- Bustos shows a technique to recover 20% of an AI's context window by clearing redundant history and restoring base prompts.
- The Model Context Protocol (MCP) is used to bridge AI ideation with codebase implementation through tools like Excalidraw.
- Entropy management involves a systematic daily and weekly process to clean personal knowledge bases in Obsidian.
- A spawn skill allows developers to jump from a master session into a dedicated sub-session while preserving relevant context.
- Local SSH and CLI workflows are presented as more direct paths to 10x engineering than resource-heavy cloud IDEs.
References
PeopleKhabib Nurmagomedov (x.com/TeamKhabib) · Andrew Ng · Max Welling · Mitchell Hashimoto · Rob Dial (coachwithrob.com) · Gergely Orosz (newsletter.pragmaticengineer.com) · Thomas Bustos (thomasbustos.substack.com) · Lawrence Moroney · Sharon Joe · Ali Abdelaziz · Glover Teixeira · Dan Gable · Armon Dadgar · Naval Ravikant
ToolsCuspAI · Ghostty · Tmux · Model Context Protocol (MCP) · GradientTape · Claude · Grok · Obsidian · TensorFlow · PyTorch · Excalidraw · Vagrant · Vault
PapersAI and Machine Learning for Coders · The Almanack of Naval Ravikant · The Invisible Life of Addie LaRue