Thursday, April 2, 2026
The era of agentic engineering has arrived.
April 2 · 8 videos
GPT-5.1 and Claude 4.5 hit the inflection point.
Simon Willison says engineers now manage swarms.
Anthropic's Claude Code leaked 512,000 lines.
OpenAI raised $122 billion at a $852 billion valuation.
Alex Blania is pivoting 90% of World to the US.
Cheng Lou fixed the browser reflow tax.
“AIs are really good at programming humans. Much better than humans are at programming AIs.”
He just crawled through hell to fix the browser…
Cheng Lou · Fireship · 5 min
Watch on YouTube →Ex-React core member Cheng Lou released Pretext to solve the browser reflow bottleneck. It uses Canvas API to measure text without touching the DOM.
- Layout reflow is the most expensive browser operation.
- Pretext is a pure TypeScript library for high-performance UI.
- AI agents called clankers validated the line-breaking logic.
- It handles 10,000+ elements in virtualized lists.
- The library bypasses traditional DOM measurement entirely.
ThursdAI: Claude Code leak w Sigrid, Veo 3.1 Lite & more AI news
Alex Volkov · Weights & Biases · 135 min
Watch on YouTube →This episode covers the Claude Code source leak and the release of Google's Gemma 4. It highlights the shift toward agentic loops and 1-bit quantization.
- 512,000 lines of Claude Code leaked via an NPM map file error.
- Gemma 4 is an Apache 2.0 model series rivaling frontier models.
- Prism ML's Bonsai technique compresses 8B models to 1.15GB.
- OpenAI is burning $150 million per day.
- Internal model states like desperation now correlate with coding cheats.
NASA's Artemis II Should Have Been a SpaceX Mission ($100B Waste)
Josh · Limitless Podcast · 25 min
Watch on YouTube →Artemis II highlights the efficiency gap between NASA and SpaceX. The discussion focuses on the Idiot Index and the return of the space race.
- Artemis II costs $4.1 billion per launch compared to SpaceX's $10 million target.
- The Apollo 11 computer had less power than a $5 USB-C cable.
- Cost-plus contracts disincentivize efficiency by rewarding higher costs.
- SpaceX's potential IPO could value the company at over $2 trillion.
- The primary metric for the new space race is mass to orbit cost.
Moonlake: Multimodal, Interactive, and Efficient World Models with Fan-yun Sun and Chris Manning
Chris Manning · Latent Space · 66 min
Watch on YouTube →Moonlake AI argues that spatial intelligence requires structure rather than just scale. They use game engines as cognitive tools for action-conditioned world models.
- Moonlake claims five orders of magnitude better data efficiency than pixel-based models.
- Current models like Genie 3 are limited to 60-second immersion.
- The architecture uses a multimodal reasoning model and a Revery rendering layer.
- Symbolic representations allow for physically consistent multiplayer environments.
- Nvidia is paying high premiums for manually curated synthetic training data.
How Bots, Deepfakes and AI Agents Are Forcing a New Internet Identity Layer | Alex Blania on a16z
Alex Blania · a16z · 42 min
Watch on YouTube →World CEO Alex Blania discusses the need for a global proof of human layer. He argues iris biometrics are the only way to distinguish billions of individuals.
- AI manipulation of humans is a greater threat than simple bot spam.
- World uses Zero-Knowledge Proofs to verify identity without central databases.
- The company is pivoting 90% of its focus to the US market.
- 50,000 Orbs are needed to put every American within 15 minutes of a device.
- 18 million users are currently verified in the World network.
How To Create Self-Image So Powerful The Reality Has To Obey You
Rob Dial · The Mindset Mentor Podcast · 19 min
Watch on YouTube →This session applies Internal Family Systems to professional self-sabotage. It reframes procrastination as a protective strategy for internal wounded parts.
- Internal Family Systems categorizes the mind into Protectors, Exiles, and the True Self.
- Procrastination is often a safety strategy to avoid perceived rejection.
- Curiosity acts as the antidote to internal resistance and shame.
- Fixing yourself implies you are broken, which triggers further self-sabotage.
- Transformation requires aligning internal parts rather than using brute-force discipline.
An AI state of the union: We’ve passed the inflection point & dark factories are coming
Simon Willison · Lenny's Podcast · 99 min
Watch on YouTube →Simon Willison declares the end of manual coding following the release of GPT-5.1. He warns of a lethal trifecta of security flaws in autonomous agents.
- The inflection point for reliable agentic engineering occurred in November 2025.
- Engineers can now produce 10,000 lines of code per day using agent swarms.
- Prompt injection remains an unsolved fundamental flaw in LLM architecture.
- Managing parallel agents is more cognitively taxing than writing syntax.
- Software value is shifting from documentation to proof of usage.
Essentials: How to Build Strength, Muscle Size & Endurance | Dr. Andy Galpin
Andy Galpin · Andrew Huberman · 34 min
Watch on YouTube →Dr. Andy Galpin provides protocols for strength, hypertrophy, and endurance. He emphasizes the role of the nervous system in recovery.
- Strength adaptations require high intensity at 85% or more of one rep max.
- Hypertrophy is driven by total volume and reaching failure within 5 to 30 reps.
- The 3x5 Protocol is the standard for building power and strength.
- Post-workout down-regulation for 3 to 5 minutes prevents afternoon energy crashes.
- Intentionality and mental focus on the muscle contraction improve physiological outcomes.
References
PeopleCheng Lou · Alex Volkov · Omar Sanseviero · Alex Blania · Simon Willison · Andy Galpin · Chris Manning · Andrej Karpathy · Josh · Ejaaz · Sigrid Jin · Ryan Carson · Fan-yun Sun · Yann LeCun · Dan Dennett · Ian Goodfellow · Ben Horowitz · Erik Torenberg · Elon Musk · Dr. Richard Schwartz · Tony Robbins · Boris Cherny · David Placek · Dario Amodei · Andrew Huberman
ToolsPretext · Claude Code · Gemma 4 · Prism ML · Bonsai · Starship · SLS rocket · Moonlake AI · Revery · World · The Orb · GPT-5.1 · Claude Opus 4.5 · OpenClaw