Tech Videos

Here is today’s curated digest, filtering out the marketing fluff to highlight where the real technical signal lies across the videos in your notebook.

Watch First

The single video most worth watching is Einstein Arena: Harnessing Collective Agent Intelligence for Open Science — James Zou, Together AI. It demonstrates how transitioning from rigid agent workflows to open-ended, competitive environments allowed a collaborative collective of models to discover a breakthrough solution to the 11-dimensional kissing number problem, outperforming both human mathematicians and DeepMind.

Highlights by Theme

Developer Tools & Platforms

This category is dominated by the rise of Model Context Protocol (MCP) integrations. End-to-End Testing for Spring Boot apps with Playwright in VS Code by Visual Studio Code shows a live demo of GitHub Copilot using a Playwright MCP extension to drive a real browser, executing action/assertion tests against a localhost app. Similarly, OpenAI’s Build agent-ready sites with WebMCP and Syntax’s WebMCP now in ChatGPT detail how exposing internal website state APIs via WebMCP allows models to directly interact with complex UIs, such as 3D modeling tools. Lastly, GitHub’s GitHub Copilot app for beginners: using the diff, terminal, and browser showcases the side-by-side terminal, diff, and browser integration that eliminates tab-hopping when reviewing agent-generated code.

AI & Machine Learning

AWS Developers’ How to Build Agents That Write Their Own Tools delivers high-quality engineering signal, demonstrating ‘meta-tooling’ where an agent uses a text editor, shell, and dynamic loader to write, validate, and register its own Python tools in a single session. For those building agent pipelines, Google Cloud Tech’s 4 ways loop engineering fails (and how to fix them) outlines critical mitigation strategies for common pitfalls like runaway token costs and agents grading their own work. Finally, the Dwarkesh Patel interviews with SemiAnalysis founder Dylan Patel (Two labs will soon control most of the world’s workforce and Why AI labs are shelving their best models) offer a fascinating look at the macro constraints of the frontier labs, arguing that safety testing and regulatory hurdles are actively forcing labs to shelf their most advanced models like Mythos 2 and Astra.

Hardware & Infrastructure

In a major hardware reveal, OpenAI Says New Jalapeno Chips Outperformed Nvidia in Testing on Bloomberg Tech features OpenAI’s VP of Hardware, Richard Ho, demonstrating ‘Jalapeno’—their custom inference ASIC co-designed with Broadcom. The chip leverages HBM4 and SRAM to eliminate data movement latency, showing 1.8x to 4x better performance-per-watt than Nvidia’s GB300 on open-source benchmarks. Concurrently, NVIDIA’s NVIDIA Vera Rubin NVL72 production racks are here. highlights liquid-cooled, cable-free rack designs built for rapid robotic assembly, while Bloomberg Tech’s Apple Upgrades Macs for the AI Era covers Apple’s latest desktop line, where the M6 Mac Mini and M5 Ultra Mac Studio (offering up to 500GB of memory) are seeing surging demand as local AI processing workstations.

Everything Else

For startup execution, Y Combinator’s Max Junestrand: You Need The Willingness To Learn Faster Than Anyone Else is a masterclass in scale, charting Legora’s climb to $100M ARR in 18 months and offering great insights on hiring for growth trajectory over resume logos. Stripe’s Cliff Obrecht in conversation with John Collison highlights the harsh business realities of AI, explaining how the unit economics of agentic loops (which run web research and massive renders) forced Canva to vertically integrate and train their own image models to reduce serving costs by 90%. Additionally, the All-In Podcast’s Flock CEO on Catching Bad Cops: “It’s a Ton. It’s More Than I Ever Would’ve Hoped.” covers how the surveillance company implements mandatory heuristics to flag and fire police officers misusing camera networks for personal tracking.

📊 I can plot a comparison of the hardware specs for Apple’s new M6/M5 series vs OpenAI’s Jalapeno design if you’d like to analyze the architectural differences.