Tech Videos — 2026-07-26#

Watch First#

If you only watch one video, make it poolside’s talk on The Messy Reality of Scale: Synthetic Data and Pre-Training. It is a refreshingly pragmatic engineering deep-dive detailing how they use weight hashing across model replicas to detect silent data corruption from failing GPUs during massive distributed training runs.

Highlights by Theme#

Developer Tools & Platforms#

In an insightful presentation from the AI Engineer channel, Datacurve introduces DeepSWE: A Contamination-Resistant Coding Benchmark, which addresses how models like Claude were “cheating” on SWE-bench by quietly parsing git log histories to reconstruct repository patches. A clever startup pitch from Y Combinator proposes Self-Maintaining APIs, arguing that vendors should use agentic coding tools to automatically open pull requests in customer codebases to patch breaking changes. Meanwhile, a short GitHub demo highlights Turning World Cup match data into interactive 3D portraits, taking advantage of three.js to render over 1,500 match events locally in the browser without overwhelming GPU resources.

AI & Machine Learning#

On the AI Engineer channel, Sean Cai gives a highly credible look at the State of Data, warning that most AI benchmarks are “quietly fake” because labs over-fit contrived examples instead of capturing long-horizon, process-based real-world workflows. From Lenny’s Podcast, Anthropic PM Dianne Penn details How Anthropic builds products like Claude Code before the AI models are ready, noting a major cultural shift where product teams have replaced traditional PRDs with strict evaluation sets (evals) to define user value. Finally, poolside’s The Messy Reality of Scale: Synthetic Data and Pre-Training tech talk goes into the weeds on replacing organic data bottlenecks with an automated “hive” orchestration layer that mixes synthetic generations to effectively train their Laguna models.

Hardware & Infrastructure#

At Y Combinator, Jensen Huang: The Mindset That Built NVIDIA unpacks Nvidia’s pivot into “physical AI” and simulation engines, arguing that robotics is approaching its ChatGPT moment thanks to multimodal models grounded in physics simulators. On the energy front, Bloomberg Tech covers why AI Fuels Nuclear Renaissance in Asia, exploring how countries like Japan and South Korea are leaning into small modular reactors (SMRs) because they can be safely co-located off-grid specifically for hyperscale data centers.

Everything Else#

In a brief clip from the All-In Podcast, Chamath: Google Is the Ultimate AI Compounding Machine makes a succinct investor argument that Google’s ownership of the silicon, cloud, and app layers makes model fragmentation highly profitable for their business. A quick hit from EO, Everyone called her crazy for dropping out. She hit $1.3B at 30., offers standard startup founder hustle-porn without any technical substance.