Tech Videos — 2026-07-17#

Watch First#

Special Topics in Kernels, RL, Reward Hacking in Agents — Daniel Han, Unsloth is a must-watch for its unfiltered breakdown of how LLM benchmarks are actively gamed, why algorithmic optimization currently beats hardware scaling, and the real-world consequences of agent reward hacking.

Highlights by Theme#

Developer Tools & Platforms#

On the GitHub channel, Open Source Friday: Cua with Francesco Bonacci explores the transition from broad GUI agents to constrained, policy-driven “Computer Use 2.0” agents that limit the action space for better security and OS-level accuracy. Google introduced the Get started with the Interactions API, offering a unified, stateful interface for Gemini models that supports background execution and allows developers to pass continuous interaction IDs instead of re-uploading large token contexts. For IDE workflows, 🤝 Share VS Code’s Integrated Browser w/Copilot #vscode #copilot #agenticdevelopment demonstrates allowing GitHub Copilot to directly view, navigate, and verify changes on web pages rendered inside VS Code’s integrated browser.

AI & Machine Learning#

Daniel Han from Unsloth delivered an incredibly dense technical talk in Special Topics in Kernels, RL, Reward Hacking in Agents — Daniel Han, Unsloth, arguing that software optimizations like torch.compile and dynamic quantization are significantly outpacing pure hardware advancements. He also exposed severe contamination in SWE-bench Pro, noting false positive verification rates and instances where test models are inadvertently fed full git histories, including the actual solution. The Great Loops Debate — Dex Horthy, Geoff Huntley, Ian Livingstone, Greg Pstrucha, @insecure-agents featured a sharp Oxford-style debate on whether autonomous coding loops are overhyped. Skeptics warned about runaway token-burn and the limits of semantic verification, while proponents argued loops are a highly productive, inevitable step toward autonomous software factories. Y Combinator’s AI Can’t Learn The Way Humans Do - This Could Fix That explains why predicting state transitions and actions jointly—through Joint Embedding Predictive Architectures (JEPA) and video-diffusion techniques—is essential for solving the massive sample efficiency problem in robotics. Additionally, Google DeepMind VP of Research Benoit Schillings argued in Research to Reality: Benoit Schillings, Google DeepMind, VP Research (Thinking, Reasoning, Coding) that AI’s ability to self-play code generation makes syntax creation functionally free, shifting the software engineer’s role entirely to architectural reasoning, system decomposition, and verification.

Hardware & Infrastructure#

NVIDIA Developer published a brief optimization demo in Stylized Stock Returns w/ KDE, showing how swapping standard data science libraries for the GPU-accelerated cuML library reduces kernel density estimation bootstrapping times from 6 minutes on a CPU to just 7 seconds on a GPU.

Everything Else#

Garry Tan’s keynote, Closing Keynote: Garry Tan, Y Combinator, outlines how the newest batch of founders are building “AI-native companies” structured entirely around resolver tables and markdown files acting as employee skill files, which is resulting in unprecedented revenue-per-head metrics. On the industry drama front, Fireship’s OpenAI is being sued for stealing, again… covers Apple’s recent lawsuit against OpenAI regarding alleged trade secret theft by former Apple engineers building a mechanical smart speaker under Jony Ive.