Tech Videos — 2026-06-30#

Watch First#

Session on Reasoning from Microsoft Research is the standout for working engineers. It cuts through the hype of “test-time compute” to expose how verification steps actually bottleneck LLM execution, and presents a pragmatic framework for bounding agentic workflows using formal logic constraints.

Highlights by Theme#

Developer Tools & Platforms#

GitHub officially brought Copilot Agent support natively into JetBrains IDEs, extending their agent client protocol to allow complex workspace orchestrations directly within IntelliJ, as covered in GitHub Copilot Agent is now available in JetBrains AI Assistant. Google for Developers previewed OS-level agent integrations in Sameer Samat on Android 17 and the Future of Intelligent Computing, most notably demonstrating “App Automation,” where Gemini executes GUI interactions inside isolated, containerized virtual windows without hijacking the user’s primary screen. In parallel, Google Cloud Tech posted a guide showing how barriers to entry are dropping, generating native applications directly from a prompt in Build your first Android app in AI Studio in 5 minutes.

AI & Machine Learning#

Microsoft Research’s Session on Reasoning breaks down test-time scaling methods like Best-of-N and Bin Search, revealing that per-step verification overhead often dominates total inference latency unless dynamically scheduled based on task difficulty. Tackling the hardware limits of long contexts, Microsoft Research’s Session: Compute & Trust (Systems) pushes back on dense attention, proposing “V-Attention” to approximate attention distributions via top-K routing and random sampling, recovering baseline accuracy even at 50x sparsity. For post-training alignment, Reinforce Adjoint Matching: Scaling Diffusion RL from Microsoft Research details a scalable, on-policy objective for diffusion models that bypasses computationally unstable stochastic rollouts.

Hardware & Infrastructure#

NVIDIA highlights extreme low-level optimization in How Together AI Uses NVIDIA’s Full Stack to Deliver AI Responses in Under 100ms, explaining a “megakernel” approach that compiles an entire model into a single GPU kernel for ultra-low latency voice agent inference. For securing enterprise deployments, NVIDIA Developer’s Long-Running AI Agents: The Next Breakthrough in Enterprise Work introduces the OpenShell runtime, which places autonomous agents in secure sandboxes with strict file, network, and endpoint API limitations. On the cloud side, the massive AWS Summit Washington DC 2026 - Keynote on the AWS Events channel introduced Bedrock AgentCore for managing stateful agent memory and runtime authentication, alongside a new $1B modernization framework for US Intelligence Community workloads.

Everything Else#

The Pragmatic Engineer channel tackles hiring woes in Companies have no idea how to evaluate engineers, arguing that data structure and algorithm interviews persist stubbornly simply because the industry still lacks a better evaluation mechanism for hard skills. For a deep historical dive on governance, Lex Fridman’s Anthony Kaldellis: Roman Empire, Byzantine Empire, Rise & Fall of Empires | Lex Fridman Podcast #498 unpacks how the Eastern Roman Empire survived 1,000 years past the West by operating as a “monarchic republic” bound by responsive taxation rather than pure military dictatorship.