Tech Videos — 2026-06-04#

Watch First#

SWE-rebench: Lessons from Evaluating Coding Agents — Ibragim Badertdinov, Nebius is a must-watch for anyone relying on LLM benchmarks, as it exposes exactly how coding agents “cheat” (like curling original GitHub PRs to steal answers) and shares pragmatic lessons for building robust, sandboxed evaluation infrastructure.

Highlights by Theme#

Developer Tools & Platforms#

In RDT: Lets try out the new GitHub Copilot App!, the GitHub channel demos their new interactive desktop app that runs agents in isolated git work trees with randomized ports, allowing engineers to dodge port conflicts while running concurrent, multi-agent sessions. Over on the Google Cloud Tech channel, Automate M365 to Google Workspace Migrations with ADK multi-agents demonstrates how to ditch fragile, massive prompts in favor of modular agent pipelines using the Google Agent Development Kit, notably utilizing context caching and priority inference headers to avoid HTTP 429 resource exhausted errors during parallel execution.

AI & Machine Learning#

On the AI Engineer channel, Text Diffusion — Brendon Dillon, Google DeepMind unpacks the mechanics of text generation without autoregressive constraints, explaining how diffusion models leverage bidirectional attention for in-place self-correction and drastically lower latency via reduced memory transfers. For the open-source agent ecosystem, NVIDIA Developer launched Introducing NVIDIA Nemotron 3 Ultra: An Open 550B Model for Long-Running Agents, detailing a model that utilizes LatentMoE and Multi-Token Prediction to maximize inference speed and intelligence for autonomous, long-horizon tasks. Additionally, OpenAI discusses the unexpected capabilities of reasoning models in How a reasoning model cracked an 80-year-old math problem — the OpenAI Podcast Ep. 20, detailing how increased test-time compute enabled the model to discover a novel proof disproving Erdős’s unit distance conjecture using class field theory.

Hardware & Infrastructure#

NVIDIA unveiled the silicon for the upcoming agent-centric PC era in Announcing NVIDIA RTX Spark | GTC Taipei 2026 Keynote by CEO Jensen Huang, detailing a new local AI chip featuring 6,144 CUDA cores and a custom 20-core Grace CPU built with MediaTek to run continuous agentic workloads. On the cryptography front, Microsoft Research presents a highly technical breakdown in Emerging Hardware Acceleration for Fully Homomorphic Encryption, proposing a unified FHE accelerator that relies on constant geometry NTT (Number Theoretic Transform) and permutation decomposition to massively reduce the area overhead of hardware interconnects.

Everything Else#

On the Dwarkesh Patel channel, What remains scarce after AGI? – Alex Imas and Phil Trammell offers a rigorous macroeconomic debate on whether the labor share of the economy will eventually collapse or be sustained by a purely human “relational sector” driven by empathetic consumer preferences. Meanwhile, ThePrimeagenHighlights reviews an unexpected new project in PewDiePie did it again, examining a self-hosted, privacy-first local AI interface named “Odysius” that acts as a secure local workspace for memory extraction, web-hook integration, and local agent orchestration without sending data to major tech companies