Tech Videos — 2026-05-26#

Watch First#

Frontier AI at Home — Alex Cheema, EXO Labs Alex Cheema cuts through the AI hype to focus purely on local hardware inference, explaining the memory-bandwidth bottlenecks of auto-regressive decoding and demonstrating how to cluster Apple Silicon and RTX GPUs using Thunderbolt 5 RDMA to run 1-trillion parameter models locally.

Highlights by Theme#

Developer Tools & Platforms#

Google’s I/O Keynote leaned heavily into “agents building agents” marketing with their Anti-gravity 2.0 orchestration platform, showing off voice-driven LoRA fine-tuning and automated Android migrations. For a more grounded take on the Model Context Protocol (MCP), the Visual Studio Code channel demonstrates how to explicitly scope and sandbox MCP toolsets so agents do not burn through context windows or execute arbitrary terminal commands unchecked. Over on Fireship, the channel highlights some excellent grassroots engineering like CUDA Oxide, which allows developers to write GPU kernels in pure Rust compiled straight to PTX, entirely sidestepping legacy C++ memory footguns.

AI & Machine Learning#

In an excellent pragmatic talk at AI Engineer, Brandon Walsenuk from Unblocked explains why naive RAG fails for AI agents due to “satisfaction of search,” where agents lazily stop looking as soon as they find a single partial answer. He argues that a functional context engine actually requires a social graph of your engineering team to properly resolve codebase conflicts and truthiness. On the research side, Microsoft Research presents Matching features, not tokens: Energy-based fine-tuning of language models, offering a rigorous alternative to SFT and RLHF that fixes compounding distribution shifts in long-sequence generation by enforcing alignment in a learned feature space. Finally, Marc Andreessen is friends with AI Vampires features ThePrimeagen rightly mocking VC claims that AI is creating “vampire” developers operating at 20x productivity, noting that the real productivity gains mostly come from bypassing slow corporate bureaucracy rather than superhuman output.

Hardware & Infrastructure#

Continuing the themes from the EXO Labs talk, Cheema highlights the massive latency costs of tensor parallelism over standard networks and shows how utilizing RDMA drops inter-node communication from 300 microseconds down to single digits. For a lower-level look at compute constraints, The “Trick” that Compilers Use for Long Division - Computerphile details how modern compilers avoid the 30-100 clock cycle penalty of hardware division. The compiler optimizes division out of your code entirely, converting it into a mathematically equivalent but vastly faster sequence of multiplications and binary bit-shifts.

Everything Else#

In a sharp critique of tech leadership culture, Chamath Rips Cloudflare CEO’s Layoff Memo: “Shut the f**k up. You suck at this.” excoriates Matthew Prince for a PR-disaster memo that callously labeled laid-off middle managers as unnecessary “measurers”. In consumer tech, MKBHD reviews the new screenless Fitbit Air in The Truth About the “Whoop Killer”, observing that Google is aggressively undercutting Whoop’s predatory $200/year subscription model simply to harvest as much health data as possible.