Tech Videos

Watch First

Building Turbopuffer: Gergely Orosz (@pragmaticengineer ) × Simon Eskildsen (CEO) is the single video most worth watching. Simon Eskildsen (former Shopify infra lead) breaks down how he built a high-performance, S3-backed vector database, proving that simplicity, first-principles “napkin math”, and a raw VM with an Nginx cache can replace over-complicated, overpriced managed solutions and win over scaling companies like Cursor. This video is essential for senior engineers because it cuts through modern infrastructure hype to show how to systematically reason about database state, hardware constraints, and the cold unit economics of cloud services.

Highlights by Theme

Developer Tools & Platforms

• The GitHub channel demonstrates the gh-stack extension in How to modify and submit stacked PRs with the GitHub CLI, showing how engineers can programmatically view, modify, fold, and submit stacked pull requests, with downstream branches automatically retargeting main bottom-up upon squashing and merging.
• For IDE integration, the Visual Studio Code channel introduces native, on-device voice dictation in 🗣️ Speak to VS Code with the new Voice Dictation integration #coding #vscode #voicedictation, leveraging a local model powered by “Microsoft Foundry local” that handles dictation directly inside both the editor and inline chat frames.
• Under the Linux Foundation, J.R. Storment’s discussion in From tokenmaxxing to tokenomics for your AI agents highlights the rapid adoption of Anthropic’s Model Context Protocol (MCP) as an open-source standard for agent-to-tool integration, establishing it as a critical pillar for upcoming standard agentic frameworks.
• For a lighter look at coding capabilities, the Syntax channel in We tried to out-code Sherlock Holmes attempts to collaboratively solve an interactive escape room using Claude to reverse-engineer GraphQL endpoints and write string-deciphering scripts, demonstrating that human-driven triangulation of stop counts remains far more efficient than waiting for black-box AI code generation to parse complex visual layouts.

AI & Machine Learning

• On the Y Combinator channel, Waymo Co-CEO Dmitri Dolgov delivers a masterclass on bridging the physical-to-digital AI gap in Waymo Co-CEO Dmitri Dolgov: The Demo Is Only 1% Of The Work, explaining how they bypassed high-consequence error rates and physical latency using a “structure-augmented end-to-end” foundation model that leverages visual language models for slow-path semantic reasoning while relying on raw fused active sensing (cameras, lidars, and radars) for split-second instincts.
• Pragmatic builders will also find interest in Bloomberg Tech’s coverage of Alibaba’s Qwen 3.8 Max in Alibaba’s New Model, Amazon’s AI Win and Apple’s Next Chapter | Bloomberg Tech 8/3/2026, a 2.4-trillion-parameter open weights model that claims par with top-tier US proprietary labs on complex agentic computer use and research reproduction benchmarks, pricing input tokens at just $2 per million.
• Regarding the hard security challenges of autonomous systems, NVIDIA’s Securely Deploy Autonomous Agents in Your AI Factory | AI Factory Insider Ep. 3 outlines a reference architecture using “OpenShell” — an open-source sandboxing environment that secures agents by enforcing policy and tools outside the sandbox to prevent models from self-modifying their permissions. The talk also highlights how confidential computing is being deployed on GPUs to protect sensitive customer data and model weights during runtime execution.
• On macroeconomics, Dwarkesh Patel’s Why smarter AI models could drive up compute prices 10x offers a highly analytical look at why AI pricing won’t fall for advanced models; he argues that because smarter models can monetize the same unit of compute far more effectively, premium H100/GB200 instances will stay extremely scarce and experience massive price hikes, forcing lower-value applications out of the market entirely.

Hardware & Infrastructure

• In AWS CEO Says AI Business Is ‘Just Massive’ on Bloomberg Tech, AWS CEO Matt Garman details how cloud Capex will exceed $220 billion as their massive $25 billion run-rate AI business continues to shift from training to inference, with their custom Trainium 3 chips seeing heavy demand and selling out through 2027 due to 20-30% cost savings over Nvidia GPUs.
• Addressing the immense energy constraints of this compute scale, Bloomberg Tech’s Valar Atomics Raises $1 Billion to Power the AI Era features CEO Isaiah Taylor discussing how they secured a massive Series B round at a $6 billion valuation to build vertically integrated nuclear reactors, which recently achieved a technical milestone by directly powering an Nvidia Blackwell AI chip on a dirt lot.
• This power distribution bottleneck is further analyzed in The Startup Turning Home Batteries Into Grid Infrastructure, where Base Power CEO Zach Dell explains their business model of deploying 3x oversized, company-owned home batteries to serve as distributed grid infrastructure, helping utilities increase grid utilization and buffer massive peak demands driven by electrification and AI data centers.
• Finally, Simon Eskildsen of Turbopuffer highlights a looming bottleneck: as laboratories ramp up massive reinforcement learning (RL) clusters to train models to search and write code, the demand is triggering a quiet but severe global CPU and NVMe SSD shortage, with hyperscalers actively fighting over server allocations.

Everything Else

• In Product managers exist because of scale. They are by no means a neccessity. (Lenny’s Podcast), the host and guest argue that product management is a modern trade created to manage sheer organizational scale, warning that over-abstracting engineering and design from direct execution severely atrophies their decision-making muscles.
• Highlighting tech culture, Life at Google’s Welcome to Sail Tower: Working at Google’s newest Austin office showcases their new Texas sail-shaped facility featuring regional-themed floors, designed as part of Google’s $40 billion Texas infrastructure investment and community-oriented AI program support.
• For historical perspective, a deep series of interviews on the Lex Clips channel with historian Gary Gallagher covers the brutal, uncodified reality of Civil War guerrilla warfare in Missouri, post-abolition racial and anti-immigrant prejudices, and the sharp difference in overt religious sentiment between Confederate leaders like Lee and Union generals like Grant.

📈 I can set up a custom performance-to-cost matrix comparison for the newly released open weights models if you want to see exactly how their token economics map to standard cloud instances.