Tech Videos — 2026-07-31#

Watch First#

What’s Next After RLHF? — Diogo Almeida, TypeSafe AI is a provocative must-watch from an OpenAI alum who argues RLHF fundamentally optimizes for human preference and “assistance,” falling short of the robust, verifiable execution required for true software “automation”.

Highlights by Theme#

Developer Tools & Platforms#

The AI Engineer channel features fighting slop with slop — Vaibhav Gupta, Boundary, an excellent talk on using BAML to enforce type-safety and determinism in LLM pipelines, effectively letting agents build a partial C compiler while bypassing flaky, regex-based parsing. For IDE integration, Slack introduced The Official Slack MCP Plugin | Slack, allowing IDEs like Cursor to directly interact with Slack canvases and messages via the Model Context Protocol. Additionally, Google for Developers posted Voice Agent observability with LangSmith, addressing the tricky engineering problem of tracing speech-to-speech models like Gemini Live, specifically around accurately logging interrupted audio outputs.

AI & Machine Learning#

The shift toward reasoning and reinforcement learning dominated recent discussions. In Data Quality Is the Compute Multiplier — Ari Morcos, DatologyAI, DatologyAI shows how rigorous data curation and synthetic “rephrasing” bends scaling laws, achieving Qwen 3.5 4B performance with 145x less training compute. In the robotics space, Google released Introducing Gemini Robotics 2, utilizing Vision-Language-Action (VLA) foundation models to tackle dexterous manipulation and whole-body spatial reasoning. Reinforcing this architectural shift, Arcee AI argues in The Base Model Is Dead — Varun Singh, Arcee AI that pre-training is no longer about accumulating world knowledge, but purely about building an atomic skill prior to survive massive RL post-training stages.

Hardware & Infrastructure#

On the silicon front, AWS is reaping the benefits of its custom hardware strategy. According to Amazon’s AI Story Is a ‘Game Changer,’ Says Mizuho, Amazon’s in-house AI chips hit a $25 billion run rate, improving cloud margins by bypassing Nvidia’s premium. Meanwhile, the All-In Podcast in Chip Stocks Crash, $20B Fund Margin Called, Frontier Labs: SLOW DOWN AI, Mamdani’s Grocery Stores breaks down the recent 20% crash in semiconductor indices, attributing it to forced liquidations of highly-leveraged AI funds and retail margin calls in South Korea rather than failing compute fundamentals. Finally, Bloomberg reports in Moonshot’s Kimi Built With Nvidia Compute that Chinese AI startup Moonshot trained its models using 20,000 Nvidia H200 chips supplied by Alibaba, exploiting loopholes in U.S. export controls by accessing compute in Southeast Asia.

Everything Else#

For a macro perspective on the startup ecosystem, Y Combinator hosted Patrick Collison: “What If You Succeed?”. The Stripe CEO pushed back against the narrative that AI will inevitably centralize market power, noting that new business formation on Stripe has doubled year-over-year and median time-to-revenue is accelerating, pointing toward a more decentralized technical economy.