Nvidia: what investors told people to watch and read

Chamath Palihapitiya promoted the new All-In episode with just "pod!!!!!!" @chamath. The Besties take apart Jensen Huang's exit from Nvidia. Topics include the spy allegations and who leaked the memo, the market impact and the AI trade, China hawks and export controls, the effect on open source, and who should be Nvidia's next CEO. A final segment covers Anthropic's IPO road show @theallinpod. Episode

People were also asked to read and sign letters, not just read them:

• Amjad Masad: "Read the Stand with Jensen letter and consider adding your name" @amasad. Letter
• Garry Tan signed the same campaign: "Jensen deserves due process, not whatever is happening right now" @garrytan.
• Brad Gerstner (Altimeter) shared a letter sent to Nvidia's board. He said it sets out "the minimum acceptable standards the Board owes shareholders" during the crisis @altcap. Post

Jason Lemkin shared a new 20VC x SaaStr takeaways piece, "Jensen Resigns, Sequoia Pulls Back to the Bay, and Anthropic's IPO Draws 6x Demand." He pulled out three points: Anthropic's book is 6x oversubscribed, Sequoia is winding down evergreen, and "the AI trade only cracks if demand cracks" @jasonlk. Article

Garry Tan: his own harness-eval repo

Tan shared his open-source g-harness-evals repo "and what it found" @garrytan. A Chinese-language summary he quoted lays out the results. The test held the model and tasks fixed and swapped only the agent harness: Claude Code, Codex, OpenClaw, Hermes Agent and QM. On 20 tasks, pass rates bunched between 71% and 79%, and the confidence intervals overlapped heavily. Cost per solved task varied much more: $0.44 for OpenClaw, $0.22 for Claude Code and $0.17 for Codex @mylifcc.

The repo also tested Tan's own gstack skill pack. The skill text loaded every time, but the model ran the skill's steps 0 times in 46 runs, and those runs cost twice as much @mylifcc. Tan discloses that QM is a YC company, YC invested in Hermes, and gstack and gbrain are his personal projects @mylifcc. Repo

Sarah Guo: an inference engine built from scratch

Guo shared @saltyph's project with the comment: "the future is generated, continually learning inference engines in open source!" @saranormous. It's an inference engine for DeepSeek V4.1-Flash written in plain Rust and CUDA, with no PyTorch or SGLang. The author reports about 1,740 tokens/sec decode on 4×B300 and a 1M-token prefill in about 12 seconds. The code is on GitHub @saltyph. Post

Eric Schmidt: Genesis

At a CHM Live event marking SRI's 80th anniversary, the interviewer asked whether Genesis went out of date before it was published. Schmidt co-wrote the book with Henry Kissinger Computer History Museum. His answer: "Henry laid out the next 20 years in a way that no one else did." He said Kissinger saw the arrival of a nonhuman intelligence as an epochal change, one that raises questions of coexistence, what it means to be human, and alignment, and that this was "all written before the common understanding" Computer History Museum. Video

Shorter picks

• Tim Sweeney pointed to his GamesBeat interview with Dean Takahashi, which also features Epic EVP Marcus Wassmer. It covers "breakthroughs in game development efficiency" and the future of engines and tools @TimSweeneyEpic +1. Interview
• Jack Dorsey pointed to @Hello_World's "Buzz Plugins Megathread," calling plugins "huge, and where the magic is" @jack. The thread presents Buzz as a "network harness" in which each team controls its own workflows, tools and rules without waiting on a roadmap @Hello_World. Thread
• Paul Graham quoted Joshua Reynolds on self-directed learning: "Few have been taught to any purpose who have not been their own teachers" @paulg.