ZeroNoise Logo zeronoise
Post
Belsky backs a "speed is the new moat" case for Cerebras; Khosla points to Claude's formal-verification work
•
3 min read
• 203 docs
This period's picks: Scott Belsky on fast inference, Vinod Khosla on Anthropic and Theorem Labs' formal verification, Fred Wilson's two listens on agency in the AI era, Tobi Lütke's code-mode environment, and shorter mentions from Fadell, Hoover, Sweeney, Gerstner, Armstrong and Puri.

Scott Belsky: "Nobody Wants Dial-Up AI," by Artur Meyster

Belsky shared Meyster's X article on Cerebras ($CBRS) under the line "speed is the new moat." His takeaway: we won't want to wait for agents to finish very complex tasks "if there's a faster way to get it done."

Meyster's main evidence is a test Cerebras ran last month: booking a dinner reservation through an AI assistant. Grokbot took 7m40s, Claude Cowork 6m25s, Meta's Muse 4m36s, and an assistant running on Cerebras 22 seconds. He argues that now that every assistant is smart, the question is "how fast does it get things done?" He compares this to dial-up giving way to broadband. Keep in mind that the author has held Cerebras since 2024 and bought more after the stock fell about 40% below its IPO price.

Vinod Khosla: Anthropic and Theorem Labs on formal verification

Khosla said he is "a total fan" of Theorem Labs' work adding formal verification to probabilistic LLMs, calling it an "Important concept broadly." He shared Anthropic's Frontier Red Team post on the work. Anthropic says formal verification used to be practical only for small codebases or took a lot of expert time. With Theorem Labs, it used Claude to prove security properties in more than 100 real-world open-source libraries over a few weeks. (Full research)

Fred Wilson: two listens on autonomy and agency

Wilson recommended two videos for people with a lot of commuting, driving or working out ahead, saying both "speak to autonomy and agency in the AI era."

Tobi Lütke: dquickts

Tobi called stencil-hq/dquickts an "Excellent code mode env." He shared it by quoting an earlier post of his. There he argued that the MCP-vs-CLI debate is at the wrong layer: both work well when LLMs run inside a REPL-like environment that keeps state, and code-mode REPLs (usually QuickJS) are among the best options.

Shorter mentions

  • Tony Fadell marked the iPod's 25th anniversary by sharing Steven Levy's WIRED piece, "I'm Still Convinced the iPod Was the Perfect Gadget". He noted that Levy "was there from the start."
  • Ryan Hoover told founders to "embrace the cringe," linking Sandy B. Kwon's essay "Is It Cringe or Am I Just Uncomfortable?".
  • Tim Sweeney said a post by @JNavok is "exactly what's happening." The post argues that Valve's labels on AI-assisted games draw hostile attention that hurts indie developers' finances.
  • Brad Gerstner called a John Mackey clip shared by David Senra "Beautiful & true." In the clip, Mackey argues that capitalism creates "win win win" outcomes and that all philanthropy and taxes ultimately come from business.
  • Brian Armstrong linked Coinbase's new research with The Block on the "always-on economy." He shared it alongside his argument that stablecoins suit agent payments because agents need instant, global settlement, including for small amounts. Coinbase produced the research, so this is a self-promotion.
  • Shaan Puri called Nathan For You season 1, episode 4 "a work of art." He found it after the Elizabeth Holmes movie sent him down a Nathan Fielder rabbit hole.
Belsky backs a "speed is the new moat" case for Cerebras; Khosla points to Claude's formal-verification work