Box CEO Aaron Levie
Aaron Levie argues that long-running agents can perform inference in data centers farther from users. A task that produces $500 in value and takes one extra second illustrates how the value of the work can change its latency tolerance. He envisages encrypted transport and discarding data after the task, giving inference operators more choice about location. Actual residency requirements depend on the system and applicable rules. He also says the right harness, domain expertise, parallel agent work, and low-friction testing and deployment could lift output on some coding tasks by one or two orders of magnitude. He attributes the leverage to combining existing expertise with new tools. The posts describe his observations and expectations, whose applicability depends on the task and environment.
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux says in his post that Pro tiers have no five-hour usage limits. He gives usage multiples relative to Plus: 10× for Pro 100, 20× for Pro 200, and 50× plus ultrafast access for Pro 500. He separately asks whether people would want a Pro 50 tier. The two posts describe stated allowances and solicit feedback on a possible new tier, respectively.
SPC general partner Aditya Agarwal
Aditya Agarwal argues that if new data centers become harder to build, hyperscalers with existing compute capacity gain an advantage. His premise is that demand will continue to rise, making installed capacity scarcer. He expects new-model training to slow and sees policy and local decisions indirectly shaping the pace of frontier development. The link he draws between construction constraints, existing capacity, and training speed is his industry judgment.
Builder Zara Zhang
Zara Zhang pushes back on predicting a startup’s fate from its initial idea alone. She points to familiar objections: an idea will never work, or its market is too narrow. She argues that there are many ways to execute the same idea and that different founders will take different paths. Strong execution can create a market that previously did not exist and make the outcome look obvious in hindsight.
FPV Ventures partner Nikunj Kothari
Nikunj Kothari announces a noon session in San Francisco on November 11 with Arcanum Labs co-founder Daniel Rupawalla. Topics include what data model labs buy, who supplies it, and whether the market can last. The 60-minute in-person session gives Kothari 45 minutes to interview his guest and the guest 15 minutes to ask about venture capital, with 10 to 12 observer spots and the Chatham House Rule. A separate post explicitly labeled satire portrays founders using charts without a vertical scale, feature-launch videos, and paid coverage to stoke fundraising interest, then claiming an early offer explains the missing data room. He says individual elements have real-life counterparts. The details illustrate the promotional tactics he is criticizing.
Andrej Karpathy
Andrej Karpathy turned Tristan’s list of YouTube channels into TristanTube. He says the site brings together about 47,000 videos and offers a TV mode for lean-back viewing. He reports building it with Claude Code in a single attempt in three minutes. The timing describes his experience with this build.
Peter Steinberger of OpenClaw
Peter Steinberger starts from the Asahi project’s goal of opening a closed system and criticizes new restrictions that make it harder for other projects to build on its work. He also objects to its rejection of AI. His concern is how much room an open project leaves downstream developers. The post states his position, with the project’s own text needed for the specific terms.
No Priors — Tokenization, Agentic Trading, Harmonic: Robinhood's Mission to Expand Ownership with CEO Vlad Tenev
Key takeaway: Vlad Tenev wants to put the tools for trade execution and mathematical proof in individuals’ hands. The two product lines face different design questions about authority over funds and verification of results.
Robinhood co-founder and CEO Vlad Tenev also co-founded Harmonic, a mathematical AI company. He recalls experiencing inflation as a child in Bulgaria and his grandfather preserving purchasing power in copper cookware. He uses that family experience to explain why ownership matters to him.
Harmonic combines model generation with verification in Lean. Tenev says Aristotle reached International Mathematical Olympiad gold-medal level and that people tried it on Erdős problems after release. He wants individuals and researchers to choose problems and make discoveries with the tools, and argues that labs should actively collaborate with mathematicians. He describes concentrating discoveries inside labs this way: “It’s too centralized, right?”
Robinhood’s Agentic Trading began with an MCP integration, then moved the harness and interface into Robinhood Agents. Tenev reports over 150,000 customers opening related accounts and a previously announced figure of roughly 30 million daily tool calls. Tool calls include actions such as queries. Order records are needed to count executed trades. He says users must fund a separate account, trade approval is on by default, and users can turn it off. He expects early growth to concentrate among a small number of experienced users, with exchange and market-maker infrastructure evaluated against actual load as adoption grows.
He uses stablecoins to explain asset tokenization: a traditional institution holds the underlying assets and on-chain tokens can compose with smart contracts. He describes experiments pairing stock tokens with other assets on Robinhood Chain and wants to extend the approach to more markets. For private-company exposure, he describes publicly traded funds that connect individuals to private assets, distinguishing later-stage Fund I from Fund II for seed and Series A investments. The product details, scale and outlook are the operator’s account. Asset rights, geographic availability and account permissions depend on the applicable terms.