AI Builders
Digest

The Takeaway: Traditional benchmarks fail to capture modern AI capabilities because model performance now heavily depends on test-time compute budgets rather than fixed parameters. OpenAI researcher Noam Brown, a pioneer in AI reasoning, explains that today's models can scale thinking time dramatically—from seconds to weeks—making capability a function of resources invested at inference. This shifts evaluation fundamentally: labs should plot performance against compute budgets (tokens, dollars, or time) or set explicit limits instead of single-point scores. Brown highlights how GPT-5.5 shows efficiency gains over predecessors when normalized for thinking time, and warns that safety frameworks haven't adapted to this reality, potentially underestimating risks at high budgets. He shares practical experience building poker solvers, where newer models dramatically accelerate optimization and reasoning with gentle guidance. A memorable insight: "The capability of the model is a function of how much money you put into it." Brown urges the community to break the bad equilibrium of misleading benchmark grids for more honest progress measurement.

OpenAI CEO Sam Altman highlighted the latest update to the GPT-5.5 instant model in ChatGPT, noting positive user experiences with its 'vibes.' He celebrated the team behind GPT-5.6 Sol, describing it as strongly performing especially for knowledge work involving heavy tool use and long-running agents, signaling continued rapid progress without hitting walls. On usage, Altman mentioned ongoing work toward more generous token access. These posts underscore OpenAI's momentum in pushing frontier capabilities.

Box CEO Aaron Levie assessed GPT-5.6 as real and very strong, particularly excelling in knowledge worker tasks requiring heavy tool use and long-running agents. He emphasized that AI progress shows no signs of slowing.

Every CEO Dan Shipper broke news on GPT-5.6 Sol access being temporarily restricted by U.S. government directive to select companies. He supports oversight for security but stresses the need for broad democratic access to maintain U.S. AI leadership and empower workers, builders, and students. Every aims to prepare users for effective adoption once access expands.

Vercel CEO Guillermo Rauch discussed the challenges of debugging AI agents due to non-determinism and distributed complexity across APIs. Prioritizing observability for agents on Vercel was key, receiving strong feedback. He also praised shadcn as the UI for AI.

OpenAI's Thibault Sottiaux announced a usage reset for all Codex users and shared mitigations after investigating potential issues, with no widespread user impact detected.

Anthropic's Cat Wu praised split screen as one of her favorite Claude Code desktop features for productivity.

Swyx celebrated taking over a new media lab in San Francisco, designed as a hub for engineer-creatives, complete with a surprise datacenter rack. He seeks advice on equipping it.

Peter Yang observed money moving toward services over pure software, questioning pure-play software value versus using tools like Codex/Claude with agents. He raised concerns about frontier model distillation, open source adoption, and potential innovation impacts from access gating.

Y Combinator CEO Garry Tan criticized certain model release approaches as harmful to small startup innovation, calling it a way to 'salt the ground.'

Investors and builders like Nikunj Kothari emphasized that taste comes from repeated iteration in the arena, not spectatorship. AI has potential to develop taste through variation and learning. Aditya Agarwal noted AI increasing demand for deep human connections while delegating shallow tasks to agents.