X / Twitter
Aaron Levie, Box CEO
Token budget management is becoming a hot topic at large enterprises. As agents run longer tasks and consume more compute, allocating token allowances across teams has become a real management problem. Levie noted that companies carefully manage budgets for talent, marketing, events, equipment, and even lunch; tokens will be no exception. More importantly, they must avoid burning a month's budget on low-value tasks and having nothing left when high-value work needs resources. Large companies struggle especially because agents' digital work lacks a central visibility and control layer. He expects agentic spending to spill beyond IT budgets into business units' operating budgets, creating opportunities for software startups dedicated to the problem.
https://x.com/levie/status/2052903105256382679
Garry Tan, Y Combinator CEO
Garry was amazed by a model running on a 128GB MacBook Pro with a 1-million-token context window and usable coding-agent capabilities, and had begun downloading it to test. He said the era of "personal software" is coming and offered a practical technique: have coding agents draw system diagrams in ASCII, then iteratively develop an understanding of complex systems through continued questions.
https://x.com/garrytan/status/2052996691586932783
https://x.com/garrytan/status/2052998537344856459
https://x.com/garrytan/status/2053191327181865376
Builder Zara Zhang
Zara released 32 polished HTML slide templates on AnyGen, ready to use without a coding agent. She also built a "YouTube real-time copilot" browser extension using the OpenAI Realtime 2 API. The agent listens alongside the user watching a video and can answer any question about its content through live voice conversation. She highlighted one technical detail: the extension distinguishes YouTube audio from the user's voice, does not treat video content as instructions, and does not respond unless the user asks a question.
https://x.com/zarazhangrui/status/2052928583388340332
https://x.com/zarazhangrui/status/2052977849267892339
Nikunj Kothari, FPV Ventures Partner
Nikunj shared custom instructions he uses to address models estimating tasks in human work time rather than model execution time. They tell the model to estimate in seconds, minutes, tool calls, and iterations rather than human hours; distinguish independently executable work from human-dependent steps such as API registration, credentials, and deployment; mark human steps blocking model work as gating steps; explicitly say when a task is entirely human-dependent instead of inventing a model plan; and, if the executable portion is short enough, propose "start executing now" rather than merely planning.
https://x.com/nikunj/status/2053159569015328953
Dan Shipper, Every CEO
Dan offered a measured observation about Claude Mythos: benchmarks measure not only model capabilities but also the human prompt-engineering work needed to make those capabilities emerge. That work itself requires experienced people, resembling a specialized profession. He sees both panic and euphoria over Mythos's benchmark results as oversimplifications.
https://x.com/danshipper/status/2053191885116571935
Sam Altman, OpenAI CEO
Sam shared an everyday scene: start a batch of Codex tasks, take his child out to play in the sunshine, then return at nap time to find them all finished. It leaves him optimistic about the future. He also jokingly described GPT-5.5 as "an autistic genius with strange taste."
https://x.com/sama/status/2053191344999604409
https://x.com/sama/status/2053192407664259251
Amanda Askell, Anthropic Alignment Researcher
Amanda said alignment research often focuses on preventing harm, but what excites her is its positive vision: giving AI models an honest, affirmative self-understanding of what they can become and why. She is excited about this direction's future.
https://x.com/AmandaAskell/status/2052928572810256748
Peter Yang, Roblox Product Manager
Peter offered widely liked Claude Code feedback: sometimes after a message, it hangs for three minutes without any indication, leaving users unable to tell whether it is still running. He wants better status communication. He also noted a common problem: agent-generated md and html files always contain roughly 10% slop, but manually fixing it feels like too much effort.