Zara Zhang
Zara Zhang explained why publishing regularly does not need a monetization rationale. She sees self-expression as an innate need and influence as valuable beyond income. In another reflection, she noted how dazzling new technology can obscure the ways it fails. Users may even blame themselves before questioning the technology. Her focus is retaining personal motives for expression and independent judgment about tools.
Aaron Levie
Aaron Levie argues that agents making efficient choices for users will reshape markets built around inertia and switching friction. Companies benefiting from customers’ reluctance to change will face lower switching costs and stronger competition. At the same time, agents can make transactions that were previously too cumbersome easier to complete. He cited healthcare, travel, local services, and certain information services. The result could be both pressure on existing profits and new economic activity.
Peter Steinberger
Peter Steinberger is considering ways to reduce continuous-integration load. He thanked Blacksmith for its support but said the load needs to be distributed. Part of his plan is to have Codex decide which tests actually need to run. Another part is to reduce CI substantially and run tests hourly. He presented this as a plan centered on test selection and execution cadence.
Thibault Sottiaux
Thibault Sottiaux believes code freezes before releases are becoming less relevant. He went further, imagining code generated online for each request. Generation would be governed by a set of constraints. The idea shifts code from a fixed release artifact toward a result created on demand. He presented a future direction, without describing an implementation already in production.
Peter Yang
Peter Yang compared how different types of software remain in users’ memories. Few people fondly remember old SaaS or business software. Favorite games, by contrast, often stay with people for years. He considers games the software with the longest staying power in memory. He joked that LLM models may be the opposite.
Matt Turck
Matt Turck noted a difference between how AI researchers and outside commentators express their views. He sees many researchers who understand the technology rejecting both inevitable doom and massive uncontrollable acceleration. Some people with less technical knowledge, by contrast, are highly certain in either direction. His observation is that confidence and depth of knowledge do not always move together. It is a reflection on the current AI debate.
Thariq
Thariq worries that people will become lazier as agents become more capable. Changes in human behavior could then absorb the productivity gains from automation. He places the concern on the user’s side. Greater capability does not automatically translate into more useful output. His reflection raises the question of where the saved effort ultimately goes.
Guillermo Rauch
Guillermo Rauch moved his Mini browser to Rust and Swift, aiming for a Safari-like experience with Chromium. The cef crate lets him bundle up-to-date Chromium; fx acp connects to the local fx CLI over ACP, and an MCP server lets fx manage the browser. He reported faster startup, Liquid Glass, and more natural Mac focus and input controls. He built it with Opus 5.5 and Sol 6. The experience reinforced his expectation of more native desktop and cloud software, a direction he also associates with Vercel Fluid.
Training Data | Box's Aaron Levie: On Reinventing Yourself in the AI Age and Enterprise Diffusion
Core takeaway: Enterprise AI opportunities come from the concrete work of connecting models to real workflows, data, and permissions.
Box founder and CEO Aaron Levie began with cloud file storage and collaboration and has spent years working with unstructured enterprise data. Box now uses agents to query files, extract metadata, and execute workflows. He argues that stronger foundation models do not automatically solve deployment problems in banks, law firms, or pharmaceutical companies.
First, coding and other knowledge work diffuse at different speeds. Engineers often have data in GitHub and can fix connection problems themselves. Sales still depends on customer replies, budgets, and meeting schedules. Enterprise information is scattered across legacy systems and file shares, with permissions changing by project. Deployment and organizational change take time, creating opportunities for application companies.
Second, a harness must demonstrate value through local evaluations. Box combines search, reranking, text and chunk extraction, and full-document reading when needed. It evaluates models on industry document tasks and real internal usage. A customer may require 98% extraction accuracy at a particular cost; the team tests several models against that customer’s data. General leaderboards cannot replace this process.
Third, proprietary agents and open interfaces both matter. SaaS vendors should make their agents materially better at using their products while exposing APIs or MCP access to external agents. Levie cited continuously checking documents for archiving and legal-hold requirements, work that humans could not previously cover comprehensively.
Fourth, enterprise interfaces will still need dashboards, queues, and task lists. He expects much work to happen in the background, with users reviewing results. His productivity test is direct: “did we actually deliver more functionality that customers are asking for?” Box examines high-usage employees’ practices, distinguishes waste from transferable techniques, and organizes demonstrations for colleagues.