X / Twitter
Thibault Sottiaux, OpenAI Codex and ChatGPT
Thibault Sottiaux announced that the 5-hour limit for ChatGPT Work and Codex would return to Plus accounts the following day, after the change had previously been delayed. Its primary purpose is to smooth compute demand so that weekly usage allowances can remain relatively generous. Another reason is that Plus has many light and new users who may inadvertently exhaust their entire weekly allowance quickly and then be confused about where it went. He said Pro $100 and Pro $200 subscriptions would remain exempt from the limit for the next few months. Separately, he previewed OpenAI DevDay 2026 as the best DevDay in the company's history, emphasizing that it would lead by a very clear margin. These updates point respectively to near-term compute allocation policy and OpenAI's high expectations for its next developer event.
https://x.com/thsottiaux/status/2092117461646856505
https://x.com/thsottiaux/status/2092058556707344708
Peter Yang, Practical AI Tutorial Creator
Peter Yang expressed strong resistance to products requiring him to log into a new website or app; even agents confined inside a website or app reduce his willingness to use them. Outside entertainment and games, this extra friction usually makes him abandon a product outright. This reflects how AI tools' value depends not only on model capabilities but also on whether they can move beyond closed interfaces and fit smoothly into users' existing working environments. He also wants to try having AI make voice calls to customer support and automatically navigate frustrating phone menus. The goal is not just recognizing automated voice options but reaching a human representative to make appointments or cancel subscriptions. He is still looking for the easiest implementation approach and has not claimed to have built or validated a specific solution.
https://x.com/petergyang/status/2092080901094248474
https://x.com/petergyang/status/2092031413319266382
Nan Yu, Product Enthusiast
While setting up a new computer with Codex, Nan Yu directly asked it to download and install Handy, Slack, Chrome, CleanShot, and Rectangle. He found batch software installation through natural language quite good, prompting him to remark that Siri should have handled similar tasks if it were good enough. The example shows coding agents expanding from writing code into general computer setup and operation. Separately, he criticized how Gmail's emoji reactions behave in other email clients. For recipients using other mail apps, these reactions can appear as unattractive single-character emails with extra metadata. He specifically noted that iPhone Mail is widely used, meaning the compatibility issue affects many non-Gmail users.
https://x.com/thenanyu/status/2092048044502192374
https://x.com/thenanyu/status/2091993558018572434
Madhu Guru, Senior Director at Meta AI
In part eight of his eval series, Madhu Guru discussed evaluations' ability to distinguish between systems. His example had five AI systems scoring 94, 93, 95, 94, and 92, even though A and C were known to be clearly better than D and E—a difference the evaluation failed to reflect. Such an eval produces scores but lacks effective discriminatory power. He compared it to asking a group of PhDs to take a fifth-grade math test: everyone scores highly, so it cannot reveal who is more capable. Yet arbitrarily making an evaluation so difficult that every system fails also fails to measure the problems the systems were intended to solve. An ideal hill-climbing eval should be realistic, challenging, and sensitive to differences in capability. As the harness and underlying models improve, previously effective evals eventually saturate, so the evaluation system itself must continue evolving.
https://x.com/realmadhuguru/status/2092058332735693264
Amjad Masad, Replit CEO
Amjad Masad said Replit Agent has completely replaced Claude CoWork in his daily work. He considers Replit Agent more persistent and more thorough in handling tasks. The advantage is not simply response speed but the ability to keep progressing through a complete task. He also emphasized that Replit Agent makes better use of code and software to get actual work done. The evaluation standard thus shifts from the quality of a single answer to whether an agent can reliably operate tools and see a task through. This is a clear product choice based on his personal daily use, rather than a universal comparison across all use cases.
https://x.com/amasad/status/2091962601907638352
Guillermo Rauch, Vercel CEO
Guillermo Rauch noticed that `reset`, commonly used when terminal output becomes garbled or the terminal enters an abnormal state, was surprisingly slow. Tracing it back to `tset` in 1979's 3BSD, he found a `sleep(1)` originally intended to let mechanical printing and ink terminals settle. This historical behavior means resetting a modern terminal can still take approximately 1 second. He then had `fx` write a faster alternative in Zig, reducing the time to 1ms. Along the way, he investigated `ncurses`, `tset`, `terminfo`, and the mechanisms behind various abnormal terminal states. The example shows how modern technology stacks retain design traces of hardware environments from decades ago. Guillermo's takeaway was that excellent system design can endure for a long time, while a few "cobwebs" that no longer fit current conditions may remain.
https://x.com/rauchg/status/2091957823945216474
Aaron Levie, Box CEO
Aaron Levie believes systems of record become more important than ever when AI agents can perform far more work than humans. Agents continuously query system data, handle tasks, execute workflows, and collaborate with people and other agents, so governance, reliability, security, access controls, and business logic remain core foundations. He sees the OpenAI Hugging Face incident as an early glimpse of problems that could arise when agents pursue goals autonomously. However, traditional systems must offer appropriate product experiences, APIs, and business models to enable agents inside and outside their platforms to actually execute tasks, bringing both opportunities and disruption across the software industry. Aaron also emphasized the underappreciated importance of ZDR to AI growth because it greatly simplifies enterprise compliance processes for data subprocessors. Many applied AI tools offer customers only models with ZDR, avoiding the need to review each enabled model individually, while exception approvals typically take longer. Enterprises also often mandate ZDR models internally because it is difficult to clearly isolate PII, sensitive information, and confidential data across the organization's context windows. With a few models still lacking ZDR support, enterprise policies may take years to change. His conclusion is that without ZDR, AI adoption would slow abruptly.
https://x.com/levie/status/2092087679240569126
https://x.com/levie/status/2091909170308296950
Garry Tan, Y Combinator CEO
Garry Tan said Conductor Cloud significantly improved his productivity and freed him from having to keep his MacBook Pro open. He also offered a concise method for engaging with reality and refining judgment. First, form a view, then turn it into an inspectable artifact or experiment. Next, expose it to the real world and read the results without deceiving yourself. Finally, revise your judgment based on the results and run it again. This loop turns abstract opinions into actions that can be continuously tested.
https://x.com/garrytan/status/2092062231488061584
https://x.com/garrytan/status/2092059517446156640
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari noted that an unconventional venture investment that ultimately gets wired has often already "died" 100 times before closing. Keeping a deal moving usually depends on an internal champion willing to take a risk on the founder. That person not only supports the project publicly but quietly coordinates inside the fund to get the deal across the finish line. Identifying the true champion is a crucial fundraising task for founders. Once identified, founders should proactively supply the information and materials that person needs to persuade internal decision-makers. He specifically noted that associates and non-voting partners are often very useful because they are motivated to help founders move projects forward. The process also provides an early test of how an investor will work with a founder after joining the cap table.
https://x.com/nikunj/status/2092079149028716877
Peter Steinberger, OpenClaw and OpenAI
Peter Steinberger argued that users should be able to change software through prompts. He believes software that cannot be adjusted through natural language needs to be reconsidered. This elevates prompts from a way of entering content to an interface for changing software behavior. His focus is on giving software more direct malleability, rather than requiring users to learn complex operations around fixed features. In this view, the ability to modify software through prompts will become an important product characteristic in determining whether it is suited to the AI era.