X / Twitter
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux is preparing a keynote with Romain Huet and Sam Altman. He said a large part of the preparation has involved figuring out how to explain the new developments to the public, because there is so much they plan to show in succession. The team will roll out some of it next week so users do not have to wait until the later, consolidated presentation. He also previewed more new features under development, and how they fit together, to be shown over the coming months. This teaser has not yet disclosed specific product names, feature details, or firm release dates.
https://x.com/thsottiaux/status/2101157729037586694
Practical AI tutorial and interview creator Peter Yang
Peter Yang called Meta’s Muse the best personal agent he has tried so far, and said it has already helped him save more than $800 a year on cable and phone bills. In one specific example, Muse negotiated with Comcast customer service over the phone and secured an annual bill reduction of $288. He shared the call transcript and argued that most companies’ customer service hotlines are not yet prepared to deal with agents. His video introduced his 10 favorite uses for Muse, including personalized morning news, habit tracking, and calling customer service on users’ behalf to negotiate bills. In his view, a free agent that delivers measurable bill savings offers compelling product value. He further suggested that Muse could become Meta’s next app with a billion users, though this is his personal assessment of the product’s prospects.
https://x.com/petergyang/status/2101083891507593576
https://x.com/petergyang/status/2101033599319613533
Anthropic Claude Code team member Thariq
Claude Code supports AGENTS.md starting with version 2.1.277. When a folder has no CLAUDE.md, Claude checks for and uses the AGENTS.md in that folder. This reading rule has an explicit condition; it does not unconditionally use both kinds of project instruction files together. Users can toggle this behavior in /config. Thariq explained that AGENTS.md support is built on Claude Code’s upcoming mods mechanism, which is used to customize Claude Code’s runtime framework. The support is currently a built-in mod, and in the future users will also be able to build their own project instruction mechanisms to suit their needs.
https://x.com/trq212/status/2101009392611278961
https://x.com/trq212/status/2101009393731223817
Vercel CEO Guillermo Rauch
Guillermo Rauch shared that open models’ share of token usage on Vercel AI Gateway that day may have reached a new high of 78.4%, with closed models accounting for 21.6%. He noted that the distribution typically looks different when measured by spending. Moonshot AI and DeepSeek ranked 3rd and 4th in spending that day, respectively; with Z.ai added, the three together exceeded OpenAI, which ranked 2nd. Spending here refers to the cost of model inference through various providers, primarily based in the United States, and is not equivalent to revenue flowing directly to open-weight model labs. In another post, he said Jev’s adoption had surprised him both in the data and in individual use cases. He attributed some of this enthusiasm to developers’ widespread view that AI is too expensive and too slow, creating an urgent desire to improve costs and speed so they can use AI in more places.
https://x.com/rauchg/status/2101186741042663579
https://x.com/rauchg/status/2101079472732848510
Box CEO Aaron Levie
Aaron Levie used a Box and Jev demo to illustrate how fast, low-cost models can make immediate judgments within enterprise workflows. The demo first reads an incident report from Box, then determines whether the incident is customer-facing and how severe it is. Based on the results, the system moves the file into an escalate, monitor, or review folder, and writes its assessment into a metadata template instance. He said the entire process completed almost instantly at close to zero cost. In his view, these capabilities are well suited to agents performing data classification, workflow routing, and the many tasks that require rapid judgments. The potential applications he listed included insurance claims, contract management, loan processing, security reviews, and customer log analysis. These are his projections for possible uses, rather than scenarios that were each individually validated in the demo.
https://x.com/levie/status/2101007708044574906
FPV Ventures Partner Nikunj Kothari
Nikunj Kothari created a website called Jevable that brings together demo projects built with Jev and shared on X. He presented it as a starting point for finding inspiration for Jev development, making it easy for developers to browse other people’s experiments. The site lets visitors filter demos by several categories to narrow their browsing. Developers can also submit their own projects using the “+” button on the page. After the introductory post, he published a correction mentioning typesafeai to fix an account tag in the earlier post.
https://x.com/nikunj/status/2101077053567332618
https://x.com/nikunj/status/2101101305125290367
OpenClaw and OpenAI member Peter Steinberger
Peter Steinberger shared how his team uses roboclaw: it runs the team’s server and is connected to Discord. roboclaw can interact with gpt-live and understands the sessions it is coordinating. During team meetings, people can ask it directly for context from current and past sessions. This makes session information immediately accessible while meetings are underway. In a separate follow-up, he said that computer use capability CUA also works in the environments referred to earlier, so agents do not have to rely solely on screenshots, but the supplied material did not specify the scope of those environments.