X / Twitter
Thibault Sottiaux, OpenAI Codex and ChatGPT Team
OpenAI launched Images 2.5, GPT-Live-1, Agents API, Data Agent, and ChatGPT for Financial Services in a concentrated series of releases this week, with more planned for the following week. The Astra team also addressed a batch of user-reported quality issues, including skills from older models triggering too frequently and some skills preventing models from checking their own work. An optional context management experiment that could cause tasks to stop prematurely or respond incorrectly to old messages has been disabled; the team roughly estimated that 4,000 to 5,000 users were affected. Some misconfigured engines that reduced quality for long-tail traffic were also removed. The improvements aim to make task follow-through, tracking of the latest messages, and checks during execution more reliable. OpenAI has also brought in Aidan and Sasha from the Git AI team and plans to keep open-sourcing and investing in the tool, which analyzes how coding agents contribute to codebases, helping enterprises measure Codex’s actual value to individual and team workflows.
https://x.com/thsottiaux/status/2098639827084480864
https://x.com/thsottiaux/status/2098612714704891959
https://x.com/thsottiaux/status/2098569976143806918
Peter Yang, AI Tutorial Creator
Peter Yang made a clear distinction between automation tasks based on where they run: all local scheduled tasks go in Codex, while cloud tasks are gradually moving to Grok Bot. He believes this division reduces confusion around tool boundaries and makes task ownership clearer. Meanwhile, he is skeptical of the “software factories” concept, arguing that beyond verification and testing, today’s AI still cannot continuously improve products or develop new features end to end without human involvement. When running overnight agents in practice, a single incorrect assumption early on can send all subsequent work off course and waste large amounts of tokens. He therefore asked which products or features on the market had actually been completed by a software factory with no human defining requirements or reviewing results at any point.
https://x.com/petergyang/status/2098614492066435228
https://x.com/petergyang/status/2098565668241334366
Madhu Guru, Senior Director at Meta AI
Madhu Guru attributed enterprise AI project failures to three common problems. First, companies retain the organizational approach used for incremental product development over the past fifteen years: the CEO appoints a trusted associate to form a central AI team, overlooking the experimentation, invention, and different team capabilities that AI products require. Second, companies underestimate the importance of evals and fail to treat evaluation systems as a first-class part of product development. Third, central teams build a unified platform from the periphery of the business, leaving the product disconnected from the actual workflows, context, and judgment of finance, sales, or customer service staff. The result is reluctant adoption rather than substantial productivity gains. He recommended hiring leaders who have successfully built AI products and can also earn the trust of business staff, and embedding the best AI builders directly in the functions being transformed to build solutions alongside their users.
https://x.com/realmadhuguru/status/2098448235048378456
Thariq, Anthropic Claude Code Team
Claude Code introduced plugin evals to check whether skills still work correctly after new model releases. Developers can enter a plugin directory and run `claude plugin eval init` to initialize evaluations for the plugin. Thariq also cautioned that looking only at pass or fail scores makes it difficult to interpret today’s model evaluation results correctly. He has observed many benchmark failures caused by overly strict hidden tests rather than a lack of merit in the model’s answer. In some cases, the model’s answer was even more appropriate to the problem than the predefined expected result. Evaluations therefore need to account for failure examples and judging criteria rather than merely chase aggregate scores.
https://x.com/trq212/status/2098531560643539440
https://x.com/trq212/status/2098490139798655427
Amjad Masad, Replit CEO
Replit acquired a company built entirely on Replit. Amjad Masad saw it as the start of this kind of transaction and expected more similar cases to follow. The event demonstrates the full cycle created as the platform expands from a development tool into startup infrastructure: a company can be built, operated, and commercialized on Replit, eventually becoming an acquisition target for Replit itself. He also noticed users conducting numerous Astra 3D experiments on Replit, showing how quickly new model capabilities are entering practical creative environments. Together, the two updates point to Replit’s positioning as a platform that both hosts experiments with AI-native products and may participate in their subsequent commercial consolidation.
https://x.com/amasad/status/2098548464452055437
https://x.com/amasad/status/2098316298820219058
Guillermo Rauch, Vercel CEO
Tailscale’s model router uses Vercel AI Gateway as its underlying infrastructure. Guillermo Rauch compared AI Gateway to a new generation of CDN, emphasizing that it is becoming a general infrastructure layer between model traffic and applications. Applications can connect directly to model providers, but this direct-to-origin approach is vulnerable to differences in interfaces, reliability, and providers. Teams can also build their own gateways, though development, maintenance, and cost control become more cumbersome. Vercel’s view is that as calling multiple models becomes routine, unified routing and infrastructure abstraction will gradually become default components, much like CDNs.
https://x.com/rauchg/status/2098531157230969062
Aaron Levie, Box CEO
Box can now be mounted in an agent sandbox, making it easier for agents to read and write files on their computers. This capability connects enterprise content systems directly to an agent’s execution environment, going beyond question answering or retrieval. Aaron Levie argued that as AI agents begin carrying out critical enterprise processes, they need the same foundational capabilities human employees have long relied on. File access is one such core primitive, since many business activities ultimately involve reading materials, generating content, and saving results. This Box integration therefore targets the transition from agents understanding enterprise content to directly working with it.
https://x.com/levie/status/2098478938003841123
Ryo Lu, Software Product Designer
Cursor has launched long-lived agents for large creative projects. Compared with agents that finish short tasks in one go, these agents emphasize sustained progress toward complex goals over longer periods. They are suited to software development that requires multiple rounds of execution, continuity of context, and periodic adjustments. Ryo Lu positioned the capability as serving “big ideas,” reflecting Cursor’s expansion of agents from code completion and immediate edits into long-term project collaboration. The real product test will be whether agents can reliably retain goals, handle changes, and keep users informed about execution throughout long runs.
https://x.com/ryolu_/status/2098324260867772806
Zara Zhang, Builder
Zara Zhang argued that the concept of a “one-person company” is overrated. AI certainly allows one person to do more work, but creating a new product can still be an extremely lonely experience. Founders need not only execution capabilities but also partners with whom they can discuss ideas, endure setbacks, and celebrate progress. It is easy to lose motivation without a long-term commitment shared with others. Her view reminds AI builders that productivity gains have not eliminated the value of a team’s emotional support, mutual calibration, and shared perseverance.
https://x.com/zarazhangrui/status/2098483800456179923
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari observed that credit for successful investments often becomes contested in venture capital. Large exits and substantial valuation increases are rare, and these cases are the most important elements of a VC’s track record when raising the next fund, so being credited for a successful investment carries direct financial interests. Over the next few years, some people’s names may be removed from investment histories or selectively ignored, while more people will attempt to retell the history and claim credit. This problem is particularly acute for emerging GPs, whose past investment records are their key assets when pitching future LPs, but who may not receive the recognition they actually deserve. He advised these investors to maintain strong relationships with founders, who may ultimately provide the most credible reference checks.
https://x.com/nikunj/status/2098550718923997430
Peter Steinberger, OpenClaw and OpenAI Builder
Peter Steinberger demonstrated an experiment in which Astra played Doom through CUA in an OpenClaw cloud session. The demo combined a model, a persistent runtime environment, and computer use capabilities in an observable task. The model had to continuously understand the game visuals and take its next action through the interface, making it closer to a real-time agent test than a single round of text generation. He explicitly said this was not AGI, but described its current performance as “possibly better than a fly’s brain.” The value of the experiment lies less in the game score itself than in showing how a general-purpose model can use CUA to act continuously in a dynamic visual environment.
https://x.com/steipete/status/2098527519213604889
Dan Shipper, Every CEO
Dan Shipper argued that higher benchmark scores do not adequately explain how a model performs in users’ real work. For the past three years, Every has written extensive vibe checks of new models through hands-on testing, directly observing their ability to handle the team’s actual tasks. The team is now further quantifying this approach, which has largely relied on experiential judgment, by building an internal platform that lets everyone create personal benchmarks based on their daily work. This direction preserves the relevance of real tasks while adding a more consistent basis for comparisons across models and versions. Every is also recruiting writers who can work with the latest models, tools, and teams, showing that model evaluation and in-depth content remain core areas of investment.
https://x.com/danshipper/status/2098481799047647715
https://x.com/danshipper/status/2098476917448118321
Podcasts
No Priors — Coinbase’s Everything Exchange: Agentic Finance, Stablecoins, and Tokenization with CEO Brian Armstrong
Key takeaway: To become genuine economic participants, AI agents need more than reasoning and execution capabilities. They also need financial accounts and learning mechanisms that let them independently hold funds, pay for services, and accumulate organizational knowledge from each human correction.
Brian Armstrong is the co-founder and CEO of Coinbase and a co-founder of the anti-aging and longevity company New Limit. He described Coinbase’s next phase in three directions: bringing stocks, commodities, crypto, derivatives, perpetual futures, and prediction markets together in Everything Exchange; expanding stablecoin payments; and building Agentic Finance.
The key advantage of stablecoins is the ability to complete global transfers in less than 1 second at a cost below 1 cent. Traditional card transactions typically include a fixed fee of about 30 cents plus an additional percentage fee, making 1-cent payments uneconomical. Coinbase has observed that approximately 76% of agent e-commerce transactions are below 30 cents, with common uses including buying paid data, calling specialized agents, and obtaining search or financial information. This makes micropayments more than a niche use case: they could become the basic settlement method for services called between agents.
Armstrong’s view is: “We don’t want AI to be unbanked.” Coinbase is exploring two types of accounts at the same time. One is an agent account linked to a human identity with segregated funds, allowing users to authorize an agent to trade or make payments on their behalf. The other is a self-custodial wallet used by an agent, which can hold stablecoins and even raise funds or issue tokens. Agents could also be equipped with stablecoin-backed credit cards to support traditional payment scenarios such as hotel bookings and AWS resource purchases.
X402 is a payment protocol incubated by Coinbase and handed over to the Linux Foundation, with companies including Google, Cloudflare, and AWS participating in related work. Coinbase Business allows companies to accept stablecoin payments from both humans and agents. Armstrong expects that agents may eventually outnumber humans and that the agent economy could grow larger than the human economy.
Internally, Coinbase is also building an agent harness called Toshi, along with “brains” organized by team, code repository, and individual. These record service incidents, financial controls, A/B tests, GitHub history, and the reasons PRs were accepted or rejected. When human review discovers something an agent missed, the correction must be written back to the brain rather than merely fixing the current PR. The goal is for subsequent agents to inherit that experience automatically, gradually increasing the proportion of PRs accepted on the first generation and turning “recursive self-improvement” into an auditable organizational feedback loop.