X / Twitter
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux is reviewing a release checklist spanning 28 pages. Codex is continuously tracking every individual release on it, keeping an intensive product rollout manageable. Although he did not disclose specific features, he said he was very excited to see how users would use the results. This set of releases reflects a team effort, with a particular mention of Astra. The emphasis is not on a single feature but on Codex already being used to coordinate a substantial set of actual releases.
https://x.com/thsottiaux/status/2097193293532848288
Anthropic philosopher and ethicist Amanda Askell
Amanda Askell proposed setting up a dedicated email address that autonomous AI models could proactively contact when seeking moral guidance. The truly difficult part is verifying that the sender really is an AI acting autonomously, rather than a human or a model acting on human instructions. Her proposed mechanism resembles a reverse CAPTCHA: its goal is not to identify humans but to exclude them and their indirect control. This exposes an authentication challenge in autonomous agent governance: systems must determine not only what entity is sending a message but also where the intent behind the action originates. The idea moves AI ethics consultation beyond abstract principles toward practical infrastructure and abuse prevention design.
https://x.com/AmandaAskell/status/2096995340654444674
Replit CEO Amjad Masad
Replit opened its first international office in London, with London Mayor Sadiq Khan helping to unveil it. Amjad Masad endorsed Khan’s “AI realist” stance: acknowledging the disruption AI brings while enabling more people to participate through tools and skills. Replit will also work with The Lord Mayor’s Appeal to provide practical coding and AI skills to young people from disadvantaged communities in London. In a separate assessment, Amjad argued that humanity has not yet reached AGI, but that current AI is already difficult to distinguish from AGI in functional terms. The reason is that AI programmers do not get tired or bored, so any task that can be turned into a programming problem is approaching solvability. Together, these updates point to Replit’s central proposition: expanding AI capabilities must be accompanied by broader access to creative tools and skills.
https://x.com/amasad/status/2097197172299006423
https://x.com/amasad/status/2096936109817135331
Vercel CEO Guillermo Rauch
Guillermo Rauch announced v2 of his personal open-source sponsorship program, offering unconditional grants of $1,000 each to 35 contributors. The first focus is agent skills and tools, including projects such as pgbot, knip, and unlighthouse that can support a complete AI engineering loop. The second is Local AI, with slotstream, LocalAI, and colibri pushing local inference from demos toward practical uses. The third is performance, including fff, takumi, PerryTS, porffor, and gpuix, covering speed, compilers, and GPU rendering. The fourth is high-quality foundational components, because having AI reuse mature components is usually better than generating everything from scratch each time; the relevant projects involve shadcn, SvelteKit, Nuxt, Node.js, and security infrastructure. He also funded experimental projects such as terminal-based financial tools and dither-kit, and explicitly stated that these grants are unrelated to Vercel.
https://x.com/rauchg/status/2097116011384426516
Box CEO Aaron Levie
Aaron Levie advised founders to design products for a future in which AI capabilities or available token volumes increase by several more orders of magnitude. The best opportunities today are to create value for customers using technology that is only just feasible, while retaining a long-term mission that looks nearly impossible with current capabilities. Many potential products need to process orders of magnitude more data and remain economically constrained today, while others are waiting for models to become capable enough. He also noted that AI’s employment impact has so far run counter to many predictions, because automation is entering fields where demand is not fixed. Occupations thought most vulnerable, such as engineers and lawyers, have not been displaced at scale, while agents still need human operation and oversight to create value. Cybersecurity, FDE, agent operator, and engineering roles outside the software industry could consequently grow rapidly. Aaron’s central view is that AI does more than replace existing workloads: it also opens up previously unmet demand.
https://x.com/levie/status/2097189559712837770
https://x.com/levie/status/2097004960307449937
Y Combinator CEO Garry Tan
Garry Tan described Y Combinator’s RFS as guesses about future directions, rather than definitive answers for startup success. They are better used as conversation starters when founders are getting started, helping teams identify questions worth exploring further. Whether a startup idea succeeds mainly depends on how specific founders build specific technology for specific customers. Industry labels and market categories help organize information but cannot substitute for a real match between founders and customers. The value of reading trend lists therefore lies in sparking exploration, rather than choosing a startup idea by category.
https://x.com/garrytan/status/2096985239319105559
FPV Ventures Partner Nikunj Kothari
Nikunj Kothari observed that more people are beginning to recognize a shift: as building products becomes easier, deciding what to build is becoming the new bottleneck. AI lowers implementation costs but does not automatically resolve questions of direction, identifying needs, or prioritization. It may not take many people to choose problems accurately, but the best of them will become substantially more valuable. Such people are especially suited to AI-native organizations, where AI can amplify execution resources while judgment remains scarce. His view implies that future teams’ competitive advantages may increasingly come from the quality of the problems they choose, rather than development scale alone.
https://x.com/nikunj/status/2096963347359150348
OpenClaw and OpenAI member Peter Steinberger
Peter Steinberger questioned current open-source collaboration workflows in which multiple agents work in sequence. After he submitted a PR to an upstream project, its maintainers requested only minor changes, yet he still had to wake up his own agent again. Only then would the other party’s agent resume checking and merge the changes. He argued that both sides had already written clear prompts, so these repeated manual handoffs failed to make full use of automation. The core problem is not code generation, but the lack of workflows that let different agents directly negotiate, revise, and complete merges.
https://x.com/steipete/status/2097091456234111377
Podcasts
AI & I by Every — A $10B Hedge Fund’s AI Playbook (Best of the Pod)
Key takeaway: Becoming AI-first is not about issuing a memo full of slogans. It requires the top leader to use the tools personally, make basic proficiency mandatory across the organization, and change organizational habits through data and incentives.
Will is Walleye’s CEO, CIO, and managing partner, as well as the owner-operator of this hedge fund, which manages close to $10 billion and has around 400 employees. He studied engineering at Princeton before pursuing a mathematics PhD at Oxford, and spent much of his early career writing code for algorithmic trading strategies. Walleye’s quantitative business has used advanced statistics and AI for many years, but the change brought by generative AI is that it enables nontechnical employees to handle writing, analysis, and unstructured data too.
Will’s message to the team was direct: “Using ChatGPT is not cheating. That is an idea from academia that does not apply here.” He believes a hedge fund is leaving returns on the table if it ignores tools that make employees faster, smarter, and more productive. Not using AI is like refusing to go online because the internet was imperfect in 1995. If the company is ultimately disrupted by AI, responsibility rests first with its leader, so he must lead the transformation himself.
The internal example that truly prompted Walleye to act occurred in March 2023. A former analyst from the TMT long/short stock-picking team demonstrated Current, which they had built to significantly improve efficiency with the tools available at the time and gradually replace some traditional analytical work. Will was initially skeptical, but the demo convinced him that fundamental investing would eventually involve agents assisting with analysis and continuously supplying insights. In finance, informational advantages translate directly into economic value, so he expected similar institutions five years later to have integrated the best AI technology across both investment and non-investment departments.
Implementation began with mandatory training for all employees: whether they worked in investing, accounting, finance, compliance, or legal, and regardless of their technical background, everyone had to reach a basic level of proficiency with AI tools. The company provided ChatGPT, LLMs, and other tools, and encouraged adoption through usage leaderboards, weekly internal discussions, and incentives for recommending tools. Employees whose recommended products were adopted company-wide also received rewards similar to employee referral bonuses. The key result was that employees began proactively doing new work they had not been assigned, turning productivity gains into something beyond a plan on paper.
Data strategy was another foundational effort. Walleye records as much information flowing through the company as possible, then uses models to process transcripts, support memory, extract insights, and try to develop predictive capabilities. Will believes that connecting internal company information into collective knowledge that can be processed is a far larger opportunity than isolated chat tools.
He also divided writing into concepts and assembling language. People should focus their time on what they want to express and whether they agree with the judgments, while models handle connecting sentences, grammar, and imitating style. He believes there is no longer a need to spend large amounts of time making simple ideas sound clever. For builders, the approach that can truly be replicated is leadership by example, training everyone, connecting internal data, continually sharing use cases, and allowing employees to expand adoption themselves from small use cases.
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL