X / Twitter
OpenAI Codex team member Thibault Sottiaux
The team quietly performed a double reset: it reset current usage limits and added an extra reset to each user’s bank, available whenever they choose to trigger it. This is a conciliatory gesture from Codex amid the reality of “a $200 subscription suddenly becoming a $2,000 usage limit.” Another reminder emphasized that the Codex App, CLI, and SDK can connect to any open-source model, not just OpenAI’s own. This effectively frees the Codex tool stack from dependence on a single model vendor; anyone who wants to run it with local or open weights is allowed to do so. Overall, OpenAI is responding to recent complaints about agent subscription economics through both flexible backends and flexible allowances.
https://x.com/thsottiaux/status/2067399435009622521
https://x.com/thsottiaux/status/2067181377028538431
Linear Head of Product Nan Yu
He pointed out that half the discussion about “taste” involves people talking past one another because of the word’s meaning: taste is not the same as aesthetic taste, just as design is not the same as visual design. This is a semantic reframing. When pg talks about taste while wearing cargo shorts, he clearly is not referring to taste in clothing, but to judgment more broadly. If people define taste differently, they can talk indefinitely without aligning. This is especially useful in startup circles, where taste and design are frequently used words with highly flexible meanings; agreeing on definitions first can avoid many fruitless debates.
https://x.com/thenanyu/status/2067327901666521478
https://x.com/thenanyu/status/2067327619897446721
Replit CEO Amjad Masad
He shared a clip from an interview with Spike Jonze and, for the first time, clearly presented Replit’s collaboration with Anthropic: “Design with Claude, Ship with Replit.” Frontend prototypes are conceived in Claude, with engineering and deployment handled by Replit. This division of labor positions Claude as the upstream design IDE and Replit as the downstream shipping engine. He was also promoting the new Vibecon event. Overall, Replit is actively building chains of partnerships with other tools rather than competing head-on, a strategy similar to Vercel’s emphasis on an “agent engineering framework.”
https://x.com/amasad/status/2067386053980266542
https://x.com/amasad/status/2067363904183783833
https://x.com/amasad/status/2067363597110391230
Vercel CEO Guillermo Rauch
At Vercel SHIP London, he laid out his central argument: today, AI SDK is to agents what Next.js is to React. React handles UI rendering, but deploying a real web application still requires Next.js. Similarly, the fiercer competition at the model layer becomes, the more builders need an engineering framework for production agents, rather than another model wrapper. He provided a concrete data point: the open-source GLM 5.2 model surpassed Opus 4.8 on Vercel’s own Next.js Evals that day, offering firsthand data that puts “open source catching closed source” on an objective metric footing. He hinted that Vercel’s upcoming product launches would follow this agent framework direction.
https://x.com/rauchg/status/2067369305826574587
https://x.com/rauchg/status/2067274101702406554
https://x.com/rauchg/status/2067242482190979186
Box CEO Aaron Levie
He broke down the Applied AI playbook he has observed over recent months into four parts, addressing the familiar question of whether a moat exists above LLMs. First, build specific capabilities that bridge intelligence and workflows. Do more than serve model-output tokens: refine data collection, specialized toolsets, and human-in-the-loop interfaces for particular industries. The deeper the work, the stronger the moat. Second, build a model router that switches between frontier and inexpensive models based on task difficulty. Players who understand evals and can commercially accommodate different price points have the strongest advantage. Third, use FDEs or similar teams for implementation and change management. Most enterprises adopting agents need to clean data, rewrite processes, run evals, and sign SLAs—industry-specific dirty work. Fourth, pursue vertical GTM across languages, compliance, and industry channels. He acknowledged that the bitter lesson might eventually consume all of this, but enterprises need problems solved now and cannot wait for perfect models.
https://x.com/levie/status/2067455756795039957
Y Combinator President and CEO Garry Tan
He used arithmetic to argue the cost of banning Fable 5: 5M daily active frontier AI-coding developers worldwide, a fully loaded cost of $90 per hour, 17.8% of work migrated to Fable within 48 hours, and Fable averaging 15% faster than alternatives. The result is around $2.40 lost per developer per hour, or roughly $12M per working hour across the industry. This estimate translates a regulatory action into a concrete dollar cost of working time. He offered two other independent observations: technical founders now have business thinking, while business founders have technical thinking, so more startups will be able to work; and YC does not judge founders by age, only by whether they can use craft and care to build something people want.
https://x.com/garrytan/status/2067366749411176831
https://x.com/garrytan/status/2067308407603048774
https://x.com/garrytan/status/2067260431597723825
Independent builder Zara Zhang
Her stance on “writing with AI” is counterintuitive: it is not itself a problem; the danger is letting AI write before you have developed taste, because you will not recognize the slop it produces. Her prescription is to read extensively first to build examples of what is good, then write extensively to discover your own voice, and only then use AI as an enhancement, ensuring the output still sounds like you. In another post, she exposed the reality of vibe-coded personal tools: building one takes only a day, but finding out whether you will actually use it takes a week. She said most abandoned projects worked fine; she simply never opened them again. Products are built for an imaginary person who remembers to open them daily, presses the right buttons, and follows steps in order. Real humans are lazy and forgetful, and products should be designed for them.
https://x.com/zarazhangrui/status/2067423674689638652
https://x.com/zarazhangrui/status/2067313780724551853
FPV Ventures Partner Nikunj Kothari
He publicly opposed tranched rounds, a new financing structure, for a specific reason: tranches raise a company’s 409A valuation, forcing later-joining early employees to have options priced at an inflated valuation that even lead preferred investors would not pay—effectively making employees’ FMV options bear the cost. He offered a recognition rule: if a company claims to have raised a given amount while diluting less than 10%, it is probably a tranche structure. This is now common at seed and Series A, and even Series B rounds are starting to feature these unusual tranches. Finally, he made a prediction for the next 12 months: a wave of debut funds from emerging managers who are not top-tier names but have strong founder affinity, distinctive perspectives, and exceptional hustle. He believes the star funds of the next decade will emerge from this wave.
https://x.com/nikunj/status/2067399657639285150
https://x.com/nikunj/status/2067397092981772501
https://x.com/nikunj/status/2067378464773292066
Every CEO Dan Shipper
He revisited “against explanations,” an essay he wrote in 2023. Its core argument is that AI may shift science from “explain first, then discover” toward “make many predictions first, then supply explanations afterward.” Recent progress at Tacit makes him feel this path is materializing faster. He also announced another Every investment, backing Tacit founder ninklefitz because he supports the founder’s mission and approach. In a third post, he shared YouTube and Spotify links to his podcast with GitHub COO Kyle Daigle, the AI & I episode covered in the podcast section below.
https://x.com/danshipper/status/2067386395283345808
https://x.com/danshipper/status/2067386342661624055
https://x.com/danshipper/status/2067293009964630081
South Park Commons General Partner Aditya Agarwal
He published a list of questions SPC members are currently exploring, arguing that the signals in these questions directly reveal where the frontier will move next. He paired it with guidance on applying to SPC, encouraging people with ideas to read the full question set before deciding whether to join the community and pursue a founder path. The overall positioning presents SPC as a community of frontier questions rather than a traditional incubator, attracting people through the quality of its questions instead of a brand narrative.
https://x.com/adityaag/status/2067306245942317426
https://x.com/adityaag/status/2067306244390428893
https://x.com/adityaag/status/2067306242825949398
OpenAI CEO Sam Altman
He officially announced that Noam Shazeer was joining OpenAI. Sam said Noam was the person he had most wanted to work with since OpenAI’s founding, and that it took a full 10 years; he believes the wait was worthwhile. An accompanying joke was, “We cannot explain why Noam is so good at AI, so we can only attribute it to divine will, as we do everything else.” Noam previously led key work at Google including Attention Is All You Need and Switch Transformer, and is regarded in the industry as a researcher representing the upper limits of model architecture expertise. His arrival is a major reinforcement of OpenAI’s top architecture talent.
https://x.com/sama/status/2067427678529974740
https://x.com/sama/status/2067427421083652131
Anthropic’s official Claude account
Claude Design is now available in beta on web and desktop across all paid plans. It integrates bidirectionally with Claude Code: designs can be handed directly to Claude Code to build, or users can start in Claude Code and sync the design project over from the terminal. Export supports PDF and PowerPoint, and projects can also be sent to other tools users already use. The editor itself has been rebuilt for more reliable everyday use, with direct dragging, resizing, and alignment on the canvas. Overall, Anthropic is pushing Claude from a chat assistant toward a builder platform combining design and code.
https://x.com/claudeai/status/2067325894268428560
https://x.com/claudeai/status/2067325893001826552
https://x.com/claudeai/status/2067325891781226581
Podcasts
AI & I by Every — GitHub's COO Explains Why AI Hasn't Replaced Developers
Key takeaway: GitHub already receives 17,000,000 agent-created pull requests a month. At the current growth rate, code volume would surge from 1B commits to 14B within a year, yet developers have not been replaced. The real differentiators are people and teams who can turn agent feedback data into a continuous hill-climbing loop of self-improvement, and those who use model routers to adjust token spending to the task.
Guest Kyle Daigle is GitHub’s COO and has been at GitHub for 13 years. He now also serves as CMO for Microsoft’s entire developer division, giving him perspectives on both GitHub’s frontline products and Microsoft’s overall developer ecosystem. Kyle does not have a formal CS background: he attended art school and started coding to pay tuition. This gives him firsthand experience as someone who writes code without identifying as a developer, the fastest-growing segment of GitHub Copilot users today. Knowledge workers in legal and finance are also starting to use Copilot to build small tools for themselves. The interview took place around Microsoft Build 2026, the first Build to extensively feature external community speakers, including Peter Steinberger and Swyx. Kyle repeatedly emphasized this as a deliberate statement that software development is a team sport.
The first especially informative point concerned output data. GitHub recorded 1B commits throughout last year; a linear extrapolation would put this year at 14B, though he stressed that actual growth would not be linear and would only be higher. In March 2026 alone, agents created 17,000,000 pull requests. Kyle directly rejected the claim that agent output is all slop nobody cares about. He said people are moving beyond early adoption and remain far from the ceiling. The next norm will be Kyle working in parallel with 1 to N agents, rather than a single developer working along one track. GitHub itself is investing for the next growth wave, and he believes the curve will not “grow and then plateau, but keep growing.”
The second point concerned strategies for open-source maintainers. Open-source communities are widely overwhelmed by the agent wave. GitHub’s response is not to impose one standard on everyone, but to provide building blocks: maintainers decide whether to accept PRs, from whom, and what evidence of good faith to require from contributors. He cited Michel Hashimoto’s public vouch system as one example, but GitHub deliberately avoids imposing it on everyone because each community wants its own approach. In his words, “We don't really ever wanna be the first to create a standard or an approach.” GitHub will not be the first to set the standard; it waits for community consensus and then cements it.
The third point concerned token economics and “how to stop a $200 subscription becoming a $2,000 bill.” Kyle’s answer was not lower prices, but model routing plus frontier tuning. Microsoft Foundry’s model router can automatically choose a model at the API layer based on task intent, moving simple tasks from expensive models such as GPT-5.5 or Fable to cheaper Haiku-class models. Meanwhile, frontier tuning lets enterprises tune base models such as MAI thinking 1 with their own m365 data, including documents and chats, producing versions suited to their workflows without training from scratch. He admitted he initially thought this was a magic parlor trick and changed his mind only after seeing results: “sometimes that's where the alpha is, it's like where it feels like this is too simple to work.” Ideally, individuals and enterprises no longer manually switch models to save tokens; the system decides for them.
The fourth point was also the most counterintuitive: Kyle’s own use of agents is not coding, but “self-review.” He has a Claude called Baxter that reads everything he has written and said each day, including this interview, then produces a communications report. Rather than “what Kyle said,” it tells him, “Kyle, that recent statement was unclear / you have overused this analogy / here are some better analogies.” He cited GitHub’s early experience with Hubot: “humans are way more willing to take critical feedback from robots than other humans.” People find negative feedback from robots easier to accept than criticism from colleagues. Every seven days, he has the agent review his email, Slack, and decisions, then check whether he acted on its previous advice. This loop has convinced him that agents’ greatest benefit to individuals is not coding, but a closed loop of self-improvement.
The fifth point was that “hill climbing” is a phrase repeated endlessly inside Microsoft. Weekly reviews focus not on heroic moonshots, but on a cycle that considers thumbs up/down, user acceptance rates, and both hard evals and softer user sentiment. He warned of a common trap: if hard evals improve while user sentiment plummets, that is essentially overfitting and calls for immediate rollback rather than continued rollout. Microsoft has already produced seven MAI models through this loop and next plans to make the same hill-climbing machinery available to customers.
Kyle’s closing framing was that developer choice cannot be compromised. Microsoft builds its own models while also partnering with Anthropic, OpenAI, and Google, allowing developers to switch freely. He compared this to GitHub’s earlier decision to offer developers free private repositories: once it becomes a walled garden, users leave, and it merely becomes “the next mousetrap.” The lesson for readers is that to use agents for sustained gains, they need both model routing to control costs and their own personalization data. That is the real way to bring a $2,000 bill back down to $200.
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL