X / Twitter
AI Builder Swyx
Swyx still actively uses `/loop` and `/goal` and believes most AI users abandoned them too soon. In the g5.6/c5 period, he considers both capabilities suitable for tasks requiring steerability and autonomy together. For open-ended goals such as "a loop that generates loops," users can specify the final state without detailing the execution path. He shared a lengthy action-and-reasoning task where repeatedly setting a goal ultimately helped prevent the work from going off track. He also believes MITM agent distillation works, despite being a technically difficult method.
https://x.com/swyx/status/2083439562437673053
https://x.com/swyx/status/2083237045720465504
Thibault Sottiaux, OpenAI Codex and ChatGPT Team
Thibault Sottiaux announced a reset of Codex and ChatGPT Work usage limits. He described it as celebrating a week of efficiency gains and encouraged users to run more tasks over the weekend. The post's exaggerated target of 100,000 Luna threads clearly emphasized greater room to experiment with highly concurrent, long-running work. He subsequently specifically confirmed that ChatGPT Work was included. For teams using Codex and ChatGPT Work frequently, this provides more generous allowances for testing workflows over the weekend.
https://x.com/thsottiaux/status/2083395449814229287
https://x.com/thsottiaux/status/2083387677945036995
Nan Yu, Linear Head of Product
Nan Yu said Linear's most common automation loop is Issue, Agent, PR, Release, with approximately 30% of bugs completing the full process. Effective instructions should require the agent to thoroughly investigate root causes and gather more evidence through Datadog and Sentry MCP. Agents should change code only when the fix is highly certain; otherwise they can consume many tokens without reliable results. If crucial information such as reproduction steps is missing, the agent should comment directly on the issue asking the reporter for evidence. His central point is that agents, like people, need explicit instructions to follow good engineering practices.
https://x.com/thenanyu/status/2083230295206121807
Amjad Masad, Replit CEO
Amjad Masad shared an 8b model rated approximately 1,500 Elo. In his tests, it consistently beat frontier models and Stockfish level 0. It still won even when GPT 5.6 used high reasoning effort and response chaining. The speed difference was also substantial: it needed only 1 to 2 seconds per move versus approximately 30 seconds for the comparison model. He provided a way to try it directly, emphasizing that small models can offer both speed and performance advantages on specific tasks.
https://x.com/amasad/status/2083424608993300824
Guillermo Rauch, Vercel CEO
Guillermo Rauch positioned AI Gateway as essential infrastructure for enterprises to turn AI into an effective investment. New features include setting budgets by key, team, or project for finer cost control. The platform also provides failover, model and provider selection, and real-time observability. He believes enterprises should move beyond simply maximizing token consumption and focus on reliability, governance, and returns on investment. As software projects become agentic software factories, the Issue, Agent, PR, Release loop will become standard. Authors and maintainers will shift toward optimizing that loop, improving product quality, and defining which tasks deserve to enter it.
https://x.com/rauchg/status/2083319868766699699
https://x.com/rauchg/status/2083208578526314513
Aaron Levie, Box CEO
Aaron Levie expects the harness to become the AI stack's most important variable after model capability. He did not claim that any specific set of figures generalizes to every task, but considers the overall trend clear. A high-quality harness must divide work efficiently and route tasks to the right model at the right time. This directly affects accuracy and cost, not just user experience. As tasks grow from hundreds of thousands or millions of tokens to tens or hundreds of millions, the harness's impact increases substantially. He sees this field as very early, with enormous product and infrastructure opportunities.
https://x.com/levie/status/2083389460679373135
Garry Tan, Y Combinator CEO
Garry Tan believes both personal AI and enterprise knowledge brains need a clearly structured harness. The solution he recommends was built by his own team, which uses it daily. The project is free and open source, available directly to other individuals and organizations. Garry sees the harness as a foundational layer shared by personal intelligence tools and company-wide AI systems. The post emphasizes daily team use and open access without making additional claims about specific performance metrics.
https://x.com/garrytan/status/2083353760701833546
Builder Zara Zhang
Zara Zhang noticed that every time she publishes something that goes viral, she first has to overcome the self-doubt of "isn't this obvious?" Her conclusion is that experiences obvious to creators are usually unknown to most people. She also highlighted Claude Tag's impact on how Anthropic works, citing a figure that 65% of PRs in Anthropic's product and engineering teams now originate from Claude Tag. For non-engineering teams, she believes the most natural agent interface is a collaboration tool employees already use, such as Slack. Her own work entry point moved in six months from the terminal in January to desktop apps such as Codex in March, then to workplace collaboration tools in June. Each transition brought agents closer to how people already communicate. Her view is that agents should appear directly where users work.
https://x.com/zarazhangrui/status/2083354965482062079
https://x.com/zarazhangrui/status/2083161173563003268
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari mentioned a seldom-publicly-discussed belief in venture capital: the best founders are often running away from something. The drive may stem from an unsettled childhood, a strong need to prove oneself, or pain that keeps one's foot on the accelerator. His new article asks where founders' real motivation comes from. It also discusses a higher-intensity state most people have never entered. He presents the belief as a question for understanding founder motivation and extreme commitment, rather than treating it directly as a universal conclusion.
https://x.com/nikunj/status/2083307235619287363
Peter Steinberger, OpenClaw and OpenAI Builder
Peter Steinberger said queues used to be the better way to handle concurrent input. By 5.5, models were no longer easily confused by new input. Users can continue handing the model tasks while it works, without waiting for the current job to finish. It will handle incoming material carefully and in order. This changes work from explicitly maintaining a task queue to continuously adding and steering tasks during execution.
https://x.com/steipete/status/2083369880599015713
Dan Shipper, Every CEO
Dan Shipper said the shift in OpenAI's momentum relative to Anthropic since early spring had become quite noticeable. He called it a compelling comeback story and reiterated his view in a Wall Street Journal article discussing the two companies. He also used fictional programmer interview questions for 2027 to satirize how AI might change engineering skill requirements. Questions included the three unresolved mathematical conjectures recently solved and the prompts used, and the agent's latest inadvertent cyber felony and its mitigation. The final question returned to testing how many Rs are in "strawberry" and how many es are in "seventeen," juxtaposing frontier capabilities with basic textual judgment. He depicts a future in which human programmers must show results and explain how they direct agents and handle the resulting risks.
https://x.com/danshipper/status/2083380721607921904
https://x.com/danshipper/status/2083239700664349128
Sam Altman, AI Builder
Sam Altman shared a family use case for ChatGPT Work. Users can connect family members' calendars and tell the system about each child's interests. Each morning it generates a personalized podcast for the trip to school, potentially covering one child's soccer match that day, another child's upcoming birthday, and some news. The example combines calendars, family context, and a regular commute into continually generated audio content. He also responded "20x" to a comparison involving Moore's law, without elaborating on the calculation or specific context.
https://x.com/sama/status/2083221585792762171
https://x.com/sama/status/2083203642975502640
Official Blog
Claude Code now supports artifacts
Claude Code now supports generating work in progress as an artifact: a visual web page that updates as a session advances. Uses include PR walkthroughs, system explanations, filterable dashboards, release checklists, and incident investigation pages. Artifacts use the session's full context, including the codebase, connectors, and conversation, without requiring users to reconnect data sources or build additional infrastructure. Updates refresh in place at the same URL, with every publication creating a recoverable version record; teams can also manage existing artifacts through a gallery. In internal testing, Anthropic frequently used them for incident timelines, suspect commits, error-rate changes, and root-cause analyses, letting team members share the same continuously updated investigation view. Artifacts are initially visible only to their authors, and sharing is limited to authenticated organization members; public publication is unavailable. Administrators can manage access through organization-level switches, role scopes, retention policies, and a compliance API. The feature is currently in beta for Claude Team and Enterprise organizations, with creation from the Claude Code CLI and desktop app and viewing in the browser.
https://claude.com/blog/artifacts-in-claude-code
Podcasts
Unsupervised Learning — Ep 92: xAI Co-Founder Unpacks the Future of Model Development
Key takeaway: Coding agents' breakthrough is only a beginning. The important next step is taking AI from verifiable software tasks into scientific discovery and personal life, while distributing control and benefits more broadly.
Igor Babushkin worked on StarCraft and AlphaCode at DeepMind, participated in early reasoning research at OpenAI, then co-founded xAI and worked on Colossus and model development. After leaving xAI, he founded River AI, focusing on personal AI for businesses and consumers, alongside local hardware. His starting point is to explore new ways of building, customizing, and distributing models rather than simply copying the established pattern of "large-scale pretraining, offering an API, and charging per token."
His first key judgment is that coding agents crossed an unignorable capability threshold last November and December. Babushkin recalled that the latest generation of Claude Opus models led software engineers and AI researchers to broadly acknowledge that agents substantially reduce development difficulty. In December, he used coding agents to rewrite many projects he had previously built or always wanted to build, while also feeling the risk of losing control as capabilities rose rapidly. He summarized the situation: "We're all becoming the sorcerer's apprentice." The tools bring almost magical productivity, but users may not always control the downstream consequences.
His second judgment is that coding and mathematics broke through first because results are easily verified, not because agents are limited to these domains. Code can run tests, and mathematical theorems and proofs can be formalized with tools such as Lean, providing clear reward signals. Scientific discovery is a natural extension, but problems in materials science, fundamental physics, and rocket engines require real-world experiments. After proposing a material design, an agent needs feedback on whether the experiment succeeded. Closing the real-world data loop is therefore a major current bottleneck.
His third judgment is that the most capable models and personal AI may branch in different directions. Models for complex coding, mathematics, and research can keep pursuing maximum capability, while agents improving ordinary lives need not prove the Riemann hypothesis. Personal AI instead needs to understand users' specific circumstances in life, work, and the outside world and act on that information. Babushkin sees giving individuals tangible AI value and control over its behavior as River AI's central motivation.
His fourth judgment concerns AI safety and concentrated power. After leaving xAI in 2025, Babushkin invested in AI safety companies, but became impatient waiting for other teams to make progress and decided to start a company himself. He considers distributing AI's benefits and control the most urgent safety issue today, as model access and behavioral rules increasingly come from a few large companies. He also supports releasing models near dangerous capability thresholds but not yet capable of real harm, letting more researchers study alignment, control algorithms, and cybersecurity. Compared with distant, unpredictable model-takeover risks, he sees the more immediate questions as whether AI amplifies inequality and whether more people can share the benefits of progress.