X / Twitter
Latent Space host Swyx
He observed that few people outside Anthropic are using ultracode. The tool burns tokens at an alarming rate, but its power depends on setting up a repository so subagents can fan out fully in parallel. He compared subagents to “subroutines with judgment”: much knowledge work consists of yak-shaving nested inside more yak-shaving, with each layer requiring some intelligence to move forward. Swyx argued that such dynamic workflows extend far beyond coding to almost any work requiring judgment. That same day, he reposted Microsoft CEO Satya Nadella’s remarks on the loop as IP, emphasizing that the real moat is not choosing the best model but building a learning loop above it so human capital and token capital compound together. Swyx singled out Satya’s words: “You can offload a task or even a job, but you can never offload your own learning.” He summarized this as a bet that the frontier ecosystem matters more than the frontier model.
https://x.com/swyx/status/2066415484149633329
https://x.com/swyx/status/2066235625695850526
OpenAI Codex / ChatGPT team member Thibault Sottiaux
He shared that Codex can now see and set /goal itself. In other words, the Codex agent can generate concrete task goals from the user’s intent. He sees this as an extension of meta-prompting: previously, people wrote prompts telling agents what to do; now agents use context to decide which goals to pursue. Sottiaux emphasized an internal philosophy: “Everything we build is also a tool for agents themselves.” Engineering teams are no longer building only human interfaces, but designing every capability as an interface agents can consume too. The post received 2232 likes and 151 replies within 24 hours, reflecting growing community interest in the boundaries of agent autonomy. For developers building agent systems, “goal setting from intent” is a clear signal of the shift from prompt engineering toward agent engineering.
https://x.com/thsottiaux/status/2066270561081454989
Vercel CEO Guillermo Rauch
He announced that ChatGPT’s skills ecosystem has passed 700,000 skills, all organic and community-driven. He described it as an “open⎵ai ecosystem,” suggesting the growth curve now has the characteristics of a platform-level ecosystem. Its 389 likes and 19 retweets were not spectacular, but for builders watching platform flywheels, it is a key signal: once skills reach six figures, ChatGPT is no longer just a conversational product but a platform where third parties continuously add capabilities. Rauch is Vercel’s CEO, and his own company is betting on AI model distribution and agent deployment, making this an indirect endorsement of OpenAI’s platform strategy. For AI product builders, one question to watch is whether, when a marketplace such as ChatGPT skills reaches 700K, independent SaaS tools will be absorbed into the ecosystem or shift to other distribution channels.
https://x.com/rauchg/status/2066299732277031042
Box CEO Aaron Levie
He published two long posts today. The first built on Satya’s learning-loop argument: the companies that win will feed their unique IP, institutional knowledge, and data to AI through the right architecture, allowing human capital and token capital to compound together inside the business. He emphasized one architectural requirement: “Enterprises must be able to replace the underlying generalist model at any time without losing the accumulated expertise of a company veteran.” Future enterprise AI systems, in other words, must fully decouple the learning loop from any particular model. The second took a political and geopolitical view, arguing that open-weight models will be the biggest winners of this cycle. Once “a model can be shut off for a country on a given day” becomes an established precedent, other countries have a stronger incentive to develop sovereign AI and their own open-source model stacks. He addressed US policymakers directly: “If the United States attaches AI regulation to the model layer rather than the applied layer, the supply of open-source models worldwide will increasingly originate outside the US, gradually diluting US leadership in the AI stack.” Levie’s overall position is that the applied AI layer will capture the most incremental value over the next few years, which is also Box’s current bet.
https://x.com/levie/status/2066237607244427761
https://x.com/levie/status/2066167615618466060
Y Combinator CEO Garry Tan
He posted three times today, connecting “long-term thinking” with “young people’s skill sets in the agent era.” Reposting a discussion about autonomy in commercial AI, he stated clearly that “open source is the escape pod that lets businesses retain control of their own destiny over the long term.” If suppliers can arbitrarily change, price, or shut down closed models, a business’s lifeline is constrained, and open weights are the only controllable fallback. Another post deserves repeated attention from builders: he believes the next generation of young people who truly change the world will almost certainly be those who excel at long-running, multistage, multiteam agent tasks, at scale and high frequency across every corner of work and life. He emphasized “high volume” and “across every part of their personal and work lives,” implying that real agent proficiency is not occasionally running a workflow but having agents continuously handle hundreds of parallel tasks, 7x24. Garry also mentioned gbrain users as early examples of this approach. For founders, this is a talent profile drawn by YC’s leader: people who orchestrate complex tasks with agents will earn disproportionate returns.
https://x.com/garrytan/status/2066307697574862905
https://x.com/garrytan/status/2066269412391637050
Builder Zara Zhang
Two successive posts sharply dissected how to make a good skill. Her counterintuitive view is: “Good skills are not written; they emerge from use.” Specifically, first do the work yourself, mess it up 20 times, and keep fixing it. Then ask AI to bottle the entire process you just worked through as a skill. A shorter formulation was “you make a skill by ending with one, not starting with one,” directly reversing the sequence of “write a skill first, then use it.” This runs against typical AI prompt engineering’s approach of writing a prompt first and then tuning it. She argues that skills should originate in real, repeatedly iterated, ground-truth workflows. For people building Claude Code or custom agent toolchains, this is a practical anti-pattern to recognize: do not begin by designing SKILL.md; first get the work done, then capture the process afterward.
https://x.com/zarazhangrui/status/2066388749244854771
https://x.com/zarazhangrui/status/2066394505037926426
Builder associated with OpenClaw and OpenAI Peter Steinberger
He shared a practical in-flight workflow: Mosh paired with tmux or zellij is a lifesaver on unreliable airplane internet. He personally prefers zellij, though tmux works perfectly well, and both can be used with Mosh. Mosh is a UDP-based alternative to SSH, with reconnection optimizations for high packet loss and disconnects. Combined with tmux/zellij session persistence, it means the remote session survives network instability and command-line work can continue. The post’s 80 likes may seem modest, but for builders relying on cloud devboxes or remote Claude Code, it offers a configuration they can copy directly. Steinberger has long been experimenting with OpenClaw and OpenAI-related tool stacks, and his development environment recommendations are often a step or two ahead.
https://x.com/steipete/status/2066427449551036469
Podcasts
Training Data — LIVE: Jensen Huang on Building the Dynamo of the Intelligence Age
Key takeaway: AI has advanced from “generating content” two years ago to “working for you by the hour.” The entire computing industry is being remade from retrieval-based to generation-based, and the next $1T in capex will stack up across energy, chips, infrastructure, models, and applications.
NVIDIA founder and CEO Jensen Huang gave investors and entrepreneurs from 128 families across 60 countries and multiple industries an introduction to the AI investment landscape. He deliberately made it accessible enough that “nontechnical people could use it to decide their next investment.” Sequoia Capital’s Training Data broadcast it LIVE. The backdrop is that $1T in annual AI capital expenditure is already happening, making the question of which layer receives the money next a highly practical investment decision.
He divided AI’s evolution into two steps. ChatGPT two years ago merely “translated,” mapping one form of information into another. Today’s agentic AI adds “thinking + tool use,” which is what makes it truly valuable. He gave the example of enterprises now hiring AI for $20 to $30 per hour, calling this “the fastest-growing software business in human history.” Jensen’s assessment is that AI is no longer a cute toy but real labor billed by the token.
He made the economics of AI factories concrete: NVIDIA will produce 8 million chips this year, assembled into racks of 72 chips each. Each rack weighs two tons, contains 1.5 million parts, and costs $4M. A gigawatt AI factory costs $50B to build but can produce $300B to $400B in “intelligence” per year, giving it a rapid return on investment. Globally, $1T will be invested in the AI ecosystem this year; he guesses the future steady state will be $20T annually.
He divided the investment landscape into a five-layer cake. The first layer is energy, an opportunity legacy giants such as Siemens, Mitsubishi, and GE Vernova have not seen in a century: nuclear, wind, solar, and hydrogen can all attract funding if they generate power. The second is chips, networking, and silicon photonics. The third is infrastructure, where land, electricity, buildings, and data center operations are all in short supply. The fourth is models: beyond OpenAI and Anthropic are non-language corpora involving proteins, genes, cells, and physics. He specifically noted that the physical world’s $80T of existing industries is the larger frontier. The fifth is applications, where $100B in VC investment went last year alone, the highest annual VC investment in history.
He challenged the popular “AI takes jobs” narrative with a counterintuitive example. Ten years ago, a leading AI scientist publicly predicted that computer vision would wipe out radiology; yet 12 years later, demand for radiologists has risen. The reason is that greater productivity allows hospitals to serve more patients, makes radiology more profitable, and leads them to hire more radiologists. He used the same logic to respond to “AI will replace software engineers”: coding is not an engineer’s job; solving problems is. His most memorable line was: “You may not lose your job to AI, but you will lose it to another person who knows how to use AI.”
He ended with two messages for policymakers and parents. First, put aside the science-fiction framing of “AI apocalypse / Terminator / singularity,” which he directly called “complete nonsense.” He said engineers are making models safer every day and hallucinations have fallen almost to zero compared with two years ago. Second, actively encourage your children, company, and country to embrace AI. Over the past 40 years, only 2% of people could write C++, whereas AI for the first time enables everyone to program in ordinary language. He called it “the strongest force for closing the technology divide” in his forty-year career.