X / Twitter
Swyx, Latent Space Host
Swyx connected Sam Altman's advice to "build a business that gets better as models get better" with his own "Agent Labs" concept. He observed a clear discontinuity in Q4 2025: revenue at agent-lab companies became directly tied to underlying model performance. The clip came from Patrick Collison's Stripe Sessions. In another post, he said the team ran an internal bake-off between Exa and competing web-search products and unanimously chose Exa after only 1.5 hours. He specifically mentioned William Bryk, Exa's founder and his former landlord, suggesting that revisiting William's pre-PMF comments a year ago teaches how to recognize generational companies early.
https://x.com/swyx/status/2057119153337545096
https://x.com/swyx/status/2057180080078791036
Google Labs, Google's AI Lab
Google Labs announced that Project Genie is now available worldwide to Google AI Ultra subscribers aged 18 and over. The selling point is designing games, not playing them: choose a character, set a scene, and let the model do the rest. The demo clearly illustrates the shift from "playing games" to "generating games in minutes." It is also available to try in the sandbox at Google I/O. This marks a key transition from internal demos of generative 3D and world models into paying users' hands and could become a differentiator for the Google AI Ultra subscription.
https://x.com/GoogleLabs/status/2057179491693470166
https://x.com/GoogleLabs/status/2057218835074437573
Guillermo Rauch, Vercel CEO
Rauch offered a short teaser: this will bring AI to 42% of the web, covering every model, provider, and modality—text, images, video, and audio. Judging from the tone and figure, it appears to be a forthcoming Vercel product providing cross-provider AI access and unifying multimodal, multi-model capabilities, most likely the next AI SDK version or an expansion of v0-related products. The 42% figure presumably refers to Vercel/Next.js web penetration, connecting distribution directly to AI models.
https://x.com/rauchg/status/2057212335811620987
Aaron Levie, Box CEO
Levie wrote a long post about FDEs—Forward Deployed Engineers—emphasizing that the role will endure in the AI era rather than quickly become standardized away like deployment roles during the cloud era. His argument has three layers. First, cloud affected a relatively narrow set of developers and IT users without requiring business workflows to change; deploying agents directly reshapes daily workflows and thus involves substantial change management. Second, cloud moved slowly enough for best practices to settle; every model update can make existing agent scaffolding redundant or even obstructive. Third, this makes vendors' or partners' experience from hundreds or thousands of implementations more useful than internal teams' experience. His career advice: FDE is an excellent path for early-career technical talent and will not disappear soon.
https://x.com/levie/status/2057315272156135501
Garry Tan, Y Combinator CEO
Garry endorsed Exa, saying YC internally and all his own OpenClaw and Hermes Agents use it for agent web search, with no alternative matching its speed, reliability, and completeness of coverage. Another post was sharper: he believes the benefits of American AI are already available, but AI leadership could be lost unless public-opinion "smoke screens" are addressed, alleging that certain forces are systematically undermining the United States. This can be read in the context of his political stance as an "SF Dem accelerating the boom loop."
https://x.com/garrytan/status/2057202833251000503
https://x.com/garrytan/status/2057168328226230520
Builder Zara Zhang
Zara proposed a view of AI-native team structures: ICs should start thinking like managers about delegating to agents, setting acceptance criteria, and validating output, while managers should think like ICs and build products themselves rather than only manage people. At Google I/O, she also shared a T-shaped diagram from a talk on "developers in the AI era" and extended it to every role: deepen domain expertise vertically, broaden into adjacent skills and fields horizontally, then add a layer of "knowing how to use AI" across the top. These are two sides of the same judgment: purely execution-focused and purely management-focused positions will both shrink, while hybrid roles gain leverage.
https://x.com/zarazhangrui/status/2057324988009685208
https://x.com/zarazhangrui/status/2057267931025957348
Nikunj Kothari, FPV Ventures Partner
Nikunj used Claude to summarize SpaceX's 308-page S-1 prospectus into key notes for people without time to read it all. Another post was more emotional: he said being a founder is extraordinarily hard, like chewing glass every day; nobody cares when you fail, and success brings more problems. He has never had the courage to become a founder himself, so he wrote specifically to those persevering: "I see you. Respect." He emphasized that entrepreneurship is heavily romanticized; actually building, scaling, and sustaining a company is different. Choose a direction you would pursue "even if the whole world opposed you."
https://x.com/nikunj/status/2057242868293816569
https://x.com/nikunj/status/2057134939875991973
Dan Shipper, Every CEO
Alongside this week's major news—Anthropic acquiring developer-tools startup Stainless at a rumored $300M price, despite its largest customers being OpenAI and Google—Dan resurfaced his October AI & I interview with Stainless founder and CEO Alex Rattray, disclosing that he is a small shareholder. He extracted five MCP lessons. First, keep servers lean: few tools, precise names and descriptions, minimal inputs and outputs, plus a JSON filter to remove irrelevant fields. Second, use dynamic mode for large APIs, exposing only list endpoints, pick endpoint, and execute endpoint so the model navigates itself. Third, treat MCP as a business copilot, connecting Notion, HubSpot, and internal databases for cross-system questions such as "which customers signed up last week?" Fourth, build a company shared brain with Claude Code by storing useful inputs such as customer feedback and SQL queries in a GitHub repo, gradually creating a searchable internal archive. Fifth, MCP's future is code execution: give models two tools, code execution and doc search, rather than hundreds, letting them write and run SDK code and consult documentation when errors occur.
https://x.com/danshipper/status/2057122805657821240
Sam Altman, OpenAI CEO
Sam listed the three things OpenAI is most excited about: AGI accelerating scientific research, AGI accelerating companies, and personal AGI accelerating everyone toward their own goals. He said that the day before, OpenAI had announced a $2M investment in OpenAI credits for every YC company, and today it announced the unit distance result—a general-purpose model solving an open mathematical problem—so the focus now turns to the third goal. In a separate post, he called a general-purpose model solving a major open problem in mathematics a "kinda big milestone." He expects to say this repeatedly over the coming years, but described mixed feelings on this particular day.
https://x.com/sama/status/2057218997503086888
https://x.com/sama/status/2057203171198636251
Podcasts
AI & I by Every — Inside Stainless: The Developer Tools Startup Anthropic Just Bought for $300 Million
Key takeaway: The best current approach to MCP is not giving models hundreds of atomic tools, but just one code execution tool and one doc search tool, letting models write SDK code to complete tasks. Stainless is already betting on this direction.
Alex Rattray is Stainless's co-founder and CEO. The approximately 50-person company was a developer infrastructure provider building SDKs and API tools for companies including OpenAI, Anthropic, and Google before Anthropic acquired it on Monday for a rumored $300M. He and Dan Shipper are college friends, making the conversation more relaxed than a typical technical interview as they openly discussed "why MCP is hard to use today."
First, most current MCP implementations wrap REST endpoints one-to-one as tools. Giving a model a Stripe-style dashboard with hundreds of endpoints can burn hundreds of thousands of tokens just translating the OpenAPI spec, exceeding context capacity and confusing the model. MCP that works must be handcrafted: package "find Dan, issue a refund, and send a discount code" into one semantically clear composite tool, rather than a platter of atomic operations.
Second, he offered practices usable today: keep tool counts low; make names and descriptions precise; use a small, clear input schema; return only data the model needs. They use JQ filters for JSON. For very large APIs, Stainless introduced dynamic mode, exposing just list endpoints, get endpoint, and execute endpoint so the model locates its target in three steps. The trade-offs are slower speed and slight lossiness, but stable context.
Third, his everyday MCP use is business operations, not coding. He connects Notion, HubSpot, Gong, and a read-only internal Postgres replica to Claude Code and asks, "Which interesting customers signed up for Stainless last week?" The model joins across systems to produce a complete answer. He also shared a counterintuitive practice: have Claude Code casually save useful inputs such as customer quotes and SQL queries into Markdown files in a private Git repo, gradually building a "company shared brain." It is deliberately unstructured: "Claude Code is already very good at handling unstructured things."
Fourth is his most contrarian judgment: "I expect the code execution tool to become the most widely used tool." Rather than hundreds of atomic tools, give models only two: code execution, letting them write TypeScript to call SDK methods such as stripe.customers.retrieve and stripe.refunds.create, and doc search for when they need instructions. Pagination and chaining across endpoints happen inside the sandbox, with only the final ten lines returning to model context. The prerequisite is static typing in the SDK; otherwise nothing stops a model from hallucinating a nonexistent stripe.transactions.list.
Fifth is the security model. Alex does not think security should rely on "limiting what MCP exposes": if the underlying API permits an action, restrictions above it are self-deception. The right layer is the API, using OAuth and granular scopes. He acknowledged that OAuth scopes are painful to implement and hopes someone makes them easier. Dan repeatedly invoked early Cloud Code versus Codex to persuade him: Cloud Code began in YOLO mode, even offering dangerously skip permissions directly, so developer adoption was fast; early Codex CLI was too restrained and pair-programming-oriented, falling far behind. He wanted the same logic applied to Stainless, urging Alex to release a developer version for individuals quickly. Alex said it was underway, starting on the MCP-provider side, locking sandbox outbound networking to one API domain before gradually expanding.
His final framing suits anyone building AI products: "The future of AI is cyborg." An agent is not a black-box model but a combination of an LLM and traditional CPU code. If a one-off AI behavior proves repeatedly useful, the next step is committing it to a repo as production software. There is no boundary between a one-off agent call and long-lived production software, only a difference in "how many times it gets used."
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL