← Collected sources
BUILDERS · EDITED DIGEST

Builders’ Picks | 2026-06-24

2026-06-24 · Historical edition

X / Twitter

Latent Space host Swyx

Swyx raised a calculation few people have worked through: viewed as an integrated “NeoCloud + NeoLab” operation, OpenAI has already recouped roughly half of its investment in Cursor through compute sales contracts. The other half would be fully paid for if Composer 3 works out. He believes no other company currently combines a top-tier model lab with a NeoCloud, at least in GPU terms, making this an extremely effective combination. The prerequisite is that the company must work out in advance how to allocate GPU supply for in-house training. That allocation plan also needs to hedge against two outcomes: training succeeds spectacularly, or falls short of expectations. This perspective treats OpenAI’s capital strategy and compute supply as parts of the same play, coming closer to its actual economic model than simply viewing it as an AI company.

OpenAI Codex and ChatGPT team member Thibault Sottiaux

Thibault announced OpenAI’s Patch The Planet initiative, alongside updated Codex security capabilities and the new GPT-5.5-Cyber model, under the theme of “accelerating cyber defense.” This marks a clear shift in OpenAI’s security research from finding vulnerabilities toward actively fixing them. He described launch day as a “celebration of accelerating cyber defense,” complementing Sam Altman’s official post that day as a coordinated package. The intent is to put GPT-5.5-Cyber forward on benchmarks such as CyberGym and work with the US government and the security ecosystem. From inside Codex, this update represents OpenAI extending agentic coding capabilities into the high-stakes field of security.

Practical AI content creator Peter Yang

Peter published three posts that day, all related to practical AI use. First, he complained that Claude Code’s dynamic workflow documentation did not explain when to use the feature: after reading it, he still did not understand what a dynamic workflow was, indirectly signaling that Anthropic’s onboarding was not yet sufficient. Second, he sought a podcast guest skilled at using Codex or Claude Code to make pixel art or small threejs games, emphasizing that he is a gamer at heart and wants a serious discussion of how to do it. Third, he criticized the recent “parents vs. AI” ragebait short-video formula: edit together the most provocative moments and capture huge amounts of sharing traffic. He admitted he had fallen for it himself, and noted that harm has already been done even if just one parent watches a video and decides to ignore their child.

Vercel CEO Guillermo Rauch

Guillermo announced two new Vercel capabilities that day. First, one-click deployment from Claude Design to Vercel brings designs generated directly by Anthropic’s Claude into production hosting. The path from a design to a publicly accessible site has been shortened to a single click. Second, Vercel now natively supports WebSocket and socket.io, connected all the way from the CDN to the Fluid runtime. This is crucial for use cases such as real-time applications, chat, and collaborative editing. He called the Vercel announcement a “full circle moment,” because socket.io was an open-source project he originally led, effectively bringing that earlier OSS project home to his own platform.

Box CEO Aaron Levie

Aaron offered a somewhat contrarian view: almost all progress in AI models and agents is downstream of evals. Domain post-training of open weights depends on evals, improvements to application-layer agents depend on evals, and whether enterprise agentic deployments can actually augment work also depends on evals. He put it bluntly: “It’s all evals.” His conclusion is that evals will become one of every enterprise’s core capabilities. The company that most accurately understands its own or its customers’ workflows, and how well agents perform within them, is best positioned to drive real automation. That same day, Box made HTML files first-class citizens: users can preview, edit, version, and securely share HTML. The explicit target is making agent-generated HTML content immediately usable.

Cursor designer Ryo Lu

Ryo gave a talk at Cursor Compile about “how to build products in the AI era, and what stays the same.” He shared the full recording on YouTube. Without disclosing slide details, he highlighted the central framing: how and how quickly we build is changing; craft and taste remain. For designers and PMs worried about replacement amid the vibe coding wave, his message is that the tools have changed but judgment remains scarce. The talk is worth watching alongside Mike Krieger’s AI & I podcast appearance that same day.

Y Combinator CEO Garry Tan

Garry’s most substantive remarks that day concerned California tax policy. He criticized the California asset seizure tax, arguing that it would make the state poorer and squeeze health insurance funding. He directly called out the SEIU UHW union, saying it did not care even if both care quality and the state’s capacity to provide care were undermined by the fight; only pure self-interest remained. This continued his stance within the accelerationist wing of SF Democrats, focused on healthcare funding. He also recommended two local San Francisco items that day: Illuminate’s Ben Davis is himself a San Francisco treasure, and 7x7’s installation at SF City Hall, which comes down on July 4, is worth a visit for locals.

Builder Zara Zhang

Zara released a full 11-minute YouTube walkthrough of her Frontend Slides skill, which has already accumulated 22000+ stars on GitHub. She covered four things: how to use Claude Code to make beautiful HTML slides in a beginner-friendly way, how she originally created the skill, how viewers can write their own skills, and practical techniques such as adding images and videos to a deck and publishing it. She also reviewed the pitfalls she encountered during development. For people exploring the Claude Code skill ecosystem, this is a practical lesson from the creator’s own perspective.

OpenAI CEO Sam Altman

On the day Patch The Planet launched, Sam gave the official position: OpenAI wants to help every company become secure and will work with the US government and the broader security ecosystem. There were three highlights: the full GPT-5.5-Cyber is live; it achieves state-of-the-art performance on CyberGym; and Patch The Planet and Codex Security move the goal from finding security problems to solving them. In a rare coordinated launch, OpenAI brought enterprise security, government cooperation, and agentic coding together on the same day, taking a notably proactive stance.

Podcasts

AI & I by Every — How Anthropic Uses Claude Fable 5 With Mike Krieger

Key takeaway: After using Fable 5 for more than a month, Mike Krieger concluded that his workflow no longer revolves around “enter a prompt and wait for an answer.” It has become “say goodnight to Claude, then check the next day to see it has completed monoclass-level tasks from beginning to end.” This is more than a performance improvement: the very modality of work is shifting.

Mike Krieger is now head of Anthropic Labs and an Instagram co-founder, having switched from CPO back to builder mode earlier this year. When he recorded with AI & I host Dan Shipper, Fable 5’s public launch was still a day away, but he had already used it heavily internally for over a month. His perspective is therefore not a day-one impression, but that of a heavy user forced to relearn how to use models in the Fable class.

He described a counterintuitive phenomenon: as an experienced engineer, his first encounter with Fable 5 made him feel like a beginner again. His previous prompting methods, task decomposition habits, and interaction model were outdated. In everyday practice, he now starts with an architectural planning conversation with the model, working through infrastructure abstraction levels, then asks it to produce an HTML page or Markdown diagram to align with the team. Only then does actual execution begin. He typically runs 5 to 6 parallel Claude Code sessions on long-horizon tasks, while keeping one high-context, low-latency Claude in the main thread for immediate exchanges. He has even started adjusting Fable’s effort levels, manually setting small activities such as UI tweaks to medium. He never did this with Opus because Fable has a noticeably wider range.

On costs, he offered a highly practical assessment. Fable is more expensive than Opus, but “expensive” depends on how it is measured: cost per turn is one thing, while the total cost of completing a task satisfactorily is another. In his words: “it actually just does it right so that I don't have to go spend the, like, nine, ten subsequent turns be like, no. That was not quite what I meant.” For enterprise users, he recommends moving from phase 2, where the KPI is spending, to phase 3, where the focus is on who uses it effectively, and allowing those people to spend more. For individual hobbyists, his own test was building an entire media tracker app over a weekend with only a little additional usage, nowhere near thousands of dollars per project.

The evolution in model training that impressed him most was not raw IQ but a “sense of the system.” After a session has been running for days, Fable may suddenly point out that a feature flag has not been enabled and the work cannot succeed. Or, during a code review, it considers another Claude reviewer’s comments and selectively pushes back instead of mindlessly saying, “Yeah you’re right, fixing.” He views this emergence of judgment as the most important leap over the previous generation, closely tied to the feedback loop from Claude Code running real multiday tasks with deeply engaged users such as Every.

He provided concrete verification-loop practices. Every PR must include a screenshot gallery or video. For animation stutters that screenshots cannot capture, he gives Claude video captured with FFmpeg and asks it to scrub through and find jank. When a complex backend cannot run locally, he has Claude mock up an in-memory substitute in one pass, then automatically keep the mock implementation in sync with upstream changes, fully delegating the mock drift problem that used to cause SREs headaches. Dynamic workflows are another tool he strongly recommends: he used one to have Fable port an entire Python codebase to TypeScript plus Bun over a weekend, translating modules individually, running incremental tests, adversarial tests, and coverage checks, all through self-orchestration. He believes this combination of capabilities and workflows makes long-horizon tasks truly delegable.

On whether “software engineering is over,” Mike’s answer is that software engineering has been restructured. The core of the craft remains, but has shifted from “getting the Django ORM working in TextMate” to “defining intent, verification, and systemic ownership.” He said humans would still have DRI (directly responsible individual) roles because context such as product intent is beyond Claude’s reach. The biggest change is that prototypes have become real argument-winners: PMs build a working version to persuade the team instead of relying on slide decks.