X / Twitter
Swyx, AI Engineer Conference Organizer
Swyx shared updates from the ongoing AI Engineer conference. He said he schedules double-length track keynotes only when confident both speaker and content deserve it; today Chris Manning and abshkbh each gave double sessions. They explored sandboxing and world models, two of the hottest engineering topics. His photos showed packed rooms, and he said online audiences would be more than 1000 times larger. Running the event meant relying on secondhand feedback, but attendees reportedly loved the deeper format. Readers interested in AI engineering should watch these keynotes when recordings appear.
https://x.com/swyx/status/2072562702703046855
Peter Yang, AI Tutorial Creator
Peter Yang’s main topic over these two days was Claude Fable 5’s return. His vibe-check conclusion was direct: it remains absurdly good, a level above every other model, and he hopes GPT 5.6 catches up. He cautioned that subscribers can access Fable 5 only until July 7, a short window. He therefore made a tutorial covering 5 worthwhile use cases: identifying high-value Fable-worthy work, getting life and business advice, polishing projects for release, planning the next big thing, and refactoring a project or codebase. He showed real output rather than talking abstractly. He is also learning to read code and said he would immediately install the explain-diff skill, which looks friendly to beginners. His overall focus is translating a brief access window to the strongest model into concrete activities ordinary people can try.
https://x.com/petergyang/status/2072525669704384612
https://x.com/petergyang/status/2072470191511113732
https://x.com/petergyang/status/2072458983886205333
Google Labs Official
Google Labs announced that MusicFX and MusicFX DJ will shut down on July 31, 2026. The stated reason is to concentrate resources on GoogleFlowMusic, its established tool for creating, sharing, and remixing original music. MusicFX was an early experiment that pushed real-time AI music generation forward. The team will bring its lessons into GoogleFlowMusic, giving music projects a long-term home. Existing MusicFX users have a clear migration signal and need to move workflows before the end of July.
https://x.com/GoogleLabs/status/2072417166952136789
Amjad Masad, Replit CEO
Amjad Masad announced that Replit apps can now be sold directly on Whop. His logic is that with AI making software easier to build, Replit is shifting toward helping founders reach the market, gain their first customer, and earn their first dollar. Whop is one of the internet’s best creator-monetization platforms, and the integration gives Replit applications ready-made distribution and payment channels. This reflects the next phase of vibe-coding platform competition: helping users turn apps into businesses rather than simply generating them. For independent developers, the path from building to monetization has shortened again.
https://x.com/amasad/status/2072385092824260748
Guillermo Rauch, Vercel CEO
Guillermo Rauch posted three updates on deployment infrastructure. Most importantly, Vercel introduced a dry-run step for agent deployments. He observed agents naturally checking work before pushing, frequently running commands such as node --check, tsc --noEmit, and next build. Vercel productized this behavior so agents can rehearse before going live, reducing cost and risk. Second, he demonstrated WordPress on Vercel Fluid using Dockerfile.vercel and Active CPU, with MySQL on PlanetScale. The entire cloud deployment, including docker build, takes only 30 seconds through one vercel command. Third, he recommended ctatedev’s portless tool, saying HTTP simply does not work for local WordPress development and portless solves the problem. Together, he is expanding Vercel’s narrative from frontend hosting to a general deployment layer for the agent era.
https://x.com/rauchg/status/2072463961597878762
https://x.com/rauchg/status/2072463293654942090
https://x.com/rauchg/status/2072398926175404250
Aaron Levie, Box CEO
Aaron Levie used Devin’s new capabilities to explain why AI inference demand will grow 100-fold. Devin’s agentic mapreduce first maps relevant signals across a repository, then dispatches focused agents to bounded shards in parallel, consolidates findings into a report, and confirms severe vulnerabilities only after validating them in isolated sandboxes. Swarms can therefore process enormous datasets and perform tasks previously impossible for humans. This case concerns code security, but he sees the pattern extending across coding and knowledge work. Box customers want to analyze millions of documents for risks, insights, and relationships; unstructured data in pharmaceuticals, banking, and other industries will follow. He added that such capabilities require flexible orchestration of frontier and low-cost models because token volumes are immense, making this a core value proposition for application-layer AI companies.
https://x.com/levie/status/2072519377371459836
Garry Tan, Y Combinator CEO
Garry Tan noted a major Anthropic hire: the head of UC Berkeley EECS, Electrical Engineering and Computer Sciences, has joined the company. He called it a “Mega get” and said “Anthropic is on a tear,” describing its strong momentum. A department-head-level scholar leaving a top university for an AI lab marks another major moment in talent migration. He also reposted what he called the technology industry’s most underrated report. His perspective remains focused on talent and capital concentrating in leading AI labs.
https://x.com/garrytan/status/2072331451270606933
https://x.com/garrytan/status/2072402517397573717
Matt Turck, FirstMark Capital Venture Partner
Matt Turck examined a counterintuitive capital-market phenomenon: in 2026, with AI commanding attention, hot IPO names include AOL, Evernote, Vimeo, and Lime scooters. He focused on how Lime successfully went public despite $1 billion in debt and having publicly expressed substantial doubt about surviving the year. He identified four reasons: financial engineering used IPO proceeds to repay toxic loans and convert remaining debt into equity, cleaning up the balance sheet; Uber owns 22% and directly supplies riders, providing a strong backer and partner; Lime has actually generated positive free cash flow for 3 consecutive years while revenue grows nearly 30% annually; and it survived the micromobility shakeout as the last survivor, with Bird already fallen. His conclusion carried some feeling: never give up.
https://x.com/mattturck/status/2072462125474181623
https://x.com/mattturck/status/2072419592354529712
Zara Zhang, Builder
Zara Zhang shared two practical items. First, a PSA: Codex can switch models to GLM, with screenshots and links showing how to use other models in OpenAI’s toolchain. Second, a method for building skills: do not start with a skill; end with one. Creating a skill is the last workflow step, not the first. First run a process manually until it works smoothly, then capture the established workflow as a skill. This corrects the common mistake of writing a skill immediately. She also posted a full YouTube video explaining the approach.
https://x.com/zarazhangrui/status/2072391971721884073
https://x.com/zarazhangrui/status/2072384777785888875
https://x.com/zarazhangrui/status/2072381929366987087
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari observed that OpenAI and Anthropic’s talent vortex is intensifying: four close friends left very secure positions for the two labs over the past 2 months. The dual appeal is difficult to refuse—helping build one of the era’s most important companies while gaining a chance at liquidity during the pre-IPO phase. He inferred that starting a company now requires extraordinary conviction and ambition. Responding to Xavi, he also offered a self-deprecating VC observation: most VCs do not make big money, at least not life-changing sums, and some join labs after realizing this. As in any profession, only the very best do well.
https://x.com/nikunj/status/2072344802570756121
Peter Steinberger, OpenClaw Creator
Peter Steinberger gave Steve Yegge an unusual shoutout. He never expected to do so, he said, but Yegge had simply been early, like most visionaries, and now everyone is building factories—agent-factory-style automated pipelines. It is belated recognition of Yegge’s early software-engineering automation predictions and reflects growing consensus around industrialized agent orchestration. Peter also invited collaboration, seeking a semi-private San Francisco hack space for himself and several OpenClaw maintainers to rent for a few focused development days, welcoming recommendations by DM. The team’s upcoming SF work session is worth watching for results.
https://x.com/steipete/status/2072532278476148881
https://x.com/steipete/status/2072475858435276840
Aditya Agarwal, General Partner at South Park Commons
Aditya Agarwal observed that he occasionally meets exceptionally smart but instinctively pessimistic people in SF, often from growth-hacking or Wall Street backgrounds. His first reaction: why are you in this city? We run on optimism; are you here to be a contrarian with your sadness? The post captures a central cultural premise of SF technology: optimism is its default operating system, making pessimists outsiders. As an investor who was an early Facebook engineer and former Dropbox CTO, he represents a characteristic builder’s perspective.
https://x.com/adityaag/status/2072449611550380526
Claude, Anthropic’s Official Account
Claude posted detailed rules for Fable 5’s return. All paid plans with included usage can access Fable 5 until July 7, capped at 50% of weekly allowances. After that, users can switch models to spend the remaining allowance or continue Fable with usage credits. An accompanying blog explains the details. Another notable update concerns feedback: Claude Code users whose requests are incorrectly flagged can run /feedback; claude.ai and Cowork users can use thumbs-up or thumbs-down buttons. The account explicitly said feedback will tune classifiers and progressively reduce false positives. For heavy users, the July 7 deadline and 50% cap are key constraints for planning the week.
https://x.com/claudeai/status/2072402642836615273
https://x.com/claudeai/status/2072402640907162072
https://x.com/claudeai/status/2072402639644766602
Podcasts
AI & I by Every — The AI Workflows Behind Every's Consulting Team
Key takeaway: In the AI era you can build almost anything, but the real question is whether you should build and maintain it. Give standard processes to agents and specialist SaaS, leaving people to do what only people can.
Natalia leads Every’s consulting business and manages Claudie, an AI employee that genuinely works every day. Despite a nontechnical background, she trains executives at large companies, hedge funds, and PE firms in AI use, offering a useful window into how top operators work. Over several months, Claudie has evolved from a half-finished system needing engineers ready to rescue it into a mature system with LinkedIn and Twitter accounts, a management dashboard, and a trust-battery mechanism. It runs in a loop evaluating its performance and improving from feedback.
The experience also taught her agent boundaries. Claudie excels at standard operating procedures but still needs ongoing supervision and management. Taste and the pursuit of excellence need human direction, so the team is hiring a human operator while using the agent. More counterintuitively, she abandoned her own vibe-coded CRM to buy Attio and Asana. Maintaining data quality in custom tools is continuous work, and tracking leads over two or three months is precisely Attio’s entire business. Her words were: “In the AI era you can build anything. The question is, should what you build be yours to build and maintain?”
She calls Codex life-changing. For a nontechnical user, embedding terminal and browser directly in chat, alongside a 5.5 model so powerful that “you can feel the compute; it just wants to do heavy work,” removed the need to worry about filesystems and directory structures. Her best loop was giving Codex a goal to complete the CRM from hundreds of customer conversations and email exchanges, sleeping for 6 hours, and waking to work that would have taken weeks already done. Her email app uses her last 150 sent messages to train a ghostwriter that drafts every incoming reply, with one-click options to send, rewrite, archive into customer-specific Markdown files, or turn it into an Asana task. She said: “My inbox now knows what’s going on better than I do.”
The most moving example was spending 13 hours with Codex to build a family system caring for her 81-year-old father. It combines Google Form reports from multiple nurses and scattered WhatsApp updates into a password-protected family portal, with one-click Spanish-English switching and responsibility checklists for each family member. The nurses’ response was not admiration for a flashy tool, but that the family had become more proactive partners.
Her advice to executives has one theme: do not try to become AI-first all at once. The biggest mistake is attempting to remake the entire company in one go. Start with existing systems, document how to do one thing well in Markdown, let AI complete a small task, then add tasks one by one until a system capable of complex work emerges. Dan Shipper’s metaphor closes the discussion: knowledge work is moving from sculpture to gardening. Instead of shaping every work yourself, you create conditions for growth and intervene at crucial moments.
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL