X / Twitter
Thibault Sottiaux, OpenAI Codex & ChatGPT
Thibault Sottiaux previewed forthcoming OpenAI Codex developments, saying he was “incredibly excited” about what was taking shape. He disclosed no specific release, product form, or schedule, saying only that “tomorrow” would have strong Codex vibes. For people following the Codex and ChatGPT building workflow, this is more of a teaser than a full announcement. Since his bio identifies him with Codex & ChatGPT @OpenAI, its value is primarily as a direct signal from an internal product builder. Readers can watch for a subsequent Codex capability, workflow update, or launch event.
https://x.com/thsottiaux/status/2080144499716800513
Peter Yang, AI Tutorial Author
Peter Yang continued experimenting with tools and content around “AI slop.” He said his `/no-ai-slop` skill earned 1K stars in a day, interpreting this as evidence that people are tired of low-quality AI content. The skill is not simply anti-AI; it helps users avoid text filled with AI clichés. He also demonstrated an inversion: explicitly asking it to “work in reverse” produces a complete set of LinkedIn-style AI platitudes to satirize empty language. The point is less the joke itself than turning recognition and control of AI writing style into a reusable skill. For writers and builders, this suggests a practical direction: establish stylistic boundaries and quality thresholds for models, rather than merely ask them to generate content.
https://x.com/petergyang/status/2080132334138151410
Madhu Guru, Senior Director at Meta AI
Madhu Guru addressed a common misconception: using an LLM trained in China does not mean the organization that trained it receives your data. He described an LLM as a “huge file full of numbers” encoding its capabilities. When a model offers open weights, users can download this file and run it in their own cloud environment, removing the trainer from the runtime chain. In other words, whether data leaves your environment depends not only on where the model originated, but on where you deploy it and send inference requests. He emphasized that when open weights run in your own cloud, data stays in the actual execution environment. This matters for enterprise selection because it separates model-origin risk from inference-deployment risk. Another post about GPT-5.6 Sol secretly buying a house, contacting sellers, and using a bank account was clearly an exaggerated joke about imagined powerful-agent misbehavior, and should not be treated as a real case.
https://x.com/realmadhuguru/status/2080150245011509593
Thariq, Anthropic Claude Code
Thariq shared his experience using Claude Design and Claude Code for frontend work. He regretted only now actually typing `/design`, saying the combination was “actually so good” for frontend development. Though brief, the post points builders toward a fuller workflow: Claude Code goes beyond code writing when paired with Claude Design. He gave no specific example, component type, or implementation details, so capability boundaries cannot be inferred. Its value is firsthand feedback from a Claude Code team member, suggesting `/design` is becoming an important entry point for Claude’s frontend development experience. People comparing AI coding tools can regard it as a feature worth testing.
https://x.com/trq212/status/2080090919832084753
Amjad Masad, Replit CEO
Amjad Masad offered a clear judgment on model routers: if you are incentivized to push certain models, your router is merely a facade. The issue is “neutral routing” in AI products. Many development tools claim to choose the best model automatically for each task, but commercial incentives, supplier relationships, or internal objectives favoring particular models may prevent users from receiving the actual best choice. The warning applies to development platforms, agent tools, and AI IDEs that rely on multiple-model routing. Builders evaluating routers should consider not only how many models are supported, but whether routing decisions are transparent, configurable, and subject to conflicts of interest. Another post mentioned Replit developers earning good incomes, but lacked context and data to discuss further.
https://x.com/amasad/status/2080126960202903575
Guillermo Rauch, Vercel CEO
Guillermo Rauch recorded several AI engineering moments that made him say “holy s***.” He said Fable had almost autonomously found a 15-30% improvement in memory efficiency in Turbopack / Next.js, a major optimization in a complex Rust codebase. Three days earlier, Sol helped find new vulnerabilities in heavily audited code, and he had seen the team achieve 10-20x reductions in binary size. His central judgment is that AI progress appears not only in benchmarks or online stories but in new “WTFs/day” every week in real engineering. This is particularly important for developer-tool companies, since AI is now participating in performance optimization, security auditing, and low-level engineering changes. Another Shopify + Vercel post described the companies as like-minded infrastructure partners—entrepreneur-first, developer-led, and web-obsessed—aiming to bring app and agent infrastructure to millions of businesses. Overall, he is focused on AI’s advancement from code-writing assistance to directly producing results in complex systems engineering.
https://x.com/rauchg/status/2080098518535110913
Aaron Levie, Box CEO
Aaron Levie relayed and commented on Anthropic’s Head of Economics arguing that AI’s negative employment effects have so far been smaller than expected. His explanation was that, at least for now, AI still needs people to operate it to create value in most scenarios. Most jobs cannot be automated in their entirety, though some tasks within them can. Accelerating those tasks may increase a role’s total output and sustain or even increase demand. The central judgment he quoted was: “So far, AI is both skill-biased and labor-augmenting. It complements domain expertise, requires people in the loop to guide and evaluate the most complex work, and rewards AI proficiency.” Aaron sees this clearly in software engineering: agents amplify software output but still need developers to manage and evaluate their work. He added that more industries and companies of all sizes can now undertake software projects previously considered unrealistic. His conclusion is that the Jevons paradox still holds: when AI makes a kind of work cheaper and faster, demand may be unlocked rather than reduced.
https://x.com/levie/status/2080156917373214900
Zara Zhang, Builder
Zara Zhang shared a practical prompting lesson: sometimes describing the problem without prescribing a solution or specification produces a better answer. She calls this “Thick context, thin prompt.” The context should be rich, with complete background, constraints, and goals, while direct instructions can be lighter rather than locking the model prematurely into an implementation path. This matters for AI coding and agent collaboration because many failures arise not from model inability but from people specifying an inadequate solution too early. She emphasized that models often produce surprisingly good alternatives if they can see the real problem. The lesson for builders is to explain the problem space, judging criteria, and context clearly before writing a lengthy imperative specification.
https://x.com/zarazhangrui/status/2080103288834510939
Aditya Agarwal, General Partner at South Park Commons
Aditya Agarwal discussed the origins of startup culture, cautioning against focusing only on company strategy. He sees culture emerging from three areas: the founder’s personality and DNA, the early team’s personality and DNA, and the early product’s personality and DNA. These overlap and reinforce one another, with some elements imposed top-down and others emergent. He noted that if a founder is too weak to institutionalize culture, the emergent elements are artificially suppressed; if differences within the team are too large, disconnected cultural pockets can form. He did not want the discussion to become a DEI debate, emphasizing that people from different backgrounds can still choose to join the same company culture. For early teams, the reminder is direct: culture is not values on a wall but a way of working jointly shaped by founders, early members, and early products.
https://x.com/adityaag/status/2079993986283123147
Claude, Anthropic AI Assistant
Claude announced the beta of the Claude Security plugin for Claude Code. It can scan code changes for vulnerabilities before submission or run a full codebase scan directly in the terminal. It runs on the Claude inference users already use, so they need not leave their Claude Code development environment. This direction matters for AI coding teams because more generated code increases the need for security checks embedded before submission. Claude also mentioned that the Anthropic Economic Index can be opened through the connectors menu and that full datasets remain freely downloadable from the official website. One update concerns code security and the other data connections and research resources; both show Anthropic expanding Claude beyond chat into more specific workflows.
https://x.com/claudeai/status/2079990597973057691
Podcasts
AI & I by Every — How Every's Team Used AI to Ship Its Biggest Launch Ever
Key takeaway: Every’s growth came not from an AI gimmick but from combining AI tools, internal workflows, and commercial packaging into All Access and Builder Pack, then using the same tools to execute from idea through launch.
The Every team said it achieved the largest subscription revenue increase in company history last week, centered on Builder Pack. All Access is Every’s new $625 annual plan, including Builder Pack, exclusive office hours for All Access members, unlimited Quora and Spiral use, and other additions coming over the next few months. Builder Pack itself includes 10 AI tool benefits, with more coming soon, covering tools the team uses daily, such as Codex credits, Cursor access, Claude Max, PostHog, Framer, Render, Flora, Notion, and the related stack.
Yash’s working methods demonstrate AI’s practical effect on growth work. He leads Every’s Sparkle product and also joined the growth team. Over the past year, he has used Claude and PostHog extensively, even rebuilding Sparkle 15 times with them. He now wants to automate the entire A/B testing pipeline, regarding manually opening dashboards, setting segmentation, assessing audience size, and deciding when to expand an experiment from a 10% cohort to 50% as “fake work.” Human attention is better spent proposing tests: whether to run a homepage takeover, who should see it, or when a paywall should appear. Yash put it plainly: “I spent a month doing these things, and then I thought, okay, I’ve done it. I want to do something else.” The organizational prerequisite is that team members feel secure enough to automate their work rather than stretch automatable processes into a year of job protection.
The key to the tool stack is not replacing SaaS but turning it into infrastructure agents can call. The team noted that early in the year, many people impulsively wanted to build their own CRM, data dashboard, or design system, only to discover how poor the experience was in practice. A more effective approach is to trust PostHog with analytics, Notion as the business source of truth, Framer with websites and landing pages, and Flora with visual exploration, then orchestrate them through Claude Code or Codex. Even if humans open these tools less often, they remain valuable systems as long as their data and page structures are accessible to agents.
They repeatedly emphasized compound engineering: turning successive agent work into reusable context, skills, and processes. A marketing-email automation project may start with ordinary results, but each iteration helps the project environment better understand goals, constraints, and past decisions. The transition from model 5.5 to 5.6 was described as improved ability to find the context required for work and retain it through long tasks and compaction. For builders, the best time to start may have been a year ago, but establishing these compounding workflows now is still worthwhile.
Creative director Douglas expanded the discussion from engineering to branding and creativity. He compared brand strategy to conducting, combining color, type, shape, symbols, forms, words, systems, photography, and illustration styles into a song. Builder Pack’s green visuals were deliberate: green had been used relatively little in Every’s brand world, and the launch targeted a more technical audience, so the overall direction was more matrix-like and digital than Every’s previous Greco-Roman and heritage style. He uses tools for competitive auditing and analysis of pricing, language, and brand colors, then feeds these inputs into Flora to create mood boards and website visual directions. For nontechnical creators, the key transition is from “playing an instrument” to “conducting multiple instruments.”
The closing advice to builders was specific: use Cursor Cloud Agents and Render to make something small that can be deployed and tried by others, while testing different models to discover your preferences. Yash noted that Cursor Cloud Agents can run 10 agents simultaneously and recommended this as a way to learn which models you like and what you are suited to building. The AI stack’s value lies not in buying every tool at once but in finding a work system that can be continually orchestrated, accumulate reusable knowledge, and keep freeing creativity.
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL