X / Twitter
Swyx, AI Builder
Swyx previewed an in-person demo event for SF Personal AI engineers, focused on personal agents, at the new media lab this Thursday evening. He specifically mentioned Shlok attending, calling him one of the best builder-writers he had met this year. This is no ordinary meetup: featured speakers from the previous similar event were later acquired by Amazon’s hardware division. More interestingly, Swyx remains a daily active user of that project two years later, showing his interest in long-term retention rather than launch excitement alone. For personal-agent builders, real demos, early feedback, and offline networks remain important amplifiers. His tone also conveys his view of PAI: personal AI products with lasting vitality are those that continually enter daily use.
https://x.com/swyx/status/2077243443391422813
Thibault Sottiaux, OpenAI Codex & ChatGPT
Thibault Sottiaux repeatedly discussed usage growth for ChatGPT Work, Codex, and GPT-5.6 Sol, signaling rapidly rising demand. He said the team might soon reach 9M and asked whether to reset ChatGPT Work and Codex usage again or give users more room, showing that allowances and capacity are now part of the experience. He then offered a feedback incentive: users explaining why they like GPT-5.6 Sol or switched to it could receive $100 in Codex credits, with free tokens for the first 10k users. This combines growth, word of mouth, and feedback in one loop, using migration reasons to assess the appeal of models and tools. He also clarified that he was not announcing a reset but collecting ChatGPT Work improvement suggestions on Twitter. Overall, OpenAI is pushing Codex beyond a development tool toward a more frequently used work entry point, with growth, allowances, feedback, and product complexity becoming central challenges.
https://x.com/thsottiaux/status/2077271889626706300
Peter Yang, Practical AI Tutorials and Interviews Creator
Peter Yang previewed a video explaining how he uses ChatGPT Work, or Codex, for nearly everything on his computer. It covers a complete setup in 7 steps, from choosing the right GPT-5.6 model to managing email, calendars, and recurring tasks. The substance lies in treating Codex as part of a personal operating-system-level workflow rather than demonstrating isolated automation. For busy users, such tutorials reduce the transition cost from asking AI questions to having it execute everyday computer tasks. His explicit use of ChatGPT Work and Codex as names for tools in the same use case also reflects an ongoing transition in naming and user understanding. Another post about a Spain player lacked AI-builder information and has been skipped.
https://x.com/petergyang/status/2077196815951417649
Thariq, Anthropic Claude Code
Thariq shared a concrete Claude Code use case: after recently playing Pokemon Champions, he began using it to research battles. He asked Claude Code to write code with Smogon’s npm library, retrieve live usage statistics, and generate reports explaining matchups, breakpoints, and theoretical team builds. Rather than conventional software development, this turns a coding agent into a game-strategy analyst, combining data retrieval, scripting, statistical organization, and explanation. For builders, it shows coding agents expanding from “help me write an app” to “help me make any structured system analyzable.” Personal-interest scenarios can also provide low-risk testing grounds before agent workflows mature.
https://x.com/trq212/status/2077051280267399550
Guillermo Rauch, Vercel CEO
Guillermo Rauch announced that Vercel will release an AI Gateway dataset on AI token flows, saying it contains interesting insights. Its value for builders is the potential to reveal token flows across models, applications, and calling patterns rather than a single product’s benchmarks. He also promoted installing agentmail on Vercel: have an agent execute `vercel install agentmail`, with no signup, automatic setup, and unified billing. The direction is clear: make third-party capabilities agents need into installable, billable, low-friction infrastructure. AI Gateway supplies traffic and model access, while agentmail adds communications, both helping agents enter production. The broader point is Vercel consolidating installation, observability, and commercialization paths for the agent stack within its platform.
https://x.com/rauchg/status/2077176141790752798
Aaron Levie, Box CEO
Aaron Levie offered a key judgment on agent adoption: code is particularly suitable because it can be tested relatively quickly, by running applications manually or executing tests. Most knowledge work reveals its outcome only after real-world delivery, such as executing a stock trade, concluding contract negotiations, or delivering a sales pitch. He expects new opportunities to test other work in software-like ways, enabling more agents in enterprise workflows. The biggest problem today is that most knowledge work lacks evals to show whether changes to models, prompts, or systems improve or worsen results. His conclusion is direct: enterprises best able to evaluate their knowledge work will likely benefit most from AI. He also discussed a proposed AI standards body, distinguishing it from a regulator and suggesting it could accelerate standards and industry collaboration, provided the industry first aligns somewhat on safety risks. Both points concern the same requirement for enterprise diffusion: testable workflows and standards collaboration faster than traditional government processes.
https://x.com/levie/status/2077201458546745553
Ryo Lu, Cursor Design
Ryo Lu wrote a long reflection on “when a dream becomes a job,” especially relevant as AI enters writing, coding, design, and reasoning. He acknowledged the dream of turning a hobby into work: being paid to care about what you love, alongside people who value the same details, and making imagined things into products others can touch. The pain comes when that once-private love is taken over by deadlines, teams, customers, strategy, money, reputation, and momentum: curiosity becomes a roadmap, taste becomes decision-making, and play becomes output. AI adds deeper anxiety by approaching activities once considered closest to the inner self, including writing, coding, designing, reasoning, and taste-like decisions. He did not reduce this to job replacement, but identified a challenge to creators’ identity stories: if machines can do what I am good at, which part is truly me? His answer is that AI can accelerate output and raise the floor of craft, but cannot want things for you or decide what deserves your love. The future discipline for builders is not merely working harder, resting more, or caring less, but repeatedly returning to where the love began and preserving its source before audiences, roadmaps, and deadlines.
https://x.com/ryolu_/status/2077162119506833627
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari offered an engineering-management observation suited to the times: pre-AI engineering leaders may struggle to understand that the strongest engineers can also be extremely online. While agents work, X provides a dopamine spike during the wait. This reframes engineers’ X use from distraction to an interstitial behavior in agent workflows. Management traditionally linked focus time closely to screen behavior, but when agents execute code, tests, or research in parallel, human work rhythms become more fragmented. He predicted time on X will continue rising, especially in technology. The reminder for managers is to reconsider productivity signals rather than judge AI-native engineers through old models of online behavior.
https://x.com/nikunj/status/2077144910508257317
Dan Shipper, Every CEO
Dan Shipper offered two levels of feedback around the Codex Desktop app launch. He first shared receipts from the launch vibe check, showing Every’s ongoing observation and documentation. He then emphasized that regular Every readers would have known 6 months ago that Codex might take off, since he had become “codex pilled” early and Every had begun sustained coverage. This is less a celebration of one product than a claim about media and research organizations’ ability to track early technical shifts. Codex’s surge did not happen suddenly, he argued; it had been building through developer habits, agent toolchains, and desktop workflows. For builders, valuable trend signals often emerge among frequent users and specialist communities before mainstream consensus.
https://x.com/danshipper/status/2077196796586025327
Aditya Agarwal, South Park Commons General Partner
Aditya Agarwal raised a product tradeoff in the new ChatGPT app. He appreciated its deep feature set but saw a real cost. Previously, he queried ChatGPT Legacy 15 to 20 times daily; the new experience feels too heavyweight for this frequent, lightweight use case. The feedback does not oppose stronger features but highlights the tension between power-user workflows and quick queries. The more a product becomes a work platform, the more it may sacrifice the original ease of opening it and asking a question. AI product teams should preserve low-friction entry points, especially for small questions asked more than a dozen times each day.
https://x.com/adityaag/status/2077130899733553560
Sam Altman, AI Builder
Sam Altman said GPT-5.6 Sol’s growth is extraordinary and praised the inference team’s heroic work supporting demand. OpenAI will continue moving mountains to expand capacity, he said, but short-term hiccups may occur. The key point concerns capacity and inference infrastructure as core product-experience variables, not model capabilities. When demand exceeds a system’s comfortable operating range, users may encounter availability, reliability, or waiting-time problems rather than model shortcomings. His framing connects growth, inference teams, and expansion, showing that the real post-launch challenge is sustaining service at scale. For builders, inference capacity and reliability are competitive strengths alongside model quality.
https://x.com/sama/status/2077106587307798989
Claude, Anthropic AI Assistant
Claude introduced core capabilities and privacy commitments for Claude for Teachers. It is built for K-12 privacy, does not train models on teachers’ conversations, and protects student information through a FERPA-compliant data processing agreement. Teachers can request lesson plans, with Claude starting from state standards and high-quality curricula and connecting resources through Learning Commons. It then generates lesson plans and student-facing materials for teachers to revise and take into class. The positioning bundles standards, curriculum resources, material generation, and privacy compliance for teachers rather than offering a general chatbot. For education builders, Anthropic is reducing adoption barriers through specific institutional requirements and connections to teaching resources.
https://x.com/claudeai/status/2077047282109714488
Podcasts
Training Data — Anthropic's Katelyn Lesse & Angela Jiang: Building an Ecosystem, not a Walled Garden
Key takeaway: Anthropic Platform aims to make knowledge, execution, and coordination into composable infrastructure for developers and enterprises, rather than turn Claude into a closed application.
Katelyn Lesse and Angela Jiang lead Anthropic Platform, the external API and developer platform that also underpins Anthropic’s own products. This unusual position serves outside builders creating applications and systems with Claude while helping internal teams quickly ship “AGI pilled products.” They deliberately distinguish two north stars: speed and leverage internally, and enabling any builder to use Claude in their particular business context externally.
The most important product philosophy is using the same primitives internally and externally wherever possible. Angela noted how quickly AI product forms change: two years ago everything was thought to be chat, now it is agents, and more forms will emerge. Anthropic does not assume it has the exclusive answer; it wants a stable, powerful platform through which the market can explore. Katelyn added that the team dogfoods internally while giving outside customers early access, avoiding overfitting to internal users’ special needs.
The abstraction layer is moving from a stateless messages API toward agentic work. Katelyn recalled that when she joined roughly a year ago, the platform was primarily the messages API plus developer tools such as MCP, SDKs, documentation, and a console. As models ran longer and handled more context, customers and internal teams repeatedly solved the same problems: making agents reliable in long-running, remote settings, sometimes without humans in the loop. The platform therefore began packaging sandboxes, governance, security, and execution infrastructure into higher-level abstractions.
Angela’s hierarchy is clear: beyond knowledge and execution comes coordination. She expects increasing use of strategies, a kind of meta harness assigning different jobs to different tokens—for example, some advise while others execute. Her key statement was: “token has a job.” This grounds concepts such as agent swarms in a concrete principle: define each portion of intelligence’s responsibilities rather than simply add more agents.
An open ecosystem is reflected in modular execution choices, not just rhetoric. Katelyn said that with Claude Managed Agents, Anthropic does not insist everything run on its own infrastructure. It has introduced self-hosted sandboxes and partnerships with Modal, Vercel, Cloudflare, and Amazon’s new micro VMs, letting developers plug in their own execution environments. MCP tunnels follow the same approach, letting enterprises call MCP servers behind firewalls.
Higher-level standardization includes interoperability and safety. Angela compared AI to electricity: transformative utility comes not from one company alone but from universal access, standards, and ways to plug in. She described skills and MCP as interoperability standards at the builder layer, while the industry also needs collaboration on cyber threats, fraud, and critical-infrastructure protection at higher levels. Her position treats safety as a practical concern of companies that do not want bad actors exploiting their services, rather than abstract philosophy.
For builders, the next stage of Claude platform competition may be about intelligence per dollar and whether it can enter real products safely, modularly, and with governance, rather than model APIs alone. Enterprises need security, compliance, and access inside walled gardens; weekend developers need open, hackable experiences. Anthropic wants to serve both, which creates the challenge: maintain an opinionated agent architecture while letting the ecosystem plug in its own infrastructure, memory, sandboxes, and workflows.