X / Twitter
Swyx, AI Product Practitioner
Swyx added batch questioning to `/align-me` to reduce the interaction cost of turn-by-turn confirmation between people and AI. Inspired by Matt Pocock and Thariq Shihipar, the design lets the system explore 2 to 10 steps ahead at a time. The intuition resembles speculative decoding: anticipating possible branches improves overall speed. Swyx said this works particularly well in design exploration. It suggests agent products can treat human feedback as a scarce resource and group clarification questions to reduce repeated waits.
https://x.com/swyx/status/2088073777779515615
Boris Cherny, Anthropic Claude Code Team
Boris Cherny is experimenting with having Claude take over routine application maintenance and said the past few weeks have shown early signs of feasibility. The team set up a dedicated Slack channel where Claude Tag checks iOS, Android, Desktop, Web, CLI, and Agent SDK every day. Automated tasks include finding crashes in simulators, consolidating duplicate abstractions, identifying and removing dead code, and fixing leaky abstractions. Over a few weeks, these routines created 388 PRs, of which 180 were merged after Claude Code Review and human review. Claude can usually complete mechanical changes in one pass. When it fails, the team adjusts the routine's prompt so it improves over subsequent days. The new question has shifted from "can AI maintain applications?" to "how can we lower the cost of merging these PRs?"
https://x.com/bcherny/status/2088014489438621990
Thibault Sottiaux, OpenAI Codex and ChatGPT Team
Thibault Sottiaux recommended installing the Computer History plugin to have ChatGPT analyze and roast a day's computer usage. The example record showed Slack accounting for 48% of activity, with the user sending 253 messages while clicking "Clear" 339 times. The system also counted 1,191 Delete presses, 133 copies, 26 pastes, and more than 200 app switches. This humorous summary turns attention switching, notification burdens, and inefficient operations into observable behavioral data. He also introduced the ability to operate Google Docs, Sheets, and Slides directly inside ChatGPT, letting users write, brainstorm, and proofread through text or voice without leaving their workflow.
https://x.com/thsottiaux/status/2088133823619895712
https://x.com/thsottiaux/status/2088103609477238858
Peter Yang, AI Tutorial Creator
Peter Yang raised a product documentation question for the agent era: how can product specifications remain concise while being understandable to both people and agents? He is also considering separate sections for the two audiences. This reflects specifications' shift from team communication material to a context interface for AI execution. A separate personal experience suggests AI's practical value in healthcare may not first come from disease research. When dealing with a family health issue, he found AI more useful for understanding and handling the healthcare system's administrative processes. Both observations point to a reality: AI's value often emerges from process friction between information and action.
https://x.com/petergyang/status/2088108304274960667
https://x.com/petergyang/status/2087946170274570385
Madhu Guru, Senior Director at Meta AI
Madhu Guru counted more than 20 AI products named "Studio," including Google AI Studio, Vertex AI Studio, Copilot Studio, Agentforce Studio, SageMaker Studio, and LangGraph Studio. She used this to point out how an industry that emphasizes limitless creativity converges heavily in product naming. More substantially, she called "prompt debt" the new tech debt. A common path is adding ten rules after a model fails, ten examples after tool calls fail, and more formatting constraints after anomalous output, until the system prompt becomes a novel. As models improve, these historical constraints may stop helping and instead turn smarter models into mechanical rule-followers. She recommends cutting at least 50% of the prompt at every model update to revalidate which constraints remain necessary. For AI product teams, prompts are not just configuration to keep expanding; like code, they need regular refactoring and cleanup.
https://x.com/realmadhuguru/status/2088074515188519182
https://x.com/realmadhuguru/status/2087916590964851172
Cat Wu, Anthropic Cowork Team
Cat Wu is gathering firsthand feedback from non-engineering users for the Cowork team. Target users include professionals in marketing, sales, finance, legal, and ops. Participants can book 15-minute office hours to directly demonstrate how they use Cowork. The team hopes to identify improvements by observing specific workflows. This feedback method focuses on actual interaction processes, rather than just collecting abstract feature assessments.
https://x.com/_catwu/status/2088006642189361564
Amjad Masad, Replit CEO
Amjad Masad believes ARC-AGI-3 is close to being solved with a coding harness. He sees the crucial improvement as giving the model an environment for writing and running code, rather than simply replacing the model. His conclusion is that coding generalizes LLM capabilities across a broader range of tasks. This also highlights evaluation results' sensitivity to system scaffolding: the same model can perform very differently in different harnesses. For agent developers, tools, feedback loops, and execution interfaces may be as important as foundation-model capabilities.
https://x.com/amasad/status/2088124774824521786
Guillermo Rauch, Vercel CEO
Guillermo Rauch predicts that configuring coding AI tokens and runtime environments through a single command will become the default way to use coding agents at scale. Key requirements he listed include uptime, model choice, lower costs, observability, and ZDR. The approach focuses on compatibility with existing coding harnesses rather than requiring teams to change how they work, with both Claude Code and Codex supported. He also recommended trying free GLM 5.2 from Blackbox AI through the related service, at speeds of up to 500 TPS. Together, these posts point toward standardizing coding AI infrastructure by separating model access, routing, and governance from individual agent tools.
https://x.com/rauchg/status/2088020529039180204
https://x.com/rauchg/status/2087982033499042205
Aaron Levie, Box CEO
Aaron Levie considers "AI will eliminate engineers" a deeply mistaken assumption. AI is more like a new power tool for engineers, helping teams build products they already wanted to create but lacked the resources for. Drug discovery automation, manufacturing automation, and larger software projects still require engineers. As AI enables enterprises to undertake more work, the domain experts overseeing that work become even more important. Even as models keep improving, experts are generally better than novices at using them effectively. He expects this pattern to extend beyond engineering, with expertise becoming more valuable as AI expands the boundaries of work.
https://x.com/levie/status/2088105350201270529
Matt Turck, FirstMark VC
Matt Turck described the current startup market as two extremes. At one end are AI-native rocketships that must keep raising at ever-higher valuations to compete for capital, talent, and customers. To win, these companies may sacrifice gross margins and remain in prolonged head-to-head competition with other rapidly growing companies. At the other end are companies the market does not see as high-growth AI prospects, which may be overlooked by investors even when their businesses are good. He believes this divergence has persisted for some time and recently intensified again. His observation reminds founders that high-growth narratives bring not only fundraising advantages but also pressure to continually replenish capital and sustain growth expectations.
https://x.com/mattturck/status/2087978386195103916
Builder Zara Zhang
Zara Zhang noted that many people predicted AI coding would reduce engineers' value, yet many sought-after jobs still have "engineer" in their titles. She listed forward-deployed engineer, design engineer, product engineer, and growth engineer. These roles typically go beyond writing code, embedding engineering capabilities directly into customer, design, product, or growth problems. As AI lowers some coding costs, engineering methods can enter more parts of the business. Her view runs counter to the narrative of engineers being replaced and instead suggests engineering capabilities are being redistributed within organizations.
https://x.com/zarazhangrui/status/2088087765267386564
Nikunj Kothari, FPV Ventures Partner
Nikunj Kothari discussed whether agents should use a single controller or multiple specialized roles from a product design perspective. Grok Bot favors the latter, giving different bots their own tools, context, and goals, with users choosing the appropriate one for each task. This resembles division of labor in an organization and accommodates current limits on context windows, tool use, complexity, and cost. Another approach is closer to Jarvis: users interact with one master agent, which creates, coordinates, and reads other bots in the background. Nikunj expects products to continue starting with single tasks initially, but potentially move toward a master agent that orchestrates multiple bots over the long term. He also described using a Matic robot at home to show that useful household robots can already become part of family life. The observations concern software agents and physical robots respectively, but both emphasize that users ultimately care about reliably achieving goals, rather than how many separate components the underlying system contains.
https://x.com/nikunj/status/2087906119914340540
https://x.com/nikunj/status/2088029329624371544
Podcasts
No Priors — What Chess.com Teaches US About Superhuman Capabilities, with CEO Erik Allebest
Key takeaway: Even though computers have long surpassed humans at chess, human engagement, skill development, and community experiences can still support an enormous product market.
In 2005, Chess.com CEO Erik Allebest and a friend bought the chess.com domain at a bankruptcy auction for $56,000. He was attending Stanford Business School at the time and had met numerous investors, but was widely told chess was an uninvestable niche. Without raising money, Allebest in a product role and his friend as CTO began building a community product. After launching in 2007, they gradually became profitable through memberships for online chess learning and always expanded in line with cash revenue.
Chess.com now has approximately 10 million daily active users, 40 million to 50 million monthly active users, more than 250 million registered members, projected annual revenue exceeding $200 million, and a team of approximately 650. Growth was not a temporary spike from a single event. COVID, The Queen's Gambit, schoolchildren, short-video content, the Mittens bot, and cheating scandals created successive traffic waves, but the user baseline after each decline remained higher than before. Allebest's long-term ambition has expanded from running a chess community to involving 1 billion people in chess.
The key product strategy is continually improving the player experience, rather than raising money early to capture the market. The early team chose free play and direct browser access, using the experience they genuinely wanted as chess players as their standard. Allebest's advice to founders is similarly unconventional: "Follow your heart, work out what you want to exist in the world, and pursue it relentlessly." His team chose remote work, no fundraising, no paid user acquisition, and a market considered too small—almost the exact opposite of the startup template at the time.
Within Chess.com, AI was first used to shorten the distance from "finding a problem" to "shipping code." It assists with support automation, data analysis, specification writing, and agentic development, and the company has built GNS, an internal authentication and knowledge layer with permission controls. On the product side, it is developing a portable AI coach that aggregates player data, compares users with players at similar and higher levels, and identifies improvements needed to advance. The prototype coach can also analyze the user's past week of games, proactively offer feedback, and chat, though Allebest clearly believes it does not replace human coaches for users needing deep guidance.
Chess.com is also extending its chess experience into poker and other classic games. Its poker rating seeks to measure "how good a player really is" by combining the levels of others at the table, chips won, and the number of hands played, rather than merely looking at who puts in more money. Allebest believes ratings may ultimately become as important as money because they represent players' judgments about their own skill and worth. The broader lesson is that superhuman AI does not automatically erase the meaning of human skill. Technology can improve education and solve social problems or worsen wealth concentration; outcomes still depend on how people establish rules, distribute benefits, and take responsibility for governance.