X / Twitter
Thibault Sottiaux, OpenAI Codex and ChatGPT
Thibault Sottiaux continued collecting ChatGPT Work use cases, this time shifting from productivity to more personal effects. Inspired by direct messages about ChatGPT Work, he wanted another kind of story that would delight the team: when had ChatGPT had a “deeply positive impact” on your life or someone close to you? This is not a product announcement, but shows OpenAI actively seeking feedback beyond feature metrics. It asks about concrete changes in real life, relationships, learning, and areas beyond work, rather than simply which features users like. For builders, it is a reminder that an AI product’s value story comes not only from benchmarks, efficiency gains, and workflow integration, but also from turning points users voluntarily share.
https://x.com/thsottiaux/status/2079058139207573541
Peter Yang, AI Tutorial Creator
Peter Yang continued offering specific feedback on Codex and ChatGPT Work. Codex is helping him with a longstanding Google Adsense payment issue: despite his YouTube channel having 100K+ subscribers, he has never received a dollar from Google Adsense. The poor support experience has left him trying to resolve it through other Google contacts. Another comment directly addressed product copy: many nontechnical ChatGPT Work users, he believes, do not know what “run this chat in the cloud” means. Codex prompted him to send something to Codex Web, but the link then told him to download the app, making the flow repetitive and confusing. He was not questioning cloud execution itself, but noting that copy and navigation determine whether nontechnical users understand a feature. Overall, his posts returned practical AI adoption to two concrete issues: whether support can solve real business problems and whether nontechnical users can immediately understand product language.
https://x.com/petergyang/status/2079053957532655890
https://x.com/petergyang/status/2079053505969676404
https://x.com/petergyang/status/2079007381695172797
Cat Wu, Anthropic Claude Code and Cowork
Cat Wu shared her specific prompt for managing a calendar with Claude Cowork. She asked it to manage a week of appointments under clear constraints: keep total meeting time below 20 hours, deduplicate conflicting meetings, infer which meeting types she normally declines from previous weeks, and exclude dinners from the 20 hours. She also asked it to build a continually refined skill and consult her before updating invitations. The value is in combining personal preferences, historical behavior, hard constraints, and human confirmation into an iterative work agent rather than a one-off calendar cleanup. Agent users can especially learn from asking it to study past meeting refusals, which is more actionable than simply “optimize my calendar.” Cat ended by asking how others use Cowork, showing Anthropic gathering agent-skill patterns from real work.
https://x.com/_catwu/status/2079011428380602526
Thariq, Anthropic Claude Code
Thariq posted two updates about Claude Code and skills/system prompts. The first was a fix notice: users encountering the relevant bug should restart Claude Code, as the fix should be propagating. Although he did not explain the bug, it was explicit practical advice for affected users. The second was more notable: he is writing an article on what the team has learned and how users can apply those lessons to their own skills and system prompts. The point is that the Claude Code team is turning internal practices into reusable methods, beyond releasing features. For builders, skills and system prompts are gradually moving from prompting tricks into product-level configuration and workflow engineering.
https://x.com/trq212/status/2079103743535280508
https://x.com/trq212/status/2078901672441790818
Amjad Masad, Replit CEO
Amjad Masad explained the difficulty of building large consumer subscription software businesses through household spending patterns. Consumers mainly spend on food, rent, entertainment, phones and internet, and shopping, while companies typically buy software. Beyond a few entertainment subscriptions such as Netflix and Spotify, it is therefore hard to identify truly enormous consumer subscription software businesses. This matters especially for consumer AI products, many of which seek individual subscription payments despite users lacking the natural “software budget” found in enterprise procurement. He was not saying consumer subscriptions are impossible, but warning builders against casually applying enterprise payment logic to individuals. AI pricing must either fit an existing frequent spending category or demonstrate enough value to create a new personal budget item.
https://x.com/amasad/status/2079086360703680583
Guillermo Rauch, Vercel CEO
Guillermo Rauch considers cybersecurity one of the best benchmarks for superintelligence. He called it software engineering’s “IQ test,” because real security work includes discovering and patching vulnerabilities, reverse engineering, and exploitation—abilities that transcend particular languages, runtimes, and frameworks. Having a model generate an XYZ clone in one shot can easily impress people on X, but is not a good test. Security tasks require genuine reasoning and “corner thinking”: thinking through boundary conditions, exceptional paths, and gaps between systems. He added that the best engineers he has worked with usually have a deep security background or interest. Seeing Kimi K3 perform well on these tasks makes him more optimistic about open models.
https://x.com/rauchg/status/2078912929714356698
Aaron Levie, Box CEO
Aaron Levie discussed open weights, AI costs, and application-layer implementation in succession. First, he argued that when strong open-source alternatives trail frontier models only slightly, restricting enterprises’ access to frontier capabilities makes them less secure and competitive. Even if the US could ban open-model access entirely, other ecosystems would face no equivalent restrictions and could use those models for defense or against US companies. Regulation must therefore reconsider the fact that open-weight models are already stronger than many expected. Second, he challenged the intuition that falling AI costs reduce AI spending, expecting the opposite: cheaper AI enables more coding, more bug and security reviews, and agents processing previously unmanageable datasets. Anything reducing token costs will consequently raise inference demand for the foreseeable future. This also explains why open-source AI business models remain viable: most people run models on infrastructure rather than local devices. Third, AI’s diffusion ultimately depends on how quickly it can interact with reality. Coding adoption is fast because one person at a computer can write, test, run, and create value, whereas life sciences, sales, contracts, and jet-turbine blade design require real-world feedback. His conclusion is that model output alone is insufficient in most industries. The opportunity lies in the applied AI layer: embedding intelligence in existing workflows and managing industry-specific feedback loops.
https://x.com/levie/status/2078992778449850769
https://x.com/levie/status/2078968158006939716
https://x.com/levie/status/2078864191683969212
Zara Zhang, Builder
Zara Zhang proposed that code and software can now be disposable. She gave three examples discarded after use: a personal design playground or modal for fine-tuning a design’s look and feel, an HTML page to understand code, and a temporary dashboard to check something once. This fits today’s AI coding environment, where the marginal cost of generating one-off software is falling and not all software needs long-term maintenance. Another post addressed creator positioning: anyone starting on social media and unsure what to discuss should recall the three questions friends, acquaintances, and colleagues ask most often and turn the answers into videos or posts. She recommended speaking as naturally as when answering in person, because that is positioning in itself. An opinion repeated more than 3 times is worth making into content. One reason she began posting, she added, was fatigue from repeatedly answering the same questions offline.
https://x.com/zarazhangrui/status/2078835308905578660
https://x.com/zarazhangrui/status/2078830510177128481
Podcasts
The MAD Podcast with Matt Turck — Stripe's AI Chief: How AI Agents Will Buy, Sell, and Pay
Key takeaway: Agentic commerce is more than “AI buying things for you.” The larger shift is agents becoming economic actors that discover services, buy resources, sell output, and even run a small piece of business as micro firms.
Emily Sense is Stripe’s head of data and AI. She discussed infrastructure Stripe is already building for an AI agent economy rather than distant science fiction: how agents discover products, merchants expose catalogs, inventory, and prices, consumers authorize payments, and agents execute transactions safely. Over the past year this has become more concrete. Stripe is working with Google to let merchants sell within AI mode and the Gemini app; with Microsoft and OpenAI to put merchants into Copilot and ChatGPT discovery and purchase journeys; and with Meta to place checkout inside ads.
She described agentic commerce as a spectrum from low to high autonomy. At the low end, humans still make the main decisions while AI helps execute transactions. In the middle, users search an AI surface for “shoes for runners with flat feet” and receive an answer with a buy button. At the highest level, agents autonomously discover services, decide to buy, and complete transactions without real-time human involvement. Emily said most consumer use remains at level two: people delegate some choices to AI but are not yet at the point of booking an entire summer holiday in one go through an LLM.
The most concrete commercial signal is the partner list. Platforms include Wix, Shopify, BigCommerce, and Commerce Tools; brands include Best Buy, Coach, URBN, and Kate Spade. Emily’s judgment is that if AI surfaces have become product-discovery entry points for large numbers of users, merchants must appear there rather than optimize only traditional search, ads, or on-site commerce journeys.
On risk, she particularly emphasized token theft. She said: “In AI, fraudsters don’t necessarily need to steal money or credentials. Stealing tokens is enough.” More strikingly, in the data she has seen, over one-sixth of new registrations at AI companies constitute this kind of abuse, resembling a dine-and-dash with tokens as the stolen goods. This directly challenges billing, risk controls, and onboarding because abusers target inference resources themselves rather than fraudulent card charges.
On pricing, she believes per-seat pricing is breaking down. AI costs and value do not always grow linearly with seats, so enterprises need more granular usage controls, ROI assessment, and guardrails. Employees should indeed use high-ROI LLMs, but companies must know when costs accumulate substantially and whether corresponding returns exist.
Her final 12-month outlook was the most imaginative: agents will become multifaceted economic participants, not just buyers. They will buy, sell, provision infrastructure, combine services, and deliver results externally. She said the more interesting future is not “Emily authorizes an agent to buy for her,” but “Emily has an agent assigned to run a business, including buying some things, selling others, and making a profit.” For builders, the missing pieces are those the old human-centered commerce stack did not cover: identity, authorization, discovery, settlement, abuse prevention, and machine-oriented commercial protocols.