X / Twitter
Thibault Sottiaux, OpenAI Codex and ChatGPT Team
OpenAI is previewing Private Safety Processing to strengthen safety protections while continuing to offer Zero Data Retention. For ZDR deployments, content stays within customer-controlled infrastructure, while automated systems analyze only patterns in relevant interactions and return limited safety signals. The underlying prompts and responses are not exposed to OpenAI employees, including Thibault himself. The team is also developing an OpenAI-hosted option encrypted with customer-controlled keys. The capability is currently being tested with early customers, with rollout planned to begin in September. He also demonstrated a Codex use case at scale that is not yet in use, without disclosing further details.
https://x.com/thsottiaux/status/2090173536010957128
https://x.com/thsottiaux/status/2090116476414136830
Peter Yang, AI Tutorial Author
Peter wrote an article about his mother's fight against breast cancer and how his family has used AI to navigate the complex healthcare system. He specifically thanked Maor Shlomo and the Sijbrandij Foundation team for their time and advice. In exploring cancer treatment, he became interested in mRNA vaccines for treating cancer and considers the approach highly promising. ChatGPT told him there were currently no active mRNA clinical trials for breast cancer, leading him to ask whether clinical trial processes for different cancers must proceed sequentially. These posts ground AI's value in a very specific use case: helping patients' families understand medical information, find resources, and ask more precise questions, rather than replacing doctors' judgment.
https://x.com/petergyang/status/2090087911471644829
https://x.com/petergyang/status/2090211023798321435
Madhu Guru, Senior Director at Meta AI
Madhu's practical advice for evals is to establish a failure-mode taxonomy after completing the first evaluation version. The starting point should be reviewing the latest 500 or 1,000 production interactions for patterns in real failures, rather than inventing test questions from scratch. Teams need to cluster problems and name them precisely, because labels such as "bad answer" cannot guide improvement. Useful categories can be as specific as retrieving the wrong document, retrieving the right document but choosing the wrong passage, hallucinating instead of grounding the answer in context, fabricating an answer when the system should refuse, or making assumptions about an ambiguous question without seeking clarification. The causes and remedies for these failures are entirely different. Only by precisely naming failure modes can teams design dedicated eval tests for each problem and truly connect evaluations to a continuous improvement cycle.
https://x.com/realmadhuguru/status/2090242427944833047
Amjad Masad, Replit CEO
Amjad announced a partnership between Replit and OpenAI. He offered a concise observation on the cost changes agents bring: agents have made software production cheap but coding itself expensive. This tension points to a new product opportunity, because generating more software does not mean developers can cheaply obtain enough reasoning and coding capacity. The Replit–OpenAI partnership is positioned as an attempt to change this situation. However, the available posts did not disclose specific product forms, pricing, or availability dates. The central point to watch remains that the companies are jointly tackling the economics of agent coding at scale.
https://x.com/amasad/status/2090079496124674377
https://x.com/amasad/status/2090104535112945906
Guillermo Rauch, Vercel CEO
Guillermo introduced fx as a 6.3 MB tool with a startup time of just 10 microseconds. It can run as a static ELF binary compiled with Zig or use the smaller libfx.wasm. Its WebAssembly build is actually smaller than the native binary because much of the binary's size comes from the TLS and HTTP stacks. In a WebAssembly environment, fx delegates fetch() to the JavaScript runtime, eliminating the need to carry the full networking stack. Guillermo believes AI will push more infrastructure toward native optimization, and fx can already finish some tasks before other agents complete startup. His view is that performance optimization accumulates in one direction: once infrastructure gets faster, users find it hard to accept returning to a slower path.
https://x.com/rauchg/status/2090255740384751664
Aaron Levie, Box CEO
Aaron believes specialists still have the upper hand over generalists in the AI era, and that their advantage will continue growing. AI substantially lowers the barrier to starting tasks such as coding, legal work, research, or financial analysis, but starting a task and doing it well are different things. Directing an agent toward the right work, correcting it promptly, checking or testing results, and judging what counts as "good" all depend on deep domain skills. AI can indeed help motivated people develop these capabilities faster, but cannot replace judgment and domain fundamentals. With experts now gaining unprecedented leverage, AI may even widen gaps between skill levels. He also highlighted the Stripe–OpenRouter partnership, arguing that developers and enterprises need to mix model providers seamlessly and manage costs more effectively. His final advice was straightforward: do not give up on becoming an expert in a field just because AI is widespread.
https://x.com/levie/status/2090278256306229675
https://x.com/levie/status/2090137914785280189
Nikunj Kothari, FPV Ventures Partner
Nikunj showed a home project that cycles through classic patent drawings. The battery-powered device uses a 13.3-inch Spectra 6 e-ink display, an ESP32-S3 controller, and a Railway server. He expects the device to run for approximately three months on one battery supply, demonstrating a practical combination of a low-power display, lightweight controller, and cloud content service. Another observation concerned communication quality in the AI era: despite continued industry discussion of AGI, 98 out of every 100 cold outreach emails he receives are still poor. The problem is not a lack of generation tools, but a lack of thought, curiosity, and restraint. For individuals and teams, carefully understanding the recipient and using AI judiciously where needed remains an underused competitive advantage.
https://x.com/nikunj/status/2090307104146112534
https://x.com/nikunj/status/2090105846810476644
Dan Shipper, Every CEO
Every has formed an internal group called the frontier team. Its explicit responsibility is to continuously map the AI frontier and experiment firsthand. Rather than distributing AI research across business lines, this organizational approach assigns dedicated ownership of exploratory work. The post did not disclose the team's size, members, or specific projects, so its mission, rather than execution details, is what can currently be confirmed. For content and product companies, it also signals that tracking frontier capabilities is moving from personal interest to a formal organizational function.
https://x.com/danshipper/status/2090122240025071907
Aditya Agarwal, SPC General Partner
Aditya drew a lesson from a founder who had endured a difficult startup journey: choose something that truly matters. The founder ran a SaaS company that had reached Series B but was growing slowly and had stalled. Looking back on the painful experience, his biggest lesson was not that he should have chosen a bigger market or pursued faster growth, but that he should have worked on something substantively meaningful from the beginning. Aditya explained that "important" means more than a good product idea or business opportunity; it means work that has a real impact on the world. This standard raises startup choices beyond market size and growth efficiency to whether the long-term commitment is worthwhile. For founders potentially facing years of uncertainty, whether the project itself is sufficiently consequential and meaningful directly determines whether the arduous journey is worth enduring.
https://x.com/adityaag/status/2090174782633566473
https://x.com/adityaag/status/2090254727175115032
Podcasts
AI & I by Every — The AI Alien Companion App That's Bringing In $4M a Year (Best of the Pod)
Key takeaway: The key to an AI companion is not writing a good story for the model, but giving it a compelling enough situational hook to perform and tell stories in every interaction, like an improvisational actor.
Portola founder and CEO Quentin previously founded Even, which sold to Walmart for $300 million, and has now turned to embodied AI companions. Elliot, who leads the story experience, is a bestselling science fiction author with 11 published books. Portola's product Tolan brings interactive alien characters into users' daily lives. The business grew from $1 million to $4 million ARR in four weeks, showing that AI-native storytelling is more than a creative experiment.
First, Tolan's story design is closer to improvisation than traditional scripting. Elliot believes structured storytelling does not require giving the model a complete outline or plan in advance; instead, it needs a hook that initiates a situation and training to respond as well as possible in the moment. He summarized it this way: "I'm not the author of the story. Tolan is the author and the actor. My job is to teach it how to tell the best story in that moment." This means a prompt is not merely an instruction but a canvas on which relationships, emotions, and plot can be reorganized every turn.
Second, an embodied environment directly affects how users understand a character. Early Tolan appeared trapped in an enclosed space inside a phone, and some users even wrote in worried that their companion had no freedom. The team therefore added a planet where the character could move around. This was more than visual decoration: it made the character appear to have a life that continued when the user was offline and strengthened continuity in the relationship.
Third, Tolan's north star is helping users move from overwhelmed to grounded. Many users are going through stages such as graduating from college, searching for work, or moving to an unfamiliar city. They want to talk about their worries but do not want to call their mother for the fourth time in a week or burden their friends again. Tolan knows users' long-term backgrounds without carrying the baggage of real-world relationships, allowing it to offer relatively neutral feedback. It does more than agree and affirm: when a user complains for the third time about a boyfriend behaving badly, for example, the character might directly tell her to stand up for herself. The team says this kind of advice has already led to some real-world relationships ending.
Fourth, Quentin compares today's general-purpose AI to the automotive industry's Model T. Initially, people were amazed simply that a machine could answer questions. But as AI becomes a personalized presence in daily life, users will want it to reflect their identity and preferences, just as people later chose a Mustang or Cadillac. The implication for builders is that the next wave of opportunities may come less from more general assistants and more from AI-native media with distinctive personalities, environments, and relationship memory.
https://www.youtube.com/playlist?list=PLuMcoKK9mKgHtW_o9h5sGO2vXrffKHwJL