X / Twitter
Boris Cherny, Anthropic Claude Code Team
Boris recommended reading the latest Threat Intelligence report, arguing that as models become more capable, risks rise alongside them when appropriate safeguards and monitoring are absent. Capabilities such as coding and biological research have clear dual uses: they can help develop software or advance science, but can also be used to attack critical infrastructure or create biological threats. On AI-generated code, he advocated setting different standards depending on its intended use. Low-impact prototypes destined to be discarded can be treated as black boxes, while production code should meet stricter quality requirements than human-written code. To this end, Anthropic uses extensive lint rules, tests, Claude-driven end-to-end tests, daily fuzzing, automated code reviews, security reviews, and refactoring processes. He recommended prioritizing the latest frontier model, increasing reasoning effort, and clearly specifying codebase conventions through CLAUDE.md and skills. If quality still falls short, more human guidance is needed, or Claude should be asked to proactively pay down technical debt.
https://x.com/bcherny/status/2098281805770309686
https://x.com/bcherny/status/2098217573276131577
Thibault Sottiaux, OpenAI Codex and ChatGPT Team
Thibault introduced an API for scaling agents on demand, built on infrastructure substantially similar to the system used by ChatGPT Work, with which developers can start building in one minute. He also mentioned a data tool widely used by OpenAI’s internal teams to build dashboards and understand the business, which has become indispensable infrastructure for daily work. Meanwhile, Astra’s system capacity is under pressure. To preserve the experience and continued access of existing users as much as possible, the team will pause new subscriptions to the $200 Pro plan. Existing accounts are unaffected, other subscription plans and the API remain available, and the team is adding capacity as quickly as possible.
https://x.com/thsottiaux/status/2098238138334548260
https://x.com/thsottiaux/status/2098165551554240764
https://x.com/thsottiaux/status/2098113585683808624
Madhu Guru, Senior Director at Meta AI
In part ten of her eval series, Madhu argued that evaluating agents requires measuring every step of task completion, not just the final answer. Even if two agents both arrive at 42, the quality of their processes can differ sharply. One might find the right source, retrieve the correct document, and complete the calculation in 4 clean tool calls, while the other might make 17 calls, repeat searches, and recover from multiple errors. She recommended first defining the complete workflow, then identifying the tasks involved in each step. Next, measurement methods need to be designed for each step, deciding whether to create a standalone eval or treat it as a slice of a larger eval. Evaluation sets should cover both moderately difficult and hard tasks, and analysis should begin with execution traces before turning to the final answers.
https://x.com/realmadhuguru/status/2098064969464217720
Thariq, Anthropic Claude Code Team
Thariq shared a prompt for improving Claude’s personalized context. Users can ask Claude to conduct an in-depth interview about personal experiences it does not yet know about but that are relevant to future collaboration. Open-ended questions can use free text, while multiple-choice questions can use the askuserquestion tool. Information gathered during the interview is then saved to memory for use in future conversations. The focus of this method is not a one-time addition of background information, but having Claude proactively identify gaps and build more complete long-term context.
https://x.com/trq212/status/2098157600361861579
Google Labs, Google’s AI Tools and Experiments Team
Google Labs announced that Dreambeans is now available to all US users aged 18 and older. The product has launched on both iOS and Android and is free to use, with no subscription required. Users can now also connect the Gemini app to Dreambeans. Once connected, Dreambeans uses Gemini’s understanding of existing chats to generate more insightful and personalized daily stories. The product is positioned to deliver a fresh set of generated stories every day, tailored to the user’s context.
https://x.com/GoogleLabs/status/2098110018289803558
Amjad Masad, Replit CEO
Amjad acknowledged that AI brings many real risks and singled out cybersecurity as a major concern. He believes these risks deserve ongoing attention and action. However, he opposed classifying “extinction risk”—a scenario in which AI causes the death of all humanity—as a realistic threat. In his judgment, this risk is not on the same level as ongoing, observable problems such as cybersecurity. His position advocates focusing attention on concrete, actionable AI risks rather than centering the discussion on human extinction.
https://x.com/amasad/status/2098171265924116732
Guillermo Rauch, Vercel CEO
Guillermo proposed providing every agent with a computer in every region, reflecting the demand agent workloads create for distributed computing resources. Vercel completes approximately 10 million deployments daily and has reached a cumulative total of 2.35 billion. Billions of application deployments can coexist and be routed to through Vercel’s CDN, while the underlying global metadata store can synchronize across regions within hundreds of milliseconds. The team recently improved the system’s p99 performance by 91% while also accelerating the build-to-deploy pipeline. This optimization came amid rapid growth in agentic deployments and enormous pressure on the system, directly improving operations such as rollbacks, configuration changes, and route additions.
https://x.com/rauchg/status/2098158541932794222
https://x.com/rauchg/status/2098091056302833837
https://x.com/rauchg/status/2098066258155708851
Aaron Levie, Box CEO
After speaking with dozens of technology leaders in banking, media, information services, insurance, and consulting, Aaron summarized several common trends in enterprise agent deployment. Companies’ most immediate concerns are cybersecurity, growing numbers of vulnerabilities, and agent identity management. Meanwhile, most companies continue to deploy multiple frontier models in parallel, finding it difficult to standardize on a single model or provider. Large-scale adoption of open-weight models in major organizations remains at an early stage, and enterprise spending is concentrated among a small number of providers. Higher ROI comes from redesigning workflows so agents change the work itself, rather than simply being layered onto existing processes, but it remains unclear who should lead this organizational change. Enterprises also frequently switch technology architectures and providers, eval capabilities remain immature overall, and legacy systems with fragmented data continue to slow implementation. Box is also deepening its partnership with OpenAI, enabling users to work securely with enterprise content in ChatGPT while AI agents process data and execute workflows across systems.
https://x.com/levie/status/2098218284139311615
https://x.com/levie/status/2098135659714085281
Matt Turck, VC at FirstMark Capital
Matt published an extensive conversation with Richard Socher about recursive self-improvement and scientific progress. Starting with the slowdown in scientific progress, they discussed whether AI could reconnect human knowledge fragmented by intense specialization. They then examined how LLMs learn biology and proteins, and why next-token prediction might produce an effective world model. Later topics included simulation, verifiers, AI hallucinations, drug discovery, cancer research, energy, materials, and an AI economist. The conversation also covered scientific data bottlenecks, virtual cells, autonomous robotic laboratories, agent swarms, compute constraints, and what Recursive plans to build first.
https://x.com/mattturck/status/2098081448330674182
https://x.com/mattturck/status/2098081452839637480
Nikunj Kothari, FPV Ventures Partner
Nikunj summarized three patterns in the current early-stage venture capital market. First, many founders want to raise $50 million seed rounds. Second, many teams expect to reach $30 million in ARR next year. Third, popular seed rounds raised in stages often end up at valuations of around $300 million. He also demonstrated his rapid writing workflow: recording a voice memo during his commute, adding to it throughout the day, then spending 30 focused minutes writing between two meetings, giving it a quick read-through, and publishing immediately. He noted that Claude’s voice transcription still needs improvement.
https://x.com/nikunj/status/2098078391065018816
https://x.com/nikunj/status/2098255116751257663
Peter Steinberger, OpenClaw and OpenAI Builder
Peter agreed with a software design judgment for the age of AI coding: duplicated logic is no longer as painful as it once was, while abstraction still carries costs. Generative coding has made it easier to copy and rewrite implementations, changing the benefits of traditionally introducing abstraction layers to avoid repetition. Abstraction, by contrast, still requires developers to understand boundaries, maintain interfaces, and bear the long-term constraints of poor design. This means classic principles such as DRY may need to be reconsidered in environments where agents generate code. The central reminder is that eliminating repetition should not automatically lead to introducing abstraction; concrete implementations can sometimes be easier to generate, inspect, and replace.
https://x.com/steipete/status/2098089196800098798
Claude, Anthropic’s AI Assistant
Claude announced that Fable 5.1 Build Days would begin this week. The Claude community will host buildathons in multiple cities around the world. The events run from September 11 to September 25. Participants can arrive with a specific problem or product idea, or simply join and explore what Fable 5.1 can do. The official announcement also provided an RSVP entry point to help developers find and register for events in the relevant cities.