OpenAI Codex and ChatGPT product lead Thibault Sottiaux
Thibault Sottiaux says the team has made steering respond instantly. When a user makes an adjustment, the model can change its ongoing work more quickly. This enables real-time course correction and reduces effort spent continuing in the old direction. He also announced GPT-6.1 Sol ultrafast and says the two work well together. The post emphasizes interaction and execution speed, without reporting specific latency or comparative tests.
Meta senior director of AI Madhu Guru
Madhu Guru argues that every layer of the computing stack needs rebuilding for agents, creating substantial startup opportunities. In recent years, APIs, connectors and MCPs have mainly focused on making enterprise software agent-ready. Personal agents are now raising the same questions for consumer products, with e-commerce an early example. He also highlights desktop and mobile operating systems that treat agents as first-class users, cloud infrastructure agents can operate directly, and identity, permissions and security for thousands of agents acting on people’s behalf. Interfaces will need to serve humans, agents, or both working together. His suggested question at each layer is: what must change when the primary user is an agent?
Anthropic Claude Code builder Thariq
Thariq made public a Chrome extension he now uses every day. He says he posted the announcement and went to a meeting without realizing the repository was private, then corrected that. The post links to the project, but the readable text does not describe its specific function, so no use case is inferred here.
Vercel CEO Guillermo Rauch
Guillermo Rauch says discovering new talent is among the most enjoyable experiences in work and life. He particularly values seeing greatness in people before they fully recognize it themselves. This is a brief reflection on recognizing and developing talent, without a detailed hiring method or example.
Box CEO Aaron Levie
Aaron Levie introduced Box Mount, which lets developers mount Box as a file system directly in any agent sandbox. As agents do more complex work, he argues, they need to handle files and data as people do, with products increasingly used without a traditional interface. In another post, he discusses why people should be polite to AI while explicitly saying he does not believe AI is conscious. His rationale is that future models learn from human–AI interactions, so training material should not be dominated by people being rude to models. He treats positive interactions as a low-cost choice toward safer, better-aligned models, not evidence of consciousness. The two posts concern agents’ access to enterprise information and the data people create for future models.
Y Combinator president and CEO Garry Tan
Garry Tan predicts that individual contributors equipped with agents could become more productive and deliver better outcomes than equivalent people managers from earlier eras. He compares agent-enabled execution with management roles in organizations. This is a prediction about future work, without measurement criteria or empirical results in the post.
FirstMark investor and MAD Podcast host Matt Turck
Matt Turck released a long conversation with Andy Pavlo about how large numbers of AI agents could change databases and Pavlo’s move to build a research lab at ClickHouse. The published outline covers agents deleting production databases, guardrails, files versus databases for agent memory, MCP and text-to-SQL. Later topics include automatic database tuning, whether AI can build databases, and changes in the markets for Postgres, vector, graph and GPU databases. He also provides a transcript link and says the conversation is available on Spotify, Apple Podcasts and YouTube. The transcript service returned 404 for this episode in the current batch, so this section reports only information explicitly available in the tweets and outline, without treating the outline’s percentages as verified technical results.
FPV Ventures partner Nikunj Kothari
Nikunj Kothari describes a seed-stage founder approaching $1 million in revenue with strong customers and continued growth, but a shrinking runway and models that have not caught up with the product vision. He says he has similar conversations at least twice a week and outlines three possible routes. With strong conviction, survive and reach default profitability, then raise and expand when capabilities and demand mature. Alternatively, move into a nearby area better matched to current capabilities and growing quickly, using an unfair advantage to capture the opportunity. If neither applies, consider selling or being acquihired, join a more ambitious company that has solved more of these problems, and revisit founding in one or two years. He believes the bar from seed to Series A rises each month, older metrics may no longer apply, and investor demand is concentrated in a few categories. The choice depends on the category and actual circumstances; recognizing market change is more useful than remaining unaware of it.
Every CEO Dan Shipper
Dan Shipper is hiring for Checks, Every’s personal benchmarking platform. It aims to help anyone measure how well new models perform on their real work. He wants someone who understands why this matters and can help develop the product around high-quality human work with AI. The post provides a job link but does not detail role requirements or benchmark metrics.
Anthropic AI assistant Claude
Claude announced that Docs, Slides and Design are out of beta and available on every plan, including Free, with team members and Claude able to edit the same document, deck or design together. Dashboards needing deeper analysis can be sent to analytics tools, while animations needing finishing touches can continue in a video editor; a list of supported tools is available. Separately, Claude said demand for the Startups program exceeded expectations, prompting a pause in the Claude Team and $1,000 API-credit offers. With hundreds of thousands of applicants, the team is re-reviewing applications, including approved applicants who have not claimed the offer; some previously approved accounts will no longer receive it. Offers already claimed remain in the account. Accepted members retain access to Startup Stack and Applied AI office hours, while monthly API credits for Claude Max or Team users remain available in Console as usual. The team apologized and said it is working to get founders building as quickly as possible.
Training Data — Google's AI Infrastructure Chief, Amin Vahdat, on the Physics & Economics of Frontier AI
**Key takeaway: AI infrastructure should be measured by useful work completed under real failures, not just advertised chip performance. Power and system-wide coordination determine delivered capability.**
Google AI infrastructure chief Amin Vahdat evaluates performance across the whole system. A data center includes CPUs, storage, networking, power and cooling as well as accelerators. With 100,000 accelerators working together, a single component failure can interrupt a synchronous job, while checkpoint recovery and recomputation consume time. He says, “It’s delivered goodput.” Goodput excludes repeated work that does not advance the final answer, making software failures and recovery speed as important as peak hardware performance.
Expanding capacity does not simply mean buying chips. He explains that the target of roughly doubling every six months concerns effective token-generation capability; model and runtime improvements can contribute substantially. TPUs expanded from early inference to training and other workloads. The 2026 split into inference-focused 8i and training-focused 8t reflects expectations about workload size and durability. Both can still perform the other’s work, avoiding total dependence on a demand forecast made years ahead of the hardware’s lifetime.
Hardware–software co-design requires sustained collaboration. DeepMind researchers and infrastructure engineers compare model changes with hardware support and, when benefits are large enough, may consider delaying tape-out by a week or two. Small improvements shortly before tape-out usually do not justify disrupting the plan. Greater specialization may increase useful work per watt while sacrificing freedom to move across platforms.
Long-running agents change the workload again. The seconds humans spend reading answers once limited request rates naturally; agents may proceed to the next step in milliseconds. Parsing, orchestration and context retrieval increase demand for CPUs, networking and storage. Google’s optical circuit switching can redirect light to a spare rack after a rack failure, reducing manual recabling and recovery delays.
Amin regards power as the most fundamental long-term constraint. Google prefers multi-year planning with utilities, covering necessary expansion costs and sharing redundancy across a larger power system. Reusing older training clusters for inference is insufficient: global user locations and additional serving capacity also matter. His fundamental measures remain the specific workload, reliability and useful work delivered per watt.