Replit CEO Amjad Masad
Amjad Masad calls Replit the only collaborative AI coding system available. Separately, he links a new sandbox design and says a custom kernel build lets Replit avoid 96% of recent Linux zero-day vulnerabilities. The post does not specify the vulnerability set or measurement scope. Both collaboration positioning and security efficacy remain his product claims; 96% is not a probability of blocking every attack.
OpenAI’s Thibault Sottiaux
OpenAI’s Thibault Sottiaux argues that humans have not been particularly good at writing secure infrastructure. He envisions highly secure infrastructure and software for everyone, with turbulence along the way. He also says formal proofs for tens of millions of lines of code are no longer ridiculous and should be considered in new-system design. These are engineering judgments rather than an announcement of a completed, verified system.
Y Combinator CEO Garry Tan
Y Combinator CEO Garry Tan points to a long-horizon benchmark for PE/IB work and views solving high-finance tasks as part of the path to AGI. In another post, he discusses market access for founders: powerful AI alone is insufficient if a few distribution-controlling companies can block new products. He urges Congress to open markets so startups can compete and lower costs. He sees that week’s political results as a bipartisan opportunity to improve competition. These are his policy and industry judgments; the batch does not detail the benchmark’s tasks or results.
Anthropic’s Claude
Claude’s official account introduces Personal Finance in beta in the US. Users choose which accounts to connect, and Claude answers money questions in everyday language using their balances and transactions. Access is read-only: Claude cannot move, manage or invest users’ money. A second post says it has no ads or sponsored recommendations and is designed to help users understand their finances rather than provide investment advice. Rollout is gradual on US Pro, Max and Ultra plans, with actual availability and permissions determined by the product interface.
Former Claude Code contributor Thariq
Thariq, formerly of Claude Code, shares a personal use for agents such as Instinct: location-based reminders. His example tracks laundry and, after more than five days, reminds him 30 minutes after he gets home on a weekday, continuing until he confirms or delays. The prompt combines an interval, weekdays, a location trigger and user acknowledgment.
Peter Steinberger of OpenClaw and OpenAI
Peter Steinberger of OpenClaw and OpenAI questions whether Apple’s claims about Safari match his experience. He finds it slower, less capable and affected by unusual bugs, with some websites failing when he develops for or uses it. He says approaching Chrome’s performance takes considerable effort; the post describes personal experience rather than a standardized browser benchmark.
Vercel CEO Guillermo Rauch
Vercel CEO Guillermo Rauch calls “doing stuff for you over time” a key agent application. He cites Muse and Instinct consolidating memories at night, sending morning reminders or messaging after N days without interaction. He describes Vercel Schedules as an enhanced Cron: defined at runtime, outside the Sandbox and able to invoke compute on demand for cost efficiency. Another post introduces a SpaceXSI partnership with Vercel, saying Grok and Cursor can now build, deploy and scale applications. He outlines a loop: subscribe to X, build with Bot or Cursor, deploy on Vercel, launch and distribute on X, and monetize through X Money or Creators. Scaling and cost benefits remain his product claims, and monetization is not a guaranteed return.
Box CEO Aaron Levie
Box CEO Aaron Levie predicts 100X more agents interacting with enterprise data and systems than people do today. He sees access to context for decisions, research and execution as central to successful deployments. He says 90% of enterprise data is unstructured, spread across documents, contracts, marketing materials, research, financial plans and design files. Box’s NVIDIA partnership combines lexical and semantic search to retrieve that context. He says the retrieval respects enterprise security, access controls and data governance so agents use only permitted data; 100X and 90% remain his forecast and stated figure respectively.
Every CEO Dan Shipper
Every CEO Dan Shipper shares a link to Thesis: 2027, live from Brooklyn, without further agenda or schedule details in the post.
AI tutorial author Peter Yang
AI tutorial author Peter Yang asks Suno and Tesla to include the music app in cars by default. He imagines dictating songs by voice while driving; this is a feature wish, not an announcement of an existing integration.
SPC general partner Aditya Agarwal
SPC general partner Aditya Agarwal notes that an investor’s willingness to fund a company at a particular valuation does not make that the right valuation for the company. The post does not develop a specific financing case or valuation method.
Unsupervised Learning — Ep 97: Jev Creator on Decision Models, Benchmark Fatigue & Why the Labs Missed It
Key takeaway: Decision models matter as reliable, composable software capabilities, not merely new output shapes chasing another benchmark.
Jev creator and TypeSafe AI co-founder Diogo Almeida wants AI to become a cognitive component called by software. He distinguishes natural language for humans from small decisions for machines, which need not always be packaged as chat. Jev has not yet reached his goal. He wants developers to define their reliability requirements and report specific uneven capabilities directly.
Reliability must be grounded in real tasks. Almeida says, “If you have to do too many hacks, that’s still our fault.” Adapting a model to an application is reasonable in the short term; his longer-term aim is SQL-like or dependency-like stability, without relearning a model’s quirks after every release. He worries that new shapes, repeated benchmark optimization and overpromising consume developers’ willingness to try genuinely useful approaches.
He connects coding-agent design constraints to the KV cache. Appending history saves cost but complicates routing, context transfer to subagents and state management. Passing everything to a subtask raises input costs, while a summary may discard information needed by the next task. He imagines a shared-state observer agent producing live visualizations, task-specific context reconstruction and retrieval over prior work. These are exploratory ideas rather than delivered features.
Small decisions could also support security boundaries. He wants each tool call checked against an organization’s semantics, such as preventing one code area from calling another, instead of relying solely on a model provider’s universal rules. Application developers need not supply private data: they could describe the rough shape of a failing task and work with model teams to repair reliability gaps. He also imagines semantic dependent typing, expressing output properties in natural language to make components easier to combine.
Multimodality needs further validation. He acknowledges uneven image-model behavior and does not endorse it as production-ready, preferring to study why models fail at simple tasks first. For coding-agent productivity he offers a personal estimate of roughly 20–40%, stressing dependence on use cases, user skill, technical debt and long-term skill changes. Code has Git, rollback and human review, whereas other knowledge work may lack comparable safeguards. He advocates deployment based on decision stakes and recoverability rather than an all-in response to FOMO. His market, valuation and AI-safety views remain judgments.