Andrej Karpathy
Andrej Karpathy argues that the gap in people’s understanding of LLM capability is widening, turning from two groups talking past each other into a steep funnel. He roughly estimates that about six billion people, or 75% of the population, have barely encountered LLMs, while roughly one to two billion, around 20%, occasionally use free chat products as better search or writing aids. Further inside the funnel, around 20 million professionals in math and coding have seen agents complete complex projects that previously took weeks or months. At its narrowest end are about 5,000 people with access to internal frontier systems; he describes thousands of agents collaborating for weeks on software, cyber offense and defense, science and mathematics, while human review and understanding begin to lag. These counts and percentages are his rough estimates, not population-survey findings. He identifies three drivers: larger, harder and longer tasks expose more capability; reinforcement learning concentrates strengths in digital, verifiable and economically valuable domains; and access differs across free, paid and internal systems. Most people consequently still see an ordinary chatbot, while a small group of professionals experiences a sharply rising capability curve, all at the same time.
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux announces that the default context length for GPT-6.1 Sol in Codex and Work has increased from 272k to 500k, a rise of approximately 80%. He says the change was carefully tested to reduce compactions and help tasks retain extensive context and nuance over time. Users can still override it upward or downward, with a maximum of 1M. A related change removes long-context fees for GPT-6.1 Sol usage drawn from subscriptions. He also discusses a user poll: he expected stronger support for bringing back five-hour limits across Pro 100/200/500 as part of what he calls the “2X” option to smooth the demand curve, but the results surprised him. That post discusses an option and poll feedback, rather than announcing an implemented subscription change.
AI tutorial author Peter Yang
Peter Yang shares AI development educator Micky’s view of manual review. Micky says, “I’m not at a point where I trust the agent, especially with stuff that people are using.” His approach places manual review at the final step, after the work is done and as much evidence as possible has been collected from the agent. Through evidence-driven testing, before-and-after comparisons and greploop, the agent must explain what it completed and why the changes are sound. Micky believes this builds trust and reduces the friction developers face during manual review. Peter presents the excerpt from a four-step workflow, emphasizing evidence for review rather than eliminating human judgment.
Anthropic Claude Code team member Thariq
Thariq agrees with the general direction of Terence Tao’s view that people should not settle for opaque superintelligence. He sees communication itself as an important signal of intelligence. Products and models should therefore build understanding with humans rather than disintermediate them. His point makes human understanding part of what an intelligent product should deliver.
Box CEO Aaron Levie
Aaron Levie uses Karpathy’s funnel to explain how early AI diffusion still is. Most people use AI for simple chatbot interactions resembling next-generation search, which is useful but shows only a fraction of its capabilities. At the other end, agent swarms could become an important form of work in mathematics, cybersecurity, biology, materials science, software engineering and finance, particularly where valuable tasks can run in parallel and compute can accelerate progress. He believes very few people currently work this way because of domain suitability, cost, available harnesses and public access to frontier models. Industries beyond software and AI research also need substantial organizational and workflow change. Companies must upgrade processes, integrate and migrate data, and address compliance, governance and security. He sees this domain-specific implementation work as a major opportunity for the enterprise AI application layer.
Y Combinator president and CEO Garry Tan
Garry Tan echoes Karpathy’s description of a small group of frontier users, arguing that access to these capabilities today creates enormous opportunities. He expects prices to fall, allowing more people to experience the capabilities and bring them to a wider audience. In a separate, more specific observation, he argues that replacing LLM calls with deterministic code where possible is an obvious choice. He believes this also produces code that better matches users’ needs. The two posts address broader access to frontier capabilities and choosing suitable tools in actual implementations.
FirstMark investor Matt Turck
Matt Turck expresses optimism about the AI transition by tracing changes in technical occupations. He lists punch-card operators, Lotus Notes administrators, Flash developers and Dreamweaver webmasters, followed by DevOps workers and Tableau analysts. People are now becoming agent managers and harness engineers. He believes this transition will work out too; the post offers a historical analogy without predicting employment totals or the pace of retraining.
Builder Zara Zhang
Zara Zhang shares a brief observation about her own usage: these days, Chrome is the only place she uses Gemini. The post does not explain her reasons or compare other entry points.
Every CEO Dan Shipper
Dan Shipper built a feed that automatically tracks the books he is reading and those he has read since 2012. He also suggests creating a feed for everyone at Every to see what people are reading. The post presents a personal tool and an idea for sharing within the team, without describing its implementation.