X / Twitter
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault Sottiaux is gathering the main reasons potential users have not tried Codex. He narrowed the question to the “single blocker,” aiming to identify the most important friction in adoption. The question targets people who have considered Codex but have not actually used it. The focus is not new features, but why users have yet to take the first step. This also provides a direct feedback channel for future improvements to product understanding, trust, or onboarding.
https://x.com/thsottiaux/status/2094588317245509959
AI tutorial and interview creator Peter Yang
Peter Yang believes trust will be both the biggest barrier to widespread adoption of personal agents and its most important driver. This means an agent’s ability to perform tasks is not the only issue: users’ willingness to grant it data, permissions, and room to act is equally important. Without sufficient trust, individuals may reject adoption even when the capabilities are mature. Conversely, once a product establishes reliable expectations, trust can encourage users to delegate more responsibilities to agents. His view shifts the competitive focus for personal agents from model capabilities alone toward reliability that can be verified through long-term use.
https://x.com/petergyang/status/2094639655258091792
Nan Yu, who is about to join OpenAI’s product team
Nan Yu announced that he will join OpenAI to work on Codex and ChatGPT products. He previously led product at Linear, where he spent 4 years. He said he was proud of what he had built with the Linear team and grateful for the experience. In this new phase, he will carry his accumulated software product craftsmanship into OpenAI. The move brings a product leader who has spent years refining developer software experiences directly into building Codex and ChatGPT.
https://x.com/thenanyu/status/2094427243565269107
Meta AI Senior Director Madhu Guru
Madhu Guru believes there is enormous outsized value in product managers deeply understanding the model frontier relevant to their use cases. PMs should know what models of different sizes currently do well, rather than select models solely from general leaderboards. They also need an accurate understanding of how different models fail in specific use cases and what workarounds are available. The higher standard is for PMs to understand their own business scenarios more deeply than most frontier lab researchers and other PMs. Product roadmaps must also account for possible changes in model capabilities over the next 2 to 3 months, rather than plan solely around today’s abilities. Madhu considers this technical judgment a core responsibility of modern PMs, not optional extra knowledge.
https://x.com/realmadhuguru/status/2094591503981281503
Vercel CEO Guillermo Rauch
Guillermo Rauch proposed that the next generation of design systems could be Markdown. Vercel is using DESIGN.md to address the lack of taste and consistency in AI-generated content and to explore how large organizations can convey design judgment at scale. The approach turns design guidelines into textual constraints that both models and development processes can easily read, bringing design intent more directly into generation. He also views coding tokens as infrastructure and argues that companies should not let teams use them as if handing out an unrestricted AWS key. A single key’s resource scope could range from a t3.nano costing $3.80 per month to a p5.48xlarge costing $40,000 per month, so a lack of governance, optimization, and observability can quickly cause costs to spiral. AI Gateway provides per-key and per-user budgets for token usage, aiming to bring AI consumption into a mature infrastructure governance framework. Both positions point to a shift: AI development is moving from personal tool use into a stage requiring organizational standards, cost controls, and quality constraints.
https://x.com/rauchg/status/2094541309579235680
https://x.com/rauchg/status/2094523399280435630
Box CEO Aaron Levie
Aaron Levie believes that as foundational open-weight models improve and post-training infrastructure matures and becomes commercialized, companies with large amounts of data will gain new opportunities to train their own models. Previously, the main way to monetize large data corpora was licensing them to outside organizations for model training. Now companies can train models for their own domains without taking on the research costs and complexity of frontier labs. General-purpose frontier models will retain advantages through their broad task coverage, but individual industries and specialized domains could see far more models than exist today. He also noted that as AI security incidents increase, enterprises need more advanced AI agents to detect and prevent security problems. Frontier models currently still lead in cybersecurity capabilities, but open models are catching up quickly. Data assets, domain-specific post-training, and security agents could therefore all become important parts of enterprise AI strategies.
https://x.com/levie/status/2094650992818274514
https://x.com/levie/status/2094545525102235844
Y Combinator CEO Garry Tan
Garry Tan added a new set of evaluations to GBrain to validate its open-source retrieval layer for AI agents. He said these evaluations demonstrate that GBrain can retrieve memories at SOTA levels without involving an LLM in the reading process. The design separates memory retrieval from model reasoning, focusing on the performance of the underlying retrieval layer itself. He also added evaluations for saving memories from agent transcripts. The aim is to extract information from agents’ interaction records and continually enrich the user’s “brain.” Garry also shared the evidence behind the results so others can inspect these capability claims.
https://x.com/garrytan/status/2094462971598754010
Every CEO Dan Shipper
Dan Shipper believes anthropomorphizing AI is not inherently harmful; what matters is the purpose it serves. When anthropomorphism helps people understand, predict, and use AI more effectively, it has practical value. Problems arise when it is used to create panic, amplify fear, or make unfounded comparisons between AI and humans. He understands why critics are wary of this language, especially when claims about consciousness are used to change agents’ moral status. However, he emphasizes that anthropomorphic language can be used pragmatically without claiming AI is conscious. He thus shifts the debate from “whether anthropomorphism is allowed” to “whether this language improves understanding and use.” Another brief observation further sums up his position: pragmatists will continue to have the advantage in the AI era.
https://x.com/danshipper/status/2094406185109647580
https://x.com/danshipper/status/2094397684186481057
Podcasts
Training Data — Rich Sutton and Khurram Javed: Why AI Models Stop Learning, and How to Start It Again
Key takeaway: True intelligence cannot learn only during training. It must develop its own abstractions through ongoing action and keep updating without destroying prior knowledge.
Rich Sutton is a major founder of reinforcement learning, the author of a classic textbook on the subject, and the author of The Bitter Lesson. Khurram Javed is his former student at the University of Alberta. They have now co-founded Oak Lab to build systems capable of truly continual learning. Sutton frames the problem directly: “I’m not the strange one; the field is strange. It insists on calling this continual learning, but it is really just learning.” In his view, agents are always in a cycle of acting, perceiving, and learning; compressing learning into a single offline training phase is the abnormal state.
The first key argument is that synthetic data cannot replace real experience. Javed’s big world hypothesis holds that the world contains infinitely many things worth learning, with complexity far exceeding that of any single agent. A synthetic world comes from a finite program and can cover only a tiny portion of the real environment. For example, making a drone fly using echolocation still requires human domain experts to first decide how to generate the data and which physical details matter. Real-world factors such as motor wear, friction, and other people’s mental states also cannot be fully replicated by a simplified simulation. Relying solely on human-designed synthetic data therefore ties a system’s rate of progress to human expertise.
The second gap is the ability to learn abstractions. Agents must not only predict what happens next, but also discover for themselves which states, features, and concepts their models should use to make predictions. Javed gave the example of elite athletes, who often develop specialized concepts for subtle movements, some of which do not even have names. These abstractions come from long experience, rather than labels predefined by developers. One of Oak Lab’s research goals is for systems to form abstractions suited to planning and reasoning on their own, based on the world they inhabit.
The third challenge is catastrophic forgetting. Updating an entire model directly from a single new example can damage its prior knowledge. Settings such as Cursor, with large amounts of user data, can mitigate the problem through large batches of examples, but most personalized agents face only a single data stream. Sutton’s proposed direction is to learn a separate step size for each network weight, allowing most weights to change only slightly while new experiences update the relevant locations.
The fourth method is continual backprop. Traditional backprop receives randomness only once, at initialization, and this diversity gradually becomes depleted. Continual backprop continually introduces newly initialized random units, then uses backprop to test which new features are useful. Sutton believes combining per-weight step-size optimization with this generate-and-test mechanism could produce substantially stronger continual deep learning systems.
Sutton also acknowledges LLMs as a major breakthrough in neural networks’ language capabilities, but rejects equating fluent language use with all of intelligence. He estimates that language ability accounts for only around 20%, or a quarter, of intelligence. Oak Lab’s ultimate goal is not a single mind that learns everything, but a general design that allows multiple systems to grow into different minds through their own experiences.
https://www.youtube.com/playlist?list=PLOhHNjZItNnMm5tdW61JpnyxeYH5NDDx8