Anthropic Claude team member Boris Cherny
Boris Cherny says Claude models have made progress against prompt injection over the past few years. As agents take on more work, he considers solving the problem increasingly urgent across the industry. Anthropic is partnering with Gray Swan on a competition inviting participants to find problems. The update calls for further scrutiny of defenses and does not claim prompt injection has been solved.
OpenAI Codex product team member Nan Yu
Nan Yu cautions that taste is not the same as aesthetic taste. He calls reducing taste to preferences about appearance its shallowest definition, and uninteresting because it is so common. The post does not go on to define a deeper form of taste.
Box CEO Aaron Levie
Aaron Levie predicts that moving from back-and-forth chat to agents doing useful work in the background could consume 100 to 1,000 times as many tokens as today. Knowledge work is constrained by human information-processing capacity, including moving data between systems, reviewing documents, checking code security, analyzing logs and researching prospective clients; chat assistants speed up individual tasks, but people still limit parallel work. He envisages handing harder problems to agents, then dividing tasks among several agents, and eventually deploying arbitrary numbers of agents toward a goal. Code security, QA for every change, infrastructure checks, document-by-document risk analysis and broader audits could cover work that was never previously done, rather than merely replacing existing labor. People, workflows, triggers and other agents could launch these swarms, provided models keep improving and costs are low enough, after which compute supply becomes a constraint. In another update, he says Meta’s open-weight Muse model could help enterprises combine frontier models for orchestration and demanding tasks with inexpensive, customized post-trained models for repetitive work. He expects this to increase tokens and workloads across the stack, and also expresses long-term optimism about robots, flying cars, space infrastructure and disease treatments. The 100-to-1,000-fold consumption estimate and those future scenarios are his predictions.
AI tutorial and interview creator Peter Yang
Peter Yang and his eight-year-old child are collecting feedback on T-shirt designs. He asks parents which design they would buy for their children and which character they prefer, requesting a design number and explanation. He shows a penguin-design sample ordered through Fourthwall. He then asks a bot to monitor replies and report the winning design after 12 hours. The sequence moves from a physical sample to user choices and delegated feedback aggregation; the posts do not provide the final results.
Every CEO Dan Shipper
Dan Shipper uses a satirical passage to caution against equating identification of internal representations with understanding a model’s plans. He sketches a path to overconfidence: studying polysemanticity and superposition, finding a neuron responsive to the Eiffel Tower, and then assuming that knowing when a model is thinking about France is enough to align a superintelligence. His point is that knowing what a model recognizes does not reveal what it is planning. This critiques the limits of interpretability claims rather than reporting a new experiment.
Builder Zara Zhang
Zara Zhang quotes Ezra Klein: “Never underestimate your audience’s intelligence and never overestimate their prior knowledge.” The remark distinguishes ability to understand from existing background knowledge. She also shows The Toddler Times, a Harry Potter-style newspaper she made with pictures that come alive. She says it was fun to create, but the post does not name the implementation tools or describe the steps.
SPC general partner and Bevel Health co-founder Aditya Agarwal
Aditya Agarwal believes Goodfire could be one of the most important companies to emerge from SPC. He values its approach to interpretability and alignment as engineering problems. This is his assessment; the post does not detail technical methods or measurements.
The MAD Podcast with Matt Turck — Jev: The AI That Makes Decisions, Not Text - and Why It Changes Everything (Diogo Almeida)
**Key takeaway: Jev turns AI into composable decision components with configurable thresholds, aiming to make software automation reliable.** TypeSafe CEO Diogo Almeida previously worked on RLHF at OpenAI. He argues that fluent, confident chat models raise expectations while subtle errors still undermine background automation. Jev therefore targets machine-native intelligence embedded directly in software. First, break a large prompt into testable decisions. An email assistant can separately determine whether a message concerns sales or recruiting, or explicitly asks for the recipient’s attention, then combine these decisions in code. Each branch can be evaluated independently and errors added to tests. He does not claim Jev has completely solved email scheduling. Second, give developers probabilities and control. Missing an important investor’s email and seeing an extra ordinary message have different costs, so developers should adjust thresholds rather than accept a model’s hidden binary choice. Almeida says, “You get what you optimize for.” He sees reliability and calibration as prerequisites for delegating background work; benchmark scores do not directly establish either. Third, training innovation starts with defining the task. RLCD involves more than changing a reinforcement-learning algorithm: it includes input and output formats, the shape and balance of data, and the objective. The team creates its own data and combines open-weight models without pretraining from scratch. He views classifiers as an interface through which intelligence influences software branches, allowing complex capabilities to be composed from reliable smaller components. Fourth, lower costs should open new demand. He hopes developers will repeatedly invoke useful components and automate work previously not worth doing. TypeSafe currently prioritizes rapid iteration through an API; open weights remain under discussion without a commitment. He also argues that copying an interface is easier than achieving reliability, while judgments about competing products still require measurement. His central proposal is to make defined tasks work reliably in real software before pursuing grander visions of intelligence.