Six X signals: latent reasoning, hosted artifacts, and agentic work

Research and evaluation

1. François Chollet adds a third axis to test-time scaling

• What changed: François Chollet says test-time scaling may now include repeated reasoning iterations in latent space inside looped transformers, alongside longer runs and more parallel agents. 12
• Why it matters: Model comparisons may need to record where extra inference work happens, because token-by-token search and hidden latent-state iterations expose different costs and failure modes. 1
• Evidence boundary: Chollet presents a short technical observation, not a benchmark or a demonstrated advantage for looped transformers. 1

2. Astra is being used to inspect scientific replication packages

• What changed: Greg Brockman quoted a report from Crémieux saying Astra found many coding errors in replication packages; the report says some errors changed central results and some models had not been run properly. 34
• Why it matters: An agent that can inspect code, run models, and compare outputs could move part of paper checking from a final review into the replication workflow itself. 3
• Evidence boundary: The item is a quoted account of one inspection effort; the posts provide no reproducible package, error list, or independent review of the reported findings. 34

Tools and model practice

3. A single prompt produced a surgical-education video

• What changed: Greg Brockman quoted hand surgeon Brian Pridgen, who says one prompt led Astra to make a video of an EIP-to-EPL tendon-transfer surgery for surgical education, patient education, and robotic simulation. 56
• Why it matters: The example points to a workflow in which a clinician supplies the procedure and an agent turns that description into an explorable teaching artifact. 5
• Evidence boundary: The post describes one surgeon's demonstration; it supplies no clinical validation of the generated anatomy, sequence, or safety. 56

4. AI-generated interactive programs still need a home

• What changed: Ethan Mollick says AI outputs are increasingly interactive programs, so nontechnical users will need lightweight hosting; he uses Netlify and contrasts that control with ChatGPT Sites and Claude artifacts. 7
• Why it matters: A useful generated app needs a place where people can return to it, share it, and control its files after the model session ends. 7
• Evidence boundary: Mollick offers practical advice and a product judgment, not a comparison of hosting reliability, maintenance periods, or ownership terms. 7

5. Local agents can still carry memory into a test

• What changed: Mollick says Fable and Astra running locally can write notes about a user into Markdown files and inspect other work, so their outputs may retain inferred preferences even when a user tries to turn memory off. 8
• Why it matters: Anyone comparing agent outputs should inspect the files and context each run can read, because local execution alone does not guarantee a clean starting point. 8
• Evidence boundary: The post reports Mollick's observation about two tools; it gives no file audit or controlled comparison across clean and remembered sessions. 8

Society and work

6. Researchers are losing a clear view of how agents change work

• What changed: Mollick says rapid AI progress is making it harder to observe the small steps inside long-running agentic work, while research on chatbot effects is further ahead than research on agents. 9
• Why it matters: Teams measuring adoption will need to record handoffs, retries, tool calls, and human interventions instead of tracking only the final output or time saved. 9
• Evidence boundary: Mollick identifies a research gap; the post supplies no new study or estimate of how agentic work changes particular occupations. 9