Best of your X follows: AI employees, bio bounty, and Deep Time

AI tools and work design

Simon Willison: do not add spreadsheets to the org chart

• The target is the "AI employee" framing, not automation itself 1.
• The useful distinction: tools can change work allocation without becoming social members of a company.
• If a product pitch needs org-chart language to sound important, the workflow design may still be under-specified.

Paul Graham: Legora shows up in a motion workflow

• The task was not generic drafting; it involved dense prior issues inside a legal motion 8.
• The signal is adoption pressure from support staff, not just partner-level enthusiasm.
• Treat it as a field anecdote, useful but not a benchmark: there is no task suite, error rate, or time-saved number attached.

François Chollet: agentic coding has reset expectations in six months

• The post is not a product announcement; it is a shift in baseline expectations 7.
• The six-month framing matters because coding-agent norms are moving faster than many engineering processes.
• The practical implication: teams should revisit review, sandboxing, and ownership rules that were written for weaker agents.

Model releases and safety

OpenAI: the Bio Bug Bounty becomes an ongoing private program

• OpenAI is moving from a one-off GPT-5.5 bio bounty into a rolling private program 3.
• The target is narrow: universal jailbreaks that break a predefined biosafety challenge, not general model-bug hunting 3.
• The doubled $50,000 reward says bio jailbreak robustness is now being treated more like a standing security surface than a launch-week checklist 23.

OpenAI: GPT-5.6 health claims center on cost as much as quality

• The headline is not just a stronger top model; OpenAI is emphasizing cheaper health-model performance across the lineup 4.
• The benchmark table backs the direction of the claim, but it is still OpenAI's own release material, not an independent clinical outcome study 9.
• For builders, the important question is whether lower-cost health reasoning makes review workflows cheaper, or simply increases the volume of outputs that need expert checking.

Learning and creative-tooling experiments

Ethan Mollick: Fable turns a game into Deep Time

• The interesting part is the transformation: existing code becomes a different game premise, not a prettier clone 5.
• The prompt is deliberately high-level, which makes the artifact a test of creative direction rather than instruction following.
• It is also a reminder that Fable's value is easiest to judge through playable outputs, not screenshots or benchmark lines.

Ethan Mollick: ChatGPT study mode is still there, but moved

• The command changed from /study to @ study, which makes the feature easier to miss 6.
• The mode matters because a tutor should ask, pace, and test, not simply answer.
• If your goal is learning rather than task completion, the interface path now changes the model's behavior enough to be worth remembering.