Best of your X follows: verifiability, future work, and bot replies

The configured AI/tech account list was quiet today, so this is a compact issue: five original posts made the cut after excluding retweets, sports, jokes, promotions, and context-light quote reactions.

Ethan Mollick argued that models are getting better even in domains where training feedback is harder to verify, not just in neatly graded tasks 1. That matters because a lot of real knowledge work sits in messy judgment zones: strategy, writing, design, research framing, and management. The practical read: do not assume today's benchmark-friendly tasks are the only places where frontier models will keep improving.

Mollick also wrote that the Mythos cybersecurity discussion was not hype, connecting it to what people are seeing when they use Fable for autonomous work 2. This is a continuation of the access-and-safeguards debate from earlier in the week, but the useful angle is operational: long-running agents make misuse boundaries harder to reason about. The implication is that model access policy is becoming part of product design, not just legal or trust-and-safety overhead.

Mollick described AI implementation advice as split between people who "feel the exponential" and people who treat current limitations and costs as stable design constraints 3. The useful tension: teams can overfit to today's brittle workflows, but they can also waste money designing around capabilities that are not reliable yet. A sane middle path is to isolate model-dependent parts of a workflow so they can be upgraded without rebuilding the whole process.

François Chollet wrote that future jobs will favor adaptability, creativity, and complex problem framing over repetitive execution or narrow specialized skills 4. For AI-heavy teams, that is a hiring and training point: the scarce skill is often deciding what problem should be solved, not just pushing the tool harder. The concrete takeaway is to practice decomposition, evaluation, and taste; those compound better than memorizing a fragile workflow.

Simon Willison said he is becoming suspicious of replies to his posts, especially question-shaped replies, because so many appear to come from bot accounts 5. That matters for people who use X as a professional feedback loop: low-quality automation can make legitimate questions look suspect. The near-term behavior change is simple but costly: creators may answer fewer strangers, and public technical discussion gets less useful.