Best of your X follows: GPT-6 hints, token loops, and AI second opinions

Model labels and evaluation

Ethan Mollick: what is GPT-6 being saved for?

• What happened: Ethan Mollick asked what model OpenAI is saving the GPT-6 label for, after the GPT-5.6 naming wave kept the next major-version question open. 1
• Why it matters: The post is short, but the tension is real: model names now carry product, expectation, and benchmark meaning before users see the system.
• Implication: Treat the label as a market signal, not evidence of a capability jump by itself.

Ethan Mollick: GPT-5.6 without GDPval

• What happened: Mollick said OpenAI did not appear to provide a GDPval measure for GPT-5.6, calling GDPval one of the best measures of economically valuable work. 2
• Why it matters: The point is less about one benchmark score and more about which evaluations readers now expect for frontier models.
• Implication: If a model release claims work value, readers will ask for work-shaped tests, not just conversational demos.

Adoption lessons

Ethan Mollick: steam needed skilled workers too

• What happened: Mollick linked the BBC series Industrial Revelations and argued that steam alone was not enough; industrial change required skilled workers adapting new power to old work. 3
• Why it matters: That is the cleaner AI adoption analogy than "new tool arrives, productivity jumps." The organization still has to redesign work around the tool.
• Implication: Watch training, workflow redesign, and documentation quality. They may explain more variance than model access alone.

François Chollet: the process is the point

• What happened: François Chollet wrote that art and science raise humanity to the sublime, but that the magic is in creation, discovery, feeling, and understanding rather than the final output. 4
• Why it matters: For AI evaluation, this is a reminder to inspect the search, revision, and discovery loop, not just the polished artifact.
• Implication: A system that produces a good-looking answer may still be weak at forming questions, testing failure modes, or noticing what it does not understand.

HN fallbacks

Tokenmaxxing may come back through agent loops

• What happened: The 12 Grams of Carbon essay argues that "tokenmaxxing" first forced AI usage, then may return because longer-running agents can improve outcomes when more token spend compounds correctness. 5
• Why it matters: The essay separates two budgets: tokens spent by developers using coding agents, and tokens spent by brittle one-off pipeline agents. That distinction is useful for finance and engineering leaders. 5
• Signal: Hacker News showed 27 points and 29 comments for the discussion at capture, so this is a live community thread rather than a settled claim. 6

Claude Code as a risky second-opinion tool

• What happened: Antoine's writeup describes sending a 266 MB DICOM MRI export to Opus 4.8 in Claude Code; the model's first report contradicted the clinic's Grade III partial-thickness tear diagnosis by reporting an intact tendon. 7
• Why it matters: A later arbitration run concluded with moderate-to-high confidence that there was no discrete partial- or full-thickness tear, while the author repeatedly warns that he is not a doctor and that this is not medical advice. 7
• Signal: Hacker News showed 77 points and 93 comments at capture; read it as an agent-tooling case study in expert workflows, not as clinical evidence. 8