Sign of the Future: GPT-5.5

  • Author/Source: Ethan Mollick (Wharton), One Useful Thing (Substack)

  • Original: https://www.oneusefulthing.org/p/sign-of-the-future-gpt-55

  • Key Ideas

  • Mollick's framing for thinking about AI progress: models (Opus 4.7, Gemini 3.1, GPT-5.5), apps (chatgpt.com, Claude Code, Codex), harnesses (the tools the model can use --- image generation, code execution, web search). All three are advancing simultaneously.
  • Concrete generation-over-generation demo: build a procedurally generated 3D simulation of a harbor town from 3000 BCE to 3000 AD. Only GPT-5.5 Pro actually modeled an evolving town rather than just replacing buildings. GPT-5.5 Pro completed the task in 20 minutes vs. 33 minutes for GPT-5.4 Pro.
  • PhD-quality paper from four prompts. Mollick gave Codex (powered by GPT-5.5) a decade-old folder of crowdfunding data (STATA, CSV, XLS, Word files), asked it to generate a hypothesis, test it sophisticatedly, write a literature review, and format the paper. Then fed GPT-5.5 Pro's comments back in. The output is what he'd be happy to see from a 2nd-year PhD project --- the literature is real, the statistics are real, the methodology is sophisticated, but the hypothesis is "not that interesting" and standard causation concerns remain.
  • From one prompt to a 101-page playable RPG. Codex generated an original tabletop RPG with rules, tables, simulated playtesting, and 101 illustrated pages of layout. The setting is "interesting and novel"; the rules appear to make sense.
  • Image-gen-2 (the new OpenAI image model) can render high-quality text in images, making it usable for slides, mockups, and example websites. Mollick demonstrates with a multi-style art gallery of his "Otter test" with readable labels under each piece.
  • Jagged frontier persists. Long-form fiction still has the characteristic AI failure modes: love of the uncanny, overly complex ideas that don't pay off, weird metaphors, identical-tone dialogue, and "the name Mara." Hypothesis generation produces sound statistics on uninteresting questions.

  • Pattern over three years of newsletter writing: every few months a new model arrives, something previously impossible becomes easy, and the size of the leap is growing each cycle. The jagged frontier hasn't disappeared --- it's just much further out.

  • Summary

Mollick's GPT-5.5 review is less about the specific model than about the pattern it represents. He uses the post to reinforce his "models / apps / harnesses" frame for understanding why AI capabilities are accelerating: not just the models are getting smarter, the wrappers around them (Codex, Claude Code) and the tools they can call (image gen, code execution, sub-agents) are improving in parallel.

The most consequential demo for academics is the crowdfunding-data paper. Mollick had hundreds of anonymized files he'd never gotten around to analyzing. He asked Codex to sort them out, generate a hypothesis, test it, write a paper. Four prompts, no manual editing of the text. The literature review and statistics are real (a notable bar-clear since hallucinated citations were the standard failure mode through 2024--2025). His complaint is now at the level of "the hypothesis isn't interesting" and "I worry about causation despite the sophisticated identification" --- which is the level of complaint a senior advisor has about a 2nd-year PhD's draft. That's the real news.

The RPG demo is the same point in a different domain: agents can now produce attractive, structured, multi-format output (a 101-page illustrated PDF with rules, tables, and simulated playtests) from minimal prompting.

The fiction critique --- "weather and architecture are the same argument at different speeds" being cool once but exhausting across an entire book --- is Mollick's clearest articulation that creative-prose quality remains a bottleneck even as analytical-prose quality is essentially solved.

  • Relevance to Economics Research

Direct. The crowdfunding-paper demo is the most concrete recent evidence that 2nd-year-PhD-quality empirical economics papers can now be produced with minimal human input from a folder of mixed-format data files --- which is the situation many faculty are sitting on (decade-old data they never wrote up). For Mihail and other empirical finance researchers, this is a direct pointer toward firing up Codex/Claude Code on dormant data folders. It also reinforces Hall (summaries/ai-10x-research) and Kustov-III (summaries/academics-wake-up-3) --- the bar for "publishable empirical paper" is rising because the floor of what an agent can produce has risen.

The "models / apps / harnesses" frame is also useful pedagogically for the master-class audience: it disentangles "GPT-5.5 is smarter" (model) from "Codex got better" (app) from "image gen now does real text" (harness/tool), all of which are independently advancing.