chat · renderer · product

ChatGPT study mode and interactive math/science visuals

Socratic study mode plus about 70 interactive concept visuals where you change variables and watch graphs update.

inspireevidence: thinby OpenAIopenai.com/index/new-ways-to-learn-math-and-science-in-chatgpt ↗

ChatGPT study mode and interactive math and science visuals (OpenAI)

What it is

  • Study mode: Socratic tutoring, switched on from “Study and learn” in the tools menu. ChatGPT asks questions and gives hints rather than answers.
  • Interactive visuals: for set concepts (the Pythagorean theorem, Ohm’s law, Coulomb’s law, lens equations, compound interest), ChatGPT shows a visual where you change variables or the equation and see graphs and results update in real time.

The problem it’s solving

The same as Claude’s learning mode and Gemini’s Guided Learning: answer-giving undermines learning; explanations of how quantities relate land better when you can manipulate them.

Its path / bet

The visuals look like a catalogue, not generation: a fixed set of 70-odd concepts, each an interactive module the model invokes and parameterises. OpenAI’s page was not readable to us (403) and Engadget doesn’t say whether the modules are prebuilt, so this is our inference from the fixed, enumerated concept list. If right, it is the data-first path (the model chooses and fills a known component) inside a lab product, next to Google and Anthropic choosing per-prompt code.

How it works (concretely)

Unverified beyond the above. Study mode is a system-prompt behaviour.

Strengths

  • A catalogue of tested modules is reliable: no layout errors, no wrong formulas drawn.
  • Free on all plans, so it reaches the most students.

Weaknesses / limits

  • 70 concepts is a small fraction of what students ask about.
  • No published evidence of learning effect that we found.
  • Closed; chat only.

Relation to fictty

Inspire. If our reading is right, OpenAI made the same trade-off fictty makes: reliability from a fixed set of components the model fills in, rather than a fresh program every time. The cost is coverage, which fictty addresses by making its components general (tables, charts, layers) rather than one per concept.

Could fictty adopt it instead of building?

No. Closed product feature.

What fictty should take from it

  • The catalogue path is respectable. It’s not a lesser version of generation; a lab chose it for learning, where correctness matters most. Use that in the “data, never code” argument.
  • “Recipes” are concept modules. A fictty recipe for “a formula with sliders and a plot” bound to a computed source would be the terminal equivalent of one of OpenAI’s 70, written once and filled by any agent.

Sources