How Well Does GPT-5 Build Interactive Web Apps?

TL;DR

GPT-5 can generate substantial interactive browser applications from short prompts, including a Rubik's Cube simulator, spreadsheet and word processor clones, physics demonstrations, and games. The tests produced impressive first attempts, but larger simulations and imported data exposed errors that required screenshots, follow-up prompts, and repeated corrections.

Transcript

I have been playing with GPT5 for about the last week and I am so excited to show you all of the tests that I've put it through. Yes, it is incredibly impressive. Let me show you. And by the way, we just updated Humanity's last prompt engineering guide. This is my team's very own prompt engineering guide updated for GPT5. It is completely free. I'l... Read More

Key Insights

• GPT-5 is capable of generating complex interactive browser prototypes from compact natural-language prompts. The demonstrations included three-dimensional rendering, editable productivity interfaces, games, and physics simulations, often with functional controls and polished visual effects in the first generated version.
• The Rubik's Cube simulator supports dynamically selected cube sizes, color-coded faces, camera movement, layer rotations, scrambling, and animated solving. It successfully demonstrated 3x3x3 and 5x5x5 cubes, while the 20x20x20 build demanded more processing time and initially exposed problems with interior-layer movement.
• Screenshot feedback is useful for correcting visually observable coding errors. When the larger Rubik's Cube failed to scramble or solve correctly, the presenter supplied screenshots and specific descriptions to GPT-5, then iterated several times to improve the implementation and add nuanced features.
• The Excel-style application maintains distinct sheets and evaluates basic formulas and cell references. It also includes alignment, cell coloring, CSV import, and CSV export controls, although the tested import produced unusual information at the top before the remaining imported content appeared correctly.
• The Word-style editor delivers a broad set of front-end document controls in one generation. Users can type, select text, change formatting and fonts, add headers and lists, adjust alignment, undo or redo changes, and insert an image directly into the document area.
• The Game of Life implementation maps its evolving grid onto multiple three-dimensional surfaces. Available shapes include a sphere, plane, torus, cylinder, and Mobius form, while controls adjust grid size, speed, point size, opacity, glow, field of view, spin, color, and scale.
• The enhanced Snake game combines standard movement with configurable audiovisual effects and gameplay modifiers. Its options include glowing trails, pulsing backgrounds, particle explosions, wall wrapping, obstacles, speed boosts, reverse controls, double vision, slow motion, obstacle density, and corridors.
• The double-pendulum demonstration provides interactive controls for gravity, damping, lengths, masses, and angles. Its generated single-file program contains 466 lines, and its motions appeared physically accurate during the test, illustrating GPT-5's ability to combine mathematical behavior with an adjustable canvas interface.

Questions & Answers

Q: How well can GPT-5 build interactive web applications?

GPT-5 produced several substantial browser applications from short prompts, including an interactive Rubik's Cube, spreadsheet and word processor clones, three-dimensional cellular automata, an enhanced Snake game, and a double pendulum. Many controls and visual effects worked immediately. However, larger simulations, CSV importing, and some physics behavior exposed defects, so the results still benefited from testing and corrective prompting.

Q: Can GPT-5 create and solve a virtual Rubik's Cube?

GPT-5 generated a Three.js Rubik's Cube simulator with selectable dimensions, color-coded faces, camera controls, scrambling, layer rotations, and an animated solve button. The 3x3x3 and 5x5x5 demonstrations scrambled and returned to solved states. A 20x20x20 version initially changed only outer layers, but follow-up corrections produced a fully scrambled cube that continued solving during the other tests.

Q: How can screenshots help GPT-5 fix generated code?

Screenshots can show GPT-5 the visible state of a malfunctioning application alongside a written explanation of the failure. During the Rubik's Cube test, the presenter returned screenshots when larger cubes did not scramble or solve properly. Several rounds of image feedback and instructions improved small implementation details, corrected behavior, and supported the addition of new features to the application.

Q: What features did the GPT-5 Excel clone support?

The Excel-style front end supported multiple independent sheets, editable cells, stored cell values, simple arithmetic formulas, and references between cells. It also offered left, right, and center alignment, cell colors, CSV importing, and CSV exporting. The tested import was imperfect because unexpected information appeared at the top, but the remaining imported content was described as working.

Q: What features did the GPT-5 Word clone include?

The Word-style clone allowed users to type and highlight content, apply formatting, add headers and lists, change text alignment, select fonts, and use undo and redo. It also inserted an image directly into the document. The presenter described the application as a single-turn result, making it an example of GPT-5 generating a reasonably broad editing interface from one request.

Q: How did GPT-5 visualize Conway's Game of Life in 3D?

GPT-5 mapped Conway's Game of Life onto several adjustable three-dimensional shapes, including a sphere, plane, torus, cylinder, and Mobius form. Users could rotate, zoom, and reposition the visualization while changing the grid size and animation speed. Additional controls modified point size, opacity, glow, field of view, spin, colors, and scale, including a view from inside the sphere.

Q: What enhancements were included in the GPT-5 Snake game?

The generated Snake game added a glowing fading trail, pulsing background gradients, particle explosions when food was collected, obstacles, and special food effects. Controls could enable or disable particles, background pulsing, trail glow, wall wrapping, and obstacles. Available modifiers included speed boost, reverse controls, double vision, slow motion, obstacle density, corridors, and terrain zones affecting the experience.

Q: What limitations appeared during the GPT-5 tests?

The tests revealed several limitations despite strong initial results. The 20x20x20 Rubik's Cube first rotated only its outer layers, a 10x10 solve required screenshot-based correction, and larger rendering reduced the frame rate. The spreadsheet's CSV import displayed unexpected material, while the screenshot-derived hexagon simulation caused the ball to stick to walls instead of behaving with consistently realistic physics.

Summary & Key Takeaways

• GPT-5 created a Three.js Rubik's Cube simulator with selectable dimensions, camera controls, scrambling, and animated solving. The smaller cubes worked convincingly, while the 20x20x20 version initially rotated only exterior layers. Follow-up instructions and screenshots helped correct its behavior, although the larger simulation ran at a visibly lower frame rate.

• GPT-5 produced functional front-end clones of Excel and Microsoft Word with minimal prompting. The spreadsheet supported separate sheets, stored cell contents, basic formulas, references, alignment, colors, and CSV controls, although one import displayed unexpected material. The document editor supported typing, selection, formatting, lists, alignment, undo, redo, fonts, and inline images.

• The browser demonstrations included three-dimensional Game of Life surfaces, an enhanced Snake game, and a configurable double-pendulum simulation. These projects combined graphics, controls, animation, and state management in single-file implementations. GPT-5 also generated text quickly, but its screenshot-based rotating-hexagon recreation showed unrealistic behavior because the ball became stuck against the walls.