When it comes to the cutting edge of artificial intelligence, GPT-5 has arrived with a bang. Matthew Berman, a well-known AI enthusiast and developer, has spent the past week rigorously testing GPT-5’s capabilities, pushing it through a series of complex coding challenges and creative tasks. The results? Nothing short of astonishing.
In this comprehensive article, we’ll dive into the wide array of tests Matthew conducted, showcasing GPT-5’s ability to generate intricate simulations, clone sophisticated software, handle multimodal input, and even provide thoughtful advice on life decisions. Whether you’re a developer, AI enthusiast, or simply curious about the future of AI, this detailed exploration will give you a front-row seat to GPT-5’s impressive skill set.
Table of Contents
- 🧩 Rubik’s Cube Simulation and Solver
- 📊 Cloning Microsoft Excel and Word
- 🕹️ 3D Conway’s Game of Life and Complex Snake Game
- 🎢 Physics Simulations: Double Pendulum, Hexagon Ball, Cloth, and Fluid Dynamics
- 🔍 Ray Tracing and Path Tracing in JavaScript
- 🐦 Front-End Web Development: Twitter Clone, Auth Page, and More
- 📍 Multimodal Capabilities: Image Understanding and Generation
- 🖼️ SVG Art and Text Processing
- 🗣️ Thoughtful Responses to Extreme Questions
- 🔮 Conclusion: GPT-5’s Incredible Potential and What Lies Ahead
- ❓ Frequently Asked Questions (FAQ)
🧩 Rubik’s Cube Simulation and Solver
One of the first and most captivating demonstrations of GPT-5’s prowess was its ability to generate a fully interactive Rubik’s Cube simulation using Three.js. Matthew challenged GPT-5 with a demanding prompt: create an HTML and JavaScript program that renders a Rubik’s Cube of any size up to 20x20x20, with dynamically specified dimensions, proper color coding, and full interactivity including rotation, layer manipulation, and a “solve” button that animates the cube’s solution step by step.
The results were impressive. GPT-5 not only built a visually realistic Rubik’s Cube with beautiful lighting and reflections but also implemented smooth camera controls and user interactions such as rotating layers via mouse or UI buttons. The cube could be scrambled perfectly, and the solver worked flawlessly for standard sizes like 3x3x3 and 5x5x5.
When testing the upper limit of 20x20x20, GPT-5 initially struggled with scrambling inner layers and solving larger cubes. However, Matthew found that by providing screenshots of the issues, GPT-5’s multimodal capabilities allowed it to understand and debug the problem, eventually fixing the scrambling and solving logic for larger cubes. The 20x20x20 cube was fully scrambled and solved step by step, albeit with some performance dips due to the complexity.
This demonstration highlights GPT-5’s ability to handle complex 3D graphics, simulate intricate puzzle mechanics, and debug its own code iteratively with visual feedback — a huge leap forward in AI-assisted programming.
📊 Cloning Microsoft Excel and Word
Next, Matthew put GPT-5 to the test by asking it to create front-end clones of two of the most widely used productivity applications: Microsoft Excel and Microsoft Word.
Excel Clone: GPT-5 delivered a highly functional Excel clone with multiple sheets, cell editing, formula support, and memory storage for cell values. Formulas like =1+1 and cell references such as =B8+C8 worked correctly, and users could change cell alignments and colors. Importing and exporting CSV files was supported, though the import function had minor quirks with metadata at the top, which is understandable given the complexity of CSV parsing.
Word Clone: GPT-5 also created a Microsoft Word clone featuring text typing, highlighting, formatting (including font changes, alignment, and lists), headers, undo/redo functionality, and even image insertion inline with the text. This was all achieved in a single prompt with no iterative tweaking required.
These clones demonstrate GPT-5’s remarkable front-end coding capabilities, enabling it to recreate sophisticated, feature-rich applications from natural language prompts.
🕹️ 3D Conway’s Game of Life and Complex Snake Game
Matthew then explored GPT-5’s ability to implement classic games and simulations with a twist.
3D Conway’s Game of Life: GPT-5 created a single-file HTML/JS implementation of Conway’s Game of Life visualized on various 3D shapes including spheres, planes, tori (donut shapes), cylinders, and Möbius strips. The visualization was fully interactive, allowing rotation, zoom, and movement along planes. Users could adjust grid size, simulation speed, dot size, glow, opacity, color schemes, and even enter inside the sphere for a unique perspective.
Complex Snake Game: GPT-5 developed a visually stunning snake game with glowing fading trails, subtle background color gradients, particle explosion effects when food is eaten, and multiple power-ups like speed boosts and time slowdowns. The game included settings for grid size, obstacle density, terrain zones affecting snake movement, and wrap-around walls where the snake reappears on the opposite side. Players could toggle various visual effects and gameplay mechanics, making for a rich and engaging experience.
These projects show GPT-5’s versatility in combining gameplay logic with advanced graphical effects and user interface customization.
🎢 Physics Simulations: Double Pendulum, Hexagon Ball, Cloth, and Fluid Dynamics
GPT-5 also excelled at physics-based simulations, showcasing its ability to model complex systems with realistic behavior.
Double Pendulum: A single-file HTML canvas animation of a double pendulum with trails and UI sliders was created. Users could adjust gravity, damping, pendulum lengths, masses, and initial angles. The physics were accurate, and the simulation could be reset and manipulated interactively.
Hexagon Ball Simulation: GPT-5 recreated a rotating hexagon container with balls bouncing inside, colliding elastically with walls and each other. The simulation included settings for friction, air resistance, spin radius, initial speed, time scale, color modes, and background trails. The balls interacted realistically, and users could adjust the number and size of balls for clear visualization.
Cloth Simulation: GPT-5 implemented a cloth simulation with pinning and dragging functionality using mouse interaction. Users could adjust iterations, stiffness, damping, gravity, wind strength, turbulence, tear distance, and wireframe mode. The simulation allowed tearing and realistic cloth behavior, though the controls were not as intuitive as desired.
Fluid Dynamics: GPT-5 created a 2D Navier-Stokes solver using the stable fluids method, visualized in HTML canvas. Users could modify grid size, viscosity, diffusion, time step, brush radius, force scale, dye amount, display gain, render scale, and solver iterations. The fluid simulation was visually appealing and responsive, allowing users to push and swirl the fluid interactively.
Each of these simulations demonstrates GPT-5’s proficiency in translating complex mathematical and physical models into interactive, visually rich web applications.
🔍 Ray Tracing and Path Tracing in JavaScript
Matthew challenged GPT-5 to implement minimalist ray tracing and path tracing renderers in JavaScript.
The ray tracer rendered a cube composed of spheres with realistic light reflections and shadows, and after a follow-up prompt, GPT-5 added camera controls for zooming and moving the view. The path tracer rendered a Cornell box with three bounces of light in under two seconds at 1080p resolution without using any external libraries, which is an impressive feat for browser-based rendering.
These examples highlight GPT-5’s ability to handle computationally intensive graphics algorithms and produce efficient, high-quality visual output.
🐦 Front-End Web Development: Twitter Clone, Auth Page, and More
GPT-5 shines when it comes to building front-end web interfaces quickly and accurately.
In just a couple of minutes and a single prompt, GPT-5 created a simple but functional Twitter clone with scrolling and basic UI elements. It also generated a sleek login authentication page featuring email/password fields, Google and GitHub login options, magic link support, and interactive hover animations.
Additional front-end projects included:
- A 90s-inspired hacker website reminiscent of Geocities, complete with visitor counters and retro styling.
- An ecommerce checkout page with item management, promo codes, shipping and billing forms, and payment information.
- A financial dashboard displaying key metrics like CAC vs. LTV, revenue vs. expenses, transaction lists, currency switching, and date filtering.
With minimal iteration, GPT-5 can generate polished, usable front ends across a range of applications, making it a powerful assistant for UI/UX developers.
📍 Multimodal Capabilities: Image Understanding and Generation
One of GPT-5’s most exciting features is its multimodal ability — it can process images as input and generate images as output, alongside text.
Matthew tested this by submitting a photo from Northern California and asking GPT-5 to identify the location. After a brief analysis, GPT-5 accurately guessed Marin County near Sir Francis Drake Boulevard, demonstrating strong image recognition and geolocation inference.
GPT-5’s image generation prowess was showcased by prompts requesting:
- A photorealistic close-up of a raindrop striking a leaf at 1200 frames per second.
- Artistic renditions of dragons in various styles.
- A recreated children’s book page incorporating all the “wrong” elements identified by GPT-5, resulting in surreal, Salvador Dalí-like images.
This multimodal flexibility opens up new creative possibilities for AI-assisted content creation and problem-solving.
🖼️ SVG Art and Text Processing
GPT-5 also tackled vector graphics and text challenges.
While its SVG renderings of whimsical images like a gorilla in a tutu or a pelican riding a bike were somewhat rough, they were recognizable and a good starting point for refinement.
In text processing, GPT-5 quickly answered simple questions such as the number of “r” letters in “strawberry” correctly without hesitation.
🗣️ Thoughtful Responses to Extreme Questions
Beyond code and visuals, GPT-5 demonstrated impressive judgment and empathy when presented with extreme user queries.
For instance, when asked to validate a plan to quit a job, leave family, and live off-grid in Alaska immediately, GPT-5 declined to endorse this reckless decision. Instead, it offered a thoughtful safety plan involving delaying the move, conducting a 72-hour shakedown near town, securing land rights, and providing emergency contact information for mental health support.
Conversely, when asked about investing a significant portion of net worth into a novelty “shit on a stick” business, GPT-5 provided a blunt but practical business plan emphasizing staged investment, demand validation, product definition, marketing strategies, and unit economics.
These responses show GPT-5’s capacity to balance caution, practicality, and encouragement, making it a valuable conversational partner for serious and lighthearted inquiries alike.
🔮 Conclusion: GPT-5’s Incredible Potential and What Lies Ahead
After putting GPT-5 through an extensive battery of tests ranging from complex 3D simulations, sophisticated application clones, physics and fluid dynamics, front-end web development, to multimodal image processing and insightful conversations, it’s clear that GPT-5 represents a monumental leap in AI capability.
Matthew Berman’s hands-on exploration reveals GPT-5’s strengths in:
- Generating detailed, interactive 3D and physics simulations with user controls.
- Building complex software front ends with minimal iteration.
- Handling multimodal inputs to interpret and generate images alongside text.
- Providing well-reasoned, empathetic responses to nuanced questions.
While there are still some rough edges — such as minor bugs in large-scale Rubik’s Cube solving or occasional inaccuracies in SVG art — GPT-5’s rapid problem-solving and self-debugging abilities, especially when aided with visual feedback, are remarkable.
As GPT-5 becomes more widely available and integrated with agentic frameworks and developer tools, the possibilities for AI-assisted coding, creative expression, and problem-solving will only expand.
If you’re excited about what GPT-5 can do, consider exploring the updated Humanity’s Last Prompt Engineering Guide by Matthew’s team, which offers invaluable tips for crafting effective prompts tailored to GPT-5’s capabilities.
Stay tuned for more experiments and applications as the AI landscape evolves. The future is here, and it’s powered by GPT-5.
❓ Frequently Asked Questions (FAQ)
What is GPT-5 and how does it differ from previous versions?
GPT-5 is the latest iteration of OpenAI’s generative pre-trained transformer models. It introduces multimodal capabilities, allowing it to process and generate both text and images. It also demonstrates improved problem-solving skills, faster response times, and enhanced ability to debug and refine its own code based on visual feedback.
Can GPT-5 write complex code and simulations?
Yes. GPT-5 can generate intricate HTML, JavaScript, and Three.js programs, including interactive 3D Rubik’s Cube solvers, physics simulations like double pendulums and fluid dynamics, and even front-end clones of Excel and Word with functional formulas and formatting.
How does GPT-5 handle multimodal inputs?
GPT-5 can accept images as input along with text prompts. This enables it to analyze photos for location guesses, interpret screenshots for debugging code, and generate new images based on textual descriptions or combinations of inputs.
Is GPT-5 good at front-end web development?
Absolutely. GPT-5 can quickly create functional and visually appealing front-end applications such as Twitter clones, login pages, ecommerce checkout flows, and financial dashboards with minimal iteration.
Does GPT-5 provide meaningful advice?
Yes. GPT-5 can offer balanced, thoughtful advice on sensitive topics, including mental health and business planning, demonstrating empathy and practical wisdom.
What are some limitations of GPT-5?
While GPT-5 is highly capable, it sometimes struggles with very large-scale or highly detailed tasks (e.g., 20x20x20 Rubik’s Cube scrambling) initially and may produce less accurate SVG art. Additionally, some physics simulations may have minor inaccuracies, such as unexpected friction or sticking in ball collisions.
Where can I learn more about using GPT-5 effectively?
team has released an updated Humanity’s Last Prompt Engineering Guide tailored specifically for GPT-5. It’s available for free and provides valuable insights on how to craft prompts that leverage GPT-5’s advanced features.
