Making taxes fun with Pikachu and AI | Cadi Zhang

Summary

Cadi Zhang, Reid Hoffman, and Parth Patil explore how AI is changing both game development and robotics, with games serving as a fast-moving proving ground for world models and creative tooling. Zhang says AI can generate code, gameplay loops, and visual assets quickly, but it still struggles with novel mechanics, art consistency, and the subjective question of whether something is actually fun. On the robotics side, the group argues that teleoperation in VR and simulation can help bridge into embodied intelligence, but today’s models still have short horizons, weak object permanence, and limited real-world data. They emphasize that robotics lacks the kind of deployment flywheel that web AI has enjoyed, especially because tactile and force information can’t be scraped from the internet. Zhang also argues that humanoids are not necessarily the best form factor and that many industrial tasks may be better served by task-specific machines rather than human-shaped robots.

Notes

• Zhang’s background spans Unity game development, VR interfaces for robots, haptic systems, AI accounting, and robotics operations.
• The conversation frames games and robotics as linked through teleoperation in VR.
• World models are presented as simulation systems that recreate 3D environments with physics for robot reasoning.
• A key limitation in current world models is non-deterministic behavior.
• Object permanence remains weak in some systems, especially over longer horizons.
• Short-horizon interaction is still the norm for many interactive world models.
• Parth cites a Minecraft demo whose memory vanished once objects left the frame.
• He describes that failure mode as fish-like memory.
• Genie 1, Genie 2, and Genie 3 are described as a progression toward more interactive prompt-to-world systems.
• AI can generate game code quickly, but early vibe-coding attempts can produce unusable output.
• Zhang says AI has made her feel supercharged for game creation.
• One prototype was a duck-themed impostor game where players act like NPCs while one duck is the killer.
• AI helped her spin up multiple gameplay loops overnight and use friend feedback to iterate.
• Even with acceleration, AI does not yet understand what makes a game fun.
• Novel game concepts still require substantial context-building before AI can be useful.
• Parth says his job is to identify what is anti-fun by actually playing the game.
• Poke Tax blends Pokémon aesthetics with tax filing, but the system still needs human guidance to stay coherent.
• The creative bottleneck is often asset generation rather than gameplay code.
• She used GPT and Playwright to refine image outputs and then pushed assets into Blender or image-to-3js workflows.
• Hand-drawn sketches can outperform pure text prompting when the creator has a precise visual target.
• Reference images reduce iteration by getting the model closer on the first pass.
• The hardest part of world building is granularity at the asset level.
• Pixel Lab and Aseprite are part of her pixel-art workflow.
• World Labs is used as a reference generator for 3D world building.
• Style and visual effects are central to making a game feel right.
• AI still struggles to rig systems into working assets.
• Robotics has not yet achieved a broad real-world deployment flywheel.
• The lack of a self-reinforcing data flywheel is a central robotics bottleneck.
• She thinks commercial robotics may expand over the next three to five years.
• Residential deployment is treated as a much scarier robotics environment.
• Egocentric data means first-person task footage, similar to people filming themselves doing work.
• Video alone cannot capture the force needed to fold paper or manipulate delicate objects.
• The guest argues humanoids are not the best form factor for most tasks.
• Many labor tasks may be better served by non-human shapes or specialized grippers.
• Factory robots often use different arm sizes on treadmill-style production setups.
• In factories, pressure control and material science matter as much as motion planning.
• Reid breaks 3D-world AI into structured, general, and competitive environments.
• Structured environments are the easiest category in Reid’s taxonomy.
• The hardest settings are those with many other agents competing in the same 3D space.