Lucid

What do they actually do

Lucid trains video-based “world models” that generate playable, action‑conditioned simulations. Their first public release, Lucid v1, is a Minecraft demo you can try in the browser or run locally. They published model weights and a demo/inference repo; on a single NVIDIA 4090, the model runs around 20+ FPS and was trained on roughly 200 hours of Minecraft gameplay videos and actions (demo notes, GitHub, YC page).

Early users are researchers, hobbyists, and game/AI builders who want to experiment with interactive neural simulations, evidenced by open weights, a public demo, and community threads (GitHub, HN). The typical workflow is: ingest video+action traces (e.g., Minecraft), compress frames to a compact latent, train an action‑conditioned video model that predicts future frames, and run it autoregressively to accept player inputs in real time (demo notes, GitHub).

Who are their target customer(s)

• Indie game developers and prototypers: They spend weeks wiring engines, physics, and assets to test mechanics. They want faster iteration on interactive scenes, but today’s Lucid demo is research‑grade (short memory, fidelity issues) and typically needs a high‑end GPU (demo notes, GitHub).
• RL and embodied‑agent researchers: They need controllable simulators for training/evaluating agents without costly real‑world data. Learned video sims are promising but currently lack long‑term consistency and scale for robust agent training (YC page, dev notes).
• Small game studios / technical leads: They require deterministic, stable simulators suitable for production. Neural world models still diverge and aren’t at AAA fidelity or long‑run stability needed for shipped titles (dev notes).
• Hobbyists, modders, independent researchers: They want open, runnable models to tinker with interactive AI. Barriers are compute and immature model quality, though Lucid’s public demo/weights make experimentation possible on a high‑end consumer GPU (~20+ FPS on a 4090) (demo notes, GitHub, HN).
• Tools/content creation companies (level editors, procedural content): They want faster ways to generate playable levels and interactive scenes from examples. Current learned simulators need higher fidelity, longer memory, and easy integration before replacing manual pipelines (YC page, dev notes).

How would they acquire their first 10, 50, and 100 customers

• First 10: Personally recruit early adopters from the demo/repo community (GitHub contributors, HN/Reddit commenters) and co-develop features/workflows with hands‑on support (demo notes, repo, HN).
• First 50: Run targeted workshops and tutorials for indies, hobbyists, and RL researchers, offer short paid pilots with clear onboarding (local inference guide, sample datasets, one‑week support) while being upfront about hardware needs and limitations (developer notes, YC page).
• First 100: Close small studios, academic labs, and tooling partners via paid pilots with SLAs, simple engine adapters (Unity/Unreal) or API, and managed hosting/licensing to remove local GPU barriers, supported by case studies from prior pilots (YC page, dev roadmap).

What is the rough total addressable market

Top-down context:

Consumer gaming spend is ~US$182–189B in 2024–2025 (Newzoo), but Lucid is a tools/middleware player. Direct adjacencies are game engines/developer tools (low‑billions market per reports, e.g., Grand View) and broader simulation software (~US$23.6B in 2024 per Grand View) (engines/tools, simulation software). The global developer pool is large (~11.1M active in game development, a demand proxy) (SlashData).

Bottom-up calculation:

Near‑term, research‑grade addressable revenue could come from: (a) 15,000 indie/hobbyist users on a $200/yr plan (~$3.0M), (b) 300 research labs on $12k/yr licenses (~$3.6M), and (c) 150 small studios on $30k/yr pilots/licenses (~$4.5M). Combined, that implies roughly ~$11M/year of practical near‑term TAM, contingent on today’s model quality and integration limits.

Assumptions:

• Price points: $200/yr indie/hobbyist self‑serve; $12k/yr lab license; $30k/yr small‑studio pilot/license.
• Penetration is a tiny fraction of the ~11.1M developer base and research/studio universe given research‑grade fidelity and GPU needs.
• Engine adapters/hosting reduce integration friction enough to win initial studio/lab deals despite non‑determinism and short memory.

Who are some of their notable competitors

• Google DeepMind — Genie (Genie 2): Foundation world models that generate playable, action‑controllable environments from prompts; Genie 2 demonstrated interactive, controllable 3D worlds and agent play (blog, paper).
• NVIDIA GameGAN: Early neural game‑engine mimic that recreated a playable PAC‑MAN from videos and actions without explicit engine code—an antecedent to learned, action‑conditioned simulators (NVIDIA blog).
• Wayve — GAIA‑1/2: Generative world models for driving that use video, text, and actions to produce controllable, realistic driving scenes—targeted at sim and autonomy research (arXiv, Wayve GAIA).
• UniSim (Google DeepMind/Berkeley/MIT): A learned, action‑conditioned real‑world simulator trained on diverse datasets to enable interactive rollouts for planning, RL, and robotics; notable for transfer to real robots (project, OpenReview/ICLR ’24).
• OpenAI — Sora (adjacent): State‑of‑the‑art text‑to‑video generation; not interactive/playable but relevant as a high‑fidelity video generator and for the broader “world simulation” race (OpenAI).