AI-generated games don't lack the technology — they lack hand-tuned taste
AI can turn one sentence into a playable game, but the real moat is not generation — it is human taste and manual tuning: an LLM only copies, it does not innovate, and the editor is the second thing it has to build for the user.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
What gets supply moving is not generation but publish-as-distribution
Astrocade positions itself not as a ‘game generator’ but as an interactive-content platform: games are simply the largest category of interactive experience, and educational content, birthday cards and the like count too. Users describe what they want in natural language, AI generates the content, and one tap on publish drops it into the feed, which you swipe up and down like TikTok. It already has 100,000+ creators from more than 100 countries. Jason's read is that the skill barrier to making a game has been compressed from years of study and internships into a single prompt; but what actually gets supply moving is the product shape in which publishing *is* distribution.
— AmirUsers are not lazy; the friction was simply too high
The company's first office was the founder's home, where the fridge was full of small grab-and-go bottles of watermelon juice — gone in two days. Swapped for half-litre bottles, nobody touched them for two weeks, until a glass was placed next to them and demand came back. From this the founding team distilled Astrocade's product theory: users are not lazy, the friction is too high; take away the big obstacle of ‘you need 5 to 7 years to learn to make games’ and demand shows up on its own. The analogy explains why the platform builds generation, publishing and playtesting into a single chain.
— AmirAI cannot build an experience that is not in its training data
After two and a half years of working with creators, the team is explicit about AI's limits: AI still cannot produce a complete game. For a genre like Flappy Bird, with enormous amounts of training data, it can copy wholesale — but the moment a new experience is involved, or an existing experience has to be modified, AI starts to struggle. That is why the best games on the platform have all gone through multiple rounds of wish — their name for a prompt — iteration. This judgement pushes ‘fully automatic generation’ back to ‘human-led, AI-assisted’, and it is also why the editor has to exist.
— AmirGame numbers cannot be handed to a probabilistic model
‘AI in the loop’ does not mean fully automatic. Which upgrade tree, how to add gamification, how to set the numbers — those remain the creator's choices; a single variable changed from 25% to 50% is enough to make a game boring because it becomes unlosable. Amir's method is to let natural language handle describing structure and numeric inputs handle precise tuning: an LLM produces probabilistic output, while a game's ‘three lives, a boss with 1000 HP, a sword doing 10 damage’ requires determinism and reproducibility. This is a rare worked example of ‘dividing deterministic from non-deterministic’ inside a generative product.
— AmirThe only moat is taste, not the model
LLMs have been trained on everything, so generated output slides easily toward the average, which is why ‘it looks like every other game’ is the default outcome; getting a distinctive theme and art style requires a much more specific definition. Amir's conclusion: the only moat is taste — taste is the new programming language. The platform also plans to retrain the model on creator data so that over time it understands this particular group's preferences better. Translated into platform strategy: set direction with human taste first, then evolve incrementally on bespoke data.
— AmirPaying by views has already produced full-time creators
Astrocade already pays creators by view count, and some people make games and interactive experiences full time, earning several thousand dollars a month. Amir sees this as reproducing the market mechanism by which YouTube produced MrBeast. Jason pressed on the spot: ‘can this move from consumer to prosumer — app stores, in-app purchases?’ Amir first offered an existing use case: someone turned photos of friends into an interactive birthday card. He then returned to the revenue question, saying there are already people doing this full time at several thousand dollars a month, that the platform is sharing revenue by views, and that a new creator economy is forming the way YouTube created MrBeast.
— AmirPlay is a human universal language, not a toy business
Why aim at interactive content at ‘a billion people’ scale? Amir gives two reasons: first, impact — he wants to build consumer products, and B2B is heavily homogenised; second, play is humanity's most intrinsic universal language, and having two children of his own lets him see how primal play is in them. Jason steers the topic toward neuroplasticity: constantly playing new games and learning variant rules forces you to rebuild existing heuristics, which is excellent exercise for the brain. This view of play as an underlying operating system explains why the AI-generated-games category is not merely a toy business.
— AmirIn their own words · checked verbatim
AI is still not there yet. AI cannot make a full game. It can copy the games that they have been trained on
Amir7:45
So one of the hard parts, even with AI in the loop is every number and every tuning matters.
Amir11:01
the only barrier right now is taste. And I call actually, I say taste is the new programming language.
Amir13:03
play is so universal is a universal human language.
Amir18:23
Figures
| Company founded | 2022 | 5:27 |
| Founder's AI track record | Started in 2013; PhD at Stanford AI | 5:27 |
| Time spent co-creating with creators | Two and a half years | 7:45 |
| Old cost of a game of equivalent quality | 5 people, 6 months | 11:01 |
| Top creators' monthly income | Several thousand dollars a month | 16:16 |
Glossary
- wish
- Astrocade's name for a prompt; every time a user revises or iterates on what they want, that is one wish.
- post-training
- Continuing to train on top of a base model with specific data, to make the model better at generating interactive content.
- creator fund
- A mechanism for paying creators a share of revenue based on view count; Astrocade already runs one.
- prosumer
- Someone who is both consumer and producer — the middle state between a casual player and a professional creator.
How to listen
People building AI-native consumer products, game founders, and investors watching content-platform distribution; also engineers who want to understand the real engineering difficulty in generative games.
The opening small talk and the closing Google/Gemini ad read are skippable; the core content runs from minute 6 to 19.