Every time a new AI model drops we get the same week: breathless threads, demo videos, and somebody declaring games as we know them finished. GPT-6 Astra from OpenAI is running that playbook right now, and per Kotaku's roundup, the showpiece this time is people using it to remake existing games. Sonic. Smash Bros. You know the drill.

And the results do work. Sort of. The videos making the rounds show recognizable versions of classic games generated from prompts, with characters that walk, jump, and fight. The catch, and it's a big one, is that they run worse than the games they're copying. Janky physics, mushy inputs, levels that fall apart the moment you poke at them. It's a knockoff T-shirt of a game: the logo's there, the fabric's wrong.

I keep thinking about the NES era, when unlicensed clone cartridges flooded flea markets. Everything from blatant Mario ripoffs to cartridges that flat-out stole sprite data. The comparison people keep making in the Astra threads is exactly that: whether the culprit was a rushed dev team back then or a language model now, the ceiling for "technically functional copy of a better game" has always been low. We just automated the part where the knockoff gets made.

There's a real conversation buried in here about what this means for small developers, and I don't want to wave it off. If a publisher can prompt its way to a "good enough" clone of an indie hit in an afternoon, the people who actually innovate get squeezed. That's the part worth watching, not the party trick videos.

But the games themselves? Come on. Games aren't just the sum of their assets. Sonic feels like Sonic because of physics tuning that took Yuji Naka's team agonizing iterations to nail. Smash works because of frame data and hitbox precision tuned over decades. An AI can approximate the look overnight. The feel, the part that matters, is where it faceplants. Every demo I've seen of this stuff has that hollow quality, like a cover band that learned the notes but never listened to the song.

My take hasn't changed since the last time a studio made headlines for shipping AI-built games: the tool isn't the problem, the ambition behind it is. Someone will eventually use models like Astra to make something genuinely new instead of something vaguely familiar but degraded. That day hasn't arrived, and this week's flood of worse Sonics doesn't move it closer.

If you want to play Smash, Nintendo will happily sell you Smash. It'll feel better, I promise.