3 ms·

Great roundup. AI are great at technical optimization problems (for example, reducing the usage of electricity in a data center). For creative and subjective wo

Great roundup. AI are great at technical optimization problems (for example, reducing the usage of electricity in a data center). For creative and subjective works, it seems like a machine-human hybrid model will be the way to go for the next decade. The AI spits out some rough drafts, then the human artist picks it up from there. All of these last-mile problems would be trivial for an artist to touch up.

lumost 4y ago
I suspect that we'll see a big wave of Large Transformer models getting built into high cost professional tools. Need to write some random background NPC dialogue for a game? Need to generate a draft Cad design for "Airframe supporting a large turbojet engine with X newtons thrust"? These are problems where the high cost of inference won't be significant, and there is an expert available to revise and discard the result if its not working. I'd be curious if there is a good startup model for anything in this space, seems like you would already need to be a vendor of the underlying technical software to make a viable business (CAD,Game Engines,Creative Tools etc.)
- flywithit 4y ago
  Sounds about right based on my experience with GitHub copilot. It does make me fear a little bit. I mean I’m not experienced with cad but I could see you might train on a particular cad program, that program may have undiscovered bug X that allows a model to pass all the synthetic tests while optimizing for something like cost (which goes undiscovered because no human would ever design in such a way)… the reviewing engineers rubber stamp under pressure and then bam, 9/10 airframes fall apart after 1,000 miles. That’s the copilot analogue at least.
maroonblazer 4y ago
Yes, and that could inspire a massive creative explosion. One application I'm pining for: I'd love Apple to introduce AI into Logic Pro X (LPX) such that it could take a 4, 6, 8 or 16 bar set of tracks and generate infinite variations, from which I can pick and choose ones to embellish. Or, similarly, I have innumerable LPX projects that are just a single idea that, in the hands of someone more capable (or with more motivation) than myself, could be turned into full pieces. I'd love to be able to give LPX 2 or more of these projects and it generates transitional pieces between them. All LPX upgrades are free, but I'd happily pay for this feature.

lumost 4y ago

I suspect that we'll see a big wave of Large Transformer models getting built into high cost professional tools. Need to write some random background NPC dialogue for a game? Need to generate a draft Cad design for "Airframe supporting a large turbojet engine with X newtons thrust"? These are problems where the high cost of inference won't be significant, and there is an expert available to revise and discard the result if its not working. I'd be curious if there is a good startup model for anything in this space, seems like you would already need to be a vendor of the underlying technical software to make a viable business (CAD,Game Engines,Creative Tools etc.)

flywithit 4y ago
Sounds about right based on my experience with GitHub copilot. It does make me fear a little bit. I mean I’m not experienced with cad but I could see you might train on a particular cad program, that program may have undiscovered bug X that allows a model to pass all the synthetic tests while optimizing for something like cost (which goes undiscovered because no human would ever design in such a way)… the reviewing engineers rubber stamp under pressure and then bam, 9/10 airframes fall apart after 1,000 miles. That’s the copilot analogue at least.

maroonblazer 4y ago

Yes, and that could inspire a massive creative explosion. One application I'm pining for: I'd love Apple to introduce AI into Logic Pro X (LPX) such that it could take a 4, 6, 8 or 16 bar set of tracks and generate infinite variations, from which I can pick and choose ones to embellish. Or, similarly, I have innumerable LPX projects that are just a single idea that, in the hands of someone more capable (or with more motivation) than myself, could be turned into full pieces. I'd love to be able to give LPX 2 or more of these projects and it generates transitional pieces between them. All LPX upgrades are free, but I'd happily pay for this feature.