GET THE MICROSLOP SHIRT: https://store.gamersnexus.net/products/microslop-parody-t-shirt and support our DATA DRAGNET campaign: https://store.gamersnexus.net...
But I run LM studio and comfy ui on my on hardware just to tinker. These generative models are fascinating, but sycophantic and inaccurate.
To be fair, models you can run on your local hardware are severely limited in comparison to what big models like deepseek 4.1, glm 5.3, gpt sol, Claude opus etc can accomplish.
It’s true those models can do more, but they are still sycophantic hallucinating machines. Whose outputs need to still be rigorously checked. So much so that between the cost they charge you to run them and the time you spend. You don’t really save anything over just learning and doing it yourself generally.
I watched a Youtube video the other day of a guy using these frontier models to generate code and assets. He wanted to make his own Mario 64 like. All said and done by the end of the project that he’d done for the video. He’d spent nearly three thousand dollars for low quality assets and real jank.
So much so that between the cost they charge you to run them and the time you spend. You don’t really save anything over just learning and doing it yourself generally.
You do, however, save over running local models using your own electricity and even more time wasted on correcting its output.
GLM 5.3 Flash, Deepseek 4.1 Flash are super cheap. Gemini 4 Hydrogen is a bit more expensive, but still among the cheapest western models and now trading blows with the big expensive models.
And most of those can be used with some form of subscription that subsidizes the usage so if you make sure to spend all your allocated usage every week, you can get like 100 dollars worth of usage out of a 20 dollar sub, if not more.
Two things, that’s extremely debatable and wildly subjective. And even if it were “free” it would still be too much for myself. Just because of who’s running it. I’m a Luddite, just not anti generative AI.
It is going to continue to improve little by little. But outside fuzzing and debugging code it really isn’t a good value proposition. When it comes to making assets especially. Someone like myself is still going to have to go over it heavily and fix it. I’ve had some of these assets on my workbench. Fixing them took me nearly as long as remaking them from scratch. And that’s not going to improve any time soon. There’s a huge gulf between looking marginally usable and actually being usable.
To be fair, models you can run on your local hardware are severely limited in comparison to what big models like deepseek 4.1, glm 5.3, gpt sol, Claude opus etc can accomplish.
It’s true those models can do more, but they are still sycophantic hallucinating machines. Whose outputs need to still be rigorously checked. So much so that between the cost they charge you to run them and the time you spend. You don’t really save anything over just learning and doing it yourself generally.
I watched a Youtube video the other day of a guy using these frontier models to generate code and assets. He wanted to make his own Mario 64 like. All said and done by the end of the project that he’d done for the video. He’d spent nearly three thousand dollars for low quality assets and real jank.
You do, however, save over running local models using your own electricity and even more time wasted on correcting its output.
GLM 5.3 Flash, Deepseek 4.1 Flash are super cheap. Gemini 4 Hydrogen is a bit more expensive, but still among the cheapest western models and now trading blows with the big expensive models.
And most of those can be used with some form of subscription that subsidizes the usage so if you make sure to spend all your allocated usage every week, you can get like 100 dollars worth of usage out of a 20 dollar sub, if not more.
Two things, that’s extremely debatable and wildly subjective. And even if it were “free” it would still be too much for myself. Just because of who’s running it. I’m a Luddite, just not anti generative AI.
It is going to continue to improve little by little. But outside fuzzing and debugging code it really isn’t a good value proposition. When it comes to making assets especially. Someone like myself is still going to have to go over it heavily and fix it. I’ve had some of these assets on my workbench. Fixing them took me nearly as long as remaking them from scratch. And that’s not going to improve any time soon. There’s a huge gulf between looking marginally usable and actually being usable.