Somewhat related: I wrote up a MTG card generator based on nanoGPT a while ago that I think produces pretty good results for being 1m parameters.
The real neat thing about this is that WotC makes a few thousand new cards each year, so my training data set just grows over time and the model gets better with no effort spent on my part.
It would be interesting to come up with a use case which requires a freshly trained model and isn't just something that generic models can already, especially with 1MM context window
You can see the invention of RLHF/ChatGPT here because text generation suddenly became much more coherent and also much less interesting. You have to go back to older tech for surrealism because nobody will let you see the good stuff (the base models).
I guess I was much more interested in being able to work with an LLM to create good, synergistic Commander decks and less interested in generating custom Magic cards.
I'm sure I can dig up info on how to do this and piece it together, but I thought OP might have a guide specifically for it.
FWIW, there was a pretty popular post on HN around generating MTG cards using AI a couple years back but I believe that their approach was a fine-tune on an existing LLM.
This is neat, I have a simple CSS effect that I apply to foil cards on my Magic card marketplace site, but I have wondered what it would take to replicate all the different foil effects that a Magic card can come in to be both more varied and more true to life, in a performant way that fits into my CRUD svelte UX.
some people just mastered their domain. i pray i get this good in my area of expertise: elliptic curves.
lol im researching curves on random math sites, and learning different implementation tricks (like the subtle diff in MSB and LSB implementations of the double and add algorithm* or that sometimes you can find the weirdest endomorphisms on non-binary Koblitz curves (they're not documented AFAIK and Koblitz the OG is too old to be chasing 1-bit, 2-bit endomorphisms)
This was me in awe of the pokemon card CSS lol. It appears to have been done before AI and this makes it even more impressive tbh.
*The diff is in MSB->LSB the generator remains constant while in LSB->MSB to the generator doubles. Either way, you always know the value of the generator at every bit. I think this is a vulnerability but I just can't find out how.
This almost exact thing happened with me with a hiking app I used to make. We had a 7-day trial when you started the app, but we shipped a release that broke this and made all the paid features immediately paywalled. This led to a big increase in sales and we got rid of the trial forever.
It’s possible that our trial might have worked better if it were like modern iOS trials that start charging you after a certain period, but ours just let you use the paid features for 7 days and then lapsed, and it stifled sales. My theory was people urgently needed the paid features (mostly to download maps for a coming trip) so the trial got in the way of them paying right away.
I find using AI very educational in some respects.
I am way more knowledgeable about SQL than I have ever been, because in the past I knew so little I would lean on team members to do SQL for me. But with AI, I learned all the basics by reading code it produced for me and now I can write SQL from scratch when needed.
Similarly for Tailwind… after having the AI write a lot of Tailwind for me from a cold start in my own Tailwind knowledge, now I know all the classes, and when it’s quicker, I just type them in myself.
In GIS world, a vector tile is a chunk of geographic data (the vectors) limited to a geographic region (the tile boundaries which fit into the projected checkerboard of your map).
You use a vector tile instead of a png or jpeg tile because you don’t want an image representation of the data, you want the raw “vector” data so you can style it, search it, and do other things with it on client devices.
Forget squarespace websites. AI is such a huge time saver for complex front ends. It doesn’t write the whole thing, but you can easily go from screenshot to code you are better off using as a starter than typing it all up yourself. And it’s way easier to make css or tailwind changes to modify components than thinking really hard.
