Turn Neuro bald and make her play Balatro

[tender whale]

Turn Neuro bald and make her play Balatro.

Thank you sincerely yours truly Boss or something.

[solemn rune]

Never cook again

[spark vapor]

while also ranting about random life topics

[bold bone]

Trailer: https://www.youtube.com/watch?v=VUyP21iQ_-g
Gameplay: https://www.youtube.com/watch?v=T2J7FEbHXys

[calm glen]

Oh, I'm playing it, it's a great game

[craggy oracle]

can we also get 1000 streams of the binding of isaac while we're at it

[vivid glacier]

great idea

[bold bone]

damn a planet X reference? respect

[bold bone]

Hmm this game seems to have a lot of complexity. To check feasibility, it might be useful to compare it to Slay the Spire. Note that this is based off of that 10min gameplay + fragments of another run + some wiki stuff.

You have 4 decks (jokers, consumables, playing cards and vouchers). There are three classes of consumables, and jokers/playing cards can 'enhancements', 'additions' and 'seals' that augment their behaviour. Thus, there are a lot of ways that a deck can be built and it might be hard to design an AI that can learn to intelligently build a set of decks that work together and commit to a specific strategy.

Balatro isn't really turn based like StS, instead you have to reach a certain score using a fixed number of times you can submit a hand and a fixed number of redraws. Thus you need to do some planning and risk analysis to determine when to discard cards or when to draw cards. I assume it would be difficult to do that with AI.

The score achieved by each hand is also complex. The score of each hand card is calculated by multiplying the chips number by mult number. Some cards add to chips, some cards add to mult, so it's important that you play hands that use both the high scoring mults and chips together. Also whether each hand is a 'flush' or a 'pair' or 'three-of-a-kind' also changes the initial value of chips and mult. Therefore, the AI approach needed might be needed to somehow reason about what cards work well together. However, in StS, you can just use the 'strongest' cards (because the effect of almost all cards are independent of each other).

All the points here can be undermined if we believe there's a heuristic that works very effectively (e.g. pick 3 cards with highest chips value and 2 cards with highest mult value). Alternatively, maybe I am underestimating the power of neural networks to learn stuff.

Also, ofc the fact that Balatro has so many mechanics would make watching Neuro play it even more interesting

[vivid glacier]

bump. this is a really good idea

[heady sonnet]

Neuro doing a stream where she just pretends to be Northernlion
Neurolion

[vivid glacier]

Think about it: Neuro just ranting about various life topics, while going off on some random chatter for 5 minutes straight.

[heady sonnet]

I wonder if it would be possible to get a bunch of transcripts of Northernlions streams

and then you could train neuro on them

so she really does just act like him

[bold bone]

apparently yt-dlp can download autogenerated subtitles off videos (without downloading the video itself)

[vivid glacier]

That does not sound ethical.

[bold bone]

obvs

[heady sonnet]

i think they call that data scraping

[mellow forge]

we'll call it content

[shell timber]

That would be very very funny

[rapid lance]

I wasn't sure about this idea when it was posted but i now get the reference and am now on board with this it sounds fun

[bold bone]

necroing because this critique was based on vedal making a custom balatro AI
If Neuro's LLM can just ingest the game's wiki and play reasonably, then it sounds doable
Perhaps Neuro would need a calculator tool to do the maths right though

It would be funny to see her mess up the maths