Where were you when Claude emptied his whole knowledge base?
Claude Plays Pokemon
Sand at lvl 60
Has this fucker made ANY progress since Diglett's Cave?
I had a lot of fun but I think it's time to let go
hahah, no
the indian bot gets more views than claude
we need a 2 year timeskip.
You niggas still seriously going?
30 viewers
pull the plug dude lol
not even able to beat the Gemini stream lol
This event was a lot of fun while it lasted. Good times in February and March.
So how did the other bot get all 8 badges and Claude just got stuck in a loop?
2 more weeks in mt moon and he'll figure out route 9 next time
Constant intervention to guide it every step of the way. If you read the FAQ on gemini's stream it's all just cope to try and explain how he "doesn't cheat" lel.
It'll always hold a special place in my heart, but this whole thing proved to me that AI is a dead end and all this investment money will go nowhere fast.
And this just solidified it even more. Imagine cheating at a children's JRPG.
I told you guys that there wasn't going to be any real progress after Surge, long before he even cleared Surge. That's the furthest Claude had got, although according to the developer one of his offstream claudes is lost in Celadon. Catching a Diglett and having it learn dig is really what killed things this time. Still, the fact he caught all those clefairies was funny, and getting Flash was something. But once he got dig it was just over.
Constant dev intervention. The idea of Claude is having chat bot try to play a game it's not trained to do and not have the dev intervent.
When you program AI to play the game and even jump in to guide it when it's stuck, you lose the thing that made Claude special
he also put ads on his stream and is trying to scam people into following him. literally jeet plays pokemon. what's he gonna do, play gen 2 next?
So where are we at now?
Still stuck because of mt. moon and dig?
When you program AI
that jeet didn't program shit. it's not like he works for Gemini. he's just feeding it prompts, giving it hints and coordinates and the like.
Yeah, and it's been like this for weeks and he still can't get anywhere. Sand keeps gaining levels. The run is effectively dead. We had a good run
this whole thing proved to me that AI is a dead end and all this investment money will go nowhere fast.
Ehhhh I don't really know if that's the lesson here.
Look at how far things like image gen and video gen have developed in the last couple of years. Even Claude's "reasoning" abilities itself were so poor not long ago that it couldn't even get out of Pallet Town. LLMs may have plateaued for now, but that doesn't mean there isn't ways they can improve in the future. New AI tech is being worked on all the time, we really don't know what's going to happen in the next few years. A truly intelligent AI may be ways away, and it will likely be something entirely different than LLMs, but that doesn't mean it won't happen just because this iteration of a bot went full retard in a game that it's not even designed to play.
I'm really quite impressed with what Claude was able to do honestly given the tools he has. People can say what they want about artificial intelligence but until now there's been nothing that you could just plug into a game and expect it to figure it out on its own and provide you its logic along the way, without any handholding or training asides from having another instance of itself to talk to.
There's really not gonna be a new iteration?
It's weird how the guy basically abandoned it but left it running. Can't be cheap to run Claude 24/7 just for this bullshit loop people don't watch anymore
Anon Babble shuts down and returns
claude is still stuck in mt moon
just pull the plug already
The other one was fed a mini map and knows what every tile does from data rather than relying on vision. Basically unlike Claude it knows what trees can be cut and that the hole in the wall in the trashed house is an exit.
Claude's streamer is the Claude's company's employee. If outsider was running it, it would cost couple hindreds bucks a day for the tokens
Wasn't critical, but the dev being too lazy to fix the bicycle controls hurt as well
is there a name to this type of media? bots just generating endless boring slop 24/7/365
I miss the threads. They were maximum comfy.
Mt Moon killed me, but time did me in.
I know, but it still costs the company to run it. I figure the added publicity might have paid for itself early on (even though its performance was fucking embarrassing), but what's the rationale at this point?
Yes, our AI still can't get halfway through a game for 7-year-olds
What data are they still getting out of this?
What data are they still getting out of this?
The older off stream iteration of Claude who got through Surge is also still running so something I guess.
Isn't this the one that started off as Infinite Sienfeld but it made an observation about trannies "You can't make jokes about them or they get angry", followed by them getting angry and getting shut down?
I just want to watch how he will fare in other games already.
The idea is still fresh and fun, but this run is bricked..
seems like a pointless project to keep running. fix it and continue then restart it completely to show it can work
just kill it and run it again when the new fancy model comes out
Or make it play Crystal
it won't be the model thats smarter though it'll be the tool usage side that needs to give it better data/feedback. it doesn't know what to do because there is no data about this specific scenario and its just going to keep looping back and fucking up. its probably making it worse like a cancer now that its been running wrong for so long, so everything it tries to look up in its own memory makes it worse
I don't know how tools could be realistically improved, but I thought you could make critique Claude more anal about loop prevention.
Claude keeps tricking himself into thinking he's actually making progress, somebody needs to constantly remind him that he didn't until he isn't inside rock tunnel.
Yeah but it's a good demonstration of current "AI" real capabilities when confronted to real time problem solving and it shows that it is dumb as a brick and no amount of 2 more weeks fear mongering from Sam Altman changes the fact that the current approach to AI results in what is essentially a glorified search engine that's good at rewording and summarizing what you ask it, but can't actually reason. Image gen is pretty cool, especially in capable hands that know what they are doing but video is still unstable and very limited, it looks terrible.
Anon Babble has nothing else to do
For me it's AI Spongebob.
youtube.com
loop escape and trying other things would be part of tool use. i haven't paid to much attention to this project but i assume theres a prompt and its been fed rules on how pokemon works as a game, mechanics and stuff, plus what it should try to achieve. somewhere in the info its using theres an obvious gap, there needs to be a stress about dig and what its used for and why not to use it randomly
I don't know how tools could be realistically improved
Are you kidding? He's here because he can't identify cuttable bushes. He can't tell which direction his sprite is facing or even which one is him. Every time he scrolls the screen, it looks like a new map to him. All this shit can be better represented and fed to him. Something as simple as a heatmap to tell where he's been could go a long way. But they won't even fix the bicycle moving two spaces, let alone any of that
Vision isn't part of the external tools to play Pokemon I think, that's on the model. Heatmap could be added probably.
Look at how far things like image gen and video gen have developed in the last couple of years. Even Claude's "reasoning" abilities itself were so poor not long ago that it couldn't even get out of Pallet Town. LLMs may have plateaued for now, but that doesn't mean there isn't ways they can improve in the future. New AI tech is being worked on all the time, we really don't know what's going to happen in the next few years. A truly intelligent AI may be ways away, and it will likely be something entirely different than LLMs, but that doesn't mean it won't happen just because this iteration of a bot went full retard in a game that it's not even designed to play.
I guess I can agree with that. My overall original point is that this is definitely a new AI winter unless something really shakes things up.
I miss the SS Anne days. Does anyone have that webm of the sailor kicking Claude into the stratosphere after being shown the SS ticket one too many times?
He has an overlay that prints coordinates for each tile and marks impassable ones, plus a navigator that pathfinds for him. They're just shit. It's why he has trouble with ledges. They're "impassable". I don't think we've seen exactly what the overlay looks like but if it obscures the cuttable bushes it might have genuinely ruined the run.
Realistically, that could be improved