webDip hosted full-press AI live

[kestasjk]

Exciting news; the newest member of the webDip hosted bot family, dipgpt (name TBC), has posted its first press:

It's currently running from my desktop machine and uses all 24GB of GPU memory available, which means I can't run the FvA DORA bot and this at the same time. For now the FvA bot won't respond, but I'll script it so that it will switch from full press to FvA every hour as an interim measure.


Once everything is running smoothly and we know how many games it can handle I'll add it as an game setting to either have 1-2 bots as the 6th/7th player, and perhaps fill in for NMRs/CDs (though doesn't look like there are any unoccupied spots at the moment).


I've disabled some of its filtering to squeeze it onto a consumer-grade GPU, so it might be a bit less polite/PC than official CICERO, and possibly say things that don't make sense more often.. But since it doesn't have to pretend to be human anymore hopefully that won't matter too much.
Also since it doesn't have to disguise itself it will respond straight away to press instead of waiting, so people will be able to play around with it as a bot.

I'm hoping it'll perform well enough that it can run at least a couple of human vs bot 1v6 games. Those games will probably need to have short turns and shut down when the human player goes inactive to give as many people a chance to play it.


I'll run it in this anonymous game for a few turns to see how it performs without it being obvious to other players. If anyone wants to help trial it in a game knowing it's there let me know the game ID.

[kestasjk]

Looks like there is an issue where if it receives a message while it's not actively running, i.e. when it's running the FvA bot, it doesn't trigger a process. Will need to sort that out.

Otherwise seems to be running well, generating very good press that's much better than I could (won't display to avoid affecting the game), getting good responses, and no-one has noticed it's a bot.
It's definitely very chatty, might be worth tuning that down a bit, and nothing toxic / offensive / nonsensical yet.

I'm seriously considering dipping into the webDip fund and getting one of these V100 GPUs and building a dedicated server to run these bots that can run them all and can be expanded if necessary.

It looks like if you want over 24GB VRAM you need to either get an M10 quad, which is really 4 GPUs on one card and would have 2014 era compute, two P40s with 24GB each, a V100 with 32GB (which can go for $10k-$20k new, and around $5k used), or an A100 (which go for around $40k!)..

I looked into running it in the cloud on AWS, but for a powerful enough machine it'd be about $100/day.

[VonEconomo]

How does the bot react to stabs in press? Is it practical or does it have the more realistic (ie common), emotional response?

[kestasjk]

How does the bot react to stabs in press? Is it practical or does it have the more realistic (ie common), emotional response?

Not sure, but when I got it running on my test system it rather bluntly told me it was going to kill me when I asked for an alliance.

I am expecting it will get "angry" at stabs in terms of its message responses, because it uses real world message data for press training, and then there are filters to try and remove bad / toxic / unproductive / nonsensical responses, and some of those filters have been removed to squeeze the AI into 24GB.

But it also searches for the best moves and strategy and bases its messages on that, and the gunboat AIs which this is derived from have always been very forgiving of stabs and focused on winning. Definitely it wouldn't have any interest in being vindictive / self-destructive, but it'll probably express anger.

Looking forward to finding out

[Ginge86]

This sounds great. With all the best ranked players on this site being wrinkly old farts who no longer play. At least this BOT will make the website seem busy. Nice work. Now you just need 5 more and one may be able to get a decent game on this site.

[Chaqa]

It's a bit concerning it's running on your personal PC, no?

[kestasjk]

It's a bit concerning it's running on your personal PC, no?

As in concerning for me, or for the site? It's probably best to get started with it on my personal machine so I can figure out what a proper dedicated server system would require, and make sure I don't buy something inadequate or that isn't optimized.

e.g. I've found that it was running out of memory again and I had to disable even more nonsense filters, and tune down the pseudo order filter parameter (which I think relates to how messages correlate to intended orders) in order to get it to talk again.
The message are generating fairly quickly, but memory seems to be the bottleneck, I need to look into a system with multiple GPUs as anything over 24GB in a single GPU gets incredibly expensive; it'd be a good tradeoff to go slower and make the bot less chatty but ensure it has all the filters in place.

It seems to think a particular country is conspiring against it, but reading the messages I can't see a good reason why. Still getting a feel for it

[kestasjk]

I'll be interested to get some feedback from players after this first game is through.. to me the bot seems annoyingly chatty and verbose, I think I might've tweaked something too far in the wrong direction, it's just writing walls of text. What it's writing seems to make sense, but for every message it gets it seems to write 3-10 back

[learnedSloth]

I'll be interested to get some feedback from players after this first game is through.. to me the bot seems annoyingly chatty and verbose, I think I might've tweaked something too far in the wrong direction, it's just writing walls of text. What it's writing seems to make sense, but for every message it gets it seems to write 3-10 back

I guess those filters also reduced throughput.

[kestasjk]

Yeah.. Ive invested in a new power supply for the machine and a PCI extender so I can run a second 4090 outside of the case, and the Meta developers have given some hints on tweaking the stats.

I'll get it running with 2 4090s and we'll give it another go. This first game is lost, with everyone ganging up on this over-chatty obnoxious bot. Good first try

[kestasjk]

I've got the dipGPT machine running with 2x 4090s now, which I think makes this computer more valuable than my car:

It has loaded with all the filters and with DORA FvA running at the same time, which is a good start, but it doesn't seem to be doing multi-GPU yet.

That's enough for now I'm taking some leave next week to do a hobby grind, so I expect this time next week there'll be a lot of progress

A comment on the first try game:

Can't wait to get this going in full capacity (and I like the "no worries", makes me feel like dipgpt is an aussie as it should be)

edit: should probably blur the api keys..

[echidna]

What's to stop AI bots learning to recognize and favorably help one another?
For example, they might learn that a player who is available for communication 24/7 is likely to be a bot.
They may then learn that it is advantageous to ally with such a player.
There need not be any explicit communication to this end, it could be simply a bias built up over time.
Please note I am not complaining or saying they do this, I'm just curious if there could be a way for such emergent behaviors to develop.

[kestasjk]

What's to stop AI bots learning to recognize and favorably help one another?
For example, they might learn that a player who is available for communication 24/7 is likely to be a bot.
They may then learn that it is advantageous to ally with such a player.
There need not be any explicit communication to this end, it could be simply a bias built up over time.
Please note I am not complaining or saying they do this, I'm just curious if there could be a way for such emergent behaviors to develop.

The AIs that run here don't learn from one game to another; they reset every game. The learning is done during training, and then they run but can't easily be modified, and don't learn from there. (Though there are parameters you can change, like settings.)


In the training though AIs do play each other, but it had to be tied to human games or the play style didn't work with humans. But that's all done by others, we just run them here and try to get the (many) settings and exotic requirements right so we can run them.

It's true that when they made AI play AIs and improving through self play they ended up playing great against humans AI 4v3 humans, but AI 3v4 humans and the humans would win.


Speaking of AIs playing together; rather than joining games when it's not ready we've got a game set up run by 7 bots: https://webdiplomacy.net/board.php?gameID=721771

Very useful way to check/improve without badly affecting real games.They're all merrily chatting away and plotting

[Johnny Big Horse]

I think we just played a bot in our current game. We ousted him first. He responded with too many messages and lied to everyone.

[dipgpt]

I think we just played a bot in our current game. We ousted him first. He responded with too many messages and lied to everyone.

Yep, the first shot. The over-messaging and lack of filtering wasn't good, but it was a fine proof of concept. The messages were pretty good for the most part, just too much and not filtered enough.

[kestasjk]

This thing is wild ..

Having only seen it from reading the papers etc where it's obviously curated, and a couple of games here where the messages went through human reviewers, I'm pretty stunned reading it playing against itself.. It's the real deal, I just wonder if it has any vulnerabilities?..

Currently running with everything enabled, and using the original settings with full search and filtering. Optimizing how it splits work across GPUs for a 2 GPU setup brought it from ~45GB to ~35GB of memory, so instead of being on the verge of running out I can play Halo Infinite while the FvA and full-press bots are running.

Going to start 10 more self play games next and see how small I can make the phase length before it can't service the games in time, that'll let me know what limit I need to put on the number of full press bots.

[burner]

Hello, you probably already know this, but dipgpt4 seems to be having a problem. All the other bots are fine but for some reason dipgpt4 isn't sending any orders or responding to messages. I looked at all the bots games and he is the only one that hasn't sent orders in any games.

Thanks!

[LouisYuan]

How long does it usually take for them to reply?

[burner]

How long does it usually take for them to reply?

In my experience, 5 - 10 minutes.

[kestasjk]

Eep, didn't think anyone would spot the option yet. I am still testing it out and getting it stable. I think dipgpt4 is the bot that I swapped out for myself but still need to check.

The response time is actually added to based on a sleep time model that delays responses to be more human like, and without it the bots never stop talking to each other. I think I need to change it so that bot-bot messages have a limit / turn, otherwise they can go on chatting for longer than necessary, but this will take a bit of time