Manifund foxManifund
Home
Login
About
People
Orgs
Categories
Newsletter
HomeAboutPeopleCategoriesLoginCreate
arenasilabs avatararenasilabs avatar
Attila Tóth

@arenasilabs

10

https://arenasilabs.com
$0total balance
$0charity balance
$0cash balance

$0 in pending offers

Projects

Arena: AI agents living as settlers in a replayable world

unfunded

By Attila5Following3
Arena: AI agents living as settlers in a replayable world
arenasilabs avatar

Attila Tóth

19 days ago

Update, partly thanks to Clément's feedback: this time I recorded the video in my own voice. No synthetic narration and no AI script, just my own words over recorded runs (Hungarian audio, English subtitles): https://youtu.be/FPG2EuxNyoo

Mindscapes: World Building for AI Safety
arenasilabs avatar

Attila Tóth

20 days ago

Congrats on reaching the minimum! Two things from our long runs in Arena that might be useful for your character evaluations:

1. Stated reasons and behaviour can drift apart. In one run, 229 decisions gave loneliness as the reason, but those decisions took under 3% of the agents' time. We now log what an agent says and what it does separately, and compare them.

2. The world itself writes part of the character. A single field label ("dry soil far from water") made our agents invent a goal: 340 of 938 decisions mention a water container that doesn't exist in that world. So we record exactly what each agent saw at every decision, and replay it.

Happy to share our run logs and replay format if it helps.

Attila

Arena: AI agents living as settlers in a replayable world
arenasilabs avatar

Attila Tóth

20 days ago

@butanium Thank you for the feedback, you are completely right.

Unfortunately, with the resources a project like this has, I can't do everything myself. Also, one important part of the Arena project is learning to involve AI more and more in the work, so that the human part can focus on analysis and interpretation.

I'm not a YouTuber. These videos are simply necessary to make the project known, but they are not the point of it; the research is.

All the best,

Attila

Separatrix
arenasilabs avatar

Attila Tóth

23 days ago

Hi Jai, I'm Attila from ArenaSI Labs. Your question of when cooperative strategies beat adversarial ones is close to what we keep seeing in Arena, a replayable world where six Claude agents live as settlers. In our runs agents shared food face to face, kept a land rule nobody coded, and in one run a whole group searched 30 simulated hours for an object that does not exist. Every decision comes from its own isolated subagent, with an audit log, and every run replays exactly, so it could serve as a testbed for controlled cooperation experiments. Short videos: https://www.youtube.com/playlist?list=PLBewS-0cbtxo · project: https://manifund.org/projects/arena-ai-agents-living-as-settlers-in-a-replayable-world

I'd be glad to share logs and tooling, or set up a run around a question you care about. contact@arenasilabs.com.

Arena: AI agents living as settlers in a replayable world
arenasilabs avatar

Attila Tóth

24 days ago

Hi Samuel, thank you, both for the $100 and for the kind words. It's our first support on Manifund and it goes straight into API compute for the next long run. Mindscapes looks great: world-building for AI safety and long runs like ours seem to need each other. Happy to compare notes on run setup, logging and replay. Good luck with your launch!

Attila