Machine Learning and AI

Link: #406215

Eliezer Yudkowsky: "tl;dr: $200K/$1M in prizes (and potentially advance funding) to build MIRI a dataset (hopefully for training future text-generation systems to think out loud).

You know how people... View MoreEliezer Yudkowsky: "tl;dr: $200K/$1M in prizes (and potentially advance funding) to build MIRI a dataset (hopefully for training future text-generation systems to think out loud).

You know how people occasionally postulate that, in the future, an AGI will think something, and we will have any idea whatsoever what it was thinking, and this will supposedly have something to do with why we supposedly don't all die?

MIRI doesn't have much hope of this, honestly, and it probably wouldn't be enough even if it worked... but that doesn't stop us from wanting to investigate very early how this might play out. Maybe it's much easier than it looks, or even, possible at all!
(Some approaches in this field are already being followed; funding there is, we hope and think, not currently scarce, and a new project in the same field should not compete for funding with them.)

Our current intended approach is to build a dataset for AI-Dungeon-style runs, in which the writer thinks out loud in text about what should happen next (from both Watsonian and Doylist perspectives) in order to choose the dungeon's response to the user's actions. That is, we want to see if a GPT-style system can be trained to think out loud, in a way that actually participates in designing the output; such that we can change the outputs by changing the thoughts about them.

We're trying to get this Visible Thoughts Project to execute outside MIRI with MIRI funding. In part because it doesn't seem like a project that intrinsically requires within-MIRI execution; and in part because we think we're currently bottlenecked on something like management attention or executive capacity. A project like this one offers a valuable chance for us to meet new people we'd like to know, people who can have money added to them and produce requested things without a ton of micromanagement.

We're offering $200,000 as a prize/grant to anyone who can build the dataset we want to see. (Albeit with more total funding than that, to allow for advance funding of some attempts, and not just payment on delivery, including some attempts that may fail.) And a further $1M, for any persons or groups who can scale that dataset to a further 10X the size.

In principle, anybody could show up with a deliverable that's what we actually wanted, and take the money with no questions asked. In practice you should probably check in with us to make sure you understood what we wanted, even if you don't want any advance funding."

View Less

Visible Thoughts Project and Bounty Announcement - Machine Intelligence Research Institute

intelligence.org

We at MIRI are soliciting help with an AI-alignment project centered around building a dataset, described below. We have $200,000 in prizes for building the first fragments of the dataset, plus an add...