Update: This project now has funding (I'm working with David Krueger's lab and able to use their funding), so I won't be withdrawing funds from this, and funders should probably look elsewhere for opportunities!
[Funded] Gabriel Mukobi Summer Research
Note: This project now has funding from outside Manifund, so probably look elsewhere for opportunities!
Project summary
Gabe is requesting $5000 to pay for LLM compute to run experiments.
From Gabe's proposal on Nonlinear Network:
I’m seeing around $5000 for AI model API compute funds (GPT-3.5/4, Claude, and PaLM) for multipolar coordination failure evaluations during the Existential Risk Alliance (ERA)’s summer research fellowship.
As a rough BOTEC, I might imagine running 12 experiments 1024 data points per experiment 8192 tokens per data point * $0.045 / 1000 tokens for GPT-4 = $4530.
The amount of funding is somewhat flexible. With less funding, I’d just be able to run fewer experiments or have to use worse models (like gpt-3.5-turbo). With more funding, I’d have more room to run more complicated experiments.
I commit to giving back any extra funds remaining at the end.
I’ll be working on this project with mentorship from Alan Chan (Mila, Krueger Lab) and Jesse Clifton (CLR, Cooperative AI Foundation).
Unfortunately, ERA and these mentors do not have clear compute budgets they could allocate for my project, which is why I’m seeking funding.
That said, it’s probably not the worst if you didn’t fund this, as I might be able to get compute funds through them, it will just be more difficult.
Other sources of funding
Gabe is considering applying to an OpenAI grant program, or receiving compute from his lab
Gabe Mukobi
over 1 year ago
Thanks for this grant and for the Manifund platform! I should let you know that I'll be slightly pivoting this project to focus first on building a general sum version of Diplomacy or a similar game for language models with the intention of evaluating and improving LM cooperation (since normal Diplomacy is zero sum which leads to poor cooperative equilibria). I still want to pursue making a scary demo for multipolar risks in open-ended environments, but in the side to start as it seems more unstructured, and I hope the first thing will help lead to this second thing.
I'll probably be using similar API model resources and will write up more about this in the next week, but wanted to share in case this is getting outside the scope of what funders here want to support.
Rachel Weinberg
over 1 year ago
Approving this grant on behalf of Manifund! This project is definitely within scope and Gabe & his team are legit.
Austin Chen
over 1 year ago
Main points in favor of this grant
This grant is primarily a bet on Gabriel, based on his previous track record and his communication demonstrated in a 20min call (notes)
Started Stanford AI Alignment; previous recipient of OpenPhil fieldbuilding grant
His proposal received multiple upvotes from screeners on the Nonlinear Network
I also appreciated the display of projects on his personal website; I vibe with students who hack on lots of personal side projects, and the specific projects seem reasonably impressive at a glance
EA aligned
Main reservations
I don't feel particularly well qualified to judge the specifics of the proposed experiment myself, and am trusting that he and his colleagues will do a good job reporting the results
Process for deciding grant amount
Gabe requested $5000 for this project, but as he's planning to apply to several other sources of funding (and other Nonlinear Network grantmakers have not yet reached out), filling half of that with my regrantor budget seemed reasonable.
Conflicts of interest
None