Token donation to signal support. (Somewhat amusing juxtaposing with Neel's much larger token regrant.)
6
Investigating and informing the public about the trajectory of AI
ProposalGrant
Closes February 3rd, 2025$5,025raised
$500minimum funding
$10,000,000funding goal
Offer to donate
17 daysleft to contribute
You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
Project summary
Epoch AI is a leading research institute that develops public knowledge on AI's trajectory through rigorous research, novel benchmarking, and comprehensive data tracking. We seek general support to expand our work informing policymakers, industry leaders, and society on AI development and its implications. Our unique position combining technical depth with accessible communication has made us a trusted source for stakeholders ranging from government officials to tech leaders.
What are this project's goals? How will you achieve them?
Broadly, our mission is to conduct and disseminate rigorous research on the trajectory and impact of artificial intelligence, to enable better policy and decision-making. With additional funding, we aim to expand three key initiatives:
Data Curation: Building on our successful AI model and hardware tracking to cover AI clusters and companies, launching weekly data insights for timely analysis.
Capability Measurement: Growing our AI Benchmarking Hub through independent model evaluations and novel benchmarks like FrontierMath.
Economic Modeling: Developing a framework combining AI scaling laws with growth theory to analyze automation impacts and compute scaling.
We'll achieve these through strategic hiring of technical talent, maintaining our commitment to rigorous analysis, and expanding our successful track record of creating accessible, high-impact content.
How will this funding be used?
We're seeking $10M over two years to scale our impact:
$2M/year for the data program (tracking AI systems, hardware, clusters)
$2M/year for AI benchmarking (capability measurement, new benchmark development)
$6M for team expansion including an AI Data Lead, research engineers, and operations staff
Reaching 30 full-time employees while maintaining project-specific contractor relationships
Who is on your team? What's your track record on similar projects?
Our 23-person team (https://epoch.ai/about) has demonstrated strong execution in 2024:
Published groundbreaking benchmarks (FrontierMath) with 70+ mathematicians
Released major reports cited by tech leaders and government officials
Published in top venues (ICLM, NeurIPS 2024)
Grew audience from 2K to 15K Twitter followers
Attracted 360K website users
Successfully scaled from 10 to 23 full-time employees this year
What are the most likely causes and outcomes if this project fails?
Talent Competition: The competitive market for AI researchers could dampen our capacity to attract and retain talent
Data Access: Limited visibility into private AI development could affect our tracking capabilities
How much money have you raised in the last 12 months, and from where?
Roughly 75% of our funding comes from Open Philanthropy. The rest comes from individual donors, organizations such as Survival and Flourishing Fund, or from service agreements with collaborators on projects aligned with our mission.
offering $25
offering $5,000
Neel Nanda
6 days ago
I'm re-granting because I've been very impressed with the quality of Epoch's work. It's clearly far and away better than any competition when it comes to actually understanding what's going on with AI (and being publicly communicated), and several people I know who are part of Epoch strike me as highly competent. I'm also a fan of their new weekly newsletter. I haven't carefully checked the details of much of my work myself, but the one example where I thought I found an error (using total rather than active parameters of an MoE model), they then produced a newsletter with a compelling argument for why their way made more sense, which I was impressed by.
I see the main theory of impact of Epoch as broadly helping key decisionmakers in society (policy-makers, businesses, the AI safety community, etc) be well informed about AI: what has already happened, what is happening, what will happen. I'm a bit less confident in how well their projects do on the dissemination front. I see them go viral fairly often, though I am totally in a massive bubble. I've seen some scattered examples like Satya Nadella having Epoch graphs in a slide deck, which seem pretty promising!
Is this theory of impact important?
I think that AI is clearly a big deal, at this point, having an important impact on society, and expect this to accelerate. In addition to the big issues of misuse and misalignment, there's also just a lot of social and policy change that needs to happen to handle this well. By default, society changes slowly, and tech policy often makes little sense. High quality information is far from sufficient to fix this, but it helps!
I have some concern that this also has negative externalities by increasing the number of people who realise AI is a big deal and decide to race (in a similar sense to how, eg, scaling laws got more people realising the potential of AI). But my guess is that this isn't that big a deal. I think that "AI is a big deal" is pretty widely believed at this point, so that ship has largely sailed. Another concern is that this increases competitive pressures by making it clearer to people when they're losing the race. But on the other hand, this helps the person in first place realise it, which is high value (see eg the Cold War "missile gap"). And I expect key actors like AGI labs already have much better awareness of what's going on than most, so Epoch adds less value to them.
Overall, this isn't my area of expertise so I'm not confident, but my best guess is that Epoch is doing great work.
I'm only donating a token amount as I don't think this is my comparative advantage as a funder, but I would encourage others to donate, and want Epoch to have as much funding as they can productively use.
Oliver Habryka
6 days ago
Agree on this. I do think the strongest argument against funding Epoch for most people on Manifund is that it can and probably will be supported aggressively by existing large funders. Evidence that this isn't the case would update me a lot on more funding to Epoch being valuable.