Seems like now's a good time to share that in July I started a new role as Lead Writer on the International AI Safety Report! Working with Dr. Bengio, my co-lead @carinaprunkl, and our ridiculously talented writing team has been a blast. Excited to see our first output launch!

Come work with us! I think the seasonal fellowship is highly impactful. Fellows have gone on to a range of interesting and important roles, and credited the program with tangibly accelerating their careers. It's also a lot of fun! Great opportunity for the right person.

New paper! 📄 “Supervising Frontier AI Developers”, by GovAI Research Scholar Peter Wills. (1/3)

Check out the summary here:

You can now read a summary of this piece on the GovAI site! We discuss different motivations for accelerating the spread of AI benefits internationally – including strategic considerations – and what this could look like in practice. (Link to post in reply.)

Here’s another link to the full report: https://t.co/shQrrcCJlU And thanks again to our incredible co-authors!

🎯 Bottom line: Benefit sharing isn't separate from other international AI governance conversations: it's integral to them. It's how frontier AI states can fulfil the technology's promise while advancing their strategic priorities.

AI capabilities are advancing while international competition intensifies. But benefit sharing isn't just idealistic. It could be the key to unlocking mutually beneficial agreements that both expand access AND reduce shared risks.

But there are also some actions that could be taken today. One concrete example? Digital infrastructure gaps will predictably slow AI's benefits from reaching many parts of the world. Investing today could unlock broader benefit sharing tomorrow.

If actors want to advance benefit sharing, they could fund research to address core uncertainties around risks and benefits. They could also engage in dialogue with the full range of stakeholders, including reps from low- and middle-income countries.

These challenges will need to be carefully considered in proposals and implementation plans. For example, sharing potentially dual-use AI systems should likely be paired with oversight and accountability mechanisms.

Across these options, though, there are some considerable challenges. Benefit sharing could be: 🔄 Redundant (market forces might be enough) 🏗️ Intractable (hard to implement) ⚠️ Dual-use (could increase certain risks) 🌐 Geopolitically constrained 🪤 Leveraged for coercion

AI companies could do this through something like a Windfall Clause, for example, or countries could raise special taxes or just increase development finance

3️⃣ Sharing financial proceeds would involve simply sending a portion of their economic proceeds from AI development to beneficiaries.

2️⃣ Sharing access to AI systems would involve making AI more widely accessible. This could be accomplished through subsidies or differential pricing, and be focused on particular applications or systems

1️⃣ Sharing AI development resources would involve widening access to the inputs needed to actually develop AI, such as chips and data. For example, an actor could provide data to developers building beneficial applications, or subsidize their compute access.

Next, we discuss how benefit sharing could actually be implemented. Here again we see three options: 1️⃣ Sharing AI development resources 2️⃣ Sharing Access to AI systems 3️⃣ Sharing Financial proceeds Each has distinct advantages and disadvantages.

🌎 Achieving geopolitical aims: Benefit sharing can help advance many geopolitical goals, from incentivising participation in international risk-reduction efforts to countering the influence of adversaries abroad by accelerating adoption of your own technologies

🧑‍💻 Supporting technological self-determination: As AI systems become more useful, we will become more reliant on them. Benefit sharing can help lower-income countries develop the capacity to develop and adopt AI in accordance with their own goals and needs

👍 Spreading economic and societal benefits: By default, AI may not fulfill its promise to benefit all of humanity. Benefit sharing can help help AI development support global development goals and compensate people who lose their jobs