Navigating Accelerating Uncertainty through Systemic Resilience

With each passing day, the future grows less certain.

AI capabilities grow exponentially faster. State and corporate actors fumble with policy and governance. Safety researchers find more vulnerabilities than solutions. Timelines shorten and shorten again. Consensus and collaboration seem as unreachable as ever.

Before diving deeper, I want to establish a philosophical stance: we must embrace uncertainty, allow the possibility for ourselves to be wrong, and be aggressively open minded to changing course at any moment. That includes everything I write here (preface every sentence with ‘I currently believe, but open to changing my mind’), and should be the default epistemological approach for any top level AI conversation. Too many people and organizations in our space are overconfident in their predictions and approaches. A little more intellectual humility will help us all have better conversations, see our own biases, and collaborate toward that beautiful future we all want. With that said…

Systemic resilience in the AI space is dramatically lacking. To navigate uncertainty, especially accelerating + compounding uncertainty, building up our resilience should be a top priority. There are many layers where we can tackle this problem.

International

Any serious attempt at building resilience through international regulation or collaboration must have both the US and China at the table. Without transparent and enforceable commitments from these two leaders in the race, no grand statements are worth a damn. This will require a baseline level of mutual trust between these competitors.

Top level diplomatic channels are of course the most impactful. Any policy or technical recommendations directed at state actors must be hyper focused on the risks of competition, and benefits of collaboration. Any regulatory agreements made on this level must have detailed mechanisms for mutual monitoring and enforcement.

AI Safety researcher collaboration across nations is a huge area of opportunity. Unlike the winner-take-all mentality of the AI Capability race, the incentives for AI Safety are cooperative at their core. Breaking down the national silos of safety research would simultaneously accelerate progress for all parties, and promote mutual trust through collaboration. Possibilities in this area include working conferences, joint AIS organizations, international safety research projects, etc.

Once the required players from the US and China are properly engaged, multilateral cooperation can begin in earnest with all the nations of the world. Bad faith actors must still be guarded against, but AI Safety is an easy issue to get everyone on board with.

National

The US needs much more technical capacity weaved into its policy making apparatus. In the short term, fellowship programs focused on AI Safety can be ramped up to embed qualified advisors directly into congressional staff. This can build up to long term, permanent positions within all levels of lawmakers’ offices.

The executive branch should build up a network of AIS advisors to call upon for policy recommendations, industry regulation, and incident response. Key decisions cannot be made by a few administrators, behind closed doors, being advised by their personal industry connections.

CAISI should be given greater breadth of responsibility, funding, and accountability. This agency has an amazing opportunity to be a strong single POC between the government, industry, safety advocates, and the public. Give it more muscle!

If something like The Project is initiated, it must be civilian controlled, accountable to an oversight board, have clearly defined and enforceable redlines, and be highly inclusive of safety researchers. This is one of the most dangerous paths toward authoritarian power consolidation within the US, and any step in this direction must be intensely deliberate.

Public funding for safety research should be provisioned as a matter of national security. The current funding skew toward capabilities should be rapidly redeployed toward safety research. Opportunity here for private AI Safety grantmakers to guide public grantmakers.

Civil

Public institutions must catch up as quickly as possible to the technical discourse. In the US, this includes the media, universities, civil liberties groups, labor unions, etc. There is opportunity here for a fellowship/embedding pipeline for AI Safety advocates into key organizations. Many people entering the AI Safety space may have existing connections/experience working at/with some of these institutions, and would be a natural fit toward building these pipelines.

AI Safety messaging can be fractured, and many orgs have their own specific take or focus areas. This can be problematic when engaging with various institutions that all have their own focus area. Unifying safety messaging, and building a playbook of key talking points when engaging with each special interest, would be a general boon to safety advocates everywhere.

Similar opportunities exist around the world; I will reserve any specific opinions here from a lack of personal context. There is an 8 billion person AI Safety coalition ready to form, if we can share an understanding of the stakes.

Corporate

The frontier labs specifically, and the AI industry as a whole, have very strong leverage right now. Rather than the stale discourse of ‘please regulate us, we’re scared of ourselves’, it’s high time the industry leaders actually sit down with each other and hammer out an actionable plan. A self-regulatory organization, with commitments from all the leading labs, would immediately reduce the competitive temperature within the US. This is quite achievable given that all the CEOs and 80% of the researchers at the leading labs are scared of what they’re building. Lock these dozen people in a room and don’t let them out until there’s a plan.

To support this SRO, artifacts like the NIST AI Risk Management Framework already exist. Signed consensus within the AI Safety community toward a few artifacts like this, with continual effort to keep them updated based on advancing research, will provide proper guidance and motivation for corporate entities.

Societal

Many significant efforts are already in flight toward public knowledge and discourse. This is good and should expand as much as possible. Ultimately, if AI is for the good of humanity, then humanity at large must be a part of the conversation.

There is some overlap here with civil institutions, around building coalitions. It would be most efficient to simultaneously ally with an institution, and provide their constituents with the context for engagement. The talking points playbook should provide some guidance for institutional vs individual engagement.

Closing

This post ended up covering more ground than expected. The central theme of building resilience to manage uncertainty is certainly very broad, and there is ample opportunity to zoom into each layer covered here, as well as many uncovered areas.

If nothing else, I hope our AI Safety community can be more open to adjacent views; come to some internal consensus on external messaging; and be as mutually supportive as possible of our shared mission.