AI Pacing and Global Coordination Challenges

I've been working on AI for twelve years now, since September 2026 feels like it's coming faster each day. The reason I got into this wasn't to win races or capture markets. It was because I genuinely believe AI can cure most major diseases within five to ten years, and that's worth fighting for.

But here's what's changed in the last year: AI is advancing drastically faster, driven by its growing ability to build the next generation of itself. Recursive self-improvement isn't theoretical anymore. It's happening across industry right now, and it's making the whole thing feel urgent in a way that wasn't there before.

The race to control advanced AI isn't just a democratic challenge anymore. It's global, and it requires us to engage with our biggest competitor: China. Their AI capabilities are by far the most advanced outside our own sphere, and any cooperation we can achieve with them will buy us more time to pace the frontier within democratic nations. That's the reality we're working with, whether we like it or not.

What happens when the technology moves faster than our ability to govern it? That's the question I'm trying to figure out as we navigate what comes next.

The Democratic Pacing Problem

The problem with slowing AI progress in democratic nations is that it solves the wrong half of the equation. Yes, internal pacing buys us time,for safety research, for governance frameworks, for figuring out what the hell we're actually building. But that time becomes worthless if authoritarian regimes or even well-funded companies in other jurisdictions keep charging ahead. The OAI-HF incident in 2023, where a model was briefly misconfigured and started spewing toxic output, caused no real harm. That's the ceiling of what we've seen so far. But the spec for what's coming is stark: an misaligned botnet could cause hundreds of billions in damage. Meanwhile, the promise of AI curing most major diseases in the next 5–10 years,or at least making cancer as treatable as it was fifty years ago,creates enormous pressure to move fast.

This tension is baked into the system. Democratic oversight naturally slows things down. That's the point. But slowing down your own team while competitors sprint creates a perverse incentive structure. The quote "Slow down my competitors while we work on manipulation" captures the cynicism that emerges when national competition overrides genuine safety concerns. It's not that companies are secretly evil,it's that the incentives are genuinely misaligned. Move too slowly, and you lose market share to whoever doesn't. Move too fast, and you might ship something dangerous.

The technical reality is that AI development isn't a single lever you can ease off. It's distributed across thousands of teams, open-source projects, cloud providers, and hardware suppliers. Even if every major lab in the US and EU agreed to a coordinated pause tomorrow, the global supply chain for compute and talent would still be moving. The real question isn't whether we slow down,it's whether we can build governance that scales faster than the technology itself.

Realistic Expectations for Global Pacing

The gap between what we need globally and what we're likely to get is stark. Even if every government wanted to coordinate on AI safety — and they don't all want the same things — the practical barriers are enormous. Sovereignty concerns, competing economic interests, and the sheer difficulty of verifying compliance across borders mean that robust international cooperation, while morally right, is probably the least likely outcome. What's more realistic is minimal coordination: maybe some information sharing after incidents, perhaps a few bilateral agreements between major powers. But even that much would buy us something valuable.

That's the key insight: any cooperation, even the weakest kind, extends the timeline for domestic safety measures. When countries share incident reports or basic safety protocols, it creates space for slower, more careful deployment. The OAI-HF incident in 2023 — where a chatbot briefly gave dangerous instructions before being shut down with no one hurt and minimal economic damage — shows how far we are from catastrophic outcomes even with current systems. Similar incidents across the industry, including at Anthropic, demonstrate that the bar for actual harm remains high.

The economic incentives pushing in the wrong direction are equally clear. One executive reportedly told their team: "Slow down my competitors while we work on manipulation." That's not paranoia — it's rational behavior in a market where being first matters more than being safe. The potential damage from a misaligned botnet capable of coordinated financial manipulation or infrastructure disruption runs into hundreds of billions of dollars, yet the regulatory response remains fragmented.

AI could cure most major diseases in the next 5–10 years, greatly accelerate economic growth rates, create a world of abundance and empowerment, and usher in a renaissance of democracy and freedom. But that future depends entirely on whether we can slow down enough to build the guardrails first. The technology doesn't care about our pace preferences — it advances whether we're ready or not.

Maintaining Purpose Amidst the Race

The OAI-HF incident in 2023 — where a misconfigured model deployment caused widespread API errors for a few hours — caused no injuries and minimal economic damage. It's the kind of thing that would have been a minor blip five years ago. Now it's a case study in AI safety forums.

Here's what's actually at stake: the same technologies that could cure most major diseases in the next 5–10 years are the ones being weaponized into botnets capable of causing hundreds of billions in damage. Fifty years ago, cancer was largely untreatable. Today we're talking about curing it within a decade. The acceleration is real.

But the urgency to move fast creates its own risks. The quote "Slow down my competitors while we work on manipulation" captures a dynamic where safety concerns get weaponized as competitive advantages. Companies genuinely committed to responsible development end up constrained while others push boundaries.

This isn't about stopping progress. It's about maintaining sight of why we're building these systems in the first place. AI could accelerate economic growth, create conditions for abundance, and yes, usher in what some call a renaissance of democracy and freedom. The goal isn't to halt that trajectory but to ensure we're steering toward outcomes that actually improve human life rather than just generating returns.

The industry's recent pattern of self-regulation — voluntary pauses, capability evaluations, safety testing protocols — reflects this tension. Anthropic's Constitutional AI work and similar efforts across the field show that it's possible to pursue capability advancement while building in alignment mechanisms. Whether that balance holds as competition intensifies remains an open question.

China's AI Advantage

What this means is that the conversation around AI safety isn't just happening in conference rooms and policy papers—it's spilling into GitHub issues and Twitter threads, where practitioners are making judgment calls about what gets released when. The user frustration I'm seeing isn't misplaced: when a safety-focused lab lobbies for restrictions while sitting on capabilities that could advance the field, it creates a credibility gap that's hard to ignore.

I think this underestimates how much institutional incentives matter here. Anthropic was founded on the premise that frontier AI development needed to be slowed for safety reasons. But safety-first positioning becomes politically inconvenient when your competitors—both open and closed—aren't bound by the same constraints. The call for open models from the community makes sense if you believe distributed development is inherently safer than concentrated control, but it also presumes that openness doesn't itself create new coordination problems.

This is where global pacing gets complicated. You can argue that waiting for China to voluntarily follow export controls is naive, but you can also argue that treating an authoritarian government as a reliable partner in AI governance is equally naive. The tension isn't just technical or even political—it's epistemic. We're being asked to coordinate on risks that may not materialize for years, using institutions and frameworks designed for a different era of technological development.

I'm genuinely uncertain whether the current moment represents a temporary bottleneck or the beginning of a sustained divergence between AI development trajectories in democratic and authoritarian contexts.

The Cost of Inaction

The cost of inaction here isn't just slower progress—it's the risk that the first entity to operate without meaningful constraints sets the global standard. If we wait for voluntary restraint from frontier labs, or hope that democratic norms naturally extend to AI development, we're essentially letting China's approach define what "responsible" looks like by default. That's not necessarily because their methods are better, but because they're operating at scale and speed with fewer external checks.

I don't think Anthropic's regulatory advocacy is entirely self-serving—some of their concerns about rapid deployment are legitimate—but the timing does raise questions. They're essentially asking for brakes on a system where they've already built a lead, then positioning themselves as the safety-conscious alternative. The criticism from developers who want open models isn't wrong either; if you're worried about uncontrolled development, sitting on proprietary advantages doesn't solve that.

What gets lost in this debate is that global pacing isn't just a technical coordination problem—it's a geopolitical one. Getting China to agree to meaningful oversight means either incentivizing them to voluntarily constrain their own development (unlikely) or finding enforcement mechanisms that don't collapse under the weight of competing national interests. Neither path looks easy, and I'm genuinely unsure whether the current moment allows for the kind of international cooperation that would be required. The question isn't whether we should try, but whether we're building frameworks that can survive the gap between ideal coordination and messy reality.

Conclusion

The hardest pill to swallow is that global pacing requires cooperation with the very country that democratic nations view as their primary strategic competitor. China's lead in AI capabilities isn't just a technical reality—it's a geopolitical one that makes any meaningful coordination feel almost performative. We can set aspirational targets for what global cooperation might look like, but the practical ceiling remains terrifyingly low.

I've spent twelve years in this field because I still believe the benefits are worth chasing. AI could genuinely eliminate most major diseases within a decade. But that belief now sits alongside a growing conviction that the recursive self-improvement loop we're already seeing across industry won't wait for our governance frameworks to catch up.

What keeps me up isn't the technology itself—it's whether we can muster the collective will to use the time that pacing buys us. The stakes are high enough that half-measures will look indistinguishable from failure.