Bridge Atlas Episode 4: Alignment
Issue #72: A conversation with Alex Stokes and Emmett Shear
Welcome back to the fourth episode of the Bridge Atlas series, hosted by Christine D. Kim. Today, we’re diving into Ethereum through the lens of alignment with Alex Stokes, Ethereum Foundation Protocol Coordination Co-Team Lead, and Emmett Shear, founder of Softmax.
Christine: Let’s start with some background. Emmett, you’ve spoken before about agent alignment. Could you give our listeners a quick summary of your focus?
Emmett: Sure. My background is in internet entrepreneurship. I started Justin.tv, which became Twitch, and had a brief stint as CEO of OpenAI during a turbulent period. That experience made me realize AI had an unanswered question: what does it really mean to align models, and to what? So I founded Softmax, an alignment research company.
We run multi-agent reinforcement learning simulations to study alignment – not as a static state, but as a protocol and ongoing process. We see alignment as coordination between agents at multiple levels, with shared success as the foundation. If one wins while the other loses, there’s no real alignment. Shared success enables specialization and team coherence. But you also need enforcement to protect alignment from parasites or bad actors. Enforcement isn’t what creates alignment – it protects what already exists. Softmax focuses on building shared success and coordination, not enforcement.
Christine: Alex, tell us about your work at the Ethereum Foundation.
Alex: I’ve been at EF for a while, focusing on coordination – essentially our version of alignment. Ethereum’s decentralized ecosystem contains many actors with different interests. We want everyone “rowing in the same direction” toward a shared Ethereum success, but that’s tricky when interests diverge. My job involves thinking about protocol direction – maintaining and securing today’s Ethereum, and scaling and evolving it for the future.
Christine: Emmett, you said alignment starts with shared success, but it’s hard to get that across a large group. In Ethereum, breaking down into smaller local groups might help, but then alignment becomes relative and harder to evaluate. How do we evaluate good alignment in large groups?
Emmett: Alignment is relative – like mass or speed are relative to a frame of reference. You can be aligned with your family or company, but not necessarily with everyone. Systems often have chains of local alignment that feed into global alignment. Crypto does well at creating global shared success – Bitcoin is almost pure shared success. But crypto struggles with multi-scale alignment and specialization, because it flattens context into one global ledger. We haven’t figured out how to support boundaries for private information while keeping global coordination.
Christine: Alex, do you agree that crypto struggles with specialization?
Alex: Partially. Bitcoin has a small “alignment surface area” – one token, one function. Ethereum’s is larger and more complex. But we already see specialization via subtokens, DeFi apps, and rollups – smaller ecosystems connected to Ethereum’s base layer. These are counterexamples to the idea that crypto can’t support localized alignment.
Emmett: Yes, Ethereum’s hierarchy of boundaries is a form of alignment. But the real economy is messier, with overlapping hierarchies and constructive ambiguity. In Ethereum, ownership is unambiguous by design, which is great for decentralization but loses some of the flexibility ambiguity provides in real-world systems. Forking is the crypto equivalent of revolution, but crypto doesn’t embrace frequent forks – a rigid, fork-resistant global chain could be dangerous.
Alex: I see your point about rigidity. But I think crypto pushes power to the edges – users choose to run the software, and forks do happen. Even a “too big to fail” chain could be forked eventually, like governments that seemed unchangeable. Ethereum changes its protocol regularly, which keeps flexibility.
Emmett: Imagine Ethereum runs the global economy – every government and corporation depends on it. Forking then becomes almost impossible. My challenge to crypto is: how can we embrace forking without destroying shared success, allowing multiple interpretations of the same ledger?
Alex: This is part of Ethereum’s rollup-centric roadmap – the base layer stays neutral, while rollups provide flexibility. Decentralization and permissionlessness ensure no one can unilaterally control the chain. Mechanisms for exit – whether via forks or new chains – are crucial.
Emmett: But how is Ethereum making exit easier today?
Alex: Right now, exiting means leaving Ethereum’s L1, but there’s not yet an alternative system to exit to. That’s why rollups are important – they can offer localized ecosystems.
Emmett: Ethereum isn’t just tokens – it’s smart contracts and identities. If the whole economy runs onchain, exiting becomes extremely hard without early planning. That’s why I’m concerned: decentralized systems can get stuck in bad ways, especially if AI agents store property there.
Alex: We could design smart contracts today with exit functions. Early national ID onchain systems will likely be centralized, so they can patch bugs. Fully decentralized ones could get stuck, as you say – but decentralization also makes exit easier because there’s less centralized control.
Emmett: You gain the ability to exit centralized systems in exchange for the inability to exit decentralized ones – that trade-off might be worth it. But with AI, there’s a risk in trying to encode “the good” as a fixed meta-ethical loss function. You can’t perfectly define the good. Alignment needs openness and ambiguity, letting agents rewrite even their base values. A unified global meta-ethics in AI could be as dangerous as a rigid global blockchain protocol.
Christine: Ethereum changes core assumptions frequently – unlike Bitcoin, which has hard-coded values. Could alignment even be possible with a protocol that changes that often?
Emmett: Yes, because survival in the same reality acts as the ultimate arbiter. Systems that believe falsehoods eventually die or adapt to align with reality. Even groups with vastly different beliefs (like Catholics and Protestants) share reality and can find common ground.
Christine: I’m skeptical that man-made protocols can self-correct dynamically like that over time – whether crypto, AI, or even the internet.
Emmett: That’s why I oppose any global protocol. The real game is constant invention of new value systems and protocols. A healthy ecosystem bubbles with new protocols, with no single durable leader. The danger comes if crypto or AI ossifies into one global layer that must be “right” forever.
Christine: Final thoughts? For me, if Ethereum ossifies, it dies. Technology always needs intervention. And as a side note, I see some protocols – like the Ten Commandments – as timeless truths, though that’s outside our scope.
Alex: I’m optimistic about Ethereum because of the people involved. The tech is an extension of human values, so misalignment can be corrected. AI is riskier because it could run away from human control.
Emmett: I think crypto will grow fast, especially as AI agents adopt it. That could lead to a global chain – which is why we need to think now about maintaining exit rights if the whole economy ends up onchain. We should also think early about AI models that may surpass human intelligence and influence decisions beyond our audit capacity. My criticism comes from optimism – I believe these systems will succeed, so we should plan for winning scenarios.