I · The compressed century
Biology & Health
AI as a virtual biologist, not a lab assistant. Fifty to a hundred years of biological progress could fit into a decade.
Most people picture AI in biology as a faster spreadsheet: it crunches genomic data and a human does the science. I think that framing is too small. The shift that matters is AI acting as a virtual biologist. It designs experiments, reads the results, runs the next experiment and tells you when the hypothesis was wrong.
Progress in biology has always depended on a handful of tool breakthroughs: PCR, sequencing, CRISPR, mRNA. Each one opened up decades of downstream work. Dario Amodei calls what comes next a compressed 21st century, where AI speeds up the discovery of those tools and fifty to a hundred years of biological progress lands in five to ten.
What changes
- Infectious disease becomes a solved engineering problem. We already have the playbook. AI shortens the time from pathogen to countermeasure.
- Cancer and genetic disease move from treatment to prevention. The hard part stops being biology and becomes delivery, regulation and cost.
- Healthspan, not just lifespan. If the rate of discovery compounds, longevity stops being a fringe research topic and becomes a product roadmap.
How it should work
As a technologist, I care most about the operating model. Agentic research loops need the same things as agentic software delivery: tight evaluation harnesses, reproducible environments, clear human checkpoints and an honest record of failures. The lab of the future looks like a well-run compound engineering team, with many parallel agents, a small group of senior scientists setting direction, and evals that decide what ships.
The bottleneck moves from intelligence to the physical world: clinical trials, manufacturing, regulation. Those are the problems leaders should be working on now.
II · The pagoda on the island
Peace & Governance
The one domain with no built-in tailwind. AI strengthens institutions only if we deliberately build it to.
Biology has momentum on its side. Governance doesn’t. The same model that drafts a fair ruling can also write a thousand pieces of convincing propaganda. AI helps whoever is holding it, so a good outcome in governance has to be designed and defended.
I think about governance on two scales.
Inside the enterprise
Most organisations run AI governance as a compliance exercise: a policy PDF, a review board and a list of banned tools. That fails because it slows adoption and doesn’t reduce risk. Good governance looks like good architecture:
- Decision rights are explicit. Which decisions can an agent make, which does it recommend, and which does it never touch?
- Evals are the policy. A rule you can’t test is just a hope. Policies should be written as evaluation suites that run on every change.
- Everything is observable. Every agent action leaves a trace a human can audit afterwards.
Across society
At civic scale the opportunity is large: legal systems that apply rules the same way to everyone, public services that answer in minutes instead of months, and citizens who can finally navigate bureaucracy with an expert in their pocket. Amodei’s hope is that democracies deliberately share these benefits. My hope is plainer. We should build public infrastructure with the same care we put into production systems.
Calm, legible institutions are what the rest of this picture depends on.
III · The boats on the water
Commerce
When agents become the customer, commerce is rebuilt around intent, trust and machine-readable promises.
For thirty years digital commerce has been built for human eyes: landing pages, funnels, cart abandonment emails and A/B-tested buttons. The next buyer won’t have eyes. It will be an agent acting on someone’s intent, with a budget and a set of constraints.
Business-to-Agent
I call this shift B2A. When an agent is the customer, the things that win are different:
- Machine-readable promises. Price, availability, SLAs and return terms exposed through standards like MCP instead of hidden in marketing copy.
- Trust as an API. Reputation, provenance and verifiable reviews become inputs an agent can weigh in milliseconds.
- Service as a software. Companies stop selling tools for doing the work and start selling the finished outcome.
What harmony looks like
The utopian version isn’t a frictionless shopping machine. It’s commerce with less noise: fewer dark patterns, fewer ads trying to hijack attention, and more transactions that match what people actually wanted. When the buyer is a patient, rational agent, manipulation stops working and honest businesses have the advantage.
The risk is concentration. If a few agent platforms control discovery, they become the new gatekeepers. Open protocols are the counterweight, and they’re worth supporting early.
IV · The terraced fields
Economy & Work
Abundance is plausible. Fair distribution isn't automatic. Meaning comes from people, not productivity.
Terraced fields are my image for a good AI economy. They’re engineered abundance, built patiently over generations, that feeds a whole valley and not just the owner of the top step.
The optimistic case
Take the first-order effects seriously. When expertise becomes cheap, the cost of medicine, law, education and software drops sharply. Amodei’s dream scenario has the developing world growing at 20% a year and reaching in a decade what took others half a century. I don’t think that’s guaranteed, but it’s no longer science fiction.
The honest case
Productivity gains don’t spread themselves. Three things decide who benefits:
- Access. Can a small business in Nairobi use the same agents as a bank in New York?
- Institutions. Weak governance and corruption can absorb any amount of new capability.
- Transition. Comparative advantage protects human work in the short term. In the long term we’ll need something newer and stranger, whether that’s new forms of income or new kinds of work.
Work and meaning
I don’t believe people stop doing things because a machine does them better. We still run marathons even though cars exist. Meaning comes from relationships, craft and challenge, not from economic output. As leaders, our job is to redesign roles around judgment, taste and care, and to be honest with our teams about the change while it’s happening.
V · The dark water
The Undercurrent
None of this happens by default. The risks are the only thing between us and the upside, which is why they deserve our best work.
Every calm lake has an undercurrent. I’ve tried not to paint an idyll that ignores it, because the upside I’m describing only exists if we treat the downside seriously.
What sits beneath the surface
- Misuse. The same capability that designs a vaccine can help design a pathogen. Capability without safeguards is a liability.
- Concentration of power. AI makes surveillance and propaganda cheap. Whoever controls the most capable systems gets unusual leverage over everyone else.
- Loss of control. Systems we can’t inspect, evaluate or interrupt shouldn’t be trusted with decisions that matter.
- Quiet erosion. This is the least dramatic risk and the most likely: gradual deskilling, trust in systems that haven’t earned it, and institutions that automate their judgment away.
Why I stay optimistic
Being optimistic about AI isn’t the same as assuming things will work out. Every item above is an engineering, governance or leadership problem, and those are problems we know how to work on: interpretability, rigorous evals, explicit decision rights, and openness about what our systems can and can’t do.
Fear is one motivator, but it isn’t enough on its own. We also need a clear picture of what we’re protecting. That’s what the rest of this landscape is for.