Anthropic's Sholto Douglas and Nick Marwell: AGI in a Couple of Years

Anthropic technical leads Sholto Douglas and Nick Marwell tell Joe Lonsdale why they expect AGI within a couple of years and what could go wrong on the way. In this episode of American Optimist, the two reinforcement learning researchers answer the regulatory capture critique, set out Anthropic’s position on open source and distillation, explain why they think biology is the next frontier after code, and give their predictions for 2028. The conversation was recorded in Napa in August 2026.

TLDW

Douglas and Marwell say models as capable as any human at computer work are very likely within a couple of years, and that AI coding has already moved from autocomplete to something like a junior team member that works alone for a day or two. They name unemployment, bioweapons and cyberattacks as the real risks, argue that cyber will swing toward defense within two years while biology stays dangerous for longer, and say Anthropic would never support slowing America while China keeps going. On the business side, they argue that distillation threatens the funding of frontier research, and that each step up in intelligence is worth far more than the last, so the frontier keeps its value even as older models get cheap. They close with a case for a post-scarcity economy in the 2030s, a warning that AI will not fix education by itself, and a prediction that an AI could win a Fields Medal by 2028.

Thoughts

The strongest part of the regulatory capture answer is the example, and it is also where the answer is weakest. Marwell points to the release of Fable, and Mythos before it. He says Anthropic had the best model in the world by a wide margin and chose to work with the government on a slower rollout when it could have taken a lead of several months. That is a costly choice, and a company that only wanted to win would not have made it. But it does not settle the question critics are asking. A company can give up revenue in the short run and still benefit from a system where every model has to pass a threshold that it helped to define. Lonsdale adds a sharper point: he says the model’s cyber capabilities were reduced to get it released, which took a defensive tool away from the people who need one. Neither guest disputes that. Their better argument is about who the open weights developers are. Marwell notes that the companies pushing open weights forward are Nvidia, Amazon, Microsoft and large Chinese labs, which are not small projects that a compliance burden would crush.

The distillation section is the clearest account I have heard of how a frontier lab thinks about its own business model. A frontier model is only the frontier for months. The cost of a given level of intelligence falls roughly ten times a year, with or without anyone copying anything. So Anthropic’s income depends on always having the next model, and the next model may need a training run that costs tens or hundreds of billions of dollars. If a competitor can copy the result in a week or two for a fraction of the cost, the copier has the better economics and nobody can afford to pay for the next step. The comparison to drug patents is a good one because of where it breaks: a patent protects a drug for 20 years, while the frontier only needs protecting for a few months. The whole argument does rest on one belief, which is that each new step in intelligence is worth much more than the last. If that stops being true, a lab whose product is out of date in six months has a serious problem. They admit that this requires a lot of optimism.

The idea that deserves more attention than it gets is the lab bottleneck. Marwell says he is more worried about whether America has the physical lab space and supply chain to use AI in biology than about whether the models will be smart enough. He comes back to it near the end, when asked what would shorten his timelines: a large national effort to build labs that produce biological data for training. Most public discussion of AI progress is about chips and power. His claim is that the limit in the most valuable field may be benches, equipment and experiments, and that US labs are behind China’s on quality. That is a problem money and construction can solve, and almost nobody is talking about it.

On jobs, the guests are more candid than the setting required. Douglas cites employment in the Philippines going up, in a country whose outsourcing industry was expected to be hit first, and says software engineering employment is up too. Lonsdale takes that as good news. Marwell does not let it stand. He says this is the period where a person paired with a model is worth more than before, so companies hire more of them, and that the period ends when the model no longer needs the person. In chess that stage lasted about 30 years. They expect it to be much shorter here. That is an honest thing to say on a show called American Optimist, and it sits awkwardly beside the career advice from earlier in the conversation, which only applies if the technology takes 10 or 20 years to spread.

The most surprising claim comes late and has nothing to do with models. Marwell, whose father has worked in education for 15 to 20 years, says many people in AI are wrong to think AI will fix education. His reasons are specific. Education compounds, so a child who cannot read at grade level by the end of third grade falls behind in everything, including math. Children who think they are bad at something stop trying. And computers have so far been bad at holding a young child’s attention and good at distracting them. A tutor that knows everything does not help a seven year old who will not look at it. It is a useful check on the rest of the episode, where intelligence is treated as the answer to most things.

Key Takeaways

Douglas says that 18 months ago he typed every line of code by hand. He can now ask a model to do a day or two of work alone, and he compares it to a junior team member.
Both guests expect models that match or beat humans at everything done on a computer within a couple of years, with physical work to follow once robotics is good enough.
On the FrontierMath benchmark, a set of problems written by professors, models went from 0% to well over 40% in about a year.
Math and code improved first for two reasons: answers are easy to check, which suits reinforcement learning, and the people building the models were experts in those fields and could read what the model was doing.
Marwell expects biology to be the field where AI has the most impact over the next 6 to 24 months, and says lab capacity and quality in the US may be the limit, not model intelligence.
If AI takes 10 to 20 years to spread through the economy, the guests think the next decade belongs to generalists, and that choosing the right problem becomes the skill that matters most.
Cyber and bio are both dual use. A request to find every vulnerability in a codebase looks the same from its owner and from an attacker.
Both fields favor attackers today. The guests expect cyber to favor defenders within two years. Biology needs hundreds of billions of dollars of infrastructure before it does.
Anthropic’s stated position on open source: one capability threshold that open and closed models both have to meet, and freedom to do anything below it.
They would only support slowing down if every party coordinated and could be trusted. They say there is no scenario where they would slow America and let China go ahead.
Distillation is hard to stop because every extra bit of access customers want, such as seeing the model’s reasoning, also makes copying easier.
The cost of a given level of intelligence falls about ten times a year. Total AI revenue is a little over $100 billion in a world economy worth tens of trillions.
Large tech companies are spending about $1 trillion on AI infrastructure this year. If that keeps doubling, Douglas thinks world GDP could start doubling in the early 2030s.
Marwell says AI revenue grew so fast that labs could spend their way past data limits he once thought would take years to get through.
Predictions: tens of thousands of humanoid robots in homes by 2028, a Fields Medal for an AI quite possibly by 2028, and a Nobel Prize very likely before 2030.

Chapters

04:59 From Typing Code by Hand to AI Junior Team Members

Douglas describes three stages in 18 months: writing all his code himself, guiding a model and correcting it every few minutes, and now handing over a day or two of work. Marwell adds that until recently models did things people could already do, only faster. He points to recent math proofs as the first signs of models going past what people have managed. He does not claim the models have ideas no human could have. He says most progress has always come from connecting existing ideas, and the models now do that.

09:26 AGI in a Couple of Years and How Progress Is Measured

The guests define AGI as a model as capable as any human at computer work, and say it is very likely within a couple of years. Anthropic’s mission, in their words, is to get the world through the risks so it can have the benefits. They explain the two ways models improve: training on large amounts of text, and reinforcement learning, where the model works problems and is checked. Douglas says it helps to think of a model as an entity and not a tool, because it is sent out to act for you without being watched.

14:21 Biology Is Next, and Labs Are the Bottleneck

Marwell calls AI the most important technology in the life sciences in our lifetime. He expects the benefits to be spread widely across society and not only show up as adoption numbers. His worry is physical. The US needs more lab space, and its labs trail China’s on several measures of quality. He calls rebuilding the American lab and supply chain system one of the major infrastructure projects worth taking on.

16:50 Career Advice: The Decade of the Generalists

Asked what a new graduate outside Silicon Valley should do, the guests first say Anthropic is concerned about unemployment. They then describe a slower path, where AI takes 10 or 20 years to spread and many careers are built on helping it spread. In that case one person has the power of a thousand person company, and the people who do best are the ones who can see which problems are worth solving for their community or their market.

20:11 Bio and Cyber Risk: Offense Versus Defense

The risks they list are unemployment, and bio and cyber misuse that they say is happening this year. Both are dual use, which makes them hard to police. Their view is that a capability should not be released until bad actors can be kept from using it. Cyber should become safer as everyone uses AI to attack their own systems and patch them first. Biology will not become safe without very large spending and advanced robotics, which they place in the 2030s.

25:08 The Regulatory Capture Critique, Open Source and China

Lonsdale puts the charge directly: that Anthropic scares people in order to get rules it will control. Marwell answers that open weights models are built by some of the largest companies in the world, and that the slow Fable rollout cost Anthropic money. Douglas says Anthropic has slowed itself more than anyone, mentions Demis Hassabis’s proposal for shared safety thresholds among closed labs, and says he learned machine learning on open source models and supports them. On China, the ideal is cooperation, and a one-sided American pause is ruled out.

33:05 Distillation and the Economics of the Frontier

Distillation means training a cheaper model on the outputs of a better one. The guests say it cannot move the frontier forward because it only copies, and that it removes the income that pays for the next training run. They then explain why the frontier does not turn into a commodity. Tab autocomplete was the first useful step in AI coding. Agentic coding in the terminal with Claude Code was the next, and it was worth vastly more. A full software engineer, and then work like curing cancer, come after that.

40:12 Post-Scarcity, the Compute Ramp and Jobs

Douglas describes the goal as a world where the cost of everything, housing included, falls to the cost of energy, and where diseases are cured. He says the hard questions are how to get there and how to share the gains so they do not only go to people who already own capital. Compute for AI has doubled or tripled every year for four or five years. Marwell warns that today’s rising employment reflects a stage where people are still needed beside the model. He also says the timing is unlucky: AI is arriving in a world divided between the US and China, and the most important chips are made on an island between them.

49:00 Tradition, Philanthropy and Why AI Will Not Fix Education

Lonsdale asks whether people in San Francisco will remake the world without regard for family, faith and tradition. The guests expect the opposite: as work matters less, tradition, ritual and local community matter more. Douglas says philanthropy is a large topic inside Anthropic, and mentions giving to a program that aims to end viral disease and a colleague funding research on labor force effects. Marwell then makes his case that AI can help teach the right thing at the right time but will not solve early education.

55:03 Predictions for August 2028

Asked what would make them say things went faster than expected, they name data centers in space, the number of humanoid robots in homes, and how fast data can be produced in the fields that matter. Douglas expects early home robots doing laundry and basic cleaning, perhaps tens of thousands by 2028. Asked whether AI should work on character and mental health, he says he would prefer that people solve those. He does not expect an AI to win a Pulitzer by 2028, thinks a Fields Medal is quite possible, and a Nobel Prize very likely before the end of the decade.

Notable Quotes

“Where we’re at now is I can ask models to do a day or two days of work independently and drive progress basically like a junior team member.”
Sholto Douglas, on how programming changed in 18 months

“I think AI is going to be the most important technology in the life sciences, certainly of our lifetime and maybe ever.”
Nick Marwell, on why biology is the next frontier

“It’s not a strategy of regulatory capture for economic interests.”
Nick Marwell, answering the charge that Anthropic wants rules that favor it

“So far we’ve slowed ourselves down much more than anyone else.”
Sholto Douglas, on the claim that Anthropic uses fear to slow its rivals

“We believe that closed and open source models should have to meet the same bar.”
Sholto Douglas, on Anthropic’s position on open source

“Each marginal unit of intelligence that you’re capable of is worth exponentially more than the unit that came before it.”
From the distillation discussion, on why the frontier keeps its value

“The number one predictor of how well a student will do in math is whether they can read at grade level at the end of third grade.”
Nick Marwell, on why AI will not simply fix education

Watch the full conversation with Sholto Douglas and Nick Marwell on Joe Lonsdale’s American Optimist.

Anthropic’s Sholto Douglas and Nick Marwell: AGI in a Couple of Years