In 2019, I attended a talk by Prof James Evans that fundamentally changed how I think about scientific breakthroughs.
While you can’t predict when or where a breakthrough may occur, you can increase their likelihood.
Here’s one thing you can do today: support small teams.
Here’s why.
The core finding
In 2019, James Evans and his collaborators Lingfei Wu and Dashun Wang published a landmark finding in Nature:
Large teams develop science incrementally. They’re excellent for scaling existing ideas.
Small teams disrupt science. New ideas and opportunities are more likely to come from individuals and small teams.
Intuitively, this makes sense. Newton saw the apple fall from the tree. Bell Labs and the Cambridge LMB had default team sizes of 1-3 people. Gmail and GPT were both one-person passion projects. Startups disrupt incumbents.
The data supports this dramatically: for every person you add to a research or technical team, the chance the team disrupts the frontier decreases exponentially.
Why size matters
In the years since, I’ve observed many research teams of different sizes. As teams grow, I’ve noticed they:
- Require more agreement on what to do next.
- Become less likely to scrap their latest work for a better idea.
- Struggle to transfer fuzzy ideas between minds – it’s hard to find the words.
This affects all levels of decision-making, from creating a new code base because you’ve suddenly realised a cleaner architecture, to changing direction or field entirely.
The fewer people you have to convince that your wild untested idea is worthwhile, the better.
Exploit vs explore
Both small teams and large teams are essential to science and technology, but they suit different goals.
Large teams are best for:
- Exploitation
- Discipline
- Building on established work
- Scaling promising developments
Small teams are best for:
- Exploration
- Originality
- First-principles thinking
- Rebellion against existing paradigms
The core dynamic is exploit vs explore. If your small team hits on a big idea, then it’s time to scale.
Size affects how teams engage with prior work
Citation patterns show that teams synthesise the existing literature differently:
- Small teams cite papers that are older (even decades earlier) and less popular. This suggests first-principles thinking or even the creation of new fields.
- Large teams cite papers that are more recent and more popular. This helps them amplify findings: they are speaking to existing, larger communities.
This creates a paradox: Large teams have the audience but focus on developments in established areas. Small teams generate breakthrough ideas but struggle to get people to listen – even if their ideas eventually have more long term success.
The funding implications
Your funding strategy should match your goals:
- Want to optimise an existing approach? → Fund big, networked teams.
- Want to discover a new approach? → Fund many small, decentralised teams.
You may have heard of “hill-climbing” – an optimisation technique in maths and computer science:
Imagine you’re dropped randomly into a national park in the dark. How do you find the highest peak? One option: only take steps that are flat or uphill. You’re guaranteed to find the top of a hill, but probably not the highest peak in the park. You need a method to find the highest peak before you start climbing.
Large teams climb existing hills. Small teams search and scatter across the landscape.
By exploring more widely, small teams discover new hills worth climbing. Then large teams can scale those discoveries.
Crucially, you want your teams to be decentralised and non-networked – it’s easy to fund lots of small teams who all share the same ideas or mental models.
Hill climbing is excellent, as long as you’re on the right hill.
You have to have a taste for risk
I chatted with James afterwards, and he shared this observation: the personality types at DARPA differ from those at NIH.
- NIH has people who understand where to put money to scale promising developments.
- DARPA has more risk takers, more of the venture capital mindset.
To fund the science that will supplant current ideas, you need a deep personal appetite for risk.
The skills are different, just like in the capital markets. We have angel investors, VCs, growth capital, private equity, investment banks etc – they each attract different types of people with different strengths.
Bad news: small teams are declining
Large teams are flourishing while small teams are systematically shrinking—precisely the opposite of what we need for breakthrough science.
It’s common to see funding calls unintentionally push applicants towards larger teams, by asking for multi-institutional collaborations or comprehensive interdisciplinary teams.
AI: a future case study?
Consider current AI research: the majority of time, talent and money is flowing into iterating the existing paradigm (deep learning) – despite well-documented limitations such as high training costs, limited context windows, catastrophic forgetting and hallucinations.
More importantly, the current approach continues to lack capabilities clearly essential to robotics:
- Generalisation beyond training data to novel situations.
- Continual real-time learning from environmental feedback.
- Sample efficient learning from a handful of real-world examples.
- Energy-efficient edge computing.
It doesn’t seem likely these capabilities will suddenly appear with more data and compute. This looks suspiciously like hill-climbing.
It’s a common story: hype and commercial pressure reduce the appetite to fund potentially disruptive new approaches, even when they may be most needed.
The bottom line
To fund breakthrough science, we should fund more small, decentralized, non-networked teams. Especially when they challenge the prevailing consensus.
This isn’t easy. The current consensus is usually robust and compelling. The obvious problems get patched. The root problems tend not to be obvious, and their solutions even less so.
But the story of humanity is the story of uprooting consensus.
It may feel risky to fund small teams who challenge the mainstream, but it’s the best way to find the new ideas that eventually take over.
Old consensus will be replaced with new consensus. Again.
Prof Evans’ talk covers a lot more than what’s here, so I recommend giving it a watch.
Here’s where he starts talking about this study (10 mins)
Further reading:
Prof. James Evans & team Fengli Xu and Lingfei Wu, have a 2022 PNAS article developing this thinking. It explores how flat teams drive scientific innovation.
Photo by Marco Montero Pisani on Unsplash