Science is learning from start-ups (with Adam Marblestone)
Spencer Greenberg speaks with Adam Marblestone about Focused Research Organizations (FROs) as a new model for scientific research, strategies for field building in science, and the beneficial applications of AI, particularly large language models, to accelerate scientific progress and address societal challenges.
Deep Dive Analysis
16 Topic Outline
Introduction to Focused Research Organizations (FROs)
FROs vs. Traditional Academic Research Models
Examples of Existing FROs: E11bio and Cultivarium
Applying the FRO Model Beyond Biology: Fusion and Climate
Addressing the Funding Gap in Scientific Innovation with FROs
Scale and Funding Mechanisms for FROs
How FROs Shift Incentives Relative to Academia
The Rationale Behind Time-Limited FROs
Organizational Structure and Leadership in FROs
Understanding Field Building in Science
Case Studies of Stagnated Fields: Cryonics and Ocean Iron Fertilization
Sociological Bottlenecks and Strategies for Field Building
Promising Areas for New Field Development
The Impact of Large Language Models (LLMs) on Scientific Research
Automated Theorem Proving and Formalized Mathematics
Beneficial AI Applications: System Two Recommenders and Contextualization Engines
5 Key Concepts
Focused Research Organization (FRO)
A special-purpose, non-profit organization with a fixed and specific mission to create a tool, system, data set, or other advancement that benefits the scientific process. It is organized somewhat like a startup, with a finite duration, typically 5-7 years, rather than aiming to be a permanent institution.
Field Building in Science
The process of counteracting non-scientific, sociological factors that can stall scientific progress, and actively increasing the probability that good ideas blossom into new fields. This involves establishing a critical mass of researchers, resources, journals, peer review processes, standards, and benchmarks for a research area.
System Two Recommender
A type of recommender system designed to optimize for a user's deeper, longer-term goals, higher values, or what they would want upon reflection, rather than immediate gratification or attention. It aims to provide recommendations that align with a 'higher self' and can be customized to mitigate biases.
Contextualization Engines
A variation of search or information retrieval systems that optimizes for providing a comprehensive, multifaceted, and relatively unbiased context on a given query. Instead of just the most relevant or click-optimized result, it aims to offer a global picture of content, context, and different views, similar to a Wikipedia article.
Automated Theorem Proving
The use of AI systems, potentially enhanced by large language models, to generate and verify mathematical proofs within formalized languages. This approach could accelerate mathematical discovery by reducing the search space for proof steps and ensuring the correctness of complex logical arguments.
8 Questions Answered
An FRO is a special-purpose, non-profit organization with a fixed, specific mission to create a tool, system, data set, or other advancement for science, structured like a startup with a finite duration.
Unlike traditional academia where individual professors secure grants for their labs and trainees pursue distinct thesis projects, FROs involve larger, coordinated teams with substantial funding focused on a specific engineering-like goal for a fixed period, rather than individual publications or training.
FROs are designed with a finite duration (e.g., 5-7 years) to maintain a sense of urgency and clarity of purpose, focusing on completing a specific mission rather than becoming permanent institutions, which can dilute the mission and increase costs.
FROs are well-suited for capital-intensive, engineering-intensive projects that aim to create public goods for science (like a brain mapping system or a method to grow diverse microorganisms), where the value is broadly catalytic rather than primarily market-driven or easily parcelled into individual academic papers.
FROs fill the gap for projects that require the capital and team structure of a deep tech startup but are focused on creating public goods, which don't typically generate venture capital returns, and are too large or coordinated for individual academic grants.
LLMs, and large data prediction systems generally, can accelerate scientific progress by handling 'type one skills' (tasks with clear success metrics and dense reward signals, like automated theorem proving) and 'type two skills' (tasks with ample human-generated demonstration data, like generating scientific papers or hypotheses).
AI can be used to develop 'system two recommenders' that optimize for a user's long-term goals and values rather than immediate attention, and 'contextualization engines' that provide comprehensive, multifaceted, and unbiased context on a topic, improving epistemic understanding.
Sociological factors, such as premature commercialization, association with fringe ideas, or ethical concerns, can give a field a bad reputation, deterring mainstream scientists and funding, thereby stalling progress even if there's a legitimate scientific core.
14 Actionable Insights
1. Form Focused Research Organizations
Create special-purpose, finite-duration (5-7 years) non-profit organizations with a fixed mission to build a specific tool, system, or dataset for science. This addresses a gap where projects need startup-like capital and team structure but aim to create public goods, not just market-driven products.
2. Diversify Science Funding Models
Explore and advocate for varied funding and organizational structures beyond the traditional grant-based academic model, such as FROs or ARPA-like agencies. This can address systemic bottlenecks and enable a wider range of high-impact scientific projects.
3. Embrace Time-Limited Projects
Structure scientific initiatives with a finite duration and clear mission (e.g., 5-7 years) to maintain urgency and focus on delivering a specific capability or public good. This prevents mission creep and the dilution of purpose often seen in permanent institutions.
4. Adopt Startup Management Practices
Implement project management, deliberate structured meetings, and efficient hiring/people management techniques from startups within scientific endeavors. This helps FROs operate with agility and precision, focusing on execution and goal attainment.
5. Facilitate Scientist Entrepreneurship
Design scientific projects (like FROs) to develop professionalized teams and skills that can lead to spin-off companies, industry integration, or new ventures. This provides alternative career paths for scientists and leverages their expertise for broader impact beyond academia.
6. Counteract Sociological Barriers
Be aware of how premature commercialization, exaggerated claims, or association with fringe topics can prevent legitimate scientific inquiry. This helps avoid fields being rejected by the mainstream, leading to a weak evidence base due to lack of research.
7. Prioritize Data-Driven Research
Focus on generating fundamental data and measurements with less agenda-driven hypotheses, especially in politicized or controversial areas. This builds a robust scientific foundation without getting entangled in specific, potentially polarizing, applications.
8. Cultivate Critical Mass for Fields
Actively work to gather a sufficient number of researchers, resources, journals, peer review processes, standards, and benchmarks for emerging scientific areas. This is crucial for validating breakthroughs and enabling progress that individual researchers cannot achieve alone.
9. Adopt Empowered Program Manager Model
Implement a model where technically strong program managers actively define milestones and incentivize researchers to pursue specific directions, similar to the DARPA model. This can effectively “pull” researchers and bootstrap new fields.
10. Utilize Field Strategists
Employ individuals (field strategists) to identify systemic bottlenecks, problems, and obstacles within a scientific field and create roadmaps for its progress. This provides a structured approach for philanthropists or other funders to support field development.
11. Formalize Mathematics for AI Proofs
Convert mathematical proofs into formalized programming languages (e.g., Lean) to enable automated theorem proving. This creates a verifiable reward signal for AI training and allows LLMs to predict next steps, accelerating mathematical progress and secure software development.
12. Develop “System Two” Recommenders
Design recommender systems that optimize for users’ long-term goals, reflective preferences, and higher values, rather than immediate engagement or clickbait. This empowers users to customize their information flow to counteract biases and align with their desired future self.
13. Create Contextualization Engines
Develop AI systems (e.g., using LLMs) that, upon query, provide comprehensive, relatively unbiased, and multifaceted context on a topic, similar to a Wikipedia article. This aims to improve epistemic understanding by offering a global picture rather than optimizing for user relevance or clicks.
14. Apply ML to Societal Problems
Frame societal problems in a machine learning context by identifying how to generate training data and feedback loops for improvement. This allows for the application of powerful AI tools to address challenges like misinformation or aligning systems with human values.
6 Key Quotes
A focused research organization is basically a special purpose, non-profit organization that has a very fixed and specific mission to create some kind of tool, system, data set, or other advancement that benefits the scientific process.
Adam Marblestone
It's kind of driven by novelty, rather than driven by the sort of working backwards from a functional purpose in the same way that, you know, an industrial effort will be working backwards...
Adam Marblestone
It almost sounds like a mini Manhattan project, where you're like, kind of saying, okay, we all we've got this common objective, we've got to bring together all these great minds, and we're going to just focus on and get it done. Do you think that's a fair analogy?
Spencer Greenberg
The broader question we're trying to get at here is creating a conversation essentially about, are we in kind of too much of a almost like monotheistic world? There's only one way of doing organizing research. What if we have a much more kind of polytheistic kind of structurally diversified approach to funding and organizing research?
Adam Marblestone
Unless you have a certain number of researchers working on something, unless you have a certain critical mass of resources, journals, peer review processes, standards, benchmarks, and all that, it's very hard to judge progress...
Adam Marblestone
I'm excited about it, and I don't know what's going to happen next. I have a pretty wide range of outcomes kind of floating in my mind, having not fully been able to really grok all the implications.
Adam Marblestone