Basis: Designing a New Kind of AI Research Organization

Executive Summary

Basis is a nonprofit applied artificial intelligence research organization, registered as a 501(c)(3) public charity, with two simple, mutually reinforcing goals.

The first is to build a universal reasoning engine. This means to establish the underlying mathematical principles of intelligence that are independent of any specific problem — what it means to infer, to learn, to make decisions, to understand, and to explain; and to construct a new generation of AI software realizing these principles.

The second is to advance society’s ability to solve intractable problems. This means expanding the scale, complexity, and breadth of problems that we can solve today, and even more importantly, accelerating our ability to solve problems in the future.

We believe meeting these goals requires a new kind of research organization unique in its technology, process, output, and structure:

AI technology that embraces complexity with simplicity

Basis is building technology that can reason about and operate in environments of the kind of boundless complexity and detail that reality has to offer. To do this, we will focus on how to represent and discover models of phenomena in the world at unprecedented fidelity and scale, incorporating all available knowledge of all kinds, be it large or small amounts of data, interactions and experiments, or the wealth of tacit knowledge accumulated by human experts.

One might think that solving complex problems requires complex and complicated technology. We believe the opposite is true: that there are fundamental principles of reasoning that confer an ability to solve problems in general. We aim to leverage ideas from programming languages, compilers and databases to uncover these principles in their most general, reusable, composable forms, and ideas from deep learning to power generic, scalable approximate algorithms for applying them.

Driving intellectual advances by solving real problems

Our approach to research revolves around solving problems. These problems are drawn from a small set of focus areas that are of significant importance to broader society; require organizing knowledge and people in qualitatively new ways through software; and are rich enough to demand domain-general advances in our technology.

Working with collaborators to identify real-world needs ensures that Basis remains impactful and effective in tangible ways. At the same time, requiring that the problems we take on lead to general advances in artificial intelligence and to new kinds of applications is equally important. It ensures that Basis learns in a compounding way, becoming progressively more capable, and able to solve harder and more impactful problems as time goes on.

Open source code as research output

Basis aims to develop a lingua franca of artificial intelligence: a unified, growing body of open-source software that serves as a bidirectional interface with the rest of the world.

In one direction, our software will allow Basis employees as well as the broader community of scientists of intelligence to consolidate their knowledge and coordinate their work. In the other direction, our software will support integrating and disseminating reasoning technology to address a much larger set of problems than Basis employees could solve alone.

Collaboration-driven organizational structure and culture

The name Basis symbolizes the position we intend to occupy. Instead of an insular, siloed institute, Basis aims to act as the nexus and foundation for the intellectual grand challenge of understanding human minds, and the human challenge to build a better society.

In concrete terms, many of the projects Basis engages in will be driven by external collaborators from academia, government, the social sector, and industry, and mediated by open source software.

Exceptional team

Basis has assembled a growing team of founders, advisors and employees from the MIT, Harvard and Columbia communities who are uniquely capable of executing on this ambitious research program.

For more information about our team, please visit our team page.

Introduction

Underpinning our human ability to solve problems of remarkable depth and breadth is a set of general faculties of intelligence. These empower us to predict possible futures, root out hidden causes, imagine hypothetical worlds, invent novel theories, learn from experience, and ultimately make good decisions in a world fraught with uncertainty and complexity.

A central premise of artificial intelligence research since the earliest days of the field (arguably since Leibniz’s Enlightenment-era quest for a calculus rationator) is that these human faculties can be decomposed into universally applicable systems of problem-agnostic formal reasoning that operate over structured representations of problem-specific knowledge. Elucidating these systems and implementing them in software is thus not only of enormous intrinsic scientific value, but if carefully pursued, will help solve some of the hardest problems facing science and society.

Yet building general intelligence means going beyond reasoning engines that just learn patterns from data or manipulate abstract symbols to machines that truly reason in the broadest possible sense. That is, machines that can come to conclusions about the real world, in all of its complexity and ambiguity. Machines that are both logically self-consistent and consistent with data, and can link heterogeneous knowledge across many levels of scale and abstraction. Machines that can distinguish causation from correlation and cause from effect, and introspect on their own reasoning to explain how and why they arrive at the conclusions that they do.

Much tangible recent progress on this front has come in the form of the deep learning revolution in machine learning, granting the ability to discover complex patterns from large amounts of data. However, machine learning systems still tend to struggle with uncertainty, causality, generality, robustness, and distinguishing fact from fiction, and even the most advanced machine learning systems pale in comparison to the breadth and flexibility of human intelligence.

Meanwhile, researchers across many disciplines of natural and artificial intelligence are increasingly converging on a shared theory of reasoning as “computational rationality”, in which formal representations of knowledge and actions are manipulated by approximate application of the laws of probability and performing this manipulation is itself cast as a decision-making problem of the optimal allocation of scarce computational resources. These classical components have since been augmented by a mature theory of causal and counterfactual reasoning and the emergence of powerful, practically successful algorithmic frameworks for linear algebra and reasoning shared across logic, databases, programming languages, statistics, physics and many other fields. However, assembling this consensus into a truly universal theory of reasoning remains a seemingly insurmountable technical challenge, in part because each field has its own bewildering array of jargon and hand-crafted algorithmic tricks.

These two trends, which might appear contradictory at first glance, are in fact deeply connected, as demonstrated by the widespread and effective real-world use of probabilistic programming languages that rely on ideas and even software from machine learning (including Pyro and Omega, created by the authors of this document).

In fact, they point unmistakably toward a new generation of AI systems that combine them more directly, using causality, logic, probability and programming language theory to describe a disparate array of learning and reasoning problems, extensible and expressive machinery for reducing complex reasoning operations to (very highly abstracted) “linear algebra” computations, and current machine learning technology as a ubiquitous low-level computational tool for approximating parts of these computations.

This in turn suggests the tantalizing ultimate possibility of a universal reasoning engine - a transformative general-purpose technology that would allow any group of any size to safely bring to bear the sum total of global knowledge on any reasoning or decision-making problem, no matter its scale or complexity, blurring the boundaries between independent tasks, between individual agents, between data and computation, between observation and interaction, and even between simulations and physical reality.

Why Basis?

As it stands, researchers do not yet know how to build reasoning technology that is general enough to help solve the full spectrum of applied problems that we face. At the same time, solving any one problem does not necessarily provide insight into others. How should we proceed?

The critical missing step is to learn principles that generalize beyond the specific problem that was solved, and hence, improve our ability to solve problems in general. Then, each subsequent problem is easier to solve, and eventually, previously intractable problems become tractable.

In every other field of computer science that has produced truly general-purpose technology, such as microprocessor design, optimizing compilers or relational databases, researchers have found that software is an effective medium for carrying out this kind of learning. The recipe is straightforward: identify general principles of intelligence and encode them in software, use the software as a foundation, albeit an incomplete one, to solve carefully chosen practical scientific and societal problems that cannot be addressed with existing technology, distill the knowledge gained from the process into an extension to the software, and repeat, until the software itself becomes the best and most complete representation of the underlying theory.

Our team has previously followed this process to success in developing Pyro and Omega, two leading examples of artificial intelligence systems based on a new paradigm of reasoning under uncertainty called probabilistic programming. Pyro builds on the technology behind the deep learning revolution to expand the scale and kind of reasoning problems that we can solve, and is used widely across academia and industry. Omega modernizes classical theories of causality in order to automate the counterfactual “what-if” reasoning that underlies our human ability to think about cause and effect, imagine hypothetical scenarios, and construct explanations.

Despite the success of these projects, they are incomplete. Their limitations in large part stem from the environments and constraints under which they were developed. Artificial intelligence researchers in typical corporate and academic environments often do not have the time, resources, or incentives to distill the results of their work into reusable open-source software. In addition, when A.I. researchers work on applied problems it is often to demonstrate a new method or solve a business need, and rarely drives fundamental research itself.

To build the kind of intelligence we believe is necessary, and to disseminate it widely for broad social benefit, we not only need new technology, but also new environments to support and drive its progression.

Theoretical Programs

Basis’ core research program focuses on developing the emerging theoretical paradigm of “causal computational rationality” into a systematic discipline to explain and engineer general intelligence. We aim to identify important problems within each component, develop algorithms for solving them, and generalize both the problems and their solutions as far as possible along relevant conceptual dimensions. In parallel, we will develop engineering workflows and best practices in order to handle progressively more complex applications over time.

The primary output of this research program is a coherent body of open-source software, a distinguishing feature of Basis. We will also publish technical reports and peer-reviewed papers at top A.I. venues explaining our research, though these will be given less emphasis than our software.

Knowledge

We aim to represent knowledge as programs in general-purpose programming languages equipped with constructs for reasoning about causality and uncertainty. These languages should be expressive enough to directly encode all of the rich features of conventional languages, like loops, recursion, branching, data structures, and function abstraction and reuse. They should also be compatible with other existing knowledge representations like relational databases, deep neural networks, continuous-time dynamical systems, term-rewriting systems, and process calculi, as well as representations yet to be discovered.

Reasoning

We will aim to develop a unified semantics that coherently mixes different forms of reasoning, especially logic, probabilistic inference, optimization, decision-making, abstraction, counterfactual reasoning, causal estimation, and more exotic causal concepts like “actual causality” and causal explanation. To enable these high-level forms of reasoning, we will develop shared intermediate representations and algorithms for primitive inference steps that operate over our expressive knowledge representations. Following the causal computational rationality blueprint, these algorithms will be approximately optimal with respect to computational and other resource constraints across all tasks and knowledge encountered by a system, trading-off between approximation error and computing resources, parallelism and sequential time, and customization and automation. They will make extensive use of learned components as a universally applicable strategy for approximation.

Learning

We aim to develop a unified understanding of the full range of learning and generalization tasks in the current A.I. literature, including learning from unstructured data, meta-learning in collections of similar models, and learning from direct interaction with the outside world. Wherever possible, we will connect algorithms based on these learning concepts into the more general reasoning framework, and hence share much of the mathematical and computational infrastructure.

Engineering

Alongside work on the components themselves, we aim to address practical issues that can hamper their utility as engineering tools. We will maximize the integration of our theoretical program with existing programming languages, compilers, developer tools, and software ecosystems. For example, compatibility with existing automatic differentiation and numerical software is particularly important since we are unlikely to build either of these ourselves. We will also codify informal workflows and best practices from our day-to-day work, develop new tools or interfaces that make using them easier or faster, and gradually attempt to scale them up alongside our technology to enable A.I. “programming in the large”.

Challenges

Basis’ theoretical programs are broad enough that any attempt to approach them directly would falter under the sheer number of possible research questions.

Instead, Basis will execute on our A.I. research program by assembling a carefully curated portfolio of domain-specific challenge problems. Each challenge has an overarching vision, concrete objectives, and is associated with a primary domain of expertise, but also two additional distinctive features.

First, solving a challenge should require us to organize knowledge and people in qualitatively new and better ways through software. This means working closely with expert collaborators in domains outside the narrow range typically considered by AI researchers and delivering new kinds of results beyond academic publications or progress on fixed benchmarks.

Second, solving a challenge should require or motivate general advances in reasoning technology that are applicable beyond the challenge domain.

These challenges will allow Basis to:

1. Advance the theoretical programs
2. Provide solutions to real-world problems
3. Attract collaborators and employees by raising awareness.

Challenges will be proposed, evaluated, and taken on.

Proposals

Challenge proposals will be developed jointly by Basis staff and a diverse mix of people from external academic institutions, nonprofit organizations, governmental departments, and corporations. We do not expect to solicit proposals from the public like a funding agency.

In practice, the challenges we select initially will come from two or three focus areas where the team has some preexisting expertise and relationships with potential collaborators. Over time it may be preferable to develop more in-house expertise and deepen relationships with promising external collaborators in other focus areas.

Evaluation Criteria

Some criteria for evaluating potential challenges are:

• Theoretical impact: Would a solution require us to develop new methods that significantly advance our theoretical programs, and help us to solve other challenges?
• Project impact: Would addressing the challenge be significant, useful, or important on its own terms? Is there a clear mechanistic link between technical improvements and project success?
• Team fit: Are we and our collaborators able to execute? Are we essential to the success of the project? Do our collaborators have sufficient resources and bandwidth? Are there clear project leaders and points of contact on both sides?
• Risk: Would taking on the challenge incur too great or too little technical or organizational risk in the context of our portfolio? Is success overly tied to a single theoretical development? Is it a prudent commitment of our resources given the opportunity costs?
• Growth potential: could work start small and over time translate progressively larger amounts of resources like money, data, headcount or compute into growing project impact? Would the solution organize and align large amounts of previously disparate knowledge and activity in qualitatively new ways?

The first criterion of “theoretical impact” is notable as a key point of differentiation for Basis versus many existing AI labs, which tend to view applied projects as products or as demonstrations of completed work that maximize revenue or visibility without regard to the potential for direct technological feedback into their research programs.

Projects and Deliverables

Challenges will be decomposed into one or more projects to develop specific technical capabilities and concrete application deliverables. Concrete deliverables will vary by challenge and project but will include decisive results and answers to specific questions, statistical reports and summaries, software artifacts, designs, and scientific papers. We will make the deliverables as specific and measurable as possible during the proposal stage to understand and mitigate risk, while also prioritizing feedback and adaptation mechanisms over detailed plans that do not reflect the inherent uncertainty of fast-moving research projects.

Example Challenges

Each challenge that Basis takes on will advance one of a small number of focus areas. In this section, we describe example challenges and projects that meet the criteria above and for which we already have active collaborations.

Emulating Intuitive Scientific Discovery

Challenge: Children resemble professional scientists when they learn about the world around them—they make predictions, construct hypotheses, revise their beliefs in light of evidence, devise and enact experiments, and consolidate their knowledge into simpler and more general theories. A profound idea from cognitive science suggests that this resemblance is more than skin-deep; perhaps the same general faculties which power the formal scientific process are present in children from a very young age. How can we develop machines that engage in this kind of intuitive scientific discovery?

Project: Basis is approaching the challenge as one of active program and language synthesis, developing systems that actively build models of real and simulated environments. These systems construct programs that explain their observations, as well as the languages in which those programs are expressed.

Making Impossible Measurements of Cell and Tissue Dynamics

Challenge: Many of the biggest questions in biology today concern changes over time — e.g. why does a healthy cell develop into a malignant tumor? Unfortunately, our most powerful experimental technologies can only be applied once to a sample, usually destroying it in the process, so many experiments interrogating changes over time are impossible to perform directly.

Project: We aim to learn dynamical models of cell or tissue multi-omic state by integrating many kinds of incomplete causal knowledge, such as known mechanistic signals of temporal change, high-quality reference datasets, and experimental perturbations. We can use the models to partially identify outcomes of impossible experiments (like taking snapshots of the same sample twice).

Citymaking through Participatory Modeling

Challenge: Residents, community groups, and policymakers all have a stake in making decisions that affect different aspects of the city, but today’s print-centric citymaking processes can be exclusionary. How can technology give agency to a broader set of stakeholders, and help them more effectively communicate their knowledge, values, and questions regarding city policies to each other?

Project: We aim to build a digital representation of the city that incorporates existing knowledge in a participatory manner and allows people to explore different scenarios—that is, a city model. Working with local collaborators, we aim for the model to give more agency to underserved constituencies in citymaking processes like rezoning, street life, and budgeting.

Building a Rational, Automated Robot Design Agent

Challenge: While human engineers can create sophisticated robots through intuitive design processes, existing automated approaches to robot design optimization remain limited in their capabilities. Current methods either require massive computational resources or are constrained to simple continuous parameter spaces, making it difficult to explore innovative robot morphologies that could better solve real-world problems.

Project: We are developing R-ADA (Rational Automated Design Agent), an AI system that combines the strengths of human design intuition with computational power. R-ADA uses large language models to control CAD/CAM environments, employs probabilistic programming for reasoning about uncertainty, and leverages Bayesian inference to bridge the simulation-reality gap. This system aims to generate optimal robot designs by breaking down complex design tasks into manageable steps while considering factors like manufacturability, maintenance, and real-world performance.

Collaborative Intelligent Systems

Challenge: Many of humanity’s greatest accomplishments and failures have been determined by our ability or failure to collaborate. As global collaborative systems face unprecedented human-induced changes, there is an urgent call for stewardship akin to the measures demanded by climate change. Beyond humans, it is essential to understand collaborative behaviors across different species, and how properties of an ecosystem affect, and are affected by, the behavior and nervous systems of each animal.

Project: We aim to create software tools for understanding and reasoning about collaborative intelligent systems. This involves integrating and synthesizing knowledge across different disciplines, species, and spatiotemporal scales. We’re developing innovative computational methods to study dynamic systems across different levels of abstraction. We aim to produce accessible open-source software tools that empower scientists, policymakers, and the public to make more informed decisions about collaborative intelligent systems.

Software Platform

Basis will distill the general principles we discover into a domain-agnostic open-source software platform, growing it into the universal reasoning engine from the introduction. The technical details of this platform are contingent on choices that can only be made after operations start, but the founding team has an exceptional track record of delivering similar projects, most notably Pyro and Omega.

Committing to open-source is consistent with Basis’ mission. It expands the set of people who have access to the knowledge and technology that we develop. Some of these people will go on to extend and deploy these in ways we could not have imagined.

Focusing on open-source software also has practical advantages. It offers a unifying scheme for branding and marketing, a clean legal and functional interface for external collaboration, a tool for making our researchers and collaborators more productive and pooling work across projects, an internal separation of concerns that reduces the burden on current and potential employees, and a source of richer and more quantitative feedback on research impact than can be obtained from publications alone.

We will produce three types of software:

1. Applications representing the entire software portion of individual challenges
2. Modules encoding general principles applicable across multiple challenges and domains
3. Infrastructure that ensures modules can be defined very generally and composed with one another

Basis will develop, maintain, and host the core infrastructure under a permissive open-source license. Applications will be developed jointly with challenge collaborators and ideally maintained by them in the long run.

Applications

Applications are solutions to challenges, and each one’s structure will be tailored to its particular domain and expert collaborators. For example, some applications will implement complete solutions to single domain-specific problems, while others may implement workflows that can be specialized across families of problems or may serve as demonstrations that evaluate and disseminate new technologies.

Each application, whatever form it takes, will motivate either the creation of new modules or significant contributions to existing ones. Any domain-general reasoning technology developed within a challenge as part of an application will be factored out, generalized, and consolidated into new or existing platform modules, making it useful to future challenges.

Modules

Modules are units of software that implement a theoretical concept. They are applicable to multiple ongoing and future challenges spanning different domains, are technically substantial but tightly scoped in functionality, and compose with one another via the core infrastructure.

A module may implement a particular inference or learning algorithm, some functionality common to many different concrete problems, or more abstract functionality that enhances other modules. The collection of modules will grow continuously as we investigate new problems.

New modules will typically be created during or after the development of an application when their utility and function is clear, only rarely before, and almost never in the absence of a relevant challenge. Existing modules will also be actively refactored, consolidated, split up, or deprecated over time.

Infrastructure

The core infrastructure is the fabric by which modules are composed together. While the number of modules will grow rapidly, the core infrastructure remains relatively stable, only rarely changing substantially.

A central conviction underlying our research directions is that knowledge can be represented as programs and reasoning and inference as transformations of those programs. Hence, the core infrastructure will consist of computational representations of models as well as algorithms that manipulate those models.

The exact nature of the core infrastructure will develop to meet several different technical demands and constraints but may include new programming languages, transformations to programs in existing languages, compiler technology, program analyses, specifications, and so on, as well as any DevOps tooling that can be shared across projects.

External Collaboration

Thus far we have described Basis’s research goals, a challenge-based strategy for achieving them, and a software-first approach to executing that strategy. If Basis were a conventional research organization generating intellectual property or a company developing software entirely for use by a single customer base, there would be little more to say about its operations.

However, Basis is different. Our software is not merely intended as an intellectual artifact. It is also a unified product for a unique, structurally underserved market spanning three large distinct segments, described below, whose incentives and capabilities overlap with complementary parts of our mission.

We refer to individuals in these segments as “collaborators” rather than “customers” or “users” because we anticipate working in collaboration with them to create value in the respective areas, and because we will tend toward deep engagement with a smaller number of high-value collaborators versus serving as many as possible from a distance. Our software will mediate and enhance these interactions.

Basis is by design well-positioned to serve this market, especially if we can quickly concentrate research talent in-house and deliver key early technical wins. Moreover, if we are successful in addressing the needs of individual collaborators in each segment, over time we expect to catalyze interactions across segments and advance our mission. These interactions will feed a flywheel of economies of scale, new technical capabilities, and market growth, serving as a durable competitive moat and accelerating our research.

Domain Experts

Domain experts include academic scientists, policy-makers, and generally individuals or organizations with both expertise and a vested interest in the success of a challenge project. This segment is already large and will continue to grow as the machine learning revolution sweeps through more application domains, especially in academia where it includes nearly all areas of computational science.

In collaboration with Basis technical staff, domain experts will propose candidate projects, review project proposals, and ultimately play a leading role as an external lead and point of contact on active challenge projects.

Close, sustained collaboration with domain experts is critical to our strategy of using challenges to advance our core research program. They will have domain-specific expertise that is not represented within Basis, and will help ensure that the project is impactful and tethered to the actual requirements.

By choosing to work with us, scientific collaborators stand to benefit from our technology and software engineering expertise. Our theoretical program will bias the selection of challenges toward ones that are not achievable without methodological breakthroughs and are inaccessible to our collaborators or their competitors without our help. We will also contribute substantial computing resources beyond the reach of most academic labs.

External Researchers

External researchers are experts in topics aligned with our theoretical program, like approximate probabilistic inference or programming language semantics, and will typically be drawn from academic research groups. Driven by intellectual excitement over A.I. and insatiable tech industry demand for computer science and especially machine learning expertise, this segment is growing at an astonishing pace — as an example, the number of attendees at NeurIPS, the leading academic machine learning conference, has increased more than 10x in the last decade.

External researchers will act in an advisory role providing technical guidance, writing, contributing code, and designing experiments. In some cases for specific projects, they will provide specialized technical expertise, but more generally they will produce new platform modules or major improvements to existing ones, as well as papers at top journals or conferences about the methods we develop.

Working with Basis will not only reduce the burden researchers face translating ideas into publications but also expand the scale of projects they can pursue. It can be difficult to develop and maintain high-quality software in an academic setting; external researchers will be better equipped to execute on or disseminate some of their best ideas by building on functionality in our platform. Some projects may create opportunities for interactions between multiple collaborators who would otherwise find it difficult to find opportunities to work together, like programming language and machine learning researchers.

Working with external researchers will allow Basis to leverage expertise that would be too expensive or in some cases impossible to attain otherwise. External researchers will also help overcome any scarcity of talent by recommending their students or colleagues to work with Basis.

Open Source Community Members

As an organization focused on the development of open-source software, Basis will interact extensively with other individuals and communities in the surrounding software ecosystem whenever doing so would advance the core organizational mission.

Open source ecosystem collaborations with Basis will likely be structured in one of three ways:

1. Working with developers of competing or upstream packages on shared infrastructure.

   Collaborators on this kind of project would benefit from critical infrastructure components that no single party could have afforded to develop individually (cf. PyTorch Distributions).

   This would increase our long-term ability to focus on research rather than lower-level engineering and improve the quality of our platform and application software.

2. Helping developers of downstream software integrate our platform into their packages

   These collaborators would benefit from our platform’s unique technical capabilities, which will continue improving as our platform grows.

   This kind of project may make us more attractive to potential employees or collaborators by increasing the reach of their work (cf. Funsor and Pyro/Stan-Pyro compiler) or providing exclusive access to new userbases and distribution channels (cf. Pyro and Terra).

3. Sporadic, ad-hoc interactions with individuals who are not affiliated with Basis or one of our official projects but who for their own reasons wish to contribute code and other resources to our open-source repositories.

   These individuals are a familiar feature of most significant open-source software projects and have a diverse array of motives, from curiosity to education and resume-building to fixes or enhancements necessary for their own work built on our platform.

   Beyond the intrinsic value of the contributions themselves, these interactions have the potential to benefit us further by developing over time into deeper and more formal employment or collaborative relationships, serving as a unique talent pipeline.

Open-source collaborators will generally benefit from the ability to influence large and influential projects. They will also get access to and the attention of domain experts who might otherwise be inaccessible.

Basis will promote and celebrate our open-source collaborations, building broad community support, while also remaining vigilant to avoid distractions that can accompany such collaborations.

Organizational Structure

The model laid out in this document is a recipe for addressing the opportunities in the introduction successfully and at scale. This section describes aspects of an organizational structure designed to execute that mandate.

Mission Statement: Building a universal reasoning engine to advance society’s ability to solve intractable problems.

Cultural Values

The people of an organization determine its culture. In our hiring, management, and in guiding our overall strategy, we aim to be:

Ambitious

Reaching our goals requires more than incremental progress. It means taking risks and embarking on ambitious, and sometimes audacious projects.

Committed

We give the projects and people we take on the time and attention they need to be successful.

Deliberative

We continuously evaluate what we’re trying to do, why we’re trying to do it, and what impact our actions could have on individuals and society at large.

Constructive

All things equal, we prefer to build, demonstrate, invent, and improve, rather than just critique.

Joyful

We recognize that this is a human endeavor that is only worthwhile if it can be done in a way that is fulfilling and joyful.

Internal Organization

Basis’ unique mission lends itself to a distinctive organizational structure. Its viability is contingent on funding for AI research remaining relatively abundant compared to scarce research and engineering talent (which seems quite likely, at least over the next 3-5 years) and ultimately on our ability to deliver on our ambitious promises. It is also subject to continuous change as we ramp up our research operations and find out what does and does not work.

We anticipate that most employees will be staff researchers and engineers with long-term contracts at competitive salaries unconnected to specific funding sources. The internal team will also consist of interns and contractors, who will be afforded largely the same roles and responsibilities as full-time employees except that their work will generally be connected to a single particular project.

Major financial and personnel decisions, including hiring and firing, will ultimately be made by the directors, who will also remain responsible for fundraising. Funds will be centrally administered and considered fungible unless earmarked by their source for specific projects or purposes. These will be decoupled to the greatest extent possible from day-to-day work and major technical decisions related to challenges and theoretical programs so that the organization can stay focused on its research mission.

We will not aim to establish fixed research groups with permanent, full-time managers. Instead, we will organize around temporary projects, which will in turn roughly mirror the structure of our software platform and challenge portfolio. Projects will be overseen by employees as technical leads. All internal employees will be able to propose a project, but smaller selected project-specific review panels will have the responsibility of making a recommendation of whether to accept. Projects will only be accepted if they can be guaranteed sufficient funds and staff to succeed or fail on their own merits.

This organizational structure makes for a distinct form of incentive-aligned research career development for employees in which they are able to reap the rewards of success as their projects and responsibilities grow while limiting their exposure to the financial risks, publication pressures and administrative grind of fundamental research funding.

We will be hybrid — some employees will work fully remotely, others fully in-person, and the majority will do both. The modern workforce expects and benefits from a great amount of flexibility in terms of when and where they work. At the same time, solving hard problems often requires deep collaboration that currently can only be done effectively in person.

Organizational Evolution

If we are successful, Basis will develop into an organization that is superlative in our ability to build general machine reasoning, unique in its real-world impact on solving problems, and distinct in structure to anything in existence today.

However, we are not interested in novelty for its own sake. Rather we aim to take a first-principles approach, designing Basis not based blindly on current standard practices, but rationally towards our core mission and in accordance with our core human values.

Model Institutions

What distinctive long-term consequences can we expect if Basis is successful in its mission? To explore this question, we offer several analogies to other transformative institutions.

One point of comparison is Johns Hopkins, the first real US research university, where modern American medical education and research were created essentially from scratch in a remarkably short time around 1900. If we are truly successful in discovering principles for engineering intelligent systems, we will have a unique capacity for training new researchers and engineers to understand and apply those principles faster and better than other companies or universities. This will greatly expand our potential labor pool and hopefully, lead to more widespread adoption of our technology.

Also relevant are global climate models and the distributed, international consortia of tens of thousands of scientists that develop them. If we are successful in developing tools and workflows for progressively larger and more complex challenges, it will become possible for decentralized research communities to rapidly encode many other field-spanning bodies of scientific knowledge in shared, self-consistent, self-adjusting causal models.

Mozilla is another point of comparison, a rare example of an organization that continues to have an outsized influence on society in a positive manner using a deep technological foundation and without sacrificing its core principles. Mozilla is also unusual and inspirational in its organizational structure, putting open-source software first and connecting with and capitalizing on a broad and international community of professional and enthusiast developers.

Finally, consider Flagship Pioneering, an unusual biotechnology-focused investment firm whose portfolio includes Moderna, the mRNA therapeutics company that developed the first successful Covid-19 vaccine. Flagship is interesting because it is less a venture capital firm than a factory for manufacturing startups. Nearly all of its companies are started in-house by Flagship employees as research projects that become proto-companies after a well-defined, systematic technical exploration phase and eventually either killed or spun out if they are deemed sufficiently de-risked for external investment. It is possible to imagine Basis eventually doing something similar with its challenge projects: systematically searching for opportunities (for profit or otherwise) uniquely suited for Basis’ core reasoning technology, staffing and exploring them with internal funding, and eventually spinning out subordinate startups or focused research organizations.

Contributors

Research: Zenna Tavares, Eli Bingham, Emily Mackevicius

Illustration: Doug John Miller