SRI International's DARPA SAIL-ON research on AI for open-world novelty — a federal contract (USAspending)

1What happens outside the training data

Most of today's AI performs well as long as the situations it faces fall within the range of its training data. When it encounters kinds of situations it was never trained for — what researchers call "novelty" — its performance can break down sharply, and it may make confident mistakes without realizing it is wrong.

This contract belongs to DARPA's SAIL-ON program (the Science of Artificial Intelligence and Learning for Open-World Novelty), a research effort aimed at overcoming the brittleness of AI built on "closed-world" assumptions, in which the rules and options are fixed in advance, so that systems can keep functioning in an "open world" where the rules and assumptions can change partway through.

2Building principles, not an application

What matters here is that the goal is not to build a specific application product, but to establish the underlying "scientific principles" and "general engineering techniques and algorithms." In other words, rather than ad hoc fixes that work in only one domain, the aim is to develop reusable, cross-cutting foundations: mechanisms for noticing the unknown, for representing how a situation has changed, and for adjusting behavior in response.

Investment in this kind of basic research tends to have wide social ripple effects, because its results are shared broadly as papers, methods, and concepts that can serve as a starting point for much later research and implementation.

3Why DARPA funds high-risk work

DARPA, the agency that issued this contract, is a Department of Defense organization that concentrates funding on high-risk, high-impact research at the stage just before practical application. The recipient, SRI International, is an independent research institute that has long conducted commissioned research for government and industry, and it takes on the role of carrying out this kind of foundational program.

Through the public procurement framework of a federal contract, one can trace exactly which research themes received taxpayer funding, through which institution, and at what scale — and that traceability is the cross-cutting significance of this record. The specific deliverables or technical outcomes produced under this contract are not stated in the source data, so they are not described here.

4What the basic research is putting in place

The intent to build not an applied product but the ground beneath it reads concretely when broken into parts.

1. 1Notice the unknownDetect situations not anticipated in training
2. 2Represent the changePut what changed into a form that can be handled
3. 3Adjust behaviourChange how it acts to match
4. 4Make it reusableEstablish it as basic technology usable across fields

Today's AI performs well within the range of situations in its training data while collapsing, or confidently erring without noticing, when it meets the unknown. Investment in basic research carries wide effect because results are shared as papers, methods and concepts and become the starting point for much later work.