Four questions I am taking into 2027
My expectations for hybrid AI and quantum systems, written in August 2026 with a few clear signs that could make me change my mind.
It is easy to remember a prediction as wiser than it was. By the time the outcome is clear, we have often softened the part that was wrong. I am writing these expectations down in August 2026 so I can return in August 2027 and read the actual words. I would like to know which ideas helped me understand the year and which ones got in the way.
I am watching four related questions: how quantum results are described, what buyers need to know about autonomous systems, what larger models leave unresolved and which hybrid components become useful in ordinary work.
1. Will results tell us where they work?
I expect useful applied quantum work to describe the conditions in which a method is competitive, alongside where it stops helping. That gives another team a starting point for a real comparison. A broad advantage headline is harder to use without knowing the classical reference behind it.
I would reconsider if the most prominent applied results in 2027 still centre on broad advantage claims followed by prolonged arguments about the comparison. That would suggest the reporting habits have changed less than I expect.
2. Will buyers ask how a decision can be explained?
Capability will still matter, but I expect more attention to the record a system leaves. If a recommendation affects a consequential decision, someone may later need to understand why it was made. That makes traceability a practical purchasing question as well as an engineering concern.
If decision provenance remains an optional extra in relevant procurement by late 2027, this expectation will have moved too quickly. I want to judge that against actual requirements, not only conference language.
3. What will a larger model leave for the system to solve?
I expect uncertainty, verification and responsibility to remain partly architectural questions. Better models can help, but the surrounding software still chooses what to check and who may act. This is an emphasis of my book, and also an assumption I need to keep open to evidence.
A scale-driven result that substantially improves calibrated uncertainty on changed data, without an accompanying architectural change, would make me reconsider. This is the expectation I would be happiest to revise: it would remove some difficult work.
4. Will useful hybrid components look fairly ordinary?
I expect the components that find a place in working systems to have defined jobs: a belief update, a sampling routine or an optimisation step. Their value may be easier to describe in a narrow setting than in a general claim about intelligence.
A general-purpose quantum decision system deployed in a consequential domain during 2027, with records that withstand independent review, would challenge that expectation. Until then, I am interested in the small contributions we can measure and learn from. I will return to these notes next year.
- Write down the expectation
- Say what could change it
- Return to the evidence