Anno Machinae '27
A field guide to the year
in artificial intelligence
AI research is advancing faster than most working professionals can follow. Anno Machinae selects twelve areas where laboratory results are within two years of becoming products, and publishes a scientific account and a practitioner's account for each. Free to read online, available in print at cost, out in mid-2027.
What is Anno Machinae?
Anno Machinae is an annual review of AI research, chosen on a single criterion: how close the work is to becoming something useful in the world. The 2027 edition covers twelve such areas, all of them at or near the point where laboratory results are turning into products.
For each topic, there is an account from a researcher working in the area and one from an industrial practitioner in a sector the research is approaching. Both are asked to write for an intelligent reader who is not a specialist — which is, in the end, a more exacting standard than writing for colleagues alone.
The 2027 edition is going to be free to read online. A printed edition will be available at cost.
Each contributes a single account of one topic — what the research currently shows, where the genuine open questions sit, and what the field is likely to look like in three to five years.
Each comes from a sector that this research is moving towards. Their account is not a forecast — it is an assessment of what is already beginning to shift, and what it would be sensible to do about it.
The 2027 Topics
The 2027 edition covers twelve topics, four of them at length and eight in shorter form. Each is covered by a researcher and a practitioner, writing independently of one another.
Primary Chapters
-
01
Intelligence Beyond Tokens
Language models predict the next token. A substantial and growing body of work argues that this is the wrong basic unit for systems that need to reason rather than merely complete — and proposes architectures that treat larger structures, such as concepts or plans, as the primary thing to model. This chapter covers that work, and what it would mean for the existing software ecosystem if it gains ground.
-
02
Acting, Adapting, Evolving
Once a model is deployed, it stops learning from the world it has been put into. The research covered here examines ways to close that gap — reinforcement from real-world feedback, online adaptation methods, and the engineering questions that arise when a system is expected to improve in production rather than simply run.
-
03
The Self-Improvement Loop
Improving a model has always required some form of human input — labels, preference judgements, or generated examples. The research covered here explores what happens when that dependency is loosened: models that produce their own training signal through synthetic data, self-play, and process reward models. The practical question is what changes about the economics of model development once annotation is no longer the constraint.
-
04
Hidden Geometries of How AI Learns
What a neural network learns has a precise geometric structure, and understanding it goes some way towards explaining why models generalise well in some settings and fail in predictable ways in others. This chapter covers the relevant mathematics and what it implies about model behaviour at deployment — including which failure modes can be anticipated in advance.
Further Chapters
- 05Recurrent Nets, Spiking Neurons and Attention-Free Flows
- 06The Long Document Problem
- 07World Simulation as Environment and Testbed
- 08Thinking Harder at Inference Time
- 09Making AI Safe, Honest and Measurable
- 10Visual and Multimodal Grounding
- 11Fast, Small and Deployable
- 12AI Meets Society
The People
Behind It
A small number of contributor slots remain in the 2027 edition. If you are a researcher active in one of these areas, or an industrial practitioner in a sector this research will reach, we would like you to write for it. Each contribution is a single piece — either the scientific account or the practitioner's account for one topic.
Misha Krymov
Founding DirectorAn architect turned entrepreneur, Krymov has built ventures spanning AI, enterprise software, and the built environment — among them ARMA AI, a Miami-based AI R&D studio, and Sleepbox, a modular micro-hotel concept. A former Research Fellow at MIT, he is based in Miami.
Lisa Bylinina
Editor-in-ChiefA linguist working at the intersection of theoretical linguistics and natural language processing. Assistant Professor in Computational Linguistics at Utrecht University's Institute for Language Sciences, where she researches semantics, AI, and how language and machines meet.