Professor of Language, Communication and Computation at Utrecht University.
a.p.j.vandenbosch@uu.nl
visiting Utrecht University, Faculty of Humanities
Trans 10
3512 JK Utrecht
The Netherlands
I study how computers learn to handle language.
My group builds the kind of AI behind chatbots like ChatGPT, and we ask:
can we make it greener, more open about where its answers come from,
and easier for society to govern?
Below you'll find my research, projects, software and the people I work with.
For the formal record, see my CV and
publications page.
Shorter, less polished thoughts go on my blog.
Upcoming book: Het geheim van de pratende machine
My new book (in Dutch),
Het geheim van de pratende machine: Hoe chatbots ons betoveren met woorden
(The secret of the talking machine), appears in October 2026 with
Uitgeverij Prometheus.
It explains for a general audience how chatbots work: brilliant, helpful imitators
that predict the next word without understanding it. The story runs from the
Second World War to today's language models.
New publication: Memory-based speculative decoding
Speculative decoding speeds up a large language model by letting a small
"draft" model guess ahead. In this paper the draft model is Olifant, our
memory-based model, instead of a neural network. On legal text, generation becomes
up to 2.4× faster and uses 2.5× less energy per token, with identical output.
The three-minute animation below shows how it works.
Towards Green Text Generation: Memory-Based Speculative Decoding.
Antal van den Bosch. To appear in the Proceedings of INLG 2026.
Code: github.com/antalvdb/olifant-speculative.
What I work on
At heart, my research is about one deceptively simple task: predicting the next word.
It is the trick at the core of every modern language model, including ChatGPT, and a good
next-word predictor turns out to capture an astonishing amount of language, reasoning and culture.
My work combines machine learning, teaching computers from examples,
with language. The models we build are useful in their own right, and they also
speak to questions in linguistics, psychology and neuroscience about how humans handle
language. I especially enjoy projects that cross disciplines.
Olifant: a lighter alternative to ChatGPT-style AI
Today's chatbots run on Transformers, a powerful but power-hungry architecture.
With Maarten van Gompel, Peter Berck, Ko van der Sloot and a growing team of students,
I am reviving an older, simpler idea, memory-based language models, and showing
that it can do much of the same work for a fraction of the energy.
Training our models uses roughly 1,000× less electricity than training
a Transformer of similar capability, and answering questions costs 10 to 100× less.
They scale well with more data and, unlike mainstream language models, they can
show which training examples an answer is based on. That makes their behaviour
easier to inspect and trust.
We released the software as Olifant
(Dutch for "elephant"), described the method in a paper,
and share demos and downloadable models.
From the Olifant paper: estimated CO2
emissions (measured with CodeCarbon) of language models predicting the next words in a
500,000-token text. Dashed lines show everyday emissions for comparison: a washing-machine run,
ten minutes of driving, a litre of milk. Two of our models stay well below the washing machine;
larger GPT-style systems climb steeply.
Try it yourself: for any sentence you type, the demo below shows which earlier
examples Olifant draws on to predict the next word. There is also a
text generation demo
and a collection of Olifant models
on Hugging Face.
Advising on AI policy
Because I build these systems, I am often asked what to do about them.
With Fabian Ferrari
and José van Dijck I advise
governments and academic bodies on regulating generative AI, bringing the technical
and the governance side to the same table. Read more in the
UU news item
or our Nature Machine Intelligence paper.
The course introduces the Transformer, the T in GPT and the architecture that
has reshaped AI since ChatGPT appeared in late 2022. At its core it does one thing:
predict the next word. So how does it pass exams, or appear to reason?
Where does it fail, and where is the field heading?
Public talks
Newly appointed professors in the Netherlands and Belgium traditionally give a
public inaugural lecture. I have done so three times: in Tilburg (2008), at Radboud
Nijmegen (2012), and at KU Leuven KULAK (2023),
where I held the
Francqui Chair on Language and AI.
All three circle the same idea: predicting the next word gets you a long way.
In 2008 I already argued that next-word predictors keep improving with more training text,
with every tenfold increase in data buying a steady, predictable gain in accuracy.
That observation (Heaps' Law,
for the technical name) quietly underlies today's boom in large language models.
Het volgende woord, Tilburg University, October 2008. Some animations on YouTube; below, a short one with Dutch sentences in which the next word is all but fixed.
Positions and memberships
Since September 2022 I am faculty professor at the
Faculty of Humanities,
Utrecht University.
Since 2026 I am also the faculty's vice-dean of Research and Impact.
Between July 2023 and June 2026 I chaired the Social Sciences and Humanities domain of
NWO, the Dutch Research Council, and was a member of its Executive Board.
Our projects also produce software, released as open source where we can.
Some packages run as webservices
with a friendly interface, and several are part of national and European research infrastructure.
A few highlights:
Olifant: the memory-based
language model described on the left. It builds on a long line of earlier work, including
WOPR (2010), which already predicted next words for fun with far less data and much smaller computers.
With Peter Berck, Maarten van Gompel, Ko van der Sloot, Teun Buijse and Ainhoa Risco Paton.
Frog: an all-in-one tool for Dutch text that tags parts of speech,
finds word stems and parses sentence structure.
With the Frog development team.
T-Scan
(also a web tool): analyses Dutch text
for features related to reading difficulty.
Timbl: the Tilburg Memory-Based Learner,
the classic toolkit behind much of our memory-based work.
With Ko van der Sloot, Walter Daelemans and Jakub Zavrel.
More software and digital infrastructure
Natural language processing
Valkuil.net and
Fowlt.net: context-sensitive spelling correctors
for Dutch and English.
Colibri Core: efficient n-gram and skip-gram modelling. With Maarten van Gompel.
Mbt: memory-based
part-of-speech tagger. With Ko van der Sloot, Jakub Zavrel and Walter Daelemans.
WOPR in 2010 generating "word salad": a tiny precursor of ChatGPT.
Digital research infrastructure
CLARIAH: Common Lab Research Infrastructure for the Arts and Humanities.
Nederlab: brings together digitised Dutch texts from the Middle Ages to today in one searchable interface (funded by NWO).
FutureTDM: a Horizon 2020 action on text and data mining.
TwiNL: a Netherlands eScience Center project with Erik Tjong Kim Sang.
ISHER: Integrated Social History Environment for Research.
Past projects
Show research projects I previously led or co-led
ADNEXT: Adaptive Information Extraction over Time, part of the COMMIT programme.
Ubisoft released the brain-training game
My Word Coach (Dutch)
for Nintendo DS and Wii. I did the background work on the Dutch localisation
together with Walter Daelemans (Antwerp).
April 2008.
