AGI-26 keynote

Alexander Lerchner

Alexander Lerchner is a Senior Staff Scientist at Google DeepMind, with a record that runs from mean-field theories of cortical networks to the disentangled representations and embodied agents his lab is known for.

“Learning an interpretable factorised representation of the independent data generative factors of the world without supervision is an important precursor for the development of artificial intelligence that is able to learn and reason in the same way that humans do.”

— opening of β-VAE (2017), on which he is last author

One question, from neurons to agents

Read the archive in order and the subject is always the representation: what should a system carry about the world, and what makes that carrying good? The systems change — cortical populations, latent variables, objects, agents — and the question does not.

  1. Cortex, in equations

    2004–2006

    The record opens in computational neuroscience: mean-field theories of balanced cortical networks, worked out so that firing rates and firing correlations can be determined self-consistently "without being restricted to specific neuron models". The question underneath is already a representational one — what does a population of neurons encode, and how stably.

    Mean field theory for a balanced hypercolumn model (2006) ↗
  2. Factors of variation

    2016–2017

    The same question in a deep network. β-VAE adds a single adjustable weight to the variational objective and finds that the latent space separates into independent, interpretable factors — motivated by the argument that learning "an interpretable factorised representation of the independent data generative factors of the world without supervision is an important precursor" for AI that learns and reasons as humans do.

    β-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework (2017) ↗
  3. Concepts

    2017

    If the world runs on "a relatively small set of coherent rules, such as the laws of physics or chemistry", the regularities they produce should be discoverable without supervision and recombinable afterwards. SCAN learns hierarchical compositional concepts on that conjecture; DARLA uses disentangled vision so an agent "learns to see before learning to act", and transfers to domains it was never trained on.

    SCAN: Learning Hierarchical Compositional Visual Concepts (2017) ↗
  4. Saying what it means

    2018

    A programme this influential eventually has to define its own term. This paper notes there is "no generally agreed-upon definition of disentangling" and proposes one grounded in symmetry transformations — an unusual move: a lab formalising the concept its own results depend on.

    Towards a Definition of Disentangled Representations (2018) ↗
  5. Objects

    2019–2021

    Factors become things. MONet decomposes a scene into entities without supervision, on the argument that "the ability to decompose scenes in terms of abstract building blocks is crucial for general intelligence"; SIMONe extends it to video, inferring objects and viewpoint jointly.

    MONet: Unsupervised Scene Decomposition and Representation (2019) ↗
  6. Agents in worlds

    2021–2025

    Then the representations get a body. Alchemy builds a structured task distribution for meta-reinforcement learning; the SIMA project trains agents to follow free-form language instructions across many commercial 3D games, tackling what it calls "a key challenge for creating general AI".

    SIMA 2: A Generalist Embodied Agent for Virtual Worlds (2025) ↗
  7. The decade ahead

    2026

    And the newest record steps back from systems to consequences. A fourteen-author report that opens by observing that human-level AGI "has moved from far-fetched speculation to being a concrete next-decade target for many of the largest AI organisations", and asks what follows.

    From AGI to ASI (2026) ↗

At a conference called AGI-26

The 2026 report in this corpus, From AGI to ASI, is written by fourteen authors including Shane Legg, Marcus Hutter, Joel Z. Leibo, Iason Gabriel and Allan Dafoe. It opens by noting that over the last decade, building human-level AGI “has moved from far-fetched speculation to being a concrete next-decade target for many of the largest AI organisations”, and that achieving it “would have profound and far-reaching impacts on human society”.

It is the only document in this archive that argues about the destination rather than building a piece of the road, and it is worth reading next to the rest — two decades of work on what a representation ought to be, by someone now co-signing a report on what comes after human-level.

Find the report ↗

Reading the author lists

Almost everything here after 2016 is co-authored, and Lerchner is usually in the last position — the senior slot in this field’s convention. The recurring names are Loïc Matthey, Christopher Burgess, Irina Higgins, Matthew Botvinick, Rishabh Kabra and Arka Pal. Two of the most recent entries are the work of a hundred-strong “SIMA team”.

The site states his position on every selected paper, computed from the record, because on a corpus like this one the position is most of the information.

Browse all 53 records →

Selected work

17 entries drawn from the archive, titled exactly as the records read, with his position in the author list computed from each record. Everything else is in the corpus.

  1. 2026
    From AGI to ASI
    arXiv (Cornell University) author 3 of 14
  2. 2025
    SIMA 2: A Generalist Embodied Agent for Virtual Worlds
    arXiv (Cornell University) author 3 of 66
  3. 2024
    Scaling Instructable Agents Across Many Simulated Worlds
    arXiv (Cornell University) author 49 of 94
  4. 2021
    Formalising Concepts as Grounded Abstractions
    arXiv (Cornell University) author 2 of 7
  5. 2019
    MONet: Unsupervised Scene Decomposition and Representation
    arXiv (Cornell University) last of 7 authors
  6. 2019
  7. 2018
    SCAN: Learning Hierarchical Compositional Visual Concepts
    UCL Discovery (University College London) last of 10 authors
  8. 2018
    Understanding disentangling in $\beta$-VAE
    arXiv (Cornell University) last of 7 authors
  9. 2018
    Towards a Definition of Disentangled Representations
    arXiv (Cornell University) last of 7 authors
  10. 2017
    beta-VAE: Learning Basic Visual Concepts with a Constrained Variational Framework
    International Conference on Learning Representations last of 8 authors
  11. 2017
    DARLA: Improving Zero-Shot Transfer in Reinforcement Learning
    arXiv (Cornell University) last of 9 authors
  12. 2016
    Early Visual Concept Learning with Unsupervised Deep Learning
    arXiv (Cornell University) last of 8 authors
  13. 2006
    Response Variability in Balanced Cortical Networks
    Neural Computation first of 6 authors

Background

Now
Senior Staff Scientist, Google DeepMind — the affiliation given on his own Google Scholar profile, which lists his interests as artificial intelligence, neuroscience, philosophy of mind and philosophy of AI.
In this corpus
Records from 2004 to 2026, in two distinct phases: mean-field computational neuroscience with John Hertz and colleagues, and — from 2016 — representation learning and agents at DeepMind.
Lines of work
Disentangled representations (β-VAE, SCAN, DARLA, and the definitional paper); object-centric scene representation (MONet, SIMONe, PARTS, and the iterative-variational-inference work on multi-object representation); and the Alchemy and SIMA agent projects.

About this archive

This is an independent archive of Alexander Lerchner’s publication record, built for Society of Minds Aligned and AGI-26. Records span 2004–2026 and are drawn from OpenAlex; preprints and versions of record are both kept, so the archive is a reading index, not a bibliometric count. It is the smallest corpus in this constellation and does not pretend otherwise.

The affiliation and research interests above come from his own Google Scholar profile. Nothing here is asserted that the corpus or that profile does not support — in particular, the site makes no claim about his views beyond what the papers he co-authored say.

Not written by, reviewed by, or endorsed by Alexander Lerchner. If you are Alexander and something here is wrong, the feedback control at the bottom of the page reaches us directly — and the claim bar will hand you the site.

Are you Alexander Lerchner? Or know Alexander Lerchner? We'd like to hand this room over. →
Join the constellation