Anders Sandberg

Anders Sandberg is a Swedish computational neuroscientist and futures researcher whose work consists largely of putting numbers on questions that are usually discussed without them: how much compute a brain emulation would need, how long it would take to colonize a galaxy, how much confidence a risk estimate can carry when the model itself might be wrong. He co-authored the standard technical roadmap for Whole brain emulation and spent nearly two decades at Oxford's Future of Humanity Institute.

§Career

Sandberg's doctorate, completed at Stockholm University in 2002, was on attractor neural network models of memory — the mathematics of how a network of units settles into stable states that behave like stored patterns. That background explains his position on emulation: he approaches the question as a modeller who has built neural systems, not as a philosopher arguing from the outside.

He chaired the Swedish transhumanist association during the 1990s and co-founded a Stockholm think tank before joining the Future of Humanity Institute in 2006, where he worked with Nick Bostrom, Toby Ord, Eric Drexler and others until the institute closed in 2024. He subsequently moved to the Institute for Futures Studies in Stockholm. He is a prolific public communicator, and much of his output takes the form of estimates published as blog posts, technical reports and talks rather than as conventional papers.

§Whole brain emulation

The 2008 report written with Bostrom is the reference document for the emulation question.¹ Its method is to refuse to argue about whether emulation is possible and instead ask what would have to be true at each level of description, then estimate the requirements under each assumption.

Its premise, stated as a premise rather than defended, is Substrate independence: that what matters about a brain is the organization of its causal structure rather than the material implementing it.

The roadmap defines a ladder of models: from a coarse brain-region simulation, through spiking neural networks, compartment models with realistic dendrites, molecular-level channel dynamics, and down to whole-molecule and quantum descriptions. Each rung carries an estimate of storage, compute and scanning resolution. The estimates span many orders of magnitude, and the report is explicit that the correct rung is unknown — this is the central finding, not a hedge. It also identifies scanning throughput rather than compute as the likely binding constraint, a judgement that Connectomics has broadly borne out: reconstructing a cubic millimetre of cortex remains a multi-year effort for a large collaboration. The same analysis makes structural Brain preservation a coherent research target, since a preserved connectome is a scannable object even if nobody can yet scan it at scale.

What the roadmap actually establishes

Not that emulation will work. It establishes that the requirements are finite and estimable given a level of description, and that the disagreement between optimists and sceptics is almost entirely about which level suffices. That reframing is the report's contribution.

Sandberg has separately written on the ethics of emulation — the welfare of a running emulation, the implications of copying, and the research ethics of experiments on emulated brains, which is Machine consciousness with the moral stakes made unavoidable. He is consistently more cautious in public than the Mind uploading discourse that cites him.

§Quantifying the long term

A recurring pattern in Sandberg's work is taking a speculative claim and computing its physical requirements. With Stuart Armstrong he calculated the energy and engineering cost of launching self-replicating probes to every reachable galaxy, concluding that intergalactic settlement is within the physical means of a civilization not much more capable than a plausible near-future one — which sharpens rather than resolves the Fermi paradox.² With Drexler and Ord he argued that the paradox partly dissolves under correct handling of uncertainty: multiplying point estimates in the Drake equation discards the distribution, and propagating the actual uncertainty places substantial probability mass on humanity being alone.³

He has written on the thermodynamic limits of computation, on "Jupiter brains" as the physical ceiling for a single computational structure, and on the aestivation hypothesis — the proposal that an advanced civilization might wait for the cosmic background to cool before computing, because computation is cheaper at low temperature. With Ord and others he formulated the argument that model uncertainty dominates very low probability estimates: a calculation giving a one-in-a-billion chance of catastrophe is worthless if the chance the calculation is wrong is one in a thousand.⁴

§Enhancement and risk

Sandberg's enhancement work is mostly with Bostrom. "Converging Cognitive Enhancements" surveys the routes — pharmacological, genetic, external, and interface-based — and argues that they interact rather than compete. The "wisdom of nature" heuristic asks why evolution has not already made a proposed improvement, and treats a satisfactory answer as a precondition for expecting the improvement to be free of hidden costs; the usual acceptable answers are changed environments, evolutionary constraints, and value discordance between fitness and welfare. It is a rare piece of pro-enhancement writing that supplies a filter its own conclusions must pass. See Human enhancement and Nootropics.

On risk, the "anthropic shadow" argument he co-wrote holds that the historical record systematically understates the frequency of catastrophes severe enough to have prevented observers from existing — so the empirical base rate for some extinction-level events is biased downward by the fact that anyone is around to compute it.

§Reception and legacy

Sandberg is cited across futures studies, AI safety and philosophy, and is unusual in retaining credibility with mainstream scientists while working on subjects they generally avoid. The reason is methodological: he states assumptions, publishes the arithmetic, and marks the gap between what follows from physics and what follows from speculation. Criticism concentrates on the limits of that method — that computing the physical requirements of a scenario says nothing about its biological or social plausibility, and that the roadmap's neutral framing gives emulation more credibility than neuroscience currently warrants.

His long-announced book on grand futures, covering the physical possibilities open to a civilization over cosmological time, has circulated for years mainly as drafts and lecture material, and is easier to encounter in talks than in print. Its subject is the same one this wiki's Future of humanity article takes up, approached from the side of physical limits rather than of technology. And its central question is the one his quantitative work keeps producing and cannot answer: the physics permits enormously more than humanity currently does, so the binding constraint is not physical, and nothing in the calculations says what it is instead.

§See also

• Whole brain emulation
• Mind uploading
• Connectomics
• Existential risk
• Nick Bostrom
• Eric Drexler
• Substrate independence
• Future of humanity

§References