arXiv subjects

Toby Ord

Publications and source records attributed to Toby Ord.

17 recordsLinked to original sources

arXiv2026

The Dynamics of Intelligence Explosions

AI is increasingly being used to help with AI R&D. Under certain conditions this feedback loop might be able to produce an intelligence explosion, with rapidly escalating AI capabilities. I explore the mathematics of the most explosive possibilities, with an eye to understanding what drives the dynamics. I show that singular growth (towards a vertical asymptote) is harder to achieve than would be expected from recent economics-inspired modelling, and that there is an important but neglected class of growth rates that are faster than exponential but don't lead to a vertical asymptote. I draw out the generation time (the time to go around the feedback loop) as a neglected parameter that plays a pivotal role in determining the behaviour of any intelligence explosion --- one cannot have singular growth unless the generation time rapidly approaches zero.

cs.AI↗

arXiv2025

Evaluating the Infinite

I present a novel mathematical technique for dealing with the infinities arising from divergent sums and integrals. It assigns them fine-grained infinite values from the set of hyperreal numbers in a manner that refines the standard theories of summation and integration. This has implications in statistics (helping us work with distributions whose mean or variance is infinite), decision theory (allowing comparison of options with infinite expected values), economics (allowing evaluation of infinitely long streams of utility without discounting), and ethics (allowing evaluation of infinite worlds). There are even implications for finite cases, as the ability to handle these infinities undermines a common argument for bounded utility and the discounting of future utility.

math.GM↗

arXiv2025

Is there a half-life for the success rates of AI agents?

Building on the recent empirical work of Kwa et al. (2025), I show that within their suite of research-engineering tasks the performance of AI agents on longer-duration tasks can be explained by an extremely simple mathematical model -- a constant rate of failing during each minute a human would take to do the task. This implies an exponentially declining success rate with the length of the task and that each agent could be characterised by its own half-life. This empirical regularity allows us to estimate the success rate for an agent at different task lengths. And the fact that this model is a good fit for the data is suggestive of the underlying causes of failure on longer tasks -- that they involve increasingly large sets of subtasks where failing any one fails the task. Whether this model applies more generally on other suites of tasks is unknown and an important subject for further work.

cs.AI↗

arXiv2025

Inference Scaling Reshapes AI Governance

The shift from scaling up the pre-training compute of AI systems to scaling up their inference compute may have profound effects on AI governance. The nature of these effects depends crucially on whether this new inference compute will primarily be used during external deployment or as part of a more complex training programme within the lab. Rapid scaling of inference-at-deployment would: lower the importance of open-weight models (and of securing the weights of closed models), reduce the impact of the first human-level models, change the business model for frontier AI, reduce the need for power-intense data centres, and derail the current paradigm of AI governance via training compute thresholds. Rapid scaling of inference-during-training would have more ambiguous effects that range from a revitalisation of pre-training scaling to a form of recursive self-improvement via iterated distillation and amplification.

cs.CY↗

arXiv2024

Bounds on the rates of growth and convergence of all physical processes

The upper limit on what is computable in our universe is unknown, but widely believed to be set by the Turing machine -- with a function being physically computable if and only if it is Turing-computable. I show how this apparently mild assumption leads to generous yet binding limits on how quickly or slowly any directly measurable physical phenomenon can grow or converge -- limits that are intimately connected to Rado's Busy Beaver function. I conjecture that these limits are novel physical laws governing which rates of growth and convergence are possible in our universe.

physics.hist-ph↗

arXiv2024

Interpolation, Extrapolation, Hyperpolation: Generalising into new dimensions

This paper introduces the concept of hyperpolation: a way of generalising from a limited set of data points that is a peer to the more familiar concepts of interpolation and extrapolation. Hyperpolation is the task of estimating the value of a function at new locations that lie outside the subspace (or manifold) of the existing data. We shall see that hyperpolation is possible and explore its links to creativity in the arts and sciences. We will also examine the role of hyperpolation in machine learning and suggest that the lack of fundamental creativity in current AI systems is deeply connected to their limited ability to hyperpolate.

cs.LG↗

arXiv2023

The Lindy Effect

The Lindy effect is a statistical tendency for things with longer pasts behind them to have longer futures ahead. It has been experimentally confirmed to apply to some categories, but not others, raising questions about when it is applicable and why. I shed some light on these questions by examining the mathematical properties required for the effect and generating mechanisms that can produce them. While the Lindy effect is often thought to require a declining hazard rate, I show that it arises very naturally even in cases with constant (or increasing) hazard rates -- so long as there is a probability distribution over the size of that rate. One implication is that even things which are becoming less robust over time can display the Lindy effect.

physics.soc-ph↗

arXiv2021

Implications of a search for intergalactic civilizations on prior estimates of human survival and travel speed

We present a model where some proportion of extraterrestrial civilizations expand uniformly over time to reach a cosmological scale. We then ask what humanity could infer if a sky survey were to find zero, one, or more such civilizations. We show how the results of this survey, in combination with an approach to anthropics called the Self Indication Assumption (SIA), would shift any prior estimates of two quantities: 1) The chance that a technological civilization like ours survives to embark on such expansion, and 2) the maximum feasible speed at which it could expand. The SIA gives pessimistic estimates for both, but survey results (even null results) can reverse some of the effect. The SIA-induced shift in expectations is strong enough to be regarded as a falsifiable technological prediction, and a test of the SIA in cosmological reasoning.

astro-ph.CO↗

arXiv2021

The Edges of Our Universe

This paper explores the fundamental causal limits on how much of the universe we can observe or affect. It distinguishes four principal regions: the affectable universe, the observable universe, the eventually observable universe, and the ultimately observable universe. It then shows how these (and other) causal limits set physical bounds on what spacefaring civilisations could achieve over the longterm future.

gr-qc↗

arXiv2018

Dissolving the Fermi Paradox

The Fermi paradox is the conflict between an expectation of a high {\em ex ante} probability of intelligent life elsewhere in the universe and the apparently lifeless universe we in fact observe. The expectation that the universe should be teeming with intelligent life is linked to models like the Drake equation, which suggest that even if the probability of intelligent life developing at a given site is small, the sheer multitude of possible sites should nonetheless yield a large number of potentially observable civilizations. We show that this conflict arises from the use of Drake-like equations, which implicitly assume certainty regarding highly uncertain parameters. We examine these parameters, incorporating models of chemical and genetic transitions on paths to the origin of life, and show that extant scientific knowledge corresponds to uncertainties that span multiple orders of magnitude. This makes a stark difference. When the model is recast to represent realistic distributions of uncertainty, we find a substantial {\em ex ante} probability of there being no other intelligent life in our observable universe, and thus that there should be little surprise when we fail to detect any signs of it. This result dissolves the Fermi paradox, and in doing so removes any need to invoke speculative mechanisms by which civilizations would inevitably fail to have observable effects upon the universe.

physics.pop-ph↗

arXiv2017

Symmetric Decomposition of Asymmetric Games

We introduce new theoretical insights into two-population asymmetric games allowing for an elegant symmetric decomposition into two single population symmetric games. Specifically, we show how an asymmetric bimatrix game (A,B) can be decomposed into its symmetric counterparts by envisioning and investigating the payoff tables (A and B) that constitute the asymmetric game, as two independent, single population, symmetric games. We reveal several surprising formal relationships between an asymmetric two-population game and its symmetric single population counterparts, which facilitate a convenient analysis of the original asymmetric game due to the dimensionality reduction of the decomposition. The main finding reveals that if (x,y) is a Nash equilibrium of an asymmetric game (A,B), this implies that y is a Nash equilibrium of the symmetric counterpart game determined by payoff table A, and x is a Nash equilibrium of the symmetric counterpart game determined by payoff table B. Also the reverse holds and combinations of Nash equilibria of the counterpart games form Nash equilibria of the asymmetric game. We illustrate how these formal relationships aid in identifying and analysing the Nash structure of asymmetric games, by examining the evolutionary dynamics of the simpler counterpart games in several canonical examples.

cs.GT↗

arXiv2008

Probing the Improbable: Methodological Challenges for Risks with Low Probabilities and High Stakes

Some risks have extremely high stakes. For example, a worldwide pandemic or asteroid impact could potentially kill more than a billion people. Comfortingly, scientific calculations often put very low probabilities on the occurrence of such catastrophes. In this paper, we argue that there are important new methodological problems which arise when assessing global catastrophic risks and we focus on a problem regarding probability estimation. When an expert provides a calculation of the probability of an outcome, they are really providing the probability of the outcome occurring, given that their argument is watertight. However, their argument may fail for a number of reasons such as a flaw in the underlying theory, a flaw in the modeling of the problem, or a mistake in the calculations. If the probability estimate given by an argument is dwarfed by the chance that the argument itself is flawed, then the estimate is suspect. We develop this idea formally, explaining how it differs from the related distinctions of model and parameter uncertainty. Using the risk estimates from the Large Hadron Collider as a test case, we show how serious the problem can be when it comes to catastrophic risks and how best to address it.

physics.soc-ph↗

arXiv2004

Using biased coins as oracles

While it is well known that a Turing machine equipped with the ability to flip a fair coin cannot compute more that a standard Turing machine, we show that this is not true for a biased coin. Indeed, any oracle set $X$ may be coded as a probability $p_{X}$ such that if a Turing machine is given a coin which lands heads with probability $p_{X}$ it can compute any function recursive in $X$ with arbitrarily high probability. We also show how the assumption of a non-recursive bias can be weakened by using a sequence of increasingly accurate recursive biases or by choosing the bias at random from a distribution with a non-recursive mean. We conclude by briefly mentioning some implications regarding the physical realisability of such methods.

cs.OH↗

arXiv2003

On the existence of a new family of Diophantine equations for $\bf Ω$

We show how to determine the $k$-th bit of Chaitin's algorithmically random real number $Ω$ by solving $k$ instances of the halting problem. From this we then reduce the problem of determining the $k$-th bit of $Ω$ to determining whether a certain Diophantine equation with two parameters, $k$ and $N$, has solutions for an odd or an even number of values of $N$. We also demonstrate two further examples of $Ω$ in number theory: an exponential Diophantine equation with a parameter $k$ which has an odd number of solutions iff the $k$-th bit of $Ω$ is 1, and a polynomial of positive integer variables and a parameter $k$ that takes on an odd number of positive values iff the $k$-th bit of $Ω$ is 1.

math.NT↗

arXiv2003

The Diagonal Method and Hypercomputation

The diagonal method is often used to show that Turing machines cannot solve their own halting problem. There have been several recent attempts to show that this method also exposes either contradiction or arbitrariness in other theoretical models of computation that claim to be able to solve the halting problem for Turing machines. We show that such arguments are flawed -- a contradiction only occurs if a type of machine can compute its own diagonal function. We then demonstrate why such a situation does not occur for the methods of hypercomputation under attack and why it is unlikely to occur in any other serious methods.

math.LO↗

arXiv2003

Representations of Omega in Number Theory: Finitude Versus Parity

We present a new method for expressing Chaitin's random real, Omega, through Diophantine equations. Where Chaitin's method causes a particular quantity to express the bits of Omega by fluctuating between finite and infinite values, in our method this quantity is always finite and the bits of Omega are expressed in its fluctuations between odd and even values, allowing for some interesting developments. We then use exponential Diophantine equations to simplify this result and finally show how both methods can also be used to create polynomials which express the bits of Omega in the number of positive values they assume.

math.NT↗

arXiv2002

Hypercomputation: computing more than the Turing machine

Due to common misconceptions about the Church-Turing thesis, it has been widely assumed that the Turing machine provides an upper bound on what is computable. This is not so. The new field of hypercomputation studies models of computation that can compute more than the Turing machine and addresses their implications. In this report, I survey much of the work that has been done on hypercomputation, explaining how such non-classical models fit into the classical theory of computation and comparing their relative powers. I also examine the physical requirements for such machines to be constructible and the kinds of hypercomputation that may be possible within the universe. Finally, I show how the possibility of hypercomputation weakens the impact of Godel's Incompleteness Theorem and Chaitin's discovery of 'randomness' within arithmetic.

math.LO↗