There is something wonderfully intuitive about the idea of an adaptive landscape.

Imagine evolution as a journey across terrain, where low ground represents poor fitness, higher ground represents greater fitness, and populations move through the landscape as mutation and selection carry them towards the peaks.

The idea reaches back to Fisher and Wright in the early 1930s, and it has survived because it turns an almost impossibly multidimensional process into something we can picture. Yet there is also a danger in a good metaphor, because eventually we can become so comfortable with the picture that we forget it is a picture at all.

In Adaptive Landscapes in the Age of Synthetic Biology, Xiao Yi and Antony Dean argue that modern biology allows us to go considerably further. Rather than merely describing peaks, valleys, trade-offs and evolutionary paths, we can increasingly investigate the biological mechanisms that create the landscape in the first place, so their concern is not simply where evolution goes, but why the terrain has the shape that it does.¹

One of their examples is particularly intriguing because evolution appears to reach the top of the mountain, stays there for a while, and then does something unexpected.

A choice between growing and going

The experiment involved E. coli living under an unusual selection regime. Twice each day, bacteria capable of swimming into a capillary containing chemoattractants were used to inoculate fresh growth medium, which meant that success depended upon two desirable characteristics: the bacteria needed to grow quickly, but they also needed to move effectively towards the next supply of nutrients.

Unfortunately, both are expensive.

The same limited carbon resource has to support growth and chemotaxis, so there is a trade-off: invest more heavily in motility and there is less available for growth, while investing more heavily in growth leaves the organism less effective at reaching the next environment. The researchers could therefore construct a landscape whose important dimensions included growth and motility, and the bacteria began climbing it.

Five independent populations adapted rapidly, reaching the experimentally established trade-off frontier and then moving along it towards the adaptive peak, where they remained for several weeks.²

It looked remarkably like the metaphor made real. Evolution had climbed the mountain, and there seemed to be nowhere higher to go.

Except it wasn't finished.

Escaping the peak

Several populations subsequently crossed what had appeared to be the frontier, retaining effective chemotaxis while also improving their growth rate.

At first glance that seems impossible, because the energetic trade-off hadn't magically disappeared and the bacteria hadn't discovered some previously unknown source of energy. Instead, evolution found something more interesting: the bacteria changed when they swam.

Compared with their ancestor, the evolved bacteria reduced their motility during exponential growth, when devoting resources to reproduction was particularly valuable, but then increased motility as the population approached carrying capacity. Rather than adopting a permanent compromise between two competing demands, the organism began matching its behaviour to different phases of its environment.

Grow now, and move later.

Much of this remarkable behavioural change could be traced to a single amino-acid substitution, Arg220Trp, in the FliA transcription factor, which altered expression within an existing regulatory network and changed the timing of motility. Yi and Dean describe the result as a new physiological programme in which different phenotypes became matched to different phases of the cyclical environment.²

And that is much more interesting than merely climbing a little higher.

If our simplified optimisation problem had been fitness = f(growth, motility), then evolution had effectively revealed that the more useful description was fitness = f(growth over time, motility over time).

Time had always existed, of course, and evolution didn't invent it; what changed was the way the organism exploited it. A problem that appeared to demand a static compromise acquired another usable degree of freedom.

There is an echo here in artificial intelligence

This becomes particularly interesting when we look at another class of systems currently being pushed through enormous optimisation spaces.

Modern GPTs are trained very differently from biological evolution, and we should not confuse the two. A large neural network is typically trained using gradient-based optimisation, which calculates how changes to its parameters affect an objective and then repeatedly adjusts those parameters in a direction expected to improve performance. Evolution has no equivalent of backpropagation, but both systems nevertheless invite us to think in terms of landscapes.

A GPT contains an enormous number of adjustable parameters, and training searches for configurations that perform increasingly well against an objective. Gradient descent is exceptionally good at navigating such spaces because it has information about the local slope; in crude terms, it can repeatedly ask: from where I am standing, which way is uphill?

And this has worked spectacularly well. Larger models, more data, more compute and increasingly sophisticated training have carried neural networks astonishingly far up that particular mountain.

Yet something interesting is happening around the model, because the model is increasingly no longer the whole machine.

The intelligence around the model

Retrieval can provide information that does not reside in a model's parameters, while memory can preserve useful information beyond an individual interaction. Tools can allow the model to calculate, search, execute code or interact with external systems, and additional inference-time computation can be spent selectively on difficult problems while simple ones consume very little.

More importantly, these capabilities can be orchestrated, and once that happens the interesting question stops being simply how capable is the model? and becomes which capability should be used, under what circumstances, and when?

That distinction feels surprisingly familiar.

The E. coli did not solve its problem by becoming permanently better at everything, because the breakthrough came through regulation. Different behaviours became advantageous at different stages of the environmental cycle, and evolution found a mechanism capable of expressing them accordingly.

Something analogous is appearing around GPTs. A difficult question might justify additional computation, while another might require retrieval; a numerical problem might be delegated to a calculator or code interpreter, while a factual claim might benefit from external verification. Some information may be worth remembering, whereas most is not, and some tasks may be better decomposed and handed between specialised components rather than attacked by one model in one pass.

The optimisation problem therefore begins to change. It is no longer solely about producing a more capable static object, because it is increasingly also about orchestrating capabilities through context and time.

Optimising the climber, or changing the climb

The distinction becomes clearer if we compare neural optimisation with evolutionary computation.

Gradient descent is extraordinarily efficient when we have a differentiable objective and an architecture whose parameters can be adjusted, because it can optimise billions of parameters with a precision and speed that biological evolution could never approach. Evolutionary algorithms operate differently: they maintain populations of candidate solutions, introduce variation and preferentially retain successful variants, and because they do not require a gradient they can explore things that are awkward to express as smooth parameter adjustments, including structures, rules, architectures and combinations of components.

Nor do the two approaches have to compete. Neural networks themselves can be evolved, while evolutionary processes can search architectures or strategies that are subsequently refined through gradient optimisation.

This suggests several layers of optimisation rather than a choice between two methods. We can optimise parameters within a system, but we can also optimise the structure of the system, and beyond both lies the possibility of optimising how that system behaves under different circumstances.

The last of those brings us straight back to the bacteria, because the important adaptation wasn't simply more of a useful characteristic. It was a change in the relationship between behaviour and circumstance.

The landscape was real, but it wasn't the whole story

There is an important subtlety here, because the growth-motility trade-off was not an experimental mistake, nor did the evolved bacteria somehow invalidate the adaptive landscape. The trade-off was real, but the mechanism through which the organism encountered it could change.

By evolving a time-dependent physiological programme, the bacteria mitigated the consequences of a hard-wired energetic constraint, and a single mutation at an important point in an existing regulatory network changed the behaviour of the whole system.²

That brings us directly back to Yi and Dean's larger argument, because they want us to look beneath descriptive phenomena such as epistasis, dominance, trade-offs and adaptive peaks and ask what biological mechanisms generate them. Their seven examples trace causal chains from genotype through phenotype to fitness, and they argue that doing so transforms the adaptive landscape from a useful metaphor into a framework within which causal hypotheses can actually be tested.¹

The mountain tells us something, but understanding what made the mountain tells us considerably more, because once we understand the mechanisms that generate a landscape we can begin to understand not only where its peaks lie, but why they exist and under what conditions they might move.

Another dimension

There is an understandable tendency in optimisation to treat a stubborn frontier as evidence that we are approaching a fundamental limit, and sometimes we are. Physics remains physics, energy remains finite, computation has costs, and every real system eventually encounters constraints that cannot be wished away.

Yet sometimes the apparent limit is partly a consequence of the space in which we have chosen to describe the problem.

Growth or motility becomes growth and then motility. A model's internal knowledge can be supplemented by retrieval, while a fixed computational budget can become computation allocated according to need, and a single model can become one component in a system of tools, memory and external information.

None of those removes the underlying constraints, but each changes how the system encounters them.

And this is why the comparison with GPTs is useful without pretending that bacteria are computers or that neural networks evolve like organisms. The mechanisms could hardly be more different, yet the conceptual problem is shared: when we represent a complex system as a landscape, the dimensions we choose determine which possibilities are visible.

An adaptive landscape is therefore not reality itself, but a representation of possibilities, and representations necessarily leave things out. Yi and Dean's work reminds us that when we understand the mechanisms beneath the landscape, we may discover possibilities that the landscape alone did not make obvious.

So when a complex system appears to have reached its peak, perhaps the most interesting question is not simply whether it can climb any higher, but whether we have drawn the whole landscape.

Sometimes evolution does not find a higher peak; it finds another dimension.

Notes

1. Xiao Yi and Antony M. Dean, "Adaptive Landscapes in the Age of Synthetic Biology", Molecular Biology and Evolution, 36(5), 2019, pp. 890–907. The authors argue for redefining adaptive landscapes in terms of the biological mechanisms that generate properties such as epistasis, dominance, trade-offs and adaptive peaks. ↩
2. Yi and Dean's discussion of the chemotaxis-growth landscape draws on their earlier experimental work and describes populations of E. coli escaping an apparent adaptive peak by evolving a time-dependent physiological programme. The Arg220Trp substitution in the FliA transcription factor was largely responsible for the altered behaviour. See Yi and Dean (2019), particularly the section "Chemotaxis–Growth: New Behavior Enables Escape from an Adaptive Peak" and Figure 8. ↩