Accelerating change

Accelerating change is the claim that technological capability does not merely improve but improves at an increasing rate, so that intervals between comparable advances shorten over time. In its strong form it treats acceleration as a property of technological evolution as such, extrapolable across domains and forward in time. That strong form is the engine of most long-horizon forecasting in this wiki's subject matter, and it is also the weakest link in most of it.

§The claim

Three distinct assertions travel under this heading, and conflating them causes most of the confusion.

The descriptive claim is that particular technical metrics have followed exponential curves over decades. This is true of several and false of many.

The general claim is that acceleration is a feature of technology as a whole, because each advance provides tools for the next. Ray Kurzweil's law of accelerating returns is the best-known version: evolutionary processes build on their own products, so the returns — speed, capability, cost-effectiveness — grow exponentially, and the exponent itself grows.¹

The predictive claim is that these curves can be extrapolated to date future capabilities, including Artificial general intelligence, Whole brain emulation or Longevity escape velocity. This is the contested step, and the descriptive claim does not entail it.

§Origins

Henry Adams noticed the pattern early in the twentieth century, plotting coal output and later energy use and concluding in The Education of Henry Adams that thought itself was accelerating on a trajectory that would carry the coming century somewhere unrecognisable. Similar observations recur through the mid-century in the work of Buckminster Fuller and in Stanislaw Ulam's report of a conversation with John von Neumann about an approaching "essential singularity" in the history of the race beyond which human affairs could not continue as known.

The modern version dates to Gordon Moore's 1965 observation that the number of components on an integrated circuit at minimum cost per component had doubled annually and would continue to do so, revised in 1975 to roughly two years.² Moore was describing a manufacturing economics trend over a decade of data; it became a coordinating expectation for the semiconductor industry, and then a metaphor for technological progress in general — a transfer he did not endorse.

§The evidence usually cited

Several curves genuinely have been exponential for long stretches.

Transistor density followed Moore's law for roughly half a century. Computing energy efficiency improved on a similar cadence for decades, a relationship documented by Jonathan Koomey and colleagues, with the doubling time lengthening after around 2000.³ Training compute for frontier machine-learning systems doubled roughly every six months through the 2010s, far faster than hardware improvement alone, because spending grew as well.

The steepest curve in biology is DNA sequencing cost. From 2007 the introduction of massively parallel sequencing pushed cost per genome down faster than semiconductor trends for several years, before flattening into a slower decline — a pattern that makes it an instructive case rather than a clean confirmation. Where cheap sequencing has mattered most on this wiki is downstream: Polygenic embryo screening, epigenetic clocks and Connectomics all depend on data volumes that were unaffordable two decades ago.

§Moore's law and what replaced it

Dennard scaling — the property that shrinking transistors also reduced their power density, so clock speeds could rise for free — broke down in the mid-2000s. Clock rates stopped climbing, and performance gains shifted to parallelism: multiple cores, then graphics processors, then domain-specific accelerators. Feature-size shrinks have continued but more slowly and at much higher capital cost, with extreme-ultraviolet lithography and three-dimensional packaging substituting for straightforward miniaturisation. Several analyses argue that cost per transistor stopped falling around the early 2010s, which if correct means the economically relevant version of Moore's law ended over a decade before the physical one.

The important structural point is that the exponential was sustained by increasing effort. Bloom, Jones, Van Reenen and Webb found that maintaining the constant doubling rate of Moore's law required a research workforce many times larger by the 2010s than in the early 1970s, and that research productivity has fallen at roughly five per cent a year across the US industries they examined.⁴ An exponential output produced by a super-exponential input is not evidence for a law of accelerating returns. It is evidence of an expensive and possibly unsustainable subsidy.

§Where the trend fails

The clearest counterexample is drug development. Scannell and colleagues documented that the number of new drugs approved per billion dollars of inflation-adjusted R&D spending halved roughly every nine years from 1950, a decline they named "Eroom's law" — Moore's law backwards.⁵ Half a century of molecular biology, structural methods and automation improved every input to the process while the output per dollar fell, for reasons the authors attribute mainly to the rising bar set by existing treatments and to regulatory risk aversion rather than to scientific difficulty alone. Whether machine learning bends that curve back is the claim assessed under AI drug discovery, and no drug originating from those methods has yet been approved.

Other domains show flat or reversed trends. Passenger aircraft cruise speeds peaked in the 1960s and fell after supersonic service ended. Energy consumption per capita in developed economies stopped rising decades ago. Measured "disruptiveness" of papers and patents, on one widely discussed index, declined across all major fields between the mid-twentieth century and the 2010s.⁶

Within this wiki's own subject matter the record is mixed and mostly slow. Tissue engineering promised engineered organs within a generation of its 1990s founding and has delivered a handful of simple structures. Medical nanorobots as described in the 1980s remain design studies. Gene therapy took roughly three decades from first clinical trial to durable approved products. Against that, CRISPR–Cas9 moved from a 2012 mechanism paper to an approved therapy in about eleven years, which is fast by any historical standard, and AI protein design brought single-chain structure prediction to near-experimental accuracy within roughly a decade of deep learning being applied to it.

Selection effects in the evidence

Exponential curves are easy to find retrospectively because the technologies that failed to accelerate are not the ones anyone plots. Any argument from a set of curves must state how the curves were chosen, and most popular presentations do not.

§The S-curve counterargument

The standard alternative model is that individual technologies follow logistic curves: slow start, rapid middle, saturation as physical or economic limits bind. Substitution between competing technologies has been modelled this way since Fisher and Pry's 1971 work, and the fit is generally good.

On this view an apparent long-run exponential is a chain of overlapping S-curves — vacuum tubes to transistors to integrated circuits — and the appearance of a smooth law is an artefact of aggregation. The distinction matters because the two models make different predictions at exactly the point where forecasting is most valuable. An exponential model says the next doubling is like the last. An S-curve model says a saturating technology's replacement may or may not arrive, and that the historical record contains cases where it did not.

Theodore Modis's critique of the strong version argues that the chosen milestones in canonical accelerating-change datasets are selected and weighted subjectively, and that the underlying pattern is better described by a single large logistic than by a runaway exponential.⁷

§Forecasting base rates

The final objection is empirical and unflattering. Philip Tetlock's long-running studies of expert political and geopolitical forecasting found that credentialed experts performed poorly against simple extrapolation, that confidence was inversely related to accuracy, and that the most accurate forecasters were those who updated frequently and avoided single grand theories.⁸

Technology-specific reviews find a characteristic distortion. Surveys of published predictions about machine intelligence over six decades found that they cluster fifteen to twenty-five years ahead of whenever the prediction was made, by experts and non-experts alike, which is the signature of a forecast anchored on career horizons rather than on evidence.⁹

This is why the wiki's articles on Technological singularity, Singularitarianism and Effective accelerationism treat timelines separately from mechanisms. The mechanism arguments can be assessed on their merits; the dates attached to them have a track record, and it is bad.

§Outlook

The defensible position is domain-specific and cheap to state: some metrics have followed exponential curves for decades, the curves are sustained by rising inputs, they saturate, and cross-domain extrapolation has no established basis. What that leaves unresolved is the question the strong claim was invented to answer. If capability in one domain — machine learning is the current candidate — begins to compound the productivity of research itself, the pattern of independent saturating curves would no longer hold, and the argument for deliberately sequencing which capabilities arrive first becomes urgent rather than theoretical. Whether that is happening is an empirical question about research productivity, and it will be answered by measurement rather than by extrapolation.

§See also

• Technological singularity
• Ray Kurzweil
• Artificial general intelligence
• Technology readiness level
• Singularitarianism
• Longevity escape velocity
• Future of humanity
• Effective accelerationism

§References