In this article
Learning something well takes time: integration, error, repetition, contrast, rest. The psychology of memory —Hermann Ebbinghaus in Über das Gedächtnis (1885), Robert and Elizabeth Bjork in the twentieth century— has spent 140 years documenting the pattern. Today's speed grants no such time. What gets "learned" today is half-obsolete tomorrow, so it doesn't settle —it floats. The consequence: a population that knows a lot of things by halves and almost nothing in depth. And depth was exactly what set the expert apart from the one who repeats.
The Ebbinghaus curve
Hermann Ebbinghaus, a German physiologist, ran a series of pioneering experiments in the 1880s with himself as the only subject. He learned lists of nonsense syllables —bjk, lor, gef— then measured how much he remembered at growing intervals. The results, published in Über das Gedächtnis (1885), traced for the first time a forgetting curve: freshly learned material is forgotten quickly in the first hours and days —hence the popular figure that gets repeated, a loss of between 50% and 80% in the first 24 hours, which is a later simplification and not a literal datum from Ebbinghaus, whose experiments were done on nonsense syllables and with a single subject— and then more slowly, until it stabilizes at a low level. Ebbinghaus also identified the solution: spaced repetition. Each time the material is reviewed before being completely forgotten, the next retention interval lengthens. The curve, repeated several times, flattens.
A hundred and forty years later, this line is still the basic frame of the psychology of memory. Hellyer (1962), Roediger and Karpicke (2006), Bjork and Bjork (2011) have added refinements —the role of the testing effect, the desirable difficulties, the effects of reactivation— but Ebbinghaus's central structure holds. Without spaced repetition, without rest between exposures, without contrast between situations of use, the material doesn't consolidate.
What real learning demands
Look at the components the body of literature has identified as conditions of deep learning. The first is the time between exposures: consolidation of long-term memory requires rest phases between practice, and sleep in particular plays a central role —Diekelmann and Born (2010) synthesized robust evidence on consolidation during sleep. The second is error and correction; Bjork described desirable difficulties, which slow immediate performance but consolidate long-term retention, because learning without failing produces an illusion of learning, not learning. The third is contrast of application: applying the same knowledge in different situations strengthens the representation, and the ability to transfer —which is what distinguishes operative knowledge from textbook knowledge— requires practical variability. The fourth is cognitive rest; after intense exposure you have to stop, because consolidative processing is subliminal and operates over intervals that aren't those of active study.
These four components imply one common thing: time. Not brief, discontinuous time; broad, distributed time, with active and passive phases, with repetitions spaced across scales running from hours to months.
The speed of obsolescence
Fritz Machlup, in The Production and Distribution of Knowledge in the United States (Princeton UP, 1962), introduced the concept of half-life of knowledge. The idea: in any field, a significant proportion of what's considered current today will be obsolete within a defined period. Originally the span was measured in decades; it hasn't stopped shrinking since.
The estimates have come down systematically. Peter Densen, in Challenges and Opportunities Facing Medical Education (Transactions of the American Clinical and Climatological Association 122, 2011), worked with a related but different magnitude, the doubling time of medical knowledge —the span in which the total volume of what's known doubles—: he estimated it had gone from about 50 years in 1950 to 7 years in 1980, to 3.5 years in 2010, with a projection of 73 days for 2020. A field whose knowledge doubles every few years is, perforce, a field in which what's learned ages fast. In engineering, a 1991 press article citing IEEE estimates put the half-life of skills below 5 years; for software, below 3. Anyone who's worked in tech over the last decade will recognize the orders of magnitude.
For generative AI the figure is absurdly lower. What you learned about prompt engineering two years ago no longer applies to today's model. What you learned about the limits of image generation a year and a half ago is invalidated by the next version. The workflow a user sets up in July is obsolete by September because of an API change.
The clash of the two clocks
Here's the nerve. Consolidating knowledge takes months or years. The obsolescence of knowledge, in many fields, happens in months. The temporal space between acquisition and obsolescence has become smaller than the temporal space between acquisition and consolidation. Result: knowledge becomes obsolete before it settles.
This isn't a metaphor. The person trying to master Stable Diffusion in 2022 invests three months, reaches operative mastery in the fourth, and by then the field has moved to Flux and another architecture. The prior cognitive effort doesn't consolidate, because the material it was supposed to consolidate on no longer exists as an operative referent. The person is left with an intermediate knowledge that was right for the previous version and isn't operative for the current one.
Maryanne Wolf, in Reader, Come Home (HarperCollins, 2018), described a parallel pattern in reading. The deep-reading brain, formed with long texts and the demand for sustained concentration, atrophies when habitual practice is fast, fragmented reading. It isn't individual laziness; it's that the conditions of practice have changed faster than the neural capacities that were built under the previous conditions.
Knowledge that floats
The aggregate consequence is a population that knows a lot of things by halves and few in depth. An average person in 2026 handles technical terms from several fields —transformer, embedding, RAG, prompt engineering, attention— without having consolidated the concepts to an operative level. They recognize names of art movements, authors and books without having read any of the texts directly. They cite scientific studies by ear —"I read that a study says…"— whose title and authors they couldn't reproduce. They operate the chatbot as a tool without understanding any of the underlying mechanisms or the constraints on its output.
This isn't simple ignorance. It's a specific form of semi-knowledge that produces recognition without recall, fluency without depth, the sensation of understanding without effective understanding. Daniel Kahneman would describe it as the dominance of System 1: the answer arrives effortlessly, without System 2 verifying.
The functional disappearance of the expert
There's a social figure this dynamic displaces, and it deserves a close look. The classic expert —twenty years in their field, deep operative mastery, capacity for grounded judgment, criterion formed through repetition— doesn't hold in an ecosystem where the field changes faster than the time it takes to form an expert. What's needed, functionally, isn't an expert but an adaptable operator: someone who can absorb fast the latest version of the field, operate competently with it, and unlearn it when the next one arrives.
This is personal opinion: that change isn't ethically neutral. The classic expert performed functions the adaptable operator doesn't. They kept the field's collective memory —past errors, dead ends, solutions that were tried and discarded. They had the criterion to tell the important from the noisy, because they'd lived long enough to see fashions fall. And, above all, they were a reliable reference point for society —if a journalist needed to quote someone serious, they knew where to find them.
The adaptable operator performs the immediate operative functions better than the classic expert. What they don't perform is the function of long memory. And long memory is exactly what's needed for a society to learn from its errors instead of repeating them.
The celebration of speed
The public discourse on adapting to technological change tends to celebrate speed. Lifelong learning, learn to learn, be agile, unlearn and relearn. Consultancies paint the adaptable one as the model worker of the twenty-first century. The education industry pivots toward micro-certifications of a few months, moving away from the long-degree model.
What that discourse leaves out: no deep field has ever been built at micro-certification speed. Philosophy, medicine, engineering, the liberal arts, basic mathematics —none is mastered in six weeks with an online module. The micro-certification works to add an operative layer onto a prior base; it doesn't build the base.
If the prior base stops being built —because there are no longer economic incentives to invest five or ten years in it— society ends up with many operators and little base. That's what the phrase "floating knowledge" means.
What erodes in silence
Some things are disappearing without being named well. The ability to read a long text whole, without interruption, holding attention to the end, is built with sustained practice over years, and that practice has declined within a generation, according to Wolf and others. The ability to retain dates, names and historical contexts without constantly resorting to external verification erodes because the memory of facts has been externalized into search engines and, now, chatbots: what isn't attempted to be remembered doesn't consolidate. The ability to follow a complex argument across several hours or days, holding the thread without pasting the conversation into the model for a summary, requires a specific cognitive musculature that atrophies with the use of automatic summaries. And the ability to sustain a deep field of interest over decades, enduring the phases of apparent stagnation any field has, requires a stable identity around the field, and that stability is ever rarer when the environment rewards rotation.
Each of these abilities, alone, isn't decisive. The four together, lost at population scale, define a different kind of collective intelligence.
Definitions
Forgetting curve (Ebbinghaus). Graph describing the loss of memory of freshly learned material as a function of time, with a fast initial drop that gradually tapers.
Spaced repetition. Strategy of reviewing material at growing intervals, based on the principle that consolidation requires reactivation before total forgetting.
Half-life of knowledge (Machlup). The time needed for half the knowledge current in a field to become obsolete.
Floating knowledge. State in which a person handles a field's terms and references without having consolidated the underlying representations, producing recognition without recall and fluency without depth.
Classic expert versus adaptable operator. Contrast between the figure formed in a field over decades, with criterion and long memory, and the figure competent at absorbing and unlearning successive versions of a fast-evolving field.
References
Ebbinghaus, H. Über das Gedächtnis: Untersuchungen zur experimentellen Psychologie. Duncker & Humblot, 1885. Origin of the forgetting curve and of spaced repetition; the popular figure of a 50-80% loss in 24 hours is a later simplification, not a literal datum from this work.
Machlup, F. The Production and Distribution of Knowledge in the United States. Princeton University Press, 1962. Introduces the concept of the half-life of knowledge.
Roediger, H. L.; Karpicke, J. D. «Test-Enhanced Learning». Psychological Science 17(3), 2006. On the effect of testing on retention.
Diekelmann, S.; Born, J. «The memory function of sleep». Nature Reviews Neuroscience 11, 2010. Synthesis of the evidence on memory consolidation during sleep.
Densen, P. «Challenges and Opportunities Facing Medical Education». Transactions of the American Clinical and Climatological Association 122, 2011. https://pmc.ncbi.nlm.nih.gov/articles/PMC3116346/ — source of the medical-knowledge doubling-time figures (50 years in 1950, 7 in 1980, 3.5 in 2010, projection of 73 days in 2020).
«Engineer Supply Affects America». The New York Times, 1991 —origin, citing IEEE estimates, of the figures for the half-life of engineering skills (under 5 years) and software (under 3 years). Referred to in this article via IEEE Spectrum: https://spectrum.ieee.org/an-engineering-career-only-a-young-persons-game
Bjork, R. A.; Bjork, E. L. «Making things hard on yourself, but in a good way: Creating desirable difficulties to enhance learning». In Psychology and the Real World, Worth Publishers, 2011. On desirable difficulties.
Wolf, M. Reader, Come Home. HarperCollins, 2018. On the atrophy of the deep-reading brain under conditions of fast, fragmented reading.
Brown, P. C.; Roediger, H. L.; McDaniel, M. A. Make It Stick. Belknap Press, 2014. Accessible synthesis of the literature on deep learning based on four decades of cognitive research.
Sennett, R. The Craftsman. Yale University Press, 2008. On the slow formation of craft and the value of expertise consolidated over decades.
You may also like
- Learning without remembering
- The illusion of learning
- The end of the classic expert
- Learning to unlearn
Comments0
No comments yet.
Leave a comment