Popular science, told with curiosity

Cover Story Continual Learning

Why does an AI forget what it has learned?

Teaching a neural network something new makes it erase what came before. This study shows that forgetting depends not only on the algorithm but on the order in which the tasks are shown.

Selim Kusursuz, Nadjib Atamnia and Omar Gharib 12 September 2026 6 min read 6 references

Agent-Based Modelling 7 min

The importance of being discrete: why life always wins on a surface

An equation that looks at averages says a population will die out where catalysts are scarce. A simulation built from individual agents shows the opposite: the population survives in islands that follow the catalysts.

History of Science 4 min

The unfading light of Alexandria: Hypatia and the grace of reason

In Alexandria, where geometry, astronomy and Neoplatonism met, a mind refused to surrender to the rote truths of its age: Hypatia. The trace she left is a reminder that thinking is not a privilege but a debt.

What shall we learn today?

Published by our editors in the last 48 hours, all grounded in peer-reviewed sources.

All articles

The Unconscious 7 min

The unconscious today: what research made of Freud’s biggest idea

Freud held that most of mental life happens below awareness. Experimental psychology and neuroscience now measure it — confirming part of his intuition and correcting part of it.

Developmental Psychology 6 min

Attachment theory: why our first relationships mark a whole life

Bowlby’s theory held that a child needs a reliable emotional base. Large longitudinal studies and meta-analyses now show how far that reaches into adulthood.

Defence Mechanisms 8 min

Why we get in our own way: defence mechanisms today

Defence mechanisms protect us from anxiety and hold us in place at the same time. Once dismissed as speculative, they can now be measured.

Cognitive Psychology 7 min

Cognitive biases: why our brain misleads us systematically

Our thinking runs on shortcuts that make fast decisions possible. The same shortcuts also produce predictable errors: cognitive biases.

Limits of Computation 9 min

Could a superintelligence never be wrong?

How realistic is a machine that knows everything and never errs? Through the limits set by Turing, Gödel, Landauer and Bekenstein, we look at what artificial superintelligence could achieve and what it never will.

Software-Defined Networking 8 min

Where does a network keep its brain?

In a software-defined network the switches make no decisions; they only look up a table. Using virtual networks built inside a single computer, we look at where decisions are made, why the first packet is slow, and how a network breaks.

Questions Answered

Reader questions, answered by our subject editors with sources cited.

See all

Question

Does an AI really forget what it has learned?

Yes — and more sharply than people expect. A network that learns a new task can lose the old one almost entirely, not partially. This is called catastrophic forgetting. The reason is simple: everything a network knows lives in the same set of weights, and a new task pulls those weights toward its own purposes.

1 answers Sourced Computer Science

Question

Does the order of tasks affect how an AI learns?

It does, and markedly. Show the same tasks in a different order and the final error rate changes: the heavier the sequence overall, and the sharper the gaps between consecutive tasks, the more the model forgets. In a continually learning system the curriculum matters as much as the algorithm.

1 answers Sourced Computer Science

Question

Does quarantine make an epidemic shorter?

Usually not — it makes it longer. In a simulation of a thousand people, quarantine covering seventy percent of them left the total case count unchanged and stretched the epidemic from 49 days to 58. What changed was the peak: from around fifty percent down to under twenty-five. The aim is not to end the epidemic but to keep the number of people ill at once within what hospitals can carry.

1 answers Sourced Computer Science

Question

In vaccination, does coverage matter more than speed?

Both matter, but speed is usually underrated. Reaching the same seventy-five percent coverage twice as fast dropped the number who fell ill from 144 to 35 in a population of a thousand. A broad campaign that starts late can do worse than a narrow one that starts early.

1 answers Sourced Computer Science

Published pieces

Languages

Sources cited

Questions answered

We are just starting. Our aim is two in-depth pieces a month — in five languages, open all the way to the source, and free of advertising.

Latest publications

How does a whole network fit inside one computer?

Switches, cables and servers can be built without real hardware, inside a single computer. We look at how Mininet does it, what can be tried on a virtual network, and what a laptop cannot show.

Information Retrieval 7 min

How does an AI find the right document?

An assistant reading a company's own records is only as good as the pages it is handed. This study measures how a thirty-year-old word-matching method outran a neural one.

Epidemic Modelling 5 min

What stops an epidemic: lockdown or vaccination?

Comparing quarantine and vaccination scenarios in an artificial town of a thousand gives a clear result: timing is as decisive as the type of intervention.

Speech Recognition 5 min

How does a machine know you said “open”?

How a word spoken into a microphone becomes a sequence of numbers, and then a command. Speech recognition before deep learning is still worth understanding.

Statistics 5 min

The line through the points: regression

The method that reduces the link between shelf space and coffee sales to a single number is also the first step into machine learning.

Load more articles

Featured

Proof: an eight-part beginner's guide

Teaching a network something new makes it erase what came before. Four answers to a thirty-year-old problem, and the variable everyone overlooks.

Read the piece

Catastrophic forgetting

When a neural network learns a new task and loses the old one almost entirely rather than partially. It happens because everything it knows lives in the same weights.

See the related piece →

Popular tags

What shall we learn today?

Questions Answered

Latest publications

A five-minute science digest, every Friday morning