When and how it might arrive

How superintelligence could emerge

Path 1: more resources and better methods

Since 2012 AI capabilities have grown through a combination of computing power, better data, better algorithms and more efficient training. The number of parameters alone is not a measure of capability — research on compute-optimal training showed how much the ratio of model size to data matters. Better memory, tool use, learning from verifiable results and planning may also help. Whether developing today's approaches will be enough, or fundamentally new principles are needed, is open — experts genuinely disagree.

Path 2: AI helps develop the next AI

The British mathematician I. J. Good, who worked with Alan Turing on codebreaking during the war, followed a simple thought to its end in 1965: a machine smarter than a human will also be better at designing machines. So it designs a smarter machine, which designs a smarter one still — and human intelligence is quickly left behind. He called the first such machine “the last invention that man need ever make”, adding: provided it is docile enough to tell us how to keep it under control. (Good's paper)

Today that need not mean a program rewriting its own code on the fly. It is enough for AI to speed up the whole process: designing models, preparing experiments, evaluating results. But writing code is only part of research — you have to recognise a good idea, design an honest experiment and show that an improvement holds beyond the test set. New hardware has to be manufactured and large experiments cost money.

A fast intelligence explosion is therefore a scenario with assumptions, not an automatic consequence of AI assistants existing. Gradual progress, alternating breakthroughs and stagnation, or long-term limits on some abilities are also possible.

Other paths

People consider whole brain emulation (simulating a biological nervous system on a computer) or linking humans and machines. But a map of neural connections is not yet a recipe for a human mind, and a brain–computer interface is not automatically an intelligence amplifier. Big changes may also come from a collective of people and AI without any single all-round superhuman model — for society, what a whole organisation can do may matter more than the label on its software.

When might it arrive?

We have no reliable date — and be wary of anyone who claims otherwise. First we also need to say what is being predicted: general intelligence, automation of occupations, or clear superiority over the best humans?

How to read those numbers:

• They are an aggregate of probability estimates, not a year chosen by half the respondents.
• Both questions were about feasibility, not actual adoption in the economy. Between “a machine can do it” and “machines do it everywhere” lie laws, cost, trust and plenty of human decisions.
• The authors themselves note that results depend heavily on how questions are worded — which is why two similar targets differ by 70 years.
• It is not a date for ASI nor a measured probability. (survey and methodology)

Heads of some leading labs publicly talk about much shorter horizons; other respected scientists, such as Yann LeCun, doubt that today's language models alone lead to general intelligence. Their arguments are summarised on Sceptical voices. Sensible planning therefore allows for several possibilities: a short horizon argues for fast preparation and testing, a longer one gives more time but does not remove today's problems. And even if ASI never arrives, more capable specialised systems can still change the economy and security considerably.

How to recognise real progress

When a model posts a record score, it is worth finding out what was actually measured. A benchmark is a set of test tasks — not a universal report card.

A model that is excellent at fixing short programs may still be unable to take over someone else's project, clarify an incomplete brief, fix a bug and deploy a change without breaking anything. Real work involves surprises, coordination and responsibility. For any result, look for:

• Novelty — could the model have seen the test or its solution during training?
• Conditions — did it have tools, many attempts or human help?
• Reliability — does it solve the task repeatedly, or did one attempt out of many succeed?
• Cost and time — how many resources did the result take?
• Transfer — does performance hold when the setting and brief change?

The organisation METR measures agents by how long the tasks they can complete are (in the time a human expert would need) at a given success rate, for example 50%. This “time horizon” is not the model's thinking time nor a yardstick for all professions. (METR methodology)