Per aspera ad astra: through hardship to the stars. For generations, it has also been a fairly good description of professional life. You spend years learning something difficult, acquire experience, become good at it and eventually society rewards you for possessing skills that other people do not.
OpenAI’s new Astra model suggests a rather less reassuring inversion: per Astra ad aspera. Through Astra to hardship. Not because Astra is some final form of artificial intelligence, but precisely because it is not. It is simply the latest point on a curve that is still moving.
What is striking about Astra is therefore not merely that it performs well on benchmarks. Frontier models have been doing that for years. It is the breadth of the performance that matters: advanced mathematics, software engineering, CAD, scientific computing, cybersecurity, medicine and other fields that would normally belong to entirely different specialists. There are still humans who can beat Astra in each of these areas, but that is not really the interesting comparison.
A more unsettling thought experiment is this: imagine that, to get a job in your field, you have to beat the best available AI on a benchmark that reasonably represents the central cognitive tasks of that job. How many people would still qualify? And if Astra does not yet set the bar high enough to cause serious trouble, what happens when the next model raises it, and the one after that raises it again?
The benchmark percentages themselves cannot answer the question. A model completing 96 per cent of a test correctly does not mean that it is better than 96 per cent of professionals, and benchmarks are cleaner than actual workplaces. Real jobs involve ambiguous instructions, half-documented systems, awkward customers, institutional memory and all the other forms of disorder that human organisations have spent centuries perfecting.
Still, the thought experiment exposes something important. AI does not need to become the best accountant, engineer, programmer or lexicographer in the world before it affects employment. It merely needs to become good enough that one unusually capable person, equipped with AI, can do work that previously required several competent people.
For most of modern history, competence has been enough. You did not have to generate ten original ideas before breakfast or reinvent your profession every six months. You learnt something useful, became good at it and performed the work reliably, and most organisations still depend overwhelmingly on people of exactly that sort.
The complication is that AI does not complement everybody in the same way. I can see this very clearly at home because my wife and I are both linguists and both work with dictionaries, but our preferred ways of working are quite different. She is extremely bright, but she likes depth: she would rather become very good at lexicography and continue doing lexicography than constantly acquire unrelated new skills.
That has always been an entirely sensible strategy. Learning a profession is expensive in time and effort, so once you have accumulated years of knowledge, judgement and experience, it makes sense to use them. Indeed, much of the modern knowledge economy is built around exactly that bargain: specialise, acquire expertise and eventually the difficulty of replacing you becomes part of your economic value.
I am almost the opposite. I get bored easily and enjoy moving between very different projects, sometimes several in the same week. I might be working on lexicography one day, software the next, then economics, an app idea or some historical question that has wandered in from nowhere. Historically, this has been a somewhat dubious professional trait, because employers generally have more immediate use for somebody who has spent fifteen years becoming exceptionally good at precisely the thing they need.
AI changes that calculation. My problem has rarely been finding things I would like to do; it has been the time needed to acquire all the intermediate skills required to do them. An idea for an app once meant learning a new development environment, while an unfamiliar technical problem might require days or weeks of preliminary work before I could even test whether the idea was worth pursuing.
Increasingly, much of that aspera can be delegated. I still need to decide what is worth doing, understand enough to recognise mistakes and know where I want to go next, but I do not necessarily need to become an expert in every intermediate skill first. That makes somebody with too many ideas considerably more productive, because the cost of moving into unfamiliar territory falls dramatically.
For somebody whose value lies mainly in having spent many years mastering one stable field, the same development is more ambiguous. AI can certainly help them work faster, but with every model generation it may also learn a little more of the work they spent those years learning to do. The expertise remains useful, yet it becomes less scarce.
This is why the reassuring claim that AI will simply make everybody more productive misses something important. Productivity and employment are not the same thing. If five people become twice as productive, perhaps demand expands enough for all five to remain employed, but perhaps the organisation simply discovers that three people can now do what five used to do.
Both outcomes appear in the productivity statistics as progress. Only one of them looks quite so cheerful from the perspective of the other two employees. Nor are the survivors likely to be selected randomly, because AI may prove especially complementary to people who decide what should be done: those who generate ideas, notice unusual connections, exercise judgement, take responsibility or move comfortably into unfamiliar territory.
That does not make such people better human beings. Civilisation would probably collapse within a fortnight if everybody behaved like a restless generalist and nobody simply got on with the job. Hospitals, companies, universities and governments depend on millions of people who are extremely good at doing difficult but established things carefully and repeatedly.
Labour markets, however, do not reward moral worth. They reward scarcity, and professional competence has traditionally been scarce partly because acquiring it takes years. The hardship in per aspera ad astra was not merely character-building; it was what made the resulting expertise valuable.
AI systems do not have to make the same journey, and neither will their successors. Once a capability has been learnt by a model, it can in principle be reproduced on a vast scale at very low marginal cost. The next system also begins from a very different starting point from a human novice, which is why Astra matters less as a destination than as evidence of how quickly the threshold can move.
The economically decisive moment in the AI revolution may therefore arrive well before machines become cleverer than every human at everything. It may come when competent intellectual labour becomes abundant. At that point, being a perfectly good programmer, analyst, translator, designer or adviser may no longer be enough to guarantee that somebody needs to employ you.
Retraining is the usual answer, but general AI makes that answer less reassuring than it was during earlier technological revolutions. A tractor could replace agricultural labour, but it could not retrain as an accountant. A spreadsheet could replace manual calculation, but it could not notice that graphic design was becoming a promising new occupation and follow the displaced clerks into it.
General-purpose AI can follow us. If humans retreat from one kind of work into another, there is no obvious reason why future models cannot learn the new work too. The familiar historical escape route, in which technology destroys old occupations while humans move into new ones, may therefore become an increasingly fast-moving race.
Some people will thrive in that world. Those who constantly find new things worth doing may discover that AI removes many of the practical limits on what they can attempt. Others may find that they followed every piece of sensible career advice they were given, worked hard, specialised, accumulated experience and reached the end of the difficult road only to discover that the destination had moved.
Per aspera ad astra was never a guarantee, but it was a fairly dependable bargain. Work hard enough to acquire a difficult competence, and that competence would probably have economic value.
Astra may be only one step along the way. That is precisely why per Astra ad aspera is worth worrying about.