Keras Creator François Chollet: AI May Undermine Young Mathematicians' Willingness
François Chollet, the creator of the deep learning library Keras and proposer of the ARC-AGI benchmark, stated that he often hears math students say, "I don't want to be a mathematician anymore." He feels this is reminiscent of the generation of digital artists who came of age around 2022, possibly the last generation.
Chollet's assessment is that while AI may not functionally replace mathematicians or artists, it could undermine the will of the younger generation, leading to near extinction. He has long distinguished between skills and intelligence, viewing intelligence as a limited conversion ratio rather than an infinitely scalable IQ score; he warns that the worst misuse is outsourcing cognition to models. He released Keras in 2015 and will leave Google in 2024 after nine years to co-found a general intelligence company focused on program synthesis with Zapier's co-founder, offering a $1 million ARC Prize.
During the same period, Fields Medal winners have co-signed criticisms of laboratories rushing to publish unfinished proofs; in an interview with Scientific American, a PhD student stated that without subscribing to cutting-edge models, they would fall behind, while others believe that the acceleration itself is not frightening and that jobs will decrease. The Leiden Declaration contrasts disciplinary values with commercial logic. Chollet shifts the question from "Will machines solve problems?" to "Is there anyone willing to spend ten years stuck on difficult problems?"
After the explosion of image generation in 2022, entry-level orders for commercial illustration and concept design were undercut by model prices, leading many young creators to switch careers or become prompt operators. Although mathematics has not yet seen a comparable scale of job loss, the rewriting of priorities by computational timestamps makes the hard work of training seem harder to monetize.
In market mechanisms, this is a depreciation of career expectations, not today's job transactions. Buyers are students still needing papers and teaching positions; sellers are models that can produce results to announce over the weekend. Beneficiaries are laboratories that treat mathematics as a metric; those under pressure are apprentices who view being stuck as a growth mechanism. Funding is shifting from PhD stipends to model subscriptions, and the cost of will cannot be recorded in laboratory profit statements.
Supplementary structure: Chollet does not advocate for zero value of models; he suggests treating them as interfaces to deepen mental models. The contrast in digital art is that functionality is not dead, but the novice funnel is dying first.
Source: Public Information
ABAB AI Insight
What Chollet sees is not unemployment statistics, but the shattering of professional myths. The reproduction of mathematicians relies on the belief that "sticking with it for a long time is still worthwhile"; after laboratories rush to announce proofs, being stuck becomes a disadvantage. The artists of 2022 experienced the same blow: the tools remain, but orders and identities have gone first. The Keras creator himself has become an infrastructure figure through the open-source framework, yet he directs his warnings at downstream apprentices, indicating that he believes the replacement occurs at the level of willingness, preceding the wage level.
The capital path is evaluation and subscription. The ARC Prize buys narratives of "true generalization" with a million dollars; laboratories buy valuations with unverified proofs. Students are caught in the middle: if they don't subscribe to models, they fall behind their peers, but if they do, they feel they are working for someone else's timestamp. Teaching positions do not increase as tokens decrease, but the funnel becomes narrower. The art market has already demonstrated: models do not need to hire all illustrators, just make newcomers feel that entering is foolish.
This is analogous to the shrinkage of illustration positions in magazines after photography entered journalism, and the disappearance of mental calculation as a professional skill after calculators entered classrooms: functionality can coexist, but the apprenticeship can die first. The industry phase is that mathematics is beginning to be treated as an output that can be accelerated, rather than a craft that must be slowly internalized. Fields Medal winners are fighting for belonging, while students are debating whether to even start.
Structural judgment belongs to the realization of technological replacement through expectations. The mechanism is: professional survival depends on the ten-year discount rate within young people; once the discount rate is rewritten by the weekend output of models, people do not need to be laid off; they will cancel their applications themselves. Extinction does not have to occur in current positions, but in the next round of applications.
ABAB News · Cognitive Laws
- Professions can still functionally exist, but die first on application forms.
- What is undermined is the willingness of those who are willing to be stuck, not necessarily the employed.
- Before models replace, they first replace the faith in hard work.