AI at a crossroads: the two AI winters, the 2026 job market for young engineers, and a method to work in it

The two AI winters, and the mechanism they share

Twice already, AI was promised far more than it could do, funders stopped paying, and the field came back later through a different door.

The story starts with a model of the neuron by McCulloch and Pitts in 1943, Turing's imitation game in 1950 and the 1956 Dartmouth workshop, where John McCarthy named the field. In 1958 Frank Rosenblatt's perceptron learned to recognise simple shapes, and the New York Times reported that the Navy expected a machine that would walk, talk and reproduce itself. In 1969 Minsky and Papert proved that a single-layer perceptron cannot separate the XOR function. The result was correct and narrow, but funding for neural networks dried up for about a decade.

The first winter (roughly 1974 to 1980) was a crisis of confidence among funders rather than a technical collapse. In 1973 the mathematician James Lighthill told the British Science Research Council that combinatorial explosion kept the programmes of the day on toy problems; the United Kingdom stopped funding AI in nearly all its universities. In the United States the 1969 Mansfield amendment had already pushed DARPA towards applied work, and in 1974 it cancelled Carnegie Mellon's speech understanding contract, about three million dollars a year.

Expert systems brought a second spring. XCON, written at Carnegie Mellon for Digital Equipment, configured VAX orders with about 2,500 hand-written rules; DEC put the saving at 25 million dollars a year in 1986. Japan launched its Fifth Generation project in 1982. The second winter (1987 to 1993) came from maintenance cost: every change in the business meant rewriting rules, and in 1987 cheap Sun, Apple and IBM workstations wiped out the market for dedicated Lisp machines within a year, while DARPA cut its programme. The way out was statistical: estimate the rules from data instead of writing them.

Back to contents.

The building blocks of machine learning, before any derivation

A neuron is a weighted sum passed through a bend, learning is walking downhill on an error, and a model is only trusted on data it has never seen.

The formal neuron keeps the structure of the biological one (the brain has about 86 billion of them, Herculano-Houzel 2009): inputs are weighted, summed, shifted by a bias and passed through an activation function. Without that function, stacking layers is pointless, because a composition of linear maps is still linear. Sigmoid and tanh saturate in deep networks; the ReLU, proposed for networks by Nair and Hinton in 2010, does not saturate for positive inputs and is one of three reasons, with GPUs and data, why deep networks started working around 2012.

Hidden layers solve the XOR problem, and backpropagation (popularised by Rumelhart, Hinton and Williams in 1986) computes how the error changes with every weight. Learning then follows Cauchy's 1847 recipe: step against the gradient of the loss. Too large a step diverges, too small never arrives, and an elongated valley makes the path zigzag. The course spends half a day on this in convex optimisation and regularisation.

The other family has neither neurons nor gradients: decision trees (CART, 1984), random forests that let many decorrelated trees vote (Breiman, 2001), and gradient boosting, where each new tree fits the residuals of the sum so far (Friedman, 2001). XGBoost (Chen and Guestrin, 2016) counted 17 of the 29 winning Kaggle solutions of 2015. On tables of data a tuned boosting very often beats a neural network.

Two rules hold everything together. First, a model is measured only on rows it never saw: split before you start, tune on validation, open the test set once. Stone formalised cross-validation in 1974. Second, the expected error splits into squared bias, variance and noise (Geman, Bienenstock and Doursat, 1992): too simple a model misses the shape, too flexible a model learns the noise.

The model ladder used throughout the course

1. Naive. A mean, a majority class or a small logistic regression, delivered within days. It forces the pipeline to exist and sets the baseline.
2. Better. The same data, better prepared, a regularised model and cross-validation. You measure the gap with the baseline.
3. Heavy. Boosting or a network, kept only if the gain pays for the extra cost of explanation, compute and maintenance.

The model is one box out of six in CRISP-DM (1999): business understanding, data understanding, preparation, modelling, evaluation, deployment. Anaconda's 2020 survey put data preparation at about 45% of a data scientist's time. Session 1 starts there, with least squares.

Back to contents.

Young developers and data scientists in 2026: what the figures say

Demand for data and AI skills keeps rising, but companies have mostly stopped hiring beginners, so the hardest step is now the first job.

The mechanism is a hiring freeze, not a wave of layoffs: in the occupations most exposed to AI, companies hire fewer young people while employment of experienced staff stays stable. In France, the Banque de France and LIFT note shows contract terminations falling too: few hires, few departures. Indeed Hiring Lab calls it seniority-biased technological change (July 2026): 71% of the rise in US software development postings between May 2025 and May 2026 came from senior roles.

Read the table as a paradox rather than a verdict. Ten-year projections remain favourable for data scientists, yet the BLS revised software developer growth down from about 15% in its previous series. The French note of 8 October 2026 was two days old when the deck was written, and the New York Fed figures by major come with wide confidence intervals; the deck flags both. The practical conclusion is the one the course is built on: an interchangeable junior who only produces code is the profile most at risk, while someone who can derive, check and defend a model in front of a client is not.

Back to contents.

The engineer, the project manager and the agent manager

Writing code is becoming cheap, so the scarce work moves to framing the problem, reviewing what the machine produced and deciding.

For the engineer, the bottleneck has moved from writing to review. Sundar Pichai said in April 2026 that 75% of Google's new code is generated by AI. In the Stack Overflow survey of October 2026, 48% of developers trust AI output only when they can verify it. METR's randomised trial of July 2025 found experienced open source developers 19% slower with AI tools while they believed they had saved 20%; its February 2026 follow-up found no significant effect either way. Faros AI (2026) reports that teams with high adoption open larger pull requests that wait 441% longer in review.

Projects keep failing at the same rate, which is why the project manager does not automate away. PMI's 2024 study of 10,000 professionals rated 48% of projects successful; the Standish CHAOS data for 2024 give about 31% for IT projects, flat since 2015; McKinsey and Oxford found large IT projects 45% over budget on average. The same PMI study finds that projects which define their success criteria at the start score twice as high, which is the reason session 1 begins with framing.

The third role is new. Microsoft's Work Trend Index of April 2025 described the agent boss who delegates to AI agents; the May 2026 edition found that the organisation explains about two thirds of the reported impact and the individual one third. Managing agents is therefore a management problem more than a prompting problem. Gartner expects 15% of day-to-day work decisions to be taken autonomously by 2028, against none in 2024.

Back to contents.

The bet: from pilot to production, or a bubble

The technology moves faster than companies can adopt it, and if that gap does not close, investors may lose faith as they did in 1973 and 1987.

• Pilots without return. MIT NANDA (July 2025) found 95% of organisations with no measurable P&L impact six months after a generative AI pilot, on a sample of 52 organisations that the report itself says may not represent the market.
• Use without profit. McKinsey's State of AI 2026 (published 25 August 2026): nearly nine organisations in ten use AI, but only 37% report a positive effect on earnings.
• Agent projects at risk. Gartner (25 June 2025) expects over 40% of agentic AI projects to be cancelled by the end of 2027, for cost, unclear value or weak risk control.
• Procurement. An enterprise sales cycle takes 90 to 180 days (Optifai, 2025-2026), plus security reviews; in finance, DORA has required a register of ICT contracts since 17 January 2025, and the AI Act obligations arrive in stages from 2025 to 2028.
• The bill. The four hyperscalers spent about 410 billion dollars of capital in 2025 and guided about 730 billion for 2026; Bain (September 2025) estimates that 2,000 billion dollars of annual revenue are needed by 2030 to pay for AI compute, of which 800 billion are missing. In October 2025 the Bank of England warned that the risk of a sharp market correction had increased.

The deck closes the part with two scenarios. In the first, tools pass review, data is ready and the share of companies that see an impact rises; the winners are those who can take a model from a demo into an organisation and keep it under control. In the second, pilots stay pilots and a third, costlier winter follows; the winners are those who learnt what survived the first two winters: least squares (Gauss and Legendre), likelihood (Fisher), gradient descent (Cauchy). In both cases the skill is the same, and it is what the decision threshold and reproducible ML session trains.

Back to contents.

Method first: Descartes and Pascal

Two seventeenth-century French scientists already wrote the rules of a good data project: prove before you accept, split the work, start simple, check everything.

Descartes' four precepts (Discourse on the Method, 1637), read as a project rule

1. Evidence. Accept nothing you cannot justify: neither the target nor the data. In a project, this is the framing.
2. Division. Cut each difficulty into as many parts as needed: the project becomes a list of tasks.
3. Order. Start with the simplest objects and climb: the naive model before the better and the heavy one.
4. Enumeration. Review so completely that nothing is left out: tests, cross review and a log of decisions.

Pascal adds the other half. At nineteen, in 1642, he built the Pascaline, a calculating machine for his father's tax work. In 1654 his letters with Fermat on how to split the stakes of an interrupted game founded probability theory. In the Pensées he separated the spirit of geometry, which reasons correctly from clear principles, from the spirit of finesse, which grasps what principles leave out. A data engineer needs both: rigorous computation, and judgement about what the results mean.

Back to contents.

Ringelmann's rope, the free rider and the Condorcet vote

People pull less hard on a rope when they pull together, and a group choice is fairer when every option is compared with every other one.

Maximilien Ringelmann, an agricultural engineer who ran a machine testing station in Paris, had students pull on a rope attached to a dynamometer, alone and then in groups, and published the results in 1913. The group pulled less than the sum of its members, and the gap grew with its size. The percentages that circulate were reported by Moede in 1927; Kravitz and Martin traced the original article in 1986 and noted that Ringelmann blamed coordination first, while seeing the motivational part that psychologists later called social loafing.

Economics gives the same result from incentives. Olson (1965) showed that the larger the group, the more each member gains by letting others do the work: the free rider. Holmström (1982) proved that when only a team's joint output is observable, no rule for sharing that output gives everyone the right incentive; a principal must observe and adjust. Kerr and Bruun (1983) found that effort drops when members feel dispensable, and Brooks and Ammons (2003) that assessing student teams early, often and on precise criteria reduces the effect. Small teams and individual questioning follow from this.

To rank options, the deck uses Condorcet's method (1785): compare options two by two, as in a league; an option that wins every duel is the Condorcet winner. There may be none, since three voters can produce a cycle where A beats B, B beats C and C beats A. The fallback is the sum of scores, the Borda count (Borda, 1784). The deck's illustrative example, with five voters scoring three categories from 1 to 5, shows that the two rules can disagree.

Reading the example

1. Finance against Data. Voters 1, 4 and 5 score Finance higher, voters 2 and 3 prefer Data: 3 to 2 for Finance.
2. Finance against Construction. Voters 1, 3 and 5 prefer Finance, voters 2 and 4 Construction: 3 to 2 for Finance, which wins both duels.
3. Data against Construction. Voters 2, 3 and 5 prefer Data: 3 to 2, so the pairwise ranking is Finance, Data, Construction.
4. The sum disagrees. Borda puts Data first with 15 points, because two voters gave it the top score. Winning every duel is the more robust criterion.

Back to contents.

Key takeaways

1. The two AI winters (about 1974-1980 and 1987-1993) followed the same pattern: a real capability, an unrelated promise, a growing hidden cost and a funder who stopped believing.
2. The way out of the second winter was statistical: estimate the rules from data instead of writing them by hand, which is what least squares, likelihood and gradient descent do.
3. A model counts only when measured on data it never saw, and a project climbs from a naive model to a heavy one only when the test justifies each step.
4. The 2026 figures show demand for data skills rising while beginners are hired less, in the United States (Stanford, August 2026) and in France (Banque de France and LIFT, October 2026).
5. With code becoming cheap to write, the scarce skills are framing, verification and the ability to carry a model through review into production.
6. Descartes' four precepts and Pascal's pairing of rigour and judgement still describe a sound data project; Ringelmann and the free rider explain why teams should stay small.

Glossary

AI winter

A period when funding and interest in AI collapse after inflated promises; the two main ones ran from about 1974 to 1980 and 1987 to 1993.

Perceptron

Rosenblatt's 1958 learning machine: a weighted sum and a threshold, able to separate only what a hyperplane can separate.

Expert system

A program built from hand-written if-then rules elicited from specialists, such as XCON at DEC; costly to maintain as the business changes.

Activation function

The non-linear function applied after a neuron's weighted sum (sigmoid, tanh, ReLU); without it, a deep network is linear.

Gradient descent

Repeatedly moving the parameters a small step against the gradient of the loss, so that the error decreases.

Cross-validation

Rotating the validation part of the data so that each row is held out once, then averaging the scores (Stone, 1974).

Bias-variance trade-off

The split of expected error into squared bias, variance and noise; simple models have high bias, flexible ones high variance.

Trough of disillusionment

The phase of Gartner's hype cycle after the peak, when results fall short of expectations; generative AI was placed there in August 2025.

Social loafing

The drop in individual effort when people work in a group, first measured by Ringelmann with a rope and a dynamometer.

Free rider

A member who benefits from collective work without contributing to it (Olson, 1965; Holmström, 1982).

Condorcet winner

An option that beats every other option in a head-to-head majority vote; it does not always exist.

Borda count

A ranking by the sum of the scores or ranks given by voters; it can disagree with the Condorcet winner.

The other notes of the course

One public summary per deck, in the order of the course.

Next: Kaggle, GitHub, Python: the tools

1. IntroductionAI winters, jobs and method
2. ToolsKaggle, GitHub, Python: the tools
3. Session 1Project framing and least squares
4. Session 2Maximum likelihood and logistic regression
5. Session 3Gradient descent, ridge and lasso
6. Session 4Decision threshold and reproducible ML
7. HomeworkA defensible baseline model
8. Session 5Kernels, SVM and decision trees
9. Session 6Forests, boosting, honest comparison
10. Session 7SHAP, fairness, PCA and clustering
11. HomeworkA documented champion model
12. Session 8Neural networks, drift, conformal prediction
13. HomeworkA delivered machine learning project
14. Session 9Defending a machine learning project
15. ReferenceCredit default scoring project
16. ReferenceAsset returns: XGBoost and sentiment
17. ReferencePython code for every ML model
18. ReferenceDeep learning essentials
19. ReferenceOpen data ML projects by sector
20. ReferenceData and AI engineer handbook
21. ReferenceData engineering, SRE and cloud
22. ReferencePanorama of machine learning models

All the notes, as a reading list

Attention and transformers, step by step

About these notes

Written by the lecturer from the slides of the deck, as a public summary: the key ideas, the formulas and one worked example. The exercises, their solutions, the tests and the grading stay in the deck, for the students of the course.

Revision: 16 October 2026.