Deep Learning
Deep Learning & Neural Networks
The engine underneath most modern AI: layer upon layer of simple maths units, a 'neural network', whose billions of internal dials are tuned automatically by showing it huge amounts of data. It recognises faces in your photos, translates languages and writes text. In one recent year, companies built 93 of the most notable AI models and universities built two. AI data centres already need about as much power as New York State at its peak.
Maturity
Where it feeds in
Where you have already met this
In 2025, industry built 93 of the year's notable AI models. Universities built two. AI data centres by then had 29.6 gigawatts of power capacity - roughly New York State at its peak. The maths is in every textbook. The electricity is not.
If this moved
If another architectural shift arrived, the enablers around it would decide whether anything changed. Automation of cognitive work, precision everything and autonomous mobility sit at 2, held by managerial capability, workforce trust and regulatory clarity. A better network touches none of them.
What to watch for
Grid-connection and transformer lead times where AI data centres cluster. How often AI agents fail real, structured tasks - not the scores models post on tests.
Not to be confused with
AI and machine learning as a whole. Deep learning is one family of methods inside it - the one that learns its own features from raw data instead of being handed them.
What it is
The architecture family underneath most modern AI: layered networks whose parameters are fitted by gradient descent over large datasets.
How it works
- 01Stack layers of numbersA network is layers of simple units, each passing a weighted sum of its inputs to the next. The weights start out random.
- 02Run an example throughText, pixels or sound, turned into numbers, flow through the layers and come out as a prediction.
- 03Measure the errorThe prediction is compared with the right answer, and the gap is scored.
- 04Push the blame backwardsBackpropagation works out how much each weight contributed to the error, from the last layer back to the first.
- 05Nudge every weightEach weight moves a little in the direction that reduces the error. Repeated over billions of examples, this is the training - and the electricity bill.
- 06Freeze and serveThe trained weights are fixed and the network answers new inputs. Every answer costs compute again.
State of play
- Industry produced 93 notable AI models in 2025 and academia two (Stanford HAI, AI Index 2026, 13 Apr 2026).
- Global AI compute capacity grew 3.3 times a year from 2022, reaching the equivalent of 17.1 million Nvidia H100 chips (Stanford HAI, AI Index 2026, 13 Apr 2026).
- AI data-centre power capacity reached 29.6 gigawatts in 2025, comparable to New York State at peak demand (Stanford HAI, AI Index 2026, 13 Apr 2026).
- Data centres used about 415 TWh of electricity in 2024, around 1.5% of the world's consumption, after growing about 12% a year since 2017 (IEA, Energy and AI, Apr 2025).
- Data-centre electricity use rose 17% in 2025, while supply chains for transformers and gas turbines tightened (IEA, 16 Apr 2026).
- AI agents still failed roughly one attempt in three on structured benchmarks (Stanford HAI, AI Index 2026, 13 Apr 2026).
Why it matters
It sets the gradient on what AI can be asked to do at all. Architecture changes - convolution, attention, mixture-of-experts - have repeatedly reset expectations faster than institutions could respond, which is a four-clocks problem expressed in engineering.
- It sets what AI can be asked to do at all. The megatrends it feeds - cognitive-work automation, precision everything, autonomous mobility - all sit at 2, held by people and rules, not by networks.
- The frontier has moved inside companies. When 93 of 102 notable models come from industry, the questions researched are the ones that pay.
- Its constraint is now physical. The limit on the next model is a grid connection, not an idea.
What holds it back
Power and grid connections
Training and serving large models needs data centres drawing hundreds of megawatts, and grid equipment has long waits.
Concentration of compute
The chips and the capital to run them sit with a handful of firms, so universities no longer set the agenda.
Reliability
A network that is usually right still fails often enough that safety-critical users keep a person in the loop.
Data at the frontier
The largest models have already consumed most high-quality public text, which is now a live limit.
Terms
- Neural network
- A model built from layers of simple connected units whose weights are learned from data.
- Parameter (weight)
- One of the numbers inside a network that training adjusts.
- Backpropagation
- The method for working out how much each weight contributed to an error.
- Gradient descent
- Adjusting weights in small steps in the direction that reduces the error.
- Transformer
- A network design that weighs every part of an input against every other part at once; the basis of today's language models.
- Inference
- Running a trained network to get an answer.
Steepens
| Megatrend | Weight | Role |
|---|---|---|
| Automation of Cognitive Work | large | |
| Precision Everything | moderate | |
| Autonomous Systems and Mobility | moderate |
The gradient behaves like a sticky rubber band: it is pulled hardest toward the weakest enabler. Marginal links are shown but do not pull.
If this moved one step
Computed from the map as it stands: this score raised from 4 to 5, every other score unchanged. It is arithmetic, not a forecast.
Raised one step on its own, it lifts no megatrend's weakest link; each gradient it feeds still rises a little.
The gradient also rises in
- Automation of Cognitive Work2.7/5 , 2.7/5
- Autonomous Systems and Mobility3.1/5 , 3.2/5
- Precision Everything2.7/5 , 2.8/5
What would change the gradient
Conditional, not predictive. No dates, no probabilities.
- 01An architectural shift, which cannot be scheduled
- 02Compute and energy cost per training run
- 03Data availability at the frontier, now a live constraint
Common misconception
That progress here is smooth. It is lumpy: long plateaus punctuated by architectural shifts, which is precisely why extrapolating the last two years is the error Chapter 14 dissects.
Talin, B. (2026). Deep Learning & Neural Networks. The Future Explorer, dataset v0.4.0 (2026-11-16). https://explorer.talin.digital/e/deep-learning/