Skip to content
The Future Explorerby Benjamin Talin
Enablertechnological

Large Language Models

A large language model is an AI trained on huge amounts of text to predict the next word, again and again. That simple trick turns out to let it answer questions, summarise, translate and write code: almost any task you can describe in words. Because you just type to it, it spread faster than any tool before; one chatbot reports about 900 million users a week. Yet fewer than one US firm in five uses AI anywhere in its work.

Maturity

4/ 5
measured
BasisUsed by hundreds of millions of people every week, at a price that fell more than 280-fold in two years. Firms lag far behind: fewer than one in five US firms used AI anywhere in their work.SourceE-052 (UBS/Similarweb via Reuters, 1 Feb 2023); E-104 (Stanford HAI, AI Index 2025, 7 Apr 2025); OpenAI via TechCrunch (27 Feb 2026); Bick, Blandin & Deming, NBER w32966 (Sep 2024); Bonney et al., US Census Bureau CES-WP-26-25 (Apr 2026); EIG analysis of Census BTOS (19 Dec 2025)Vintage2022–2026
ReadinessCapability 5Deployment 4Absorption 4Bound bytrustTrendimproving

Where it feeds in

5 megatrends in 3 forces

Where you have already met this

In February 2026 OpenAI said 900 million people used ChatGPT every week. Walk into an American firm and the odds that it used AI anywhere in its work were under one in five. The person at the desk has it open. The firm, mostly, does not.

If this moved

If models became reliable enough to run unsupervised, none of the five megatrends they feed would move on that alone: they sit at 1 and 2, held by managerial capability, workforce trust, regulatory clarity, epistemic trust and consumption norms. The model is not the slow part.

What to watch for

The Census Bureau's count of US firms using AI to produce goods or services - about 10% in September 2025. Error rates for models in legal and financial work, measured by someone other than the vendor.

Not to be confused with

Artificial general intelligence. A language model predicts text; AGI is the idea of a system that can do any cognitive task a person can. The first exists. The second is an argument.

What it is

Models trained to predict text, which turn out to perform a wide range of tasks that can be described in language.

How it works

  1. 01Cut text into tokensText is split into word fragments, and each fragment gets a number.
  2. 02Learn to predict the next oneThe model reads an enormous amount of text and adjusts billions of internal settings until it gets better at guessing the next token.
  3. 03Weigh the contextA mechanism called attention lets every token take account of every other token in the passage. That is the Transformer, published in 2017.
  4. 04Teach it to follow instructionsThe trained model is then tuned on examples and on people's ratings of its answers, so it responds to requests instead of just continuing text.
  5. 05Generate one token at a timeWhen you ask something, it picks a likely next token, adds it to the text, and repeats. Fluency comes from this. So do confident mistakes.

State of play

  • The price of using a model that scores like GPT-3.5 fell from $20 to $0.07 per million tokens between November 2022 and October 2024, more than 280-fold (Stanford HAI, AI Index Report 2025, 7 Apr 2025).
  • OpenAI reported 900 million weekly active ChatGPT users and 50 million paying subscribers - the company's own figures (TechCrunch, 27 Feb 2026).
  • 39.4% of Americans aged 18–64 had used generative AI by August 2024 (Bick, Blandin & Deming, NBER w32966, Sep 2024).
  • 18% of US firms used AI in at least one business function - 32% when weighted by employment - and 57% of those used it in three functions or fewer (Bonney et al., US Census Bureau, CES-WP-26-25, Apr 2026).
  • On a narrower question - using AI to produce goods or services - the share of US firms rose from 4.6% at the start of 2024 to 10% in September 2025 (EIG analysis of Census BTOS, 19 Dec 2025).

Why it matters

It is the enabler that made AI touch everyone at once. Adoption was fast because the interface was language and the switching cost was near zero - which is why its human enabler scored unusually high and its ramp was unusually steep.

  • It feeds five megatrends and holds none of them. They sit at 1 and 2, held by people and institutions, not by models.
  • It is the exception that proves the rule on ramps. The interface was language, so people adopted it in months. Firms have to reorganise, so they adopt it in years.
  • Fluency made adoption easy and checking hard. A wrong answer reads exactly like a right one.

What holds it back

Reliability

Models produce confident errors, which keeps them out of work nobody has time to check.

Liability and data

Firms hesitate to put confidential data into a model, or to answer for what it says.

Reorganisation

Most firms that use AI use it in a few functions. Getting value means changing processes, and that is managerial work.

'Not for us'

Most US firms that do not use AI say it does not apply to their business (Bipartisan Policy Center, on the Census AI supplement).

Terms

Token
A piece of text - often part of a word - that a model reads and produces.
Transformer
The model design, from 2017, that uses attention to weigh every part of a text against every other part.
Context window
How much text a model can take into account at once.
RLHF
Reinforcement learning from human feedback: tuning a model towards the answers people rate higher.
Hallucination
A fluent answer that is false or not supported by any source.

Steepens

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

What if it moved?

What would change the gradient

Conditional, not predictive. No dates, no probabilities.

  1. 01Reliability guarantees that make unsupervised use defensible
  2. 02Cost at the point of embedding rather than at the point of chat
  3. 03Liability rules, which currently keep the highest-value uses out of reach

Common misconception

That fluency indicates understanding. The book does not need to settle that question, and should not pretend to; what matters for the model is that fluency lowered the adoption barrier regardless of what sits behind it.

Talin, B. (2026). Large Language Models. The Future Explorer, dataset v0.4.0 (2026-11-16). https://explorer.talin.digital/e/large-language-models/