← kristianhilberg.tech
A field guide for the next generation

The machine that
guesses the next word

How AI actually works, why even its builders can't fully explain it — and how to choose a profession in a world where the guessing machine keeps getting better.

Part 1 — Inside the machine

One trick, repeated billions of times

A language model does exactly one thing: given some text, it predicts what comes next. Everything that looks like intelligence — answering questions, writing code, translating Norwegian — is that single trick applied over and over.

1

Text becomes tokens

Words are chopped into pieces called tokens and turned into long lists of numbers. fjord might be one token; batforerproven might be three.

2

Layers pass signals

The numbers flow through dozens of layers. In each one, every token "looks at" every other token and decides what matters — this is called attention. The model's knowledge lives in billions of learned dials called weights.

3

Out comes a bet

The final layer produces a probability for every possible next token. The model picks one, adds it to the text, and starts over. That loop is the entire conversation you're having right now.

4

Training tunes the dials

Nobody programs the rules. The model reads enormous amounts of text, guesses, gets corrected, and its weights are nudged — trillions of times — until the guesses become uncannily good.

Try it — be the model

The bars show how strongly the model would bet on each token. Tap one to add it and see the next prediction. Notice: every choice changes all the probabilities that follow.

Part 2 — The honest part

The builders grew it. They didn't design it.

Engineers fully understand the recipe: the architecture, the math, the training code. What they don't fully understand is what the finished model learned. Its abilities are spread across billions of numbers that no human wrote and no human can simply read.

A useful comparison: a gardener controls the seed, soil and watering perfectly — but didn't design the tree cell by cell. Researchers now do a kind of neuroscience on these models ("interpretability"), finding small circuits for individual concepts. It works — but it covers a small fraction of what's really going on inside.

What builders know
  • The exact architecture and math
  • What data trained it
  • How to measure and steer behaviour
  • Fragments of internal circuits
What nobody knows yet
  • Why a given answer came out
  • Where a specific "skill" lives
  • What it will do in every new case
  • The full map of its inner world
Part 3 — Choosing a profession

What the guessing machine changes about work

Here's the pattern that matters: AI is strongest wherever work is digital, repeatable, and pattern-shaped — exactly the things next-token prediction is good at. It's weakest where work is physical, where a human must be accountable, or where people specifically want a person.

Tap each band to see what that means in practice. Note that "exposed" rarely means "gone" — it usually means the job changes shape, and the people who direct AI outperform the people who compete with it.

Anything done alone at a screen, following patterns, is the machine's home turf.

Data entry & basic adminPure pattern-matching on text and numbers
Routine translationNext-token prediction across languages is the core trick
Template writingGeneric marketing copy, standard reports, boilerplate code
First-line supportAnswering common questions from a knowledge base

These jobs survive — but the daily work transforms. The junior tasks get automated; judgment, direction and responsibility remain.

ProgrammerLess typing code, more architecting and reviewing what AI produces
Lawyer / accountantAI drafts and searches; a human still signs and answers for it
DoctorAI helps diagnose; humans examine, decide and carry the trust
Designer / producerTools generate options fast; taste decides what's actually good

The physical world is messy and every job in it is different. Robots lag far behind language models here — and some work people simply want from other people.

Skilled tradesA plumber solves a different physical puzzle in every basement
Nurses & caregiversCare is presence, not information
EntrepreneursOwning the problem, the risk and the customer can't be delegated
Live performers, teachers, coachesPeople pay for a human in the room
Part 4 — The four durable qualities

What to build in yourself

Don't pick a job title — titles will keep changing. Pick qualities that stay valuable no matter how good the machines get. Any career that stacks two or more of these is on solid ground.

Embodied skill

Work done with your hands in the unpredictable physical world. Every boat engine, pipe and stage is a new problem.

⚖️

Accountability

Someone must be responsible when it matters — sign the contract, make the call, carry the risk. Machines can't be sued, trusted or elected.

🤝

Human connection

Care, leadership, teaching, negotiation, performance. Wherever the point is another person, a person stays in the loop.

🧭

Taste & ownership

Knowing what's good when there are a thousand options, and owning the thing — the brand, the catalogue, the company — instead of only selling hours.

Part 5 — Advice worth keeping

Five rules for choosing

01

Learn to direct the machine, not race it. The person who can tell AI exactly what to do — and judge whether the output is any good — multiplies themselves. That skill applies in every field.

02

Combine a digital skill with a physical or human one. An electrician who automates their admin. A nurse who understands the AI tools. Rare combinations beat pure specialists.

03

Aim to own something. Wages for routine work face pressure; assets don't. A brand, a business, a catalogue of songs, a customer list — ownership compounds.

04

Bet on fundamentals, not tools. Specific apps and languages will die. Math, writing, mechanics, psychology, and knowing how to learn — those transfer to whatever comes next.

05

Expect to change careers, and treat that as normal. The winners won't be the ones who guessed the right job at 18 — they'll be the ones who could reinvent themselves at 28, 38 and 48.