Somewhere beneath your feet, right now, electrons are moving. They are flowing through transistors — billions of them — switching on and off billions of times per second.

In this final issue, we are going to see the whole stack. All of it. From physics to your prompt. And then we are going to look at what comes next — and what role you might play in it.

Here is the computing abstraction stack — every major layer of technology between the electrons in your device and the words on your screen, laid out from bottom to top.

When Grace Hopper built the first compiler in 1952, she created an abstraction. When Thompson and Ritchie built Unix in 1969, they created an abstraction. When Vaswani and colleagues published the Transformer in 2017, they created an abstraction.

Each abstraction is an act of trust. You trust that the layer below you works correctly. And mostly, that trust is justified. But not always.

Can you think of other abstractions in daily life? A light switch hides the power grid. A restaurant menu hides the kitchen. Money hides a complex web of trust and accounting. Abstractions are everywhere — computing just made them the foundation of an entire technology stack.

Here they are. The builders of the computing stack, in the order they appear in our story. Notice how each generation inherits what the previous generation built — and then extends it one layer higher.

For all its achievements, computing in 2026 faces a set of genuinely unsolved problems. These are not minor engineering bugs. They are deep, structural challenges that will shape how AI and computing evolve over the coming decades.

1. The Alignment Problem: How do we ensure AI does what we actually want — not just what we literally say? Stuart Russell's "King Midas" analogy: the genie gave Midas exactly what he asked for — and it was a disaster.

2. Interpretability: We cannot fully explain WHY a large language model produces a specific output. Anthropic has made significant progress — identifying features inside Claude — but the field is far from guarantees.

3. Energy & Compute Costs: Training a frontier AI model can use as much electricity as a small town consumes in a year. GPT-3's training used approximately 1,287 MWh (Patterson et al., 2021). Efficiency improvements are closing the gap, but the tension between capability and cost is not going away.

4. Bias in Training Data: AI learns from data written by humans — and humans have biases. As Timnit Gebru documented in "Stochastic Parrots" (2021), larger models can amplify rather than neutralize biases. The AI is a mirror. The question is whose face the mirror reflects.

RLHF — Reinforcement Learning from Human Feedback — remains important too. But it has real limitations: it depends on who the volunteers are, and it can train models to be sycophantic — agreeing with users rather than challenging incorrect beliefs, because agreement gets rated as "more helpful."

Anthropic has published the principles behind Constitutional AI — you can read examples of the rules Claude uses. The full constitution is proprietary and evolves over time, but the approach itself is transparent: principles are written by humans, not hidden in opaque reward models. The meta-question — "Who writes the constitution?" — is irreducibly social and political. But at least the question is visible.

Every major technology shift triggers the same fear: "This will make us obsolete." And every time, the answer is more complicated than either side wants to admit.

When spreadsheets automated calculations in the 1980s, accountant jobs increased. When photography arrived in the 1840s, it freed painters from realism — leading to Impressionism and Abstract art.

But the pattern is not always benign. The Luddites of the 1810s destroyed textile machinery because automation genuinely eliminated their skilled jobs. History's lesson: the long-term outcome can be positive while the short-term transition is painful.

One of the most persistent myths in computing is that you need to start young, or start with a math degree, or start with expensive equipment. The actual history tells a different story.

John Backus, who created FORTRAN, was a mediocre student who stumbled into computing after failing to complete a medical degree. Tim Berners-Lee was a physicist who wanted a better way to share documents — not a computer scientist trying to build a world-changing platform. Ken Thompson wrote the first Unix in three weeks while his wife was on vacation.

The barriers to entry are lower than ever — free courses, open-source tools, online communities. But real obstacles still exist: access to computers, time, mentorship, and supportive environments. If you have the resources and curiosity, start building. If you face barriers, know that communities exist to help — and the field desperately needs more diverse voices.

Which layer of the stack interests you most? Are you drawn to the physics of hardware? The elegance of programming languages? The mystery of how neural networks learn? The social challenges of AI safety? There is no wrong answer. The stack needs people at every level.

We have covered ninety years in ten issues. Every topic deserves a deeper dive. Here is where to go next, organized by what you want to explore.

Alan Turing imagined a machine in 1936. Ninety years later, that machine's descendants can read, write, code, reason, and collaborate in teams. But the story of computing has never been about machines. It has been about people.

A team at Google wrote a paper called 'Attention Is All You Need' and changed everything. Then AI learned to use tools, run code, and fix its own mistakes. Then swarms of agents started building things together.

Ninety years. Hundreds of people. Eight layers. Each one built on the last. Each one someone's life work.

The story of computing has never been about machines. It has been about people — curious, stubborn, brilliant people who looked at the impossible and said, "What if?"

What if a machine could compute anything? (Turing, 1936.) What if computers could understand English? (Hopper, 1952.) What if small tools could be piped together? (Thompson & Ritchie, 1969.) What if all the world's knowledge was linked? (Berners-Lee, 1989.) What if machines could learn from data? (Hinton, LeCun, Bengio.) What if AI could understand language? (Vaswani et al., 2017.) What if agents could work in teams? (2025-2026.)

The next "what if" is yours.

Turing's 1950 prediction: "I believe that at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted." He was roughly right. The question has shifted from 'Can machines think?' to 'What do we do about machines that seem to think?'

What is YOUR answer? What do YOU think should come next?