THE FIRST WAVE
1956–1972
Things move pretty fast from there.
An idea becomes a machine. A machine learns a game. Then it starts a conversation. For a moment, the possibilities seem endless.
SCROLL TO FOLLOW THE FIRST BREAKTHROUGHS
Read the source ↗
DavidCHAPTER 002 · THE FIRST WAVE
01 / 05 SCROLL TO TRAVEL THROUGH TIME
CHAPTER 003 / THE FIRST GREAT PAUSE1974 — 1980
Huge promises were made.
But the machines weren’t ready.
THE YEARS PASSED.1974
For AI, progress felt painfully slow.
Research continued, but funding and confidence fell. Three main factors held the field back.
THREE REASONS THE MOMENTUM FADED
Expectations
meet reality.
01 / THE LIMIT
The hardware just wasn’t ready.
Computing power couldn’t support the scale of the problems researchers wanted to solve. Memory, speed, and access to machines put ambitious ideas out of reach.
02 / THE LIMIT
Too little data. Too many rules.
Learning systems lacked large, usable datasets. Handwritten rules worked in small, controlled settings, but struggled with the uncertainty and complexity of the real world.
03 / THE LIMIT
Historical context / IBM ↗We had promised too much, too soon.
Predictions ran ahead of results. When human-level intelligence failed to appear on schedule, funders pulled back. The ideas survived; the confidence did not.
CHAPTER 004 / THE IDEAS THAT CHANGED THE TRAJECTORYThree inflection points.
1986
Three inflection points.
A different future.
The comeback wasn’t a single moment. It was a chain of breakthroughs, spread across three decades.
198601 / INFLECTION
BACKPROPAGATION
Learning to learn.
Errors become feedback. Rumelhart, Hinton, and Williams popularize a way to train multilayer neural networks by sending errors backward.
Read the original research ↗201202 / INFLECTION
ALEXNET
The power of depth.
Large datasets, GPUs, and deep networks converge. AlexNet’s ImageNet victory shows how much better visual recognition can become.
Read the original research ↗201703 / INFLECTION
THE TRANSFORMER
Attention is all you need.
Attention lets models connect information across a sequence. The Transformer becomes a foundation for modern language models.
Read the original research ↗HUMAN CURIOSITY. COMPOUNDING PROGRESS.
But there was still a long way to go
to the AI systems we know today.
CHAPTER 005 / THE INTERFACE THAT CHANGED EVERYTHING
January 1, 1986
Hello, ChatGPT.
A research preview gave the public a simple way to talk to a language model. Built from GPT-3.5 and refined with human feedback, ChatGPT could answer follow-up questions, draft text, and help with code — while still making mistakes.
The November 30, 2022 launch announcement ↗ONE SHORT POST. A VERY BIG MOMENT.
“today we launched ChatGPT.”
Sam Altman’s invitation was brief. The shift it introduced was anything but.
Open the original November 30 post ↗
A NEW CHAPTER / GENERATIVE AI AT PUBLIC SCALEThe world had entered
The world had entered
the fifth generation
of AI.
A storytelling label for this chapter, not a universally accepted technical classification.
JAN 2023FUNDING
The money follows the models.
Microsoft expands its partnership with OpenAI through a multiyear, multibillion-dollar investment. AI becomes an infrastructure and business race.
Read the record ↗2023MODELS
One chatbot becomes an industry.
GPT-4, Claude, Llama, and later Gemini turn foundation models into a competitive market. Capability, access, safety, and cost become points of contention.
Read the record ↗NOV 2023BOARDROOM CRISIS
A technology boom. A governance shock.
OpenAI’s board removes Sam Altman. He returns days later after a leadership crisis. A subsequent review describes a breakdown in trust between the board and management.
Read the record ↗2024–2025THE STAKES
More capable. More consequential.
Voice, images, reasoning, and tools expand what AI can do. Copyright disputes, training-data questions, safety concerns, and the cost of computing accompany the progress.
Read the record ↗CHAPTER 006 / ANOTHER MAJOR PLAYER
Claude.
Anthropic introduced Claude in March 2023. The race was already underway; a new family of assistants gave it another serious competitor.
Introducing Claude / March 14, 2023 ↗
THE PERSON BEHIND THE NAME / 1916–2001
Claude Shannon.
Claude is named after the pioneer of information theory. Shannon’s work explained how information could be measured and reliably transmitted; his earlier work connected Boolean logic to electrical switching circuits.
He helped lay foundations for the digital age and co-authored the proposal for the 1956 Dartmouth workshop. The name brings this story back to its beginning.
The Dartmouth proposal ↗FEBRUARY 24, 2025 / CLAUDE CODE RESEARCH PREVIEWFrom answering questions
From answering questions
to working in a codebase.
❯ claude
Read the project.
Understand the request.
Edit the files.
Run the checks.
Claude Code put an agent in the terminal: reading files, making changes, and using development tools. It helped make the shift from chat assistance toward coordinated coding workflows concrete.
Anthropic’s original announcement ↗CHAPTER 007 / MEANWHILE, ONE MONTH EARLIERThen came
JAN 2025
Then came
DeepSeek.
DeepSeek-R1 arrived on January 20. Its released weights and technical report put a different question at the center of the race: how much capability could be achieved more efficiently?
DeepSeek-R1 release ↗JANUARY 27, 2025 / NVIDIAThe market
The market
felt the shock.
Investors reassessed the cost of AI and the demand for its chips. Nvidia shares fell almost 17% in a single session, erasing about $593 billion in market value.
This was a market reaction, not proof that advanced chips were obsolete. Nvidia remained central to the infrastructure powering the AI boom.
Reuters / January 27 market reaction ↗NVIDIA / ONE-DAY SHARE-PRICE CHANGE−17%Illustrative trajectory · rounded daily change, not live market data.
THE STORY IS STILL UNFOLDING.The race wasn’t over.
The race wasn’t over.
It had become global.
Models. Money. Chips. Efficiency. Trust.
Every breakthrough changed the terms of competition.
CHAPTER 008 / THE QUESTION BECOMES PERSONALWill AI
Will AI
replace me?
Or augment me?
After seventy years of asking what machines could do,
we’re asking what happens to us.
NOVEMBER 2022 → MARCH 2023 / A CHANGE IN EXPECTATIONSIn less than four months,
In less than four months,
the conversation changed.
ChatGPT’s research preview arrived on November 30, 2022. GPT-4 followed on March 14, 2023. The difference between their model generations was striking on a simulated bar exam.
GPT-3.5 / ESTIMATED PERCENTILE≈10th
Near the bottom of the comparison group.
GPT-4 / ESTIMATED PERCENTILE≈90th
Near the top of the comparison group.
Both estimates come from OpenAI’s March 2023 evaluation. They describe performance on one simulated exam, not human IQ or the ability to do an entire job.
Read the original evaluation / March 14, 2023 ↗THE PACE OF PROGRESS BECAME THE CONCERNThen came the calls
Then came the calls
to slow down.
A six-month pause.
The Future of Life Institute published an open letter calling for a pause of at least six months on training systems more powerful than GPT-4. Researchers and technology leaders backed the call for time to establish shared safety protocols.
Read the pause letter ↗AI leaders sound an alarm.
Sam Altman, Demis Hassabis, and Dario Amodei joined a separate statement calling for AI extinction risk to be treated as a global priority. This statement addressed risk; it did not call for the six-month training pause.
Read the statement and signatories ↗FAST-FORWARD TO THE PRESENTThe models
The models
keep changing.
What’s next
for us?
For me, the question is about my work, my skills, and the kind of engineer I want to become.
A PERSONAL PERSPECTIVE / THE PERSON BEHIND THIS STORYMy name is
My name is
David Okoye.
I’m a full-stack developer. But increasingly, I think that if I’m going to build a career in an industry changing this quickly, I need to think like a software engineer.
That means taking responsibility for the whole system: understanding the problem, choosing the trade-offs, and knowing whether what we built actually works.
THE CHISEL IS CHANGINGWe spent years
We spent years
mastering the tool.
Writing software once meant learning how to wield a very particular chisel. Languages, frameworks, and syntax demanded years of practice before an idea could take shape.
AI is abstracting more of that work. I can spend more of my attention deciding what to build, how its parts fit together, and whether the result serves the person using it.
WHAT THE WORK HAS ALWAYS BEEN ABOUTThe job was never
The job was never
just to code.
It was to bring an idea
from the human mind
into something that works.
The language still matters where it affects the system. But the outcome matters more than loyalty to a particular syntax.
01 / MY APPROACH
Understand what you ship.
AI can write the code. I still need to understand its behavior, question its assumptions, and explain the decisions behind it.
02 / MY APPROACH
Own the decisions.
Architecture, security, performance, and the user’s needs still demand judgment. Generating a solution is one step; knowing it is the right solution is another.
03 / MY APPROACH
Spend less energy on syntax.
I don’t need to obsess over every language’s punctuation to solve a problem. I need the foundations to read, test, debug, and improve whatever the system produces.
MY ANSWER IS STILL TAKING SHAPE.
I want AI to augment my ability
to turn ideas into reality.
David OkoyeFULL-STACK DEVELOPER / THINKING LIKE AN ENGINEER