Pacific Design/ artificial intelligence

section 15 · status: live · 6 entries · updated 2026-08-29

History of AI

The field that named itself at a summer workshop, died twice by its own promises, rebuilt itself as statistics, and detonated when three quiet ingredients finally met. Not nostalgia — a pattern library for reading the present.

live · seventy years swept in fifteen seconds — booms, winters, detonation

  1. The founding dream3 mina summer workshop names a field and mortgages its credibility
  2. Winters & thaws2 mintwice the field promised, twice the funding froze
  3. The quiet decades2 minstatistics, GPUs, and a million labeled images
  4. The deep-learning decade2 minfrom AlexNet to ChatGPT in ten compounding years
  5. The open-weights turn3 minthe release that split the field into two ecosystems
  6. Reading the present3 minusing seventy years of pattern to see this moment clearly

check yourselfAnswer before you open

Trying to recall something teaches it better than re-reading does. Have a go, then open the answer.

What actually caused the AI winters?

A repeating mechanism, not a refutation: demos generate extrapolated promises, concentrated funding follows, the demo-to-deployment gap accumulates quietly, an audit or market event snaps expectations to evidence, and concentrated funding exits together. Neural networks were 'refuted' in 1969 and run the world now. Winters & thaws →

Three ingredients met in 2012. What were they, and why is that the lesson?

GPUs from the gaming industry, ImageNet's million labeled images, and algorithms a small community kept refining through the unfashionable years. All three were public; the sum was not obvious. Revolutions assemble from unglamorous parts, and the tell is a benchmark gap too large to argue with. The quiet decades →

Is this moment a bubble?

The wrong question, and the site refuses to answer it for you. Run the mechanism: how large is the demo-to-deployment gap, how concentrated is the funding, how much is revenue versus expectation, and are the input curves still improving? Reasonable people disagree — but they should be disagreeing about measurements. Reading the present →

Why did the founders' 1956 optimism turn out to be a liability?

Because the conjecture that intelligence can be precisely described was productive, and the timelines attached to it were not. Promising human-level performance within a generation set the terms by which the field was later judged — and funding was withdrawn against those promises, not against the science. The founding dream →

What actually caused the deep-learning breakthrough of the 2010s?

Three long-running curves finally meeting: enough labeled data, cheap parallel compute in GPUs, and a stack of practical training tricks. The core ideas were decades old. That is the pattern worth carrying forward — the bottleneck is usually an input, not an insight. The deep-learning decade →