Move 37 Explained (Siraj Raval)

Where Paul Roetzer's 2025 keynote asks what Move 37 means for your job, this 2018 video asks how the machine did it — and lands, characteristically for the era, on optimism. It's a useful primer with a couple of caveats worth flagging up front.

The verdict

A clear, fast, genuinely useful mechanics explainer — CNN + tree search, then three hopeful applications. It's also a time capsule: the 2018 techno-optimism ("automation will free us into jobs we enjoy") reads very differently after the 2025 CEO-driven wave. Watch it to understand how AlphaGo worked; pair it with a 2025 source for what it costs. One factual slip and the creator's later credibility issues (below) mean: verify, don't cite.

Title card: Move 37 Explained — 0:00 The framing: "three reasons why it's so significant for our future — in terms of human jobs, health, and lifestyle."

§1 · The mechanicsHow AlphaGo actually did it

The strongest part of the video is the plain-English machinery. Go is hard because it's vast:

"A very challenging game, with more potential board positions than there are atoms in the universe."Siraj Raval — 0:43

Slide: AI experts thought a computer couldn't beat a human at Go until the year 2100 — ~0:35 The stakes: experts polled in the '90s guessed a machine Go champion was **decades — even a century — away**.

His explanation of the method is accurate and tidy: a Convolutional Neural Network trained on ~30 million expert moves to predict good moves and evaluate board positions, wrapped inside a Monte Carlo Tree Search that explores promising branches instead of brute-forcing them.

Diagram: How AlphaGo chooses its next move — Monte Carlo Tree Search — ~2:06 The neural net guides *which* branches of the game tree to search, and scores the positions it finds — the core of AlphaGo's play.

Factual note. Raval repeatedly says "AlphaGo Zero beat the world champion." That's wrong: the program that beat Lee Sedol in 2016 was AlphaGo (trained on human games, as he describes). AlphaGo Zero came later (2017), learned purely from self-play with no human data, and never played Sedol. The mechanism he explains is right; the name is conflated.

§2 · The move"Different, and better, than human intuition"

On Move 37 itself, the video's best line captures exactly why it mattered:

"Somehow, a computer program knew something about the game that we didn't. Somehow, its intuition was both different, and better, than human intuition."Siraj Raval — 3:04

And — like Roetzer — he doesn't stop at Move 37. He gives Lee Sedol his comeback, Move 78, the "Divine Move," AlphaGo's own 1-in-10,000, and uses it to define his central concept:

Board diagram: AlphaGo appears confused by Lee's innovative strategy, Move 78 — ~3:58 Move 78 — AlphaGo "plays poorly, ceding the whole right side." Human intuition finding the machine's blind spot.

"This was an example of intelligence augmentation: better algorithms lead to better performances… humans who make better decisions can create environments where algorithms fail — 'divine moves' — and learning about these failures helps us design better algorithms."Siraj Raval — 4:18

Diagram: intelligence augmentation feedback loop — ~4:26 The thesis in one loop: better AI ↔ more powerful ways of thinking ↔ better cognitive tools. Human *plus* machine, each improving the other.

§3 · The optimismThree places augmentation pays off

The back half is a tour of upside. It's where the video is most dated — and most interesting to re-read now.

Healthcare. Drug discovery is slow and brutally expensive — "at least ten years… an average of $2.6 billion… less than 12% of candidates make it through Phase 1." His fix: generative models proposing molecules for humans to test.

Diagram: drug discovery funnel from discovery to FDA approval — ~5:50 The pipeline the augmentation is meant to compress — thousands of candidates down to one approved medicine.

Slide: Generative Adversarial Networks generating molecular structures — ~6:18 GANs as the "molecule generator." *2026 note:* the real drug-discovery breakthroughs came from AlphaFold and diffusion models — the instinct was right, the specific tool aged.

Slide: Diagnosis by AI, 92.5% correct — ~6:46 AI as a "second diagnosis" alongside the doctor.

Software design. His pitch — "give high-level input to our machine, and it decides on the implementation details by itself" — is essentially a 2018 description of what coding agents became. He cites Google AutoML and, more speculatively, AI-designed blockchain consensus.

Wheel diagram: different types of consensus algorithms — ~8:10 The blockchain-consensus tangent — the weakest, most of-its-moment segment.

Screenshot: Google Magenta music demo — ~9:06 Google Magenta — AI as a creative *duet* partner, not a replacement. The "greater than the sum of its parts" framing.

A personal assistant. The most prescient and the most unrealized: a local, transparent, privately-trained assistant.

Slide: average human attention span dropped from 12 to 8 seconds — ~9:34 The problem it solves — attention scarcity and algorithmic manipulation. (The "12→8 seconds" stat is itself a debunked myth, worth noting.)

"A personal AI assistant, stored locally, trained on our data, with a transparent backend… could know us better than any human could — from our web browsing history to our heart rate."Siraj Raval — 10:07

In 2026 this is still mostly aspiration: the assistants that arrived are cloud-based and ad-adjacent, not local and transparent. He described the thing we still don't have.

§4 · The closeChina, and a very 2018 ending

Document: China's 2030 AI plan — ~10:38 Move 37 as catalyst for China's national AI plan — "to become the world leader in AI by 2030." This one has aged into a live geopolitical story.

"Automation technology will help free humans from labor-based jobs, and create new classes of jobs that we actually enjoy doing."Siraj Raval — 11:18

That's the line the seven intervening years press hardest on. It may still come true — but Roetzer's keynote is 42 minutes of evidence that the transition is the hard part, and this video waves past it.

Two takes on one move2018 vs. 2025

Watched together, the pair is the whole arc of the AI-and-work conversation:

Siraj (2018) — mechanism and optimism. Here's how the machine did it, and here's the abundance it unlocks. Strong on the "how," light on the cost.
Roetzer (2025) — meaning and urgency. Here's what it does to 100 million jobs, and here's the choice leaders face. Strong on the "so what," anchored in economics.

Same move, same Move 78 comeback, opposite altitudes. The honest read sits between them: intelligence augmentation is real (Siraj), and the market may reach for replacement before augmentation (Roetzer).

Contact sheet of the video's key visuals — The video's key screens at a glance — title, the mechanics, Move 78, and the three applications.

One more caveat. Siraj Raval was a popular AI educator in 2016–18; in 2019 he faced well-documented controversies (a paid course that drew mass refund demands, and a plagiarized research paper). None of that is in this 2018 video, but it's a reason to treat its claims as a starting point to verify, not a citation.