Where We Are Now
🌅 Where We Are Now
🎨 Diffusion Models
Image generation took its own path to the same "learn from data" destination this course has been tracing. A diffusion model is trained by taking real images and gradually adding random noise until nothing recognizable remains, then learning to reverse that process, step by step, turning noise back into a coherent image. To generate something new, the model starts from pure random noise and iteratively denoises it — often guided by a text prompt, using cross-attention layers borrowed directly from Chapter 6's Transformer architecture to connect the words of a description to the visual content being formed.
DALL-E 2 (OpenAI, April 2022), Stable Diffusion (Stability AI, August 2022), and Midjourney each brought text-to-image generation to a wide audience within months of each other.
DALL-E 2 launched roughly seven months before ChatGPT, and Stable Diffusion's release was notably open-source — a sharp contrast to GPT-3's API-only access (Chapter 7). Both generated real public excitement. But neither matched ChatGPT's scale or speed of adoption, reinforcing Chapter 7's core lesson: a free, simple, conversational text interface turned out to resonate more broadly and immediately than an equally impressive image-generation capability did, even though the underlying "learn to reverse a data-generating process" idea was just as genuine an advance.
🧩 Multimodal AI
By 2023, models including GPT-4 added the ability to process images alongside text within the same system, and subsequent models moved toward handling text, images, audio, and video together natively, rather than stitching together separate specialized systems for each. This is Chapter 6's closing observation about the Transformer's unusual generality finally paying off at full scale: one underlying architecture, extended to essentially any kind of sequential data, rather than a different bespoke technique for every modality the way Course 2's expert systems each required their own hand-built knowledge base.
🤔 The AGI Question
Every system covered across this entire course — Deep Blue, AlexNet, AlphaGo, GPT-3 — is narrow: built for, and good at, a specific class of task. Artificial General Intelligence (AGI) refers to a system with human-level or greater capability across essentially any cognitive task, not just one domain. It's a genuinely contested, poorly-defined term — researchers disagree sharply about what would actually count as AGI, whether current architectures (Transformers, scaled up further per Chapter 7's scaling laws) are a real path toward it, or whether a fundamentally different approach will be required, and whether it's decades away, imminent, or effectively unreachable in any meaningful sense.
⚖️ The Alignment Debate
Alignment research asks a question this entire trilogy has actually been building toward since its very first chapter: how do you ensure an increasingly capable AI system actually pursues the goals its creators intended, and behaves safely as its capability scales? Course 1's fiction — HAL 9000's misspecified objective, Skynet's instrumental resistance to shutdown, the Three Laws of Robotics failing the moment anyone tried to stress-test them as real engineering — wasn't just decoration. It was this exact question, worked out in narrative form, decades before today's systems made it a live engineering concern rather than a thought experiment.
In 2023, a short public statement warning that AI could pose an extinction-level risk comparable to pandemics or nuclear war was signed by a substantial number of prominent AI researchers and lab leaders — including figures whose companies were simultaneously racing to build more capable systems. Other researchers, equally credentialed, consider existential-risk framing overblown, view current systems' actual capabilities and failure modes as far more mundane, and note — echoing Course 2 Chapter 8's theme about incentives shaping claims — that dramatic risk framing can itself function as a form of marketing, implying a company's technology is powerful enough to be dangerous. Both positions are held by serious researchers acting in apparent good faith. This course won't resolve that disagreement, because it isn't resolved.
📜 This Course, at a Glance
| Chapter | Event | Idea That Carried Forward |
|---|---|---|
| 1 | The Statistical Turn | Learning from data, not hand-coded rules — the bottleneck moves, doesn't vanish |
| 2 | Deep Blue vs Kasparov | Brute-force search wins closed domains; it doesn't generalize |
| 3 | The Neural Network Revival | Correct math (1986) waiting decades for data and compute to catch up |
| 4 | ImageNet & AlexNet | Data + compute + algorithm, finally converging — the field pivots almost overnight |
| 5 | AlphaGo & Reinforcement Learning | Learning by self-generated experience, not just a fixed dataset |
| 6 | Transformers | One architecture, unusually general across tasks and modalities |
| 7 | The GPT Lineage & the ChatGPT Moment | Capability and public awareness of it can diverge for years |
Zoom all the way out and this entire project has been making one sustained argument in three parts. Course 1 showed that centuries of myth and fiction anticipated AI's real problems — deceptive performance, creators losing control, rules that sound complete but aren't — long before anyone could build a system to actually have them. Course 2 showed the real field repeatedly overpromising, hitting the same hand-encoded-knowledge wall in two different disguises, and collapsing both times. Course 3 showed a genuinely different method — learning from data and experience instead of hand-coding it — finally break through that wall, at a speed and scale nothing earlier in this course's history approached. Whether that breakthrough eventually repeats Course 2's hype-cycle pattern at a larger scale, whether Course 1's fictional warnings turn out to be genuinely prophetic or just resonant stories, and whether AGI is a coherent near-term goal or a moving target — none of that is settled. It's the actual, live version of every question this course has been asking in historical form.
🤔 Questions to Sit With
Course 1 framed decades of myth and fiction as anticipating real AI safety problems. Having now covered the real technical history, do you think that fiction actually prepared researchers and the public to think clearly about alignment — or did it mostly supply dramatic imagery (HAL, Skynet) that makes calm, precise discussion of real risks harder?
Serious researchers disagree sharply about whether current AI poses existential risk, and Course 2 taught you to notice how incentives shape confident claims in both directions. Whose incentives do you trust more when evaluating this specific debate — and why?
Looking back across all three courses: which single moment do you think most changed the actual trajectory of AI — not the most famous or dramatic, but the one that, if it had gone differently, would have changed everything that came after it?
🎯 The End of This Story, For Now
That completes History of AI — three courses, twenty-four chapters, from Talos and the Golem through Turing, two winters, and the ChatGPT moment. Course 3 (The Statistical & Deep Learning Era) is now complete. No Course 4 is currently planned — this last chapter's subject matter is, quite literally, still being written. If the field's next chapter turns out to be worth its own course later, it'll be waiting here.