Source: AI Evolution Archive Parts 1 & 2
One of the most useful frameworks I encountered early in this course was what I’d call the bottleneck theory of AI progress. The idea is simple: every major era of AI is defined not by what it can do, but by what’s currently constraining it. And each breakthrough is essentially a constraint getting removed.
The AI Evolution Archive maps this across five variables: Representation (how the system encodes the world), Objective (what it’s trained to optimise), Data (what it learns from), Compute (what hardware scale is feasible), and Deployment (cost, latency, and human acceptance). Each historical paradigm maxes out a different combination of these variables and then hits a wall.
The first AI winter hit because the representation bottleneck was unsolved — symbolic AI could encode logic but couldn’t learn from data. The deep learning revolution happened when compute and data availability both unlocked simultaneously, enabling deep hierarchical representations. AlexNet’s 2012 ImageNet victory wasn’t just an algorithm win — it was the moment a benchmark result reorganised an entire field. The GPT revolution happened when scaling laws emerged: more data plus more compute equals qualitatively different behaviour.
The current bottleneck, by this framework, appears to be deployment: not making the models smarter but making them reliably act in the world over sustained periods of time. Agentic systems, harnesses, context engineering — all of these are responses to the deployment bottleneck.
I find this framework genuinely useful because it cuts through the noise. Instead of asking “is this new model better?” you can ask “which bottleneck does this address?” That’s a much more productive question.