Source: AI Revolution Detailed — MM6451 Transformative AI, Prof. Michael Xu
Technical AI papers can look intimidating from the outside. Lots of architecture names, benchmark tables, and acronyms behaving like they pay rent. But the useful question is usually simple: what bottleneck is this trying to remove?
AI has died twice and come back twice. Understanding the pattern of those deaths and resurrections is essential context for reasoning about the current moment. The first wave 1950s–1970s was driven by symbolic AI — the idea that intelligence could be encoded as explicit rules and logical inference. The first winter arrived when the limits of rule-based systems became apparent: they couldn’t generalise, couldn’t handle ambiguity, and couldn’t learn. The second wave 1980s was driven by backpropagation and expert systems. The second winter arrived when hardware proved inadequate for the algorithms, and expert systems proved brittle outside narrow domains. The deep learning revolution 2006–2012 was different because it coincided with the simultaneous availability of large datasets the internet and sufficient compute GPUs. AlexNet’s 2012 ImageNet result is the canonical moment: decisive superiority under visible, competitive conditions. The GPT revolution 2017 onwards followed from scaling laws: more data plus more compute produces qualitatively different capability, not just quantitatively better performance. The question for the current moment: are we in a genuine paradigm shift, or at the top of another wave? The current wave has generated economic value at a scale the previous ones didn’t, and the bottlenecks that ended the first two waves have both been substantially addressed.
My takeaway: architecture matters because it decides what kind of intelligence is affordable, reliable, and usable outside the lab. The benchmark is the beginning of the story, not the ending.