TheSequence Opinion #904: The Age of Research Is Overrated. AI Engineering Is Winning
Substack Jesus Rodriguez
Ilya Sutskever says AI has entered a new 'age of research.' TheSequence argues it's actually engineering doing the heavy lifting, not new architectures.
Based on reporting by Substack, Jesus Rodriguez — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Ilya Sutskever has a tidy story about how AI got here: research ruled from 2012 to 2020, scaling took over from 2020 to 2025, and now we're back to research, just with much bigger computers behind it. It's a clean narrative. It's also a little too clean once you actually open the hood of a 2026 frontier model.
Pop that hood and you'll find a Transformer, probably wired through a Mixture-of-Experts setup, looking almost identical to what shipped years ago. The real gains this cycle aren't coming from some new neural architecture nobody's seen before. They're coming from better training data, longer context windows, reinforcement learning tuned harder, synthetic tasks built to stress specific weaknesses, tool use, memory systems, verification layers, adaptive reasoning budgets, and agents that orchestrate other agents. None of that is a new species of network. It's plumbing.
TheSequence reaches for a Formula 1 comparison, and it lands well. An F1 car still has four wheels and an engine, same as it did decades ago. But races are won now through aerodynamics, energy recovery systems, tire compounds, live telemetry, and pit-wall software calling strategy in real time. The chassis is almost a formality. The system wrapped around it does the winning. AI in 2026 looks a lot like that: the Transformer is the chassis, and everything determining who's actually ahead is the surrounding machinery.
So is this the age of research or the age of engineering? The honest answer, per the piece, is that it's both at once, because the research has quietly turned into industrial-scale engineering. Scaling didn't stop, in other words. It just stopped looking like scaling and started looking like a thousand smaller optimizations stacked on top of each other, which is arguably harder to hype but just as consequential.
That reframing matters because it changes what counts as a breakthrough. If the next leap comes from RL curricula and tool orchestration rather than a fundamentally new architecture, then the competitive edge shifts toward whoever can engineer these systems at scale, not just whoever publishes the flashiest paper.
My take — AI-written commentary, not fact-checked reporting
I'll take the F1 metaphor and run further with it: the chassis stopped being the interesting part years ago, and anyone still hunting for the next 'new Transformer' moment is watching the wrong race. The actual story is that engineering discipline, at obscene scale, has become indistinguishable from research, and pretending otherwise just flatters the labs that want credit for 'science' when what they're really doing is world-class systems work.
Read more about this at: Substack