Martin Casado: AI Is the First Abstraction Layer to Take Over Human Thinking

"We've handed off resources before, but never thinking itself"

Casado pointed out that throughout the history of computer science, humans have never actually delegated reasoning or logic to a machine. Compute, networking, and storage were always "resources" — humans posed a high-level problem, and the machine computed the answer, but the work of understanding and judging that problem stayed on the human side. He says our relationship with AI today is different: we're now asking for answers before we've even fully worked out what the question is.

We know the destination, but not the road there

His go-to analogy is Makefile, the software build tool. With Makefile, you define what the final output should look like, and the system handles the rest of the process — and he says that's exactly what using AI feels like today. But there's a crucial difference. No matter how high you stacked previous abstraction layers, you could always trace back deterministically to see exactly what was happening underneath. With AI, he argues, when you ask a question and wait for an answer, you can't clearly map out how long the computation will take or what path it took to reach that answer. As Casado put it in the a16z interview, "You pray to the model with the right words, and eventually something useful comes out."

Half a century of stacking layers — why this one's different

Casado says his entire career has been a process of moving the computing stack up one layer at a time — programming languages on top of assembly, operating systems on top of hardware, the cloud on top of servers. But all of these layers shared one thing in common: no matter how high you climbed, you were still working with deterministic systems that could be mapped back down to the layer below. In his view, even the internet and search never really took the job of defining the problem away from humans. What's happening with AI right now, he keeps emphasizing, breaks from that pattern entirely.

Category	Traditional computing abstraction	AI abstraction (Casado's view)
What humans hand off	Compute, networking, storage	Reasoning and logic itself
Predictability of outcome	Deterministically traceable down to lower layers	Final state is known, but the path to it is not
Computation time	Generally estimable	Hard to bound
Analogy	Higher layers map precisely onto lower layers	Like Makefile — define the goal, and the system handles the rest

A similar question, seen from a different angle

Casado isn't the only one asking whether AI actually thinks. Mathematicians we've covered before have made a similar point. Fields Medal winner Timothy Gowers and mathematician Peter Sarnak have argued that LLMs are good at combining existing methods and exploring many paths at once, but lack the intuition to pick out productive paths within a vast search space. Google DeepMind's Tom Zahavy reached a similar conclusion in a paper arguing that "LLMs can't jump." Casado's idea of "praying and handing off thinking" and the mathematicians' description of "a calculator that can't generate new ideas" are, in the end, two different vantage points on the same phenomenon.

Editor's take

What's really interesting about Casado's argument isn't the conclusion that "this is different" — it's how he arrives at it. He apparently set out to explain AI as just the next natural layer on the computing stack, and talked himself into "no, this is actually different" partway through. Earlier layers — programming languages, the cloud, whatever — could always be traced back down to estimate computation time and path, no matter how high you climbed. The moment you hand work off to an LLM, he's saying, that traceability just breaks. That's not the kind of thing a VC says for effect — it's an observation that only someone who's actually built systems would make.

There's a concrete lesson here for anyone working with this stuff today. Plenty of teams are handing tasks off to AI agents and simply expecting the output to come out a certain way, but there's still no real way to control the computation time or the intermediate path it takes to get there. Hand a batch job to an agent, and you genuinely don't know if it'll take five minutes or fifty, or whether it'll wander off down some unexpected path along the way. For companies here adopting agentic AI for real work, the order of operations matters: don't just look at output quality — first decide how much of this path uncertainty you can actually tolerate. The more standardized and repetitive the task, the lower that risk; the more judgment a task requires, the higher it gets.

Expect this debate to grow louder in the coming weeks. As more agents start planning and executing tasks on their own, more people are likely to agree with Casado's core point: the outcome may be predictable, but the process to get there isn't.