Turner: Calling LLMs 'Just Autocomplete' Ignores Emergence

The tired shorthand for dismissing a frontier language model is some version of "it's just next-token prediction" or "fancy autocomplete." In a March essay on Paired Ends, computational biologist Stephen Turner takes that shorthand seriously enough to argue against it, and the piece is worth reading for anyone who has watched that line end a conversation at their own company.

Turner's move is not to deny the mechanics. He grants that a model is doing token prediction, the same way he grants that you are, in his phrase, "a wet bag of atoms." The point is that neither description is a complete account of the thing. He leans on the physicist Phil Anderson's 1972 paper "More Is Different", which argued that knowing the fundamental laws of physics does not automatically give you superconductivity or biology. Each level of complexity, in Anderson's telling, needs its own vocabulary and its own rules, even though those rules are compatible with the level beneath. Turner's line is that "'it's just autocomplete' is not a serious critique. It's a profoundly unscientific thought-terminating cliche."

Why this matters outside the argument itself is that the framing your team uses for what a model is quietly decides what you build, what you measure, and what you worry about. If a system is "just autocomplete," you do not staff an evals team for behaviors that were not in the training objective. If it is a complex system with emergent properties, you do. The same essay's usefulness cuts the other way for buyers being sold on a demo, because the honest version of the emergence argument is that surprising behavior deserves study, not that surprising behavior is proof of intelligence.

The honest caveat is that this is an opinion essay by a working scientist, not a benchmark or a paper. Turner does not offer a test that distinguishes real emergent capability from measurement artifact, and he does not take on the specific technical critics who argue emergence in LLMs largely disappears under smoother metrics. What the reporting does not give you is a decision rule for a product leader.

What it does give you is a cleaner way to argue. If the next time someone at your company reaches for "it's just autocomplete," you can point at Turner's brain-is-just-atoms line and reopen the conversation, that is a small but real upgrade in how the room thinks.