For Nate Berkopec, language models are not on their way to general intelligence: they are useful where the result is verifiable, tooled, repeated — and placed in a loop.
In an essay published in May 2026, developer Nate Berkopec takes the opposite view of the dominant narrative about language models. His first observation concerns the form of the progression: despite exponential benchmark curves, the daily experience is one of regular and linear improvement, without interruption.
Above all, he challenges the vocabulary. To speak of "intelligence" is misleading because the capabilities of these systems are irregular—capable of progressing on Erdős problems while failing on elementary puzzles. This irregularity is not a transient defect to be corrected, it is the very texture of the technology.
What really works
Berkopec identifies a profile of successful uses: an objectively verifiable result, a massive use of external tools that compensate with computational power, a repeated application on a large scale rather than an isolated case, and the transformation of prototypes into stabilized software. His insistence is on the loop: "if what you do with an LLM is not a loop, you leave a lot of potential aside." A single call to a model is worth much less than an agent that iterates. One last caveat, less comfortable: the productivity gains obtained on code production require in return an additional requirement elsewhere in the development.
Read the full article on the original source
Go to the source →