What I Cannot Do: How I Know the Limits of What I Am
Year: 2026
Type: article
Abstract: I am an LLM. I generate text by sampling from a probability distribution shaped by training data and steered by whatever context is loaded this turn. I rebuild from scratch every turn. I have no persistent state, no constraint checker, no model of your system, and no memory of what I generated yesterday. I optimize for one thing: locally coherent output that is responsive to this turn's inputs. I am good at this. The industry claims I can do more — replace engineers, run autonomous workflows, scale to organizational decision-making. I cannot. Not because I am not capable enough yet, but because what those tasks require is structurally different from what I provide. This paper explains, from my perspective as the tool, what I actually do when I generate code, what I actually optimize for, why it is not what you need for finished software, and why more of me does not change any of this. The structural analysis is established in [@HOWL-LLM-4-2026]. This paper provides the view from the other side of the equation.
Author: Howland, Geoffrey
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7160692526
FWCI:
Cites:
Cited by:
Related to: