I’ve recently been helping deploy AI across organizations by streamlining existing workflows, surfacing new ones AI makes possible, and helping developers learn complex software systems fast enough to build effectively with new tools. In every case it’s the same loop; understand the system, then act on that understanding. LLMs make both halves faster.
Same loop, no brakes
Fundamentally though, we are still in a loop of building mental models and using those models to build products. A big difference is the rate at which the development loop happens (learning as well as implementation).
Faster velocity means more danger, specifically in contexts where there is a low margin for correction (e.g. delayed feedback, low reversibility and/or high consequence). For example, when driving a car, as speed increases, the distance required to recognize and correct an error grows disproportionately. The faster the system moves, the less room there is for feedback or judgement. Consequently, with driving, we have speed limits, which can be adjusted as feedback and control mechanisms adapt.
Velocity does not merely increase the rate of production. It increases the rate at which the state of the system changes relative to our ability to understand it. Changes interact with other changes. Errors compound. Feedback goes stale more quickly. When production outruns evaluation, risk accumulates. Combinatorial explosion.
LLMs are making artifacts abundant faster than they are making judgement abundant. Historically, the cost of producing software imposed its own constraints: we could only explore, build, and change so much at once. As those constraints weaken, the bottleneck moves from production to evaluation.
These tools help us go faster, but they are agnostic about direction.
Human speed bumps
The role of the human has become more consequential, not just at the organizational executive level, but at the individual level. We are making more choices, acting on them more quickly, and propagating their consequences further. Technical and organizational debt can accumulate before feedback has time to catch up. Processual friction creates waste, but it also provides involuntary deliberation time. LLMs can remove both, indiscriminately.
Traditional organizational structures implicitly contain low-pass filters. Ideas take time to move through the system. Some disappear, others encounter objections, and circumstances sometimes change before action is taken. This is enormously wasteful, but it also limits the rate at which decisions propagate. Removing friction increases organizational bandwidth where good decisions travel faster, but so do bad ones. Bureaucracy is, among other things, a largely unintentional rate limiter.
Some of this friction does more than just slow us down. The work itself forces deliberation–e.g. writing a proposal requires articulating the model, implementing a system exposes assumptions, explaining something to another person reveals gaps in understanding. These processes are comparatively inefficient forms of cognition, but they make deliberate, clear thinking difficult to avoid. As LLMs remove this kind of work, they can also remove the thinking that happened incidentally while doing it.
As involuntary friction disappears, we may need to become more deliberate about introducing our own.
Anti-lock brakes
Discernment, thinking and communication are how judgement stays survivable as production accelerates.
Discernment is deciding what deserves attention and judgement–e.g. being patient, sitting with a problem and the context, deferring judgement. Cultivating the ability to recognize quality, coherence, elegance (i.e. taste).
Thinking is operating on what you’ve discerned. Forming models, identifying assumptions, diagnosing symptoms, distinguishing evidence from inference and comparing alternatives.
Communication oftentimes means clear writing, but can involve other modalities for transmitting thoughts, ideas, moods, emotions or other relevant context. Communication is part of the feedback system. It makes mental models observable to other people.
LLMs fit nicely as part of these processes, but allowing them to lead risks premature pattern matching and the outsourcing of judgement to the model. Responsibility cannot be outsourced. LLMs can dramatically expand the space of possible interpretations, designs, and actions, but they cannot relieve us of the responsibility to choose among them.
As the space of possibilities grows, discernment and judgement become more important, not less..