Pinned post
The most exciting part of electrical and computer engineering to me was always the prospect of designing tools that adapt to us as we adapt to them.
Some of this fascination is probably due to my background in neuro, anthro, and cybernetics, but mainly I’m fascinated by how this evolution both vastly magnifies the tools’ capabilities, while also making their effective design more akin to designing effective social systems — since the machine becomes more of an actor (now ‘agent’) than tool in the traditional sense.
As the machine becomes more adaptive, effective design mandates shifting away from classical approaches to human factors design (which focus on either the tool or user experience in ~isolation) to the study and optimization of the tool-and-user as a single, deeply-integrated complex system. LLMs demand a more radical adoption of this mindset than previously, bringing concepts from the forefront of systems engineering back into the initial design of tools themselves, in ways that are quite novel. Because of the high variety of the LLM and open-ended nature of prompting, optimization of user behaviour is as much a part of the design of effective LLMs (and LLM-based tools) as the tweaking of the model itself.
Yet this area (integrated human/AI systems design) is as-yet woefully understudied and underemphasized. Many of the system design approaches (both in terms of UI/UX, but also fundamental engineering) are not just directly ported from deterministic systems, but increasingly from decades-old deterministic systems like command prompts. I believe this is a serious miss, and is actively constraining both successful adoption and maximization of the potential of the tools as they exist today.
Some of this fascination is probably due to my background in neuro, anthro, and cybernetics, but mainly I’m fascinated by how this evolution both vastly magnifies the tools’ capabilities, while also making their effective design more akin to designing effective social systems — since the machine becomes more of an actor (now ‘agent’) than tool in the traditional sense.
As the machine becomes more adaptive, effective design mandates shifting away from classical approaches to human factors design (which focus on either the tool or user experience in ~isolation) to the study and optimization of the tool-and-user as a single, deeply-integrated complex system. LLMs demand a more radical adoption of this mindset than previously, bringing concepts from the forefront of systems engineering back into the initial design of tools themselves, in ways that are quite novel. Because of the high variety of the LLM and open-ended nature of prompting, optimization of user behaviour is as much a part of the design of effective LLMs (and LLM-based tools) as the tweaking of the model itself.
Yet this area (integrated human/AI systems design) is as-yet woefully understudied and underemphasized. Many of the system design approaches (both in terms of UI/UX, but also fundamental engineering) are not just directly ported from deterministic systems, but increasingly from decades-old deterministic systems like command prompts. I believe this is a serious miss, and is actively constraining both successful adoption and maximization of the potential of the tools as they exist today.
sanity (accelerated improvement demands co-research between experienced human experts and machine technicians) prevails againView quoted post on X
every burst of sanity brings us closer to a soft landing