Most U.S. AI debates focus on speed, scale, investors, and competition. I argue the more urgent issues are governance, labor displacement, surveillance, and how algorithmic systems are already reshaping public narratives and civic infrastructure, often without democratic accountability. Curious how others here see this, especially people building or deploying AI systems.
But I argue the more urgent issue is simple misunderstanding.
LLMs were designed to summarize web pages while sidestepping copyright issues --- they are fundamentally *language* prediction engines. You don't have to take my word for it, it is right in the name --- Large *LANGUAGE* Model.
But what people seem to be expecting is a *logical* deduction engine. This is a total misapplication.
Examples are posted here on a daily basis --- "Can AI run your business?"
Short answer --- no, not now and not anytime soon. Expecting LLMs to actually *understand* or *reason* or *make decisions* is a misapplication. And one that I expect will carry significant legal liability issues once this becomes painfully clear.
Agreed, about the widespread misunderstanding of what LLMs are/aren't. Treating probabilistic models as decision engines is an error, and I think you're right about liability issues.
Especially concerning that, even with those limitations, these systems are already being deployed inside workflows that influence decisions... I assume hiring (and maybe firing down the road?), content moderation, surveillance, targeting, triage. It's often done with the assumption that "the model is good enough" or that human oversight will catch errors later. In practice, that oversight tends to erode once systems scale.
Wondering how organizations are handling that gap: what the models can do vs. what they're implicitly trusted to do once integrated into real systems.
Similar to the way management handles most things (in the USA at least) --- by going with the flow until the error and pain of doing so becomes unbearable.
The asymmetry of it all rings true: the "unbearable pain" shows up first (and sometimes instantly) for people on the receiving end of automated decisions, long before it's felt by system deployers -- a CV silently filtered out by automation, or a person physically or virtually canceled by an automated drone. By the time incentives flip for deployers, the workflows are already entrenched.
Most U.S. AI debates focus on speed, scale, investors, and competition. I argue the more urgent issues are governance, labor displacement, surveillance, and how algorithmic systems are already reshaping public narratives and civic infrastructure, often without democratic accountability. Curious how others here see this, especially people building or deploying AI systems.
America's AI Debate Is Missing the Point
Absolutely.
But I argue the more urgent issue is simple misunderstanding.
LLMs were designed to summarize web pages while sidestepping copyright issues --- they are fundamentally *language* prediction engines. You don't have to take my word for it, it is right in the name --- Large *LANGUAGE* Model.
But what people seem to be expecting is a *logical* deduction engine. This is a total misapplication.
Examples are posted here on a daily basis --- "Can AI run your business?"
Short answer --- no, not now and not anytime soon. Expecting LLMs to actually *understand* or *reason* or *make decisions* is a misapplication. And one that I expect will carry significant legal liability issues once this becomes painfully clear.
Agreed, about the widespread misunderstanding of what LLMs are/aren't. Treating probabilistic models as decision engines is an error, and I think you're right about liability issues.
Especially concerning that, even with those limitations, these systems are already being deployed inside workflows that influence decisions... I assume hiring (and maybe firing down the road?), content moderation, surveillance, targeting, triage. It's often done with the assumption that "the model is good enough" or that human oversight will catch errors later. In practice, that oversight tends to erode once systems scale.
Wondering how organizations are handling that gap: what the models can do vs. what they're implicitly trusted to do once integrated into real systems.
Wondering how organizations are handling that gap
Similar to the way management handles most things (in the USA at least) --- by going with the flow until the error and pain of doing so becomes unbearable.
The asymmetry of it all rings true: the "unbearable pain" shows up first (and sometimes instantly) for people on the receiving end of automated decisions, long before it's felt by system deployers -- a CV silently filtered out by automation, or a person physically or virtually canceled by an automated drone. By the time incentives flip for deployers, the workflows are already entrenched.