Reality of "Artificial Intelligence"

Opinion

Reality of 'Artificial Intelligence'

Practically, what is AI? Silicon valley has a convenient answer: a powerful autonomous agent that will spell the end of human white collar work as we know it. A system that can reason, plan, and eventually outcompete humans across most cognitive domains. Dario Amodei, Anthropic CEO, in Machines of Loving Grace, explores this idea quite deeply. AI systems driving massive economic expansion by completely taking over knowledge work. It’s a compelling story. It’s also one that a lot of people have a strong financial incentive to believe. That alone doesn’t make it wrong. But it should make you skeptical.

This optimistic view is in large part backed up by the idea that these systems have reasoning capability in the same way a human does: stable internal world models, memory and experience, grounded understanding… I believe an observation of the confabulated, so-called “thinking” token output of any frontier LLM is enough evidence to see how fatal of a misnomer this is.

Be that as it may, language models are genuinely capable for the easy 80% of most tasks in many economically important domains like healthcare, management, and engineering. The problem is that the easy 80% was never the valuble part. The work that makes something correct, safe, useable lives in the last 20%, exactly where AI becomes unreliable due to but not limited by hallucinations, overconfidence, temporal staleness... These are embarassingly exposed, foundational issues with LLMs that make many of the grandiose promises of AI hollow. At least for now, and likely many years to come.

The truth is this: the transformer architecture is roughly eight years old. Companies building at the frontier have existed for only a few years. And yet these systems are being pushed into education, medicine, hiring, and governance at enormous scale. Historically, there is always a lag between invention and understanding. Electricity, for example, took over a century between discovery and commercialization. That lag is where society figures out how, when, where, and by whom a new invention should be used.

I propose that after striping away the marketing and getting a clear hold on the fundamentals of LLMs, two practical uses stand out.

First, AI is an extremely powerful interface to human knowledge. The sum of human writing is now explorable in natural language at a higher resolution than ever before. A doctor can surface relevant case studies in seconds. A student without access to a great teacher can get one on demand. This is not insignificant. It is a compression (but not deletion) of the access gap that has always existed between those with expert networks and institutional resources and those without. The value here is real and already being realized.

Second, for tasks that are well-bounded, simple, repetitious, rote, AI's potential is actually not fully realized. Drafting boilerplate, summarizing structured documents, generating first-pass code from well-specified requirements... These are tasks where the ceiling of required correctness is low enough that the 80% is actually sufficient. The economic value here is also real, but it is the value of, say, a faster typist, not a new colleague.

These two uses, knowledge retrieval and rote augmentation, are, I believe, the honest foundation of what AI can deliver today. And this is the line we build from: instead of selling autonomous agents replacing human judgment, we are developing tightly scoped, polished tools that make the capable more capable. This is a narrower claim than what the broader industry is making. It is also a true one.