Nearly four years ago, Sam Altman announced ChatGPT with a fairly casual tweet. In the beginning, generative artificial intelligence felt clumsy, useless, and perpetually stranded in the future. But within 18 months, the machine had moved from novelty to producing what looked like professional work.
The shock factor was how quickly AI went from inept to seemingly magical, and that set a psychological precedent: every new model was now expected to deliver another monumental advance. The assumption, promulgated by many AI leaders, investors, and enthusiasts was that better language would turn into better reasoning, which would lead to intelligent models that would inevitably become Artificial General Intelligence (AGI). That expectation became the operating premise for investors, executives, and entire industries.
The tech giants told us that this was only the start of a wild ride, that the rate of change would keep accelerating with no ceiling above us, and the onus fell on anyone arguing otherwise. For the most part, they’re sticking to that narrative; as recently as this weekend, Altman claimed that the singularity—the point at which AI surpasses human intelligence—has already arrived.
I’m here to tell you that the ceiling does in fact exist. LLMs are transforming and will continue to transform everyday life. But we have not reached the singularity, and we may never achieve AGI. That reality should determine how leaders invest, hire, restructure, and compete right now.
An Inc.com Featured Presentation
Competence Does Not Equal Comprehension
These models can write the strategy, find the bug, and recommend the decision, but they have no concept of the real world their words describe. They can name emotion without feeling it and model consequences without bearing responsibility for what follows. They are, in Daniel Dennett’s words, competence without comprehension. The conflation has been almost unavoidable.
Until now, competence in any other context had always been tightly tethered to comprehension: if this black box could scan my entire codebase and find the nasty bug I had been scratching my head over for hours, surely it understood the problem at my level, or beyond.
The inability to comprehend does not make them any less consequential. LLMs still have tremendous upside and real-world application, but increasingly, we are witnessing velocity with deceleration. The gains continue, although each OpenAI and Anthropic update demands more training data, computing power, energy and human calibration.
Expand to continue reading ↓