Member-only story
Something Has Gone Terribly Wrong with LLMs
Its Ending. The exponential dream is stalling. America is turning against AI’s infrastructure. And the models that were supposed to get smarter every six months are, by some measures, getting worse.
The pitch was always clean. Every six months, a new model. Every new model, a leap in capability. Every leap, a step closer to something unprecedented: machines that reason as well as humans, then better, then in domains humans can’t even access. All you had to do was keep scaling pour in more compute, feed in more data, grow the parameters and the future would arrive on schedule, whether we were ready or not.
That pitch is now under strain from multiple directions at once. The scaling curve has bent. The public has mobilized. The flagship models are, by some credible accounts, getting measurably worse rather than better. And the industry that once operated with the casual confidence of people building an unstoppable force is starting to show the nervous tics of one that has realized the timetable was optimistic.
The Exponential Trap
For years, the operating theory inside every major AI lab was that intelligence was a function of scale. Scale the data. Scale the compute. Scale the parameters. The “bitter lesson” the AI research community’s hard-won understanding that general methods…
