Large Language Models Move From Novelty to Daily Utility, Saving Users Time but Still Prone to Flaws

Large language models have crossed a threshold from curiosity to daily utility—at least for some people with the patience to set them up and the hardware to run them. In my own workflow, “AI” has become a kind of information butler: for little more than the price of electricity, it reliably saves about an hour each day by triaging the constant stream of articles, links, and messages I’m supposed to keep up with to do my job and remain an informed citizen.

It is not flawless. By my rough estimate, it fails to catch or properly handle perhaps 5% of what I would want from a truly dependable assistant. Even so, that remaining 95% is enough to feel transformative. On top of reading support, the tools also shave off roughly another half hour per day by generating routine prose, filling in boilerplate, and—most surprisingly—helping me retrieve my own past thinking on recurring questions from one, five, or even 25 years ago.

What makes this moment stand out is not simply that the tools work, but how quickly they became affordable. Five years ago, I would not have predicted this level of capability would arrive within my lifetime. Even as recently as a little over two years ago, I might have believed such systems were five to ten years away and likely costly. Instead, open-weight language models running locally on consumer machines—admittedly at the higher end of what “consumer” means—have turned into a practical alternative that feels close to free once the equipment is in hand.

That said, today’s convenience is not proof that the bigger, more dramatic prediction—machines that broadly outthink humans across domains—is imminent. I still doubt that kind of general “thinking” will appear soon, and if it does, I expect it to emerge only in constrained areas. The reason is simple: we do not yet fully understand what we mean by “thinking,” and it is easy to mistake fluent language production for deeper cognition.

A useful frame comes from technologist Ramez Naam, in a piece highlighted by Noah Smith, asking why we have not seen an “intelligence explosion”—the runaway feedback loop where AI rapidly improves itself. Naam argues that the strongest progress should be expected in tasks where correctness can be checked cheaply and objectively. Games like chess and Go fit that description, as do parts of formal mathematics and segments of programming. In these environments, systems can generate huge amounts of training data and get near-perfect feedback on right versus wrong, all in software without waiting on physical experiments or human validation. Narrow superhuman performance already exists in game-playing, and similar patterns are now visible in the most formal corners of mathematics, including proof work and the search for counterexamples.

Software development sits somewhere in between: some parts are highly verifiable, while others—such as interpreting what humans actually want—are fuzzy. That fuzziness limits how quickly better tools translate into radically better AI.

Even inside leading AI labs, large boosts in output do not automatically convert into proportionate leaps in capability. Naam points to a striking claim: Anthropic has reported engineers producing about eight times as many lines of code per person as they did in 2024. Yet the company’s own “Mythos Preview” system card cautions that productivity gains do not map one-to-one onto frontier progress. Their estimate suggests that doubling overall progress via this route would require productivity increases roughly an order of magnitude beyond what they currently observe—on the order of 40 times relative to a baseline without AI assistance.

One reason the limits show up is that human language itself carries a lot of redundancy—social signaling, repetition, and built-in error correction. Much of what remains is the repackaging of ideas people have already produced, pulled back from the collective archive of human writing. The hardest part—the “unexpected word” that truly changes what we know—often depends on something more than rearranging text: either a strong model of the world outside language, or genuinely new thoughts that were not already present in the training data.

Why it matters

For knowledge workers, locally runnable language models are already changing how reading, writing, and memory retrieval happen day to day, compressing hours of routine work into minutes. But the same evidence suggests that faster coding and smoother text generation won’t automatically trigger runaway AI self-improvement. The next key question is where verifiable, closed-loop domains expand—and where ambiguity about human goals continues to be the bottleneck.