Qualcomm’s new Elite sound chip is aimed at a problem most people assumed was already solved: real Wi-Fi-grade audio in a pair of wireless earbuds, without the latency and dropout that Bluetooth has never fully shaken. It is not a flashy launch. No keynote, no head-to-head slide against a rival’s flagship. Just a chip that lets two earbuds hold a stable connection to each other and to a phone over Wi-Fi instead of a lower-bandwidth radio protocol designed in the 1990s. If it does what Qualcomm claims, the fix for years of skipping audio and lopsided battery drain was never a software patch. It was better plumbing.
That is usually how progress actually shows up: not as an announcement, but as an unglamorous chip nobody outside a supply chain newsletter will mention again for another year.
Which makes it worth sitting with a study making the rounds this week: a survey of AI researchers found that even people whose job is forecasting the field’s pace have been badly undershooting it, repeatedly, for years running. These are not casual observers. They are people close to the compute, the benchmarks, the training runs. And the finding is that their predictions have consistently lagged reality, not because the future was unknowable, but because they kept extrapolating from what was already visible.
What Gets Measured Isn’t What Moves
Here is the mechanism worth naming: forecasters, like everyone else, anchor on signals built to be seen. A splashy model release. A funding round. A benchmark chart with a steep line on it. Those function the same way a diploma on the wall or a logo on a bag does: costly, visible, and deliberately legible as proof of something. Companies spend enormous effort making sure you notice them, because being noticed is the entire point of putting one out.
But the actual curve of what a technology can do rarely bends because of the thing you were told to watch. It bends because a chip got 20 percent more efficient, because a memory bottleneck got re-architected without a press release, because inference got cheaper by a factor nobody bothered to announce. Qualcomm’s earbud chip is a small, literal version of the pattern running underneath AI all year: the visible signal is the frontier model announcement; the real driver is the infrastructure beneath it getting incrementally and compoundingly better, on a schedule nobody forecasts because it doesn’t photograph well.
That is why the experts undershoot. A forecast built on legible signals will always lag a curve driven by illegible ones. You can track every major model release for five years and still miss the field’s real trajectory, because the releases are lagging indicators of infrastructure that improved months or years before anyone announced it.
The Honest Version of Surprise
None of this means forecasting is worthless, and it would be a cheap conclusion to say the experts are simply fools. The more useful read is narrower: the people closest to a fast-moving system are also the people most anchored to its recent, visible past, because that is the only data they have direct access to. Being embedded in a field sharpens your read on what already happened and dulls your read on what happens next, precisely because the next thing tends to come from a place nobody had a camera pointed at.
The practical takeaway is not “distrust the experts.” It is “distrust the part of any forecast built entirely from what was loud.” A model that only updates on keynotes, benchmark charts, and funding headlines is measuring status signals, not the underlying capability curve underneath them. The people who catch a real shift early tend to be the ones paying attention to boring things: a chip vendor solving a connectivity problem nobody was asking about out loud, a cost curve bending in a spreadsheet nobody is screenshotting, a supply chain re-optimizing itself in the background.
Qualcomm’s earbud chip will not change anyone’s AI timeline. That is exactly the point. The things that eventually do change those timelines almost never announce themselves as such. They arrive first as small, well-made pieces of infrastructure that nobody outside the supply chain notices, and only later, once enough of them have stacked up, as a headline everyone insists they saw coming.
Leave a Reply