Research · AI Studies
A training technique that matches models three times its size, a global adoption curve that's breaking records, and a talent pipeline that's quietly reversing. Three findings, one uncomfortable pattern underneath them.
Photo by Adrien on Unsplash
Most months, the AI research that actually matters and the AI research that gets attention are two different piles. This month they overlapped more than usual. One finding is a genuinely clever piece of engineering that could change what runs on your phone by next year. The other two are less flattering: a global adoption boom that's arriving faster than the internet did, sitting right next to evidence that it's arriving very unevenly, and that the country doing the most AI investing is quietly losing the researchers who'd build the next generation of it.
A training trick that makes small models punch above their weight
Start with the genuinely good news. Research presented at this year's International Conference on Machine Learning introduced something called selective activation sparsity — a training method that teaches a model to only switch on the parts of itself that are actually relevant to whatever task it's handling in that moment, instead of lighting up the whole network for every request.
The headline result is the kind of number that's easy to skim past and shouldn't be: on reasoning benchmarks, models trained this way performed comparably to models roughly three times their size. Not close. Comparable. That's not a marginal efficiency tweak, it's the sort of result that, if it holds up outside the lab, changes the cost math for running genuinely capable AI on hardware that isn't a data center.
Photo by Daniel J. Schwarz on Unsplash
Why should a freelancer or a small team care about a training technique they'll never touch directly? Because this is exactly the kind of research that quietly shows up eighteen months later as "why does my phone suddenly run a decent AI assistant without a subscription." Efficiency research doesn't make headlines the way a new flagship model does, but it's arguably done more to expand who actually gets to use capable AI than any single product launch this year.
Generative AI adoption is breaking records, and also breaking along predictable lines
Here's where it gets more complicated. Stanford's widely-cited AI Index report, released this year, found that generative AI reached 53% population adoption within roughly three years of ChatGPT's launch — faster than the personal computer managed it, and faster than the internet itself. On its own, that's a remarkable adoption curve. Read the next line of the report, though, and the picture gets murkier.
Adoption correlates strongly with GDP per capita, and the pattern isn't subtle. Singapore sits at roughly 61% adoption. The UAE comes in around 64%. The United States — home to the companies building most of these models — ranks a startling 24th globally, at just 28.3%. Whatever "AI boom" means in the boardrooms of the labs building this technology, it apparently doesn't automatically translate into everyday adoption in the country doing the building.
Photo by Nejc Soklič on Unsplash
The US is investing more in AI than anyone — and losing the researchers who build it
This is the finding we keep coming back to, because it cuts against the story most people assume is true. The same Stanford report found the US led the world in AI entrepreneurial activity in 2025, with 1,953 newly funded AI companies — more than ten times its closest competitor. By any measure of capital and company formation, the US is not just leading, it's lapping the field.
And yet, over that same stretch, the number of AI researchers and developers relocating to the US has dropped 89% since 2017, with an 80% decline in the last year alone. Read that twice, because it's a genuinely strange combination: record AI investment, record company formation, and a research talent pipeline that's reversing at the same time. Money is flowing in. People, apparently, are not — or at least, not at anywhere near the rate they used to.
Photo by Wonderlane on Unsplash
We won't pretend to have a clean explanation for this one. Visa policy, competing offers from labs elsewhere, researchers choosing to build closer to home now that funding exists in more places than it used to — probably some mix of all three, and probably a few reasons nobody's tracking yet. What's not in dispute is the direction of the trend, and a talent pipeline doesn't reverse for one quarter and then snap back. If this holds, it's a multi-year story, not a blip.
This month's findings, at a glance
What actually connects these three findings
On the surface, an efficiency paper, an adoption survey, and an immigration statistic don't have much to do with each other. Sit with them for a minute and a pattern emerges anyway: the technical side of AI is getting more efficient and more capable at a genuinely fast pace, while the human side — who gets to use it, and who gets to build it — is moving in a much messier, less predictable direction.
Efficiency research like selective activation sparsity is, in a real sense, a partial answer to the adoption inequality problem. If capable AI can eventually run on a mid-range phone instead of requiring a data center subscription, that does more to close the GDP-correlated adoption gap than any single company's pricing decision. It won't happen overnight, and it won't happen evenly, but it's the most concrete technical lever anyone currently has pointed at the problem.
What this actually means for you
If you're running a small business or working independently, none of this changes what you should do this week. It does change what's worth watching over the next year.
Photo by Kaitlyn Baker on Unsplash
- Watch for efficiency-driven price drops, not just new model launches. Research like this typically shows up in production pricing 12 to 24 months later. If your AI costs feel high today, that number has a decent chance of dropping meaningfully within the next year or two.
- Don't assume "most advanced" means "most widely used" in any given market. Adoption figures vary enormously by country and income level. If you're selling into a specific region, check actual local adoption data rather than assuming global AI hype translates evenly everywhere.
- Keep an eye on where AI talent and AI investment are actually landing. A widening gap between where the money goes and where the researchers go is the kind of structural shift that eventually reshapes which companies and countries lead the next wave of tools you'll be choosing between.
Frequently asked questions
Will selective activation sparsity actually make it into consumer products?
It's a genuinely promising early result, but plenty of promising training techniques never make it out of the research stage at scale. Worth watching, not worth assuming as a given yet.
Why does the US rank so low on generative AI adoption despite leading in AI investment?
The research doesn't give a single definitive cause, but adoption correlating with GDP per capita cuts against a simple "US leads in everything AI" narrative. Investment and everyday usage appear to be genuinely separate metrics.
Is the drop in AI researchers moving to the US a policy story or an economic one?
Likely both, and the available data doesn't cleanly separate the two. What's clear is the trend's direction and its multi-year persistence, not a single tidy explanation for it.
The bottom line
The technology is getting better and cheaper to run, genuinely and measurably. Whether that translates into a more evenly distributed AI boom is a separate question entirely, and this month's research suggests the honest answer is: not yet, and not automatically. The efficiency gains are the most hopeful lever available for closing that gap — but a lever only works if someone actually pulls it.
Further reading
See our earlier Research roundup on why human oversight matters more as AI agents take on longer, more autonomous tasks, and our Insights coverage of this month's agentic AI pricing shift.
Findings referenced in this article are drawn from publicly available research, including ICML 2026 proceedings and the Stanford HAI 2026 AI Index Report, as of the publish date. This article does not contain affiliate links; where future articles do, they will be disclosed per our Affiliate Disclosure.