Where the New Ideas in AI Are Showing Up
I read a lot about what people are building with AI. Some of it comes through newsletters, some through GitHub, and a lot of it through people on X who build in public.
Over the last year I’ve started sorting what I see into three piles:The first pile is people building a personal version of a tool they used to pay for, so the subscription goes away. The second is people taking an existing product apart and rebuilding it as a free clone. The third, and probably the biggest, is people building because building got easy and they want to see what happens. What I don’t see much of is anything new for a broad market. Most of what looks new is tooling for AI itself (agents to manage agents, wrappers for wrappers), and I’ve been wondering whether AI is quietly killing new ideas while making it cheaper than ever to produce more of the old ones. I also know my reading habits shape what lands in front of me, so I went looking for evidence on both sides, including the evidence that I’m wrong.
Apple received 235,800 new app submissions in the first quarter of 2026, up 84% from a year earlier after submissions grew 30% in 2025 to nearly 600,000; according to Sensor Tower data reported by The Information. Sensor Tower’s analyst ties the jump to agentic coding tools like Claude Code and Codex, though that link is an inference from timing. Over the same stretch, downloads grew just 3% in 2025 and 2% in the first half of 2026. Supply is exploding and demand is flat, which fits my second and third piles without proving them.
Anil Doshi and Oliver Hauser ran a preregistered experiment in which some writers got story ideas from GPT-4 and others wrote alone. Writers with AI ideas produced stories that readers rated as more creative, better written and more enjoyable, but the AI-assisted stories were also more similar to each other . The authors call it a social dilemma: each writer is better off, and the group produces “a narrower scope of novel content.” (I wrote about the individual side of this a while back in LLMs and Creativity , and this is the part I didn’t see then.) In January, Nature published a study of 41.3 million research papers showing that scientists who use AI publish 3.02 times as many papers and collect 4.84 times as many citations, while AI adoption shrinks the range of topics science studies by 4.63% and cuts scientists’ engagement with each other’s work by 22%. The authors say AI tools “seem to automate established fields rather than explore new ones,” and they describe the result as individual scientists gaining reach while science as a whole contracts toward the areas richest in data. The study covers AI tools broadly, most of them machine learning rather than chatbots, but it’s the pattern I was worried about at a much bigger scale.
If I stopped there, I’d write the essay I expected to write, but every piece of evidence above comes from a place I can see. The App Store, GitHub and builder X are public by design. The Nature study is about published papers. When I went looking for what AI is doing in places that don’t post their work, I found a lot that doesn’t fit my three piles.
AI outside the builder feed
Insilico Medicine used generative AI to pick a new target for idiopathic pulmonary fibrosis (a lung disease with no treatment that reverses it) and then to design a molecule against that target. The phase 2a results, published in Nature Medicine in June 2025, showed patients on the highest dose gained an average of 98.4 ml of forced vital capacity over 12 weeks while the placebo group lost 20.3 ml. It was a small safety trial (71 patients, 18 at that dose), and 11 of the authors work for the company. Still, in September Insilico dosed the first patient in a 52-week phase 3 trial that plans to enroll 320 patients. Whatever happens next, a new target and a new molecule both found with AI are now in a late-stage trial. Weather is another example.
In February 2025 the European Centre for Medium-Range Weather Forecasts put its machine-learning model into daily operation next to its physics model, saying it beats physics-based models on many measures , including tropical cyclone tracks by up to 20%, and uses about 1,000 times less energy per forecast. The national weather services of its 35 member and cooperating states can use it. None of this shows up on Product Hunt, and none of it is somebody’s free clone of a paid tool.
Regulated health is the same story at a larger scale and a slower pace. The FDA says it has authorized over 1,600 AI-enabled medical devices as of September 2026. A preprint analyzing the FDA’s list counted 331 authorizations in 2025 alone , the highest year on record, and across all 1,430 authorizations since 1995, radiology accounts for 76.5% and 96.2% came through the 510(k) route, which requires a company to show its device is substantially equivalent to one already on the market. It’s real work reaching real patients, built by people with domain knowledge most app builders don’t have. It’s also mostly incremental and concentrated in imaging, where the data is easy to get. The same analysis found 9 devices in pathology and none under a psychiatry review panel. That pattern (lots of activity where the data is rich, very little elsewhere) matches the Nature finding about science, and I’ll come back to it.
The World Bank ran a randomized trial in Nigeria where students used an AI tutor with teacher support in an after-school program for six weeks. The learning gains were about 0.3 standard deviations , which the researchers put at nearly two years of typical learning, and the program outperformed 80% of the education programs in their database of randomized trials from developing countries. Girls, who started behind, seemed to gain even more.
In agriculture, Digital Green’s FarmerChat has nearly 600,000 users across 120 countries in 14 languages, and in a survey of 450 Kenyan farmers, 83% said it was the first time they’d used a service like it and 7 in 10 had applied its advice in the previous 30 days. That survey was done with the company, and the trial was short, so I’d hold both loosely. Both are old problems (a tutor for every student, an extension agent for every farmer) that nobody could afford to solve until recently. I’d call that new work for a broad market, and I missed it because the people doing it aren’t posting their GitHub repos.
So my three piles describe what builders post in public, and that’s a small and unusual slice of what’s going on. The work that looks most like real new ideas is in labs, regulated industries, schools and farms, and it moves slowly through trials, approvals and budgets.
The productivity J-curve
That still leaves a fair question from the executive side: if all this is happening, why don’t companies feel it? The best recent data says most of them don’t. An NBER survey of nearly 6,000 executives in the US, UK, Germany and Australia found 69% of firms use AI, but 89% report no impact on labor productivity over the past three years, with an estimated average boost of 0.29%.
Erik Brynjolfsson, Daniel Rock and Chad Syverson call this the productivity J-curve . Their argument is that general purpose technologies like AI require large complementary investments that are often intangible and poorly measured, so productivity growth looks understated in the early years and overstated later when those investments pay off. The intangible investment is the un-sexy part: retraining people, redesigning workflows, cleaning data, rewriting job descriptions. None of it shows up as an app, and a lot of it looks like cost with no return for a while. I find that more convincing than “AI is overhyped” or “AI is already transforming everything,” because it explains both the flat productivity numbers and the pockets of real progress. If the J-curve is right, the firms with real results will mostly be the ones that rebuilt how a process works around the tool, and the firms that dropped a chatbot on top of the old process are sitting in the 89%.
The J-curve also changes how I read my three piles. People building personal replacements for paid tools are doing a small version of exactly that intangible investment. They’re learning what the tools can do by rebuilding something they already understand. Some of those clones will stay clones. Some of the people building them will be the ones who redesign a company-wide process two years from now, because they’ve already done the reps. Imitation has always been how most people learn a new technology, and adoption by people who never invented anything is where most of a technology’s economic value ends up.
AI and convergence
Doshi and Hauser found sameness in short stories. The Nature study found it across 41.3 million papers. The FDA data shows it in where medical AI gets built: three-quarters of devices in radiology, none under psychiatry. In each case AI makes individuals more productive and pulls everyone toward the same data-rich, well-trodden ground. I’d call that convergence, and it’s a better description of the risk than “dying creativity.”
The money is converging too. CB Insights counts $226 billion going to AI companies in 2025, 48% of all venture funding and the largest share on record, while total deal count fell 17% and mega-rounds took 65% of the money. The six largest rounds of the year all went to AI companies, and those six raised 24% of all venture capital. CB Insights’ own description is that venture capital is essentially making a single bet.
Microsoft’s latest diffusion report finds 28.8% of the working-age population in the Global North used generative AI in the second quarter of 2026, against 16.2% in the Global South, and the gap is widening. The Nigerian classroom and the Kenyan farm are real, and they’re also the exception.
A preprint from the Max Planck Institute for Human Development found that using AI to generate ideas reduced the diversity of what a group produced, while diversity was largely preserved when writers came up with their own ideas and used AI only to refine them. Since it hasn’t been peer reviewed yet, I treat it as an idea to consider and keep in mind, but not as a definitive concept. If it holds up, the sameness problem depends a lot on where in the work people bring the tool in, and that’s something a leadership team can decide.
What a leadership team can do
If every team uses the same models to brainstorm strategy, write product specs and draft the roadmap, the company’s ideas will drift toward everyone else’s, because everyone else is using the same models. The Doshi and Hauser result is a warning about a company’s strategy as much as about short stories. I’d want people to bring their own ideas to the table first and use AI to pressure-test and refine them.
I’d also stop measuring AI by counting tools and pilots. My three piles are mostly tools and pilots, and so is most of what gets reported upward in a lot of organizations. I’d ask which processes we have actually rebuilt around AI, what we stopped doing as a result, and whether anyone can show the before and after. If the answer is none, the company is in the 89% and the J-curve is not going to bend on its own.
I’d send executives to different sources, too. If the work that matters is happening in regulated industries, in operations, and in markets that don’t post on X, then executives who learn about AI from the same public builders I read will see what I saw: a lot of clones and a lot of tooling. The better information is in the trial registries, the FDA lists, the World Bank evaluations and the companies quietly redesigning a claims process or a supply chain. The new ideas are there, mostly in places I wasn’t reading. The bigger risk is everyone using the same tools in the same way until their ideas start to look alike.