Such a healthy reframe — the “everyone’s using AI for everything” narrative makes a lot of people feel behind for no reason. What strikes me about the “once a week or less” group: it’s rarely about access. Most tried it once on the thing it’s worst at (a factual lookup, usually) and quietly wrote it off. The people who stick around almost always found one boring, recurring chore it quietly nails — and built out from there. I’m in the “multiple times a day” slice, but only because I stumbled into that first useful task early. Do you think the gap is mostly about exposure, or about people not knowing what it’s actually good for?
I think the Gen Z segment is a decent answer to this question since exposure has been so high and it seems like the latter (not having a daily use case).
Exactly — Gen Z is the cleanest test case, because access and exposure are basically maxed out, so whatever’s left has to be the use-case gap. My hunch is the daily habit tends to show up alongside a recurring responsibility — a job, a deadline, running a household — something with enough stakes that “good enough, faster” suddenly matters. Exposure makes it familiar; responsibility makes it useful. Do you think the daily habit forms more from a killer feature, or just from life handing someone a problem boring enough to want to delegate?
Great post! If anything, anecdotally I'm just hearing people around me using less AI, especially since they are all now dealing the real implications of job loss and career loss, not necessarily because their jobs were actually replaced by AI, but because that's what corporations are doing and people are just exhausted from the continuous doomsday rhetoric from Anthropic, OpenAI, the data center people etc.
LLMs are useful of course, and my bet is on using small models on your laptops only, just like any other Python package. Let the tool be a tool, not something more than that.
As someone building developer infrastructure, I'd agree. We deliberately avoided AI features in our initial build — webhook delivery is a reliability problem, not an intelligence problem. Retries, signing, observability — these are solved with good engineering, not models.
There's a class of backend tooling where AI adds noise more than value. Developers want predictability, not cleverness.
Nice screenshots - appreciate visual aids throughout.
Ah, you’re with the Duck! An AI advancement: long struggled with Bing’s results (used by DDG), but the marketed-as-private GPT-5 Mini was able to overcome their imprecise results and provide a couple useful links the other day. So, an unexpected AI win, if the ole slop machine can do what Microsoft can’t, and we all get better search results out of DuckDuckGo.
Imagining now about trying to train the Bing (DDG) crawler by distilling an LLM… if Bing (DDG) results have stagnated in a permanent second-tier versus the [Don’t] Be Evil search engine, supplant some results with what the LLM would have recommended instead. This technique “would have” worked recently on a traditional search for a health-related query, where bad Bing results were fed to DDG but GPT on Duck dot AI found authoritative links. Bing (DDG) results were commercially/sales-focused, GPT were major reputable organization links, so let’s get rid of the bad results and return the good results. I realize saying this, this should happen over at Microsoft unless they deem traditional search too unprofitable, or the token burn to “train” the crawler (replacing links on SERPs for oft-asked queries with a model’s recommendations… then re-do periodically unless particularly evergreen) is absurd.
Such a healthy reframe — the “everyone’s using AI for everything” narrative makes a lot of people feel behind for no reason. What strikes me about the “once a week or less” group: it’s rarely about access. Most tried it once on the thing it’s worst at (a factual lookup, usually) and quietly wrote it off. The people who stick around almost always found one boring, recurring chore it quietly nails — and built out from there. I’m in the “multiple times a day” slice, but only because I stumbled into that first useful task early. Do you think the gap is mostly about exposure, or about people not knowing what it’s actually good for?
I think the Gen Z segment is a decent answer to this question since exposure has been so high and it seems like the latter (not having a daily use case).
Exactly — Gen Z is the cleanest test case, because access and exposure are basically maxed out, so whatever’s left has to be the use-case gap. My hunch is the daily habit tends to show up alongside a recurring responsibility — a job, a deadline, running a household — something with enough stakes that “good enough, faster” suddenly matters. Exposure makes it familiar; responsibility makes it useful. Do you think the daily habit forms more from a killer feature, or just from life handing someone a problem boring enough to want to delegate?
Great post! If anything, anecdotally I'm just hearing people around me using less AI, especially since they are all now dealing the real implications of job loss and career loss, not necessarily because their jobs were actually replaced by AI, but because that's what corporations are doing and people are just exhausted from the continuous doomsday rhetoric from Anthropic, OpenAI, the data center people etc.
LLMs are useful of course, and my bet is on using small models on your laptops only, just like any other Python package. Let the tool be a tool, not something more than that.
Good report! Just one thing, AI is not intelligent!
As someone building developer infrastructure, I'd agree. We deliberately avoided AI features in our initial build — webhook delivery is a reliability problem, not an intelligence problem. Retries, signing, observability — these are solved with good engineering, not models.
There's a class of backend tooling where AI adds noise more than value. Developers want predictability, not cleverness.
Nice screenshots - appreciate visual aids throughout.
Ah, you’re with the Duck! An AI advancement: long struggled with Bing’s results (used by DDG), but the marketed-as-private GPT-5 Mini was able to overcome their imprecise results and provide a couple useful links the other day. So, an unexpected AI win, if the ole slop machine can do what Microsoft can’t, and we all get better search results out of DuckDuckGo.
Imagining now about trying to train the Bing (DDG) crawler by distilling an LLM… if Bing (DDG) results have stagnated in a permanent second-tier versus the [Don’t] Be Evil search engine, supplant some results with what the LLM would have recommended instead. This technique “would have” worked recently on a traditional search for a health-related query, where bad Bing results were fed to DDG but GPT on Duck dot AI found authoritative links. Bing (DDG) results were commercially/sales-focused, GPT were major reputable organization links, so let’s get rid of the bad results and return the good results. I realize saying this, this should happen over at Microsoft unless they deem traditional search too unprofitable, or the token burn to “train” the crawler (replacing links on SERPs for oft-asked queries with a model’s recommendations… then re-do periodically unless particularly evergreen) is absurd.
/clumsy rambling
DDG Duck Icons are Iconic, especially Pink Duck🩷