silicon valley's biggest hits make luxury available to everyone, but not in the way most people think.
uber didn't just bring private drivers to the masses. its rating system and data-driven dispatch got you a good driver every time, instead of whoever was at the front of the taxi line. airbnb connected travelers with local hosts who knew every hidden restaurant and quiet beach. spotify's recommendations got to know your taste better than the record shop guy who'd known you for years.
so the luxury being handed out isn't really the perks. it's expertise.
silicon valley's playbook is pretty clear: find deep expertise and spread it widely. but so far AI companies are selling a different dream, a digital butler that does everything and masters nothing.
the name changes every few years. siri was going to handle your phone, alexa was going to run your house, and google assistant was going to know what you meant before you did. now chatgpt, claude and gemini are making the same promise with much better language: one polite box that does everything.
we've seen this movie before. siri collects dust on millions of iphones, google's assistant has faded into android wallpaper, and alexa is great at timers and weather reports. ask siri to set a timer and it's fine. ask it to book a family trip across four calendars, two airports, a budget, the kids' bedtimes and one person who refuses early flights, and it falls apart. tech giants spent a decade chasing the perfect servant and got glorified alarm clocks.
--
the universal assistants didn't fail just because the models were weak. they failed because the goal itself doesn't make sense.
nature never builds generalists. four billion years of evolution and not one species does everything well. even humans, supposedly the ultimate generalists, succeed through collaboration and extreme specialization.
real usefulness isn't one person doing everything badly. it's the sommelier who knows every wine, or the chef who's obsessed with one cuisine.
a really useful assistant won't do everything itself. it'll be good at execution: knowing who to call, what to hand off, and when to step back.
--
this isn't a new idea. in AI's first spring, researchers built expert systems, programs meant to match human specialists in narrow domains. MYCIN helped diagnose infections, and XCON configured DEC's VAX computer orders.
they saw something important: domains have natural edges, and expertise is naturally focused.
the technology just wasn't ready. the systems couldn't learn from data or understand context, so they could only follow rigid rules. getting knowledge out of experts was expensive, rules conflicted with each other, and maintenance was a nightmare. a human expert can say "hmm, this one's different." the system needed someone to write a rule for the exception, and then another for the exception to the exception.
the dream was right, but the tools were decades too early. so we shelved expert systems and filed them under failed experiments.
now everything they needed has arrived, in the form of LLMs.
and even these models say something about specialization. look inside a language model and you don't find one giant network doing everything. you find lots of specialized sub-networks, some recognizing patterns, others generating responses, others checking for consistency.
it's experts all the way down.
but instead of leaning into that, we're chasing universal assistants again, as if specialization were a bug to fix and not the whole point.
modern agents only look similar to the old systems, to be fair. an expert system is a rule base. a tool-using agent is a model that can call other software. mixture of experts is a routing trick inside some models. and human delegation is an org chart, with all the politics and trust and blame that come with it. but they share one idea: send the task to the part that knows how to do it.
--
intelligence specializes, then specializes again. how far down that goes, and what's at the bottom, is a question for another time.
we're the proof, eight billion times over, each of us mastering our own corner of the world. no one person does your taxes, fixes your teeth and drives you to the airport. division of labor isn't a bug. it's what built civilization.
even our brains split up expertise, with different regions for vision, speech and movement, each divided again into specialized circuits. consciousness itself might just be the ultimate delegation system, routing tasks to whatever handles them best.
nature doesn't build swiss army knives. it builds eagles with perfect eyes and sharks with perfect teeth, unremarkable in every way except the one that matters.
now we can finally build AI the way nature would. a language model can really live inside a domain: a legal agent that knows case law and writes precise contracts, or a medical agent that's read every paper since 1950 and remembers all of them.
the interesting part is the handoffs. your health agent watches your vitals while your tax agent hunts for deductions. each one is great at its thing, and none of them tries to do everything.
and the LLMs themselves aren't monoliths either. the mixture of experts approach in the best models means different parts of the network each handle their slice of the task. humans become conductors, directing the agents and the specialists inside the specialists.
so the useful AI probably looks less like jarvis and more like a hospital, or a dispatcher. a task comes in and the router works out whether it's scheduling, legal, code, finance, or "please ask a human before touching anything." narrow agents do narrow work, you can see the handoffs, and humans decide who gets called in. it's a relay race where we hold the baton.
that last part matters to me. the butler version hides who did what behind one smooth voice. a good dispatcher tells you which specialist answered, what it assumed, and where it stopped.
--
we built machines in our image but forgot one of our oldest lessons: mastery means choosing what not to master. the early AI researchers saw it clearly. their expert systems weren't failures. they were just early.