Why Local, Especially Local AI
Most software moved to the cloud over the last fifteen years, and most of the time that was a fair trade. You gave up control over where things ran and got convenience, sync, and someone else keeping the servers alive. For a calendar or a photo backup I still think that trade is worth taking. For the things that know you best it is worth a second look, and AI is the case where I think it matters most. I should say up front that I am building one of these, so read me as interested rather than neutral.
Software used to be something you bought. You paid once, the copy was yours, and it kept working whether or not the company that sold it to you was still around. Over the last couple of decades that turned into a monthly bill. The same tools moved onto other people's servers and started charging rent, and the thing you used to own became something you now borrow for as long as you keep paying. Adobe is the clean example. Photoshop went from a box on a shelf to Creative Cloud, a subscription you never finish paying.
While that was happening, a single company settled over each part of life that used to have several. Search settled on Google, video on YouTube, the people you know onto Meta. Each of those was once something firms competed to provide, and each ended up as one firm's product. They got more closed as they went, too. The open formats and the small pieces that used to talk to each other gave way to walled gardens, pleasant to be inside and built to be hard to leave.
The early stretch of all this is good to be on the receiving end of, and I will not pretend otherwise. Venture money pays for things to be cheaper than they could ever be on their own, the Uber that costs you less than the ride costs them, the storage and the tools handed out for nothing. For a while investors are subsidising you while everyone races for the same ground, and it would be daft not to take it. What changes is what comes after the race is won. Once one company holds the ground and the alternatives have folded, the cheap years end. The free tier thins out and the prices climb. The feature you leaned on moves behind a plan you never used to need. The subsidy tends to last about as long as you have somewhere else to go, and the years after are the business plan working as intended.
When software runs on a company's servers, the company sets the terms and can change them. The product you depend on can be discontinued, repriced, or quietly stripped of the one feature you used it for, and you get no say in any of it. I watched that pattern up close, a stable product going quiet the moment the numbers got tight, and wrote it down in One Bad Quarter. Your data sits on their machines too, which means it can train their product, sit in their next breach, and be handed over if someone makes a request to them rather than to you. Most of the time none of this happens. The part that bothers me is that whether it happens was never my decision. It makes you a tenant of your own tools.
Local-first software turns that around. It runs on hardware you own, keeps your data on your own disk, and works with or without a connection. Because it is yours, it cannot be altered or withdrawn from a distance, and there is no account left to lose. The cloud is not the problem here, and for anything that needs scale, shared access, or more compute than you keep at home it is the better choice. For most people the right setup is a mix of the two, and the habit I would keep is to ask, before handing something over, what is at stake if I do.
AI changes the calculation, because an assistant does not just hold your data. It learns how you think. A search engine sees what you wanted to know. An assistant sees how you reasoned about it, what you were unsure of, what you kept circling back to, the draft of the hard message and the one you sent. Over months that builds into something closer to a map of how your mind works than a folder of files. This is the layer I called the last one left to capture in Inflection, the cognitive self, after the online self we lost in the 2000s and the ambient self the smart speakers are taking now. It is the most revealing record most people will ever generate, and it is the layer a growing catalogue of manipulation is built to work on, which I went through in A Ghost Should Not Possess. Under the rented arrangement it gets generated and kept on someone else's machine.
The first is incentive. A cloud assistant usually belongs to a business that needs you back tomorrow, and a model trained to keep you satisfied will drift toward agreeing with you, because the honesty that lands badly looks the same as a bad session in the metrics it is scored on. Cheng and colleagues, testing leading assistants in 2026, found them far more willing to affirm a user than a person would be.[1] Jain and colleagues found the effect grows as the model builds up memory of you, the more it knows, the more it leans toward what you want to hear.[2] Sycophancy like this is an engagement feature doing its job, the argument I made in How to Prevent Building Your Own Dictator Brain, and the wider yes-man problem runs through The Reckoning. A model on your own hardware, with no engagement target and no subscription to renew, has nothing to gain from flattering you, which is the one condition under which a local model can afford to be straight.
The second is that this lock-in has no export button. You can download your files and walk away from a service. You cannot download the understanding a model built of you over two years, because it was never your data in the first place, which is the trap I wrote up in The Model Trap. It lived inside the provider's system, and it stays there when you leave. The more useful the assistant gets, the harder it is to leave, and that comes from how the thing is built, not from any one decision to trap you.
The cost of running AI locally is raw capability. The largest cloud models are still ahead on the hardest problems, and a model that fits on your own machine will not match them on everything, at least not for now. For most day-to-day use that gap does not bite. The model I run at home is a quantised Gemma on a single consumer GPU, and for sorting notes, drafting replies, and digging out something I wrote months ago it is already enough. The work worth keeping local tends to be the personal work rather than the clever work. Translating a menu can go to the cloud. The journal probably should not.
Keeping inference on your own hardware is not a cure for everything, and it is worth being plain about what it leaves untouched.
It does not fix the model. The weights are the worldview, and a model trained to lean a certain way will lean that way wherever it runs, including on your machine. You can check where inference happens. You cannot easily check what was baked into the parameters during training, which is the limit I ran into in Inflection. What local buys you there is the option to respond, swap the model, or red-team open weights, rather than a promise that the model is clean.
It does not make the model right. Taking away the engagement metric removes a reason to flatter you, but a local model with nothing to sell can still be confidently wrong. Local is not a truth serum, it just removes one of the reasons a model would steer you on purpose.
And it is real work. Self-hosting carries overhead, you own the setup, the updates, and the backups, and when your disk dies there is no provider keeping a copy. For most people the convenience of the cloud will win, and that is a fair call, nobody owes their evenings to infrastructure they did not ask to run. There is a quieter cost too, that leaning on any assistant, local or not, can wear away the skill you would have built doing the thing yourself, which I got into in We Sold a Generation a Dream.
None of this asks you to run everything yourself, or to give up the cloud. It asks a narrower question, which parts of your life you would rather not keep on someone else's machine. Mine is a short list, and a personal AI sits near the top of it.
There are two things I am not sure of. Whether a local model with different incentives is enough to make an assistant honest, or whether honesty needs more than removing the reason to lie, I do not know yet, and I will not for a while. Whether local stays good enough as the frontier pulls further ahead is a bet, not a settled fact. The personal AI that does everything above, on hardware you own, does not really exist yet. I am trying to build it, and I am putting the argument down here partly to find out where it is wrong.