Answers
Plain answers about running AI on your own device.
These pages explain how AI can run on hardware you own. We answer 30 questions as precisely as we can, without claiming more than the facts support. You may quote these pages with attribution.
How a model runs on a device
A trained model is a finished file of numbers. Running it is arithmetic, not a lookup and not a request. These pages explain what that file contains, how it is made small enough to carry, and what the machine is doing between your sentence and its answer.
- What does it actually mean for AI to run on your device? On-device means the model runs on your hardware; which parts of the app still use a network is a separate question worth asking.
- How can a language model answer you with no internet connection? The network moves your question to a computer. The thinking is arithmetic over a file, and a file can sit on your own device.
- What are a model's weights, and why is the file so large? A model file is a table of learned numbers. Its size is the parameter count multiplied by the bytes each number takes.
- What is quantization, and what does shrinking a model cost? Quantization changes how precisely each number is written down, not how many there are, and the rounding error it leaves is paid unevenly.
- Does a model learn from what you tell it? Training writes the weights; using a model only reads them. Anything that looks like memory is text being stored and handed back.
- Why does a model forget what you said earlier? Between requests, systems may resend or summarize text; ordinary inference does not store conversation history in the base model’s weights.
The machine in your pocket
A phone is a small computer with a hard budget: memory, bandwidth, heat. These pages describe what the silicon can do, which of the three budgets runs out first, and why the question changed answers in the last few years.
- Can a phone actually run a language model? A phone runs a language model exactly as well as its memory allows: small enough to fit, and read fast enough to keep up.
- What is a phone's neural engine, and what does it do? A neural engine makes each operation cheap in power. It does not make the phone's memory bigger, and memory decides what fits.
- Why is memory the limit on a phone, not processing power? Memory availability and bandwidth often constrain local models; streaming from storage can trade speed for lower RAM use.
- Does running AI on a phone drain the battery or overheat it? Short bursts generally use less energy than sustained generation; the amount depends on the device, model and workload.
- Why have useful language models become practical on more phones? Useful local language models depend on hardware, memory, model design and compression working together.
Privacy as architecture
A promise about your data assumes the data arrived somewhere. These pages trace the path a sentence takes when it leaves a phone, read the wording that usually describes that path, and show how to check for yourself. The full argument lives on the privacy page.
- What is the difference between a privacy policy and private by design? Local processing can reduce reliance on a remote provider for that path; the app, device and updates still matter.
- What happens to a sentence after you send it to a cloud assistant? Where computation happens is one part of the path; storage, access and other network features matter too.
- If it’s encrypted, why isn’t that the same as private? Encryption in transit and at rest does not by itself settle who can access data during processing.
- Does “works offline” mean the same thing as “runs on your device”? Offline means the app still works without a signal. On-device means the work happened there — not that nothing was sent.
- How can you tell whether an app is sending your data somewhere? Offline tests and network reports can show some behavior under specific conditions; neither proves that an app sends nothing.
Owned and rented software
Software stopped being a thing you have and became a thing you visit. These pages explain the mechanism behind that, what actually decides whether a program still runs years from now, and what owning your own data requires in practice. The conviction itself is argued on the ownership page.
- What is the difference between owning software and renting it? If it needs a server, an account or a check to run tomorrow, you are renting it, whatever you paid.
- Why does software you paid for stop working? A service or platform dependency can stop working; bugs, hardware, and data problems can also cause failures.
- Why is AI almost always sold as a subscription? Local inference avoids a provider’s compute cost for each answer, but architecture does not determine a product’s license or price.
- Will an app you install today still work in ten years? Local execution reduces some service dependencies, but compatibility and maintenance shape how long an app keeps working.
- What does it take to actually own your own data? A file you hold, can read without its original program, and can still open years later is owned. Everything else is access.
What a small model can and cannot do
Compression costs capability, and it does not cost it evenly. These pages say plainly what falls away as a model gets smaller, what a model cannot know at all, and which jobs stay reliable at a size a phone can hold.
- What can a large model do that a small one cannot? Breadth buys the long tail: obscure facts, rare languages, long reasoning and vague instructions, and a small model trades that away.
- Can a model know about something that happened after it was built? A cutoff describes the coverage of training data. Newer facts can come from your prompt or local files; fetching them from an online source needs a connection.
- Why do language models make things up? The model predicts likely text, not true text; a wrong answer costs it exactly as much effort as a right one.
- How small can a model be and still be genuinely useful? There is no smallest useful model, only the smallest one that does your job. Name the job and the answer appears.
- Can a phone understand speech without sending the audio anywhere? Recognizing speech on a phone is ordinary engineering now; understanding what the words mean is the harder half.
What it means in a day
The architecture shows up in small moments. How long you wait, whether you have to sign in, what still works with one bar, and what the microphone is really doing while the phone sits face down.
- Why does a cloud assistant feel slower than a local one? Local inference avoids network and remote-queue delay; model loading, prompt length, device speed and other activity still affect response time.
- Why do AI apps make you create an account? An account may support identity-linked limits or features, but remote inference itself does not require sign-in.
- Is your phone listening to you? Wake-word detection may run locally; what happens after a trigger depends on the feature, device, and service.
- What can an AI assistant do with no signal? Offline, a phone can still compute; it just cannot fetch. The failure to watch for is the one that pretends to succeed.
Why this exists
A research company should be able to explain its subject.
These pages carry no product claims. They are here because the questions are real, the honest answers are hard to find, and the work only makes sense once you know how the parts fit together.