AI & Machine Learning

Artificial intelligence stopped being a category of its own the moment every product started shipping a chat box. What is left, once that novelty wears off, is a more interesting question: which tools actually use a model to do something that was previously impossible, and which ones bolted a text field onto an existing feature and raised the price?

The launches collected here lean towards the first group. You will find coding assistants that read a whole repository rather than the file in front of them, transcription tools that run on the machine instead of a server, research assistants that cite the paragraph an answer came from, and agents that stop and ask before they spend money on your behalf. What ties them together is that the model is doing work a person would otherwise have done by hand, not decorating work that was already automated.

The most useful distinction when comparing them is where inference happens. Tools that run locally — on a laptop, a phone, or a small board tucked behind a router — never send your data anywhere, work on a plane, and cost nothing per request once you own the hardware. They are slower, and the models are smaller. Tools that call a hosted API get frontier-quality output and pay for it per token, which means the pricing page matters as much as the feature list. A tool that feels cheap at ten documents a week can be ruinous at a thousand a day, and the launches that are honest about this tend to say so on the page rather than in a footnote.

The second thing worth checking is what happens when the model is wrong, because it will be. Good products in this category are built around the assumption of failure. They show their working, they make corrections cheap, and they keep a human in the loop for anything irreversible. A summariser that links back to the source is more useful than one that is right slightly more often, because you can verify the first in seconds. An agent that drafts an email for review is a different risk profile from one that sends it.

Pricing here is unusually varied. Some makers charge a flat monthly fee and absorb the inference cost, betting on average usage. Others pass tokens through at cost with a margin. A growing number ask you to bring your own API key, which is the cheapest option if you already have one and confusing if you do not. None of these is wrong, but they suit different people, and it is worth knowing which you are signing up for.

If you are shipping something here, the most common feedback from this community is to be specific. "AI-powered" describes almost every launch on this page and distinguishes none of them. What the tool does, who it is for, and what it costs at real volume are the things people actually want to read.