Cloudflare's Clef, and whether your robot vacuum is about to get smarter
On 1 October Cloudflare released Clef, the first AI models its own team has trained. They are open source, so anyone can download them and run them on their own machines.
What Clef does
Most of the AI people use writes back. You ask a question and it answers in sentences. Clef does not write anything. You hand it something to look at, an email or a photo, along with a short list of questions, and it returns an answer to each one with a probability attached. Is this a person or a car? Is this message urgent? Which of these five boxes does it go in?
It comes in two sizes. Clef is the accurate one. Clef-flash is the quick one, and Cloudflare puts its typical answer at 38.8 milliseconds. Jev, the rival model TypeSafe released in mid September, takes 524 milliseconds on the same test. Both Clef models can read images.
Cloudflare’s own examples are all business work: sorting customer support requests, spotting bad web crawlers, flagging dodgy websites.
The robot vacuum part
The post going around is from @jpthor on X, who says he predicted two weeks ago that every major AI lab would put out models like this, that they would read pictures and sound, and that they would end up open source and tuned for robots. His line: “Your Roomba is about to get very smart and very fast.”
Some of that has happened. Clef is open source and it reads images. It does not take sound, and nothing in Cloudflare’s announcement mentions robots.
The fit is in the shape of the job. A robot vacuum with a camera spends its day answering small questions about what is in front of it: a sock or a cable, the dog or a cushion, kitchen or hallway. Those are the questions Clef is built to answer, quickly and with a number on how sure it is.
Size is the catch. Even the small Clef is a 9 billion parameter model, and Cloudflare runs both on graphics cards in its own data centres. A robot using one today would send what it sees to a server and wait for the reply.
What it means for a house here
There is nothing to buy yet. No vacuum, mower or camera on sale in Australia runs Clef, two days after it was released.
The kind of decision it makes is the one a house camera makes every night: person or possum. Clef makes that kind of call quickly, cheaply, and it is free to download.