AI is rapidly entering the “getting real” part of its life cycle. Now that the hype is dying down, and the inevitable backlash is getting louder, two recent developments underscore the sobering fact that consistently getting good, practical output from AI tools is actually quite hard.
The Humane Ai Pin — hyped to the extreme last fall as one of the first devices built around generative AI — was universally panned by reviewers, criticized as a barely functional gadget that, even when it provides good answers, is leagues away from the ideal AI assistant depicted in the movie Her.
I haven’t used the Pin, but I was already deeply skeptical of the belief that people would spend money on a device whose sole purpose was to act as a verbal bridge to a general AI knowledge base. “Ask it whatever you want,” is the same problem as the blinking cursor in a chatbot: without a specific task, you don’t know what to ask it.
Alhouth Google Glass wasn’t powered by AI, it provided a similar experience via abbreviated versions of search results, speaking answers to fact-based questions like, “Who is the prime minister of India?” I remember that experience getting old quickly — outside of trivia night, it’s not much more than a novelty.
That said, I do think we’re moving toward a world where ambient-computing experiences — where we’re all basically talking to the Star Trek computer — will eventually become common. But we’ll get there through practical applications, not parlor tricks. For instance, a doctor or engineer using something like the Humane Pin, trained on their area of expertise, to supplement their own knowledge would be a much better use case, and target customer. Of course, that’s not as flashy, but the value of flash in the AI world is dropping fast.
Adobe’s About-face
The other development that emphasizes the renewed focus on the practical is Adobe’s announcement that it would soon allow third-party models (as opposed to its own) to power AI features in its Creative Cloud apps. Leading the list is of course OpenAI, whose mind-blowing Sora video generator could end up in Premiere Pro. Soon, users who want to use generative tools in Adobe apps will be able to choose from all kinds of models.
This is a surprising move since Adobe made a big deal about the relative legal safety of its AI tools, often pointing to the fact they were trained only on licensed and public-domain imagery. If you wanted to create or enhance images with AI, but were afraid of the legal and ethical issues inherent to models that train on “publicly available” data, Adobe’s tools presented a compelling option.
That is, until you started using them. Last fall, I extensively used Firefly, Adobe’s AI image generator, but eventually switched to OpenAI’s DALL-E for a simple reason: the output was considerably and consistently better. It turns out limiting your training data to what’s legally safe means your model will never be as good as a competitor who is willing to hoover up every JPEG they can get their hands on.
And for users, especially creative users, it’s the output that matters, and they’ll always use what gives them the best results. No one wants to get sued, of course, but most of the big AI companies offer protection from legal liability, which removes most of the risk for individuals.
With its announcement, Adobe has recognized this reality, giving users the option to swap in whatever AI model works the best for whatever they’re doing. That’s a win for users, though it’s a blow to the idea that there was an easy path to avoiding the copyright question. As a practical matter, though, Adobe has given its customers license to ignore the issue and simply find ways to make AI work for them.
It turns out finding the practical side of AI is harder than it looks, and I fully expect to see more compromises, about-faces and outright busts in the coming months. AI’s hype phase was fun, and flashy hardware and squeaky-clean training sets are great ideas. But in the real world, if those ideas don’t solve problems, they’re not worth paying for.
Leave a Reply