At last, a new way to be annoying in public
What happened: Ray-Ban and Meta announced the imaginatively-named Ray-Ban Meta, which is an updated version of the Ray-Ban Meta. These glasses have always enabled people who are bad at reading social cues to record and stream video of what they’re seeing. The new version has a voice-driven AI assistant in it, enabling you to be just like Joaquin Phoenix in Her, except instead of Scarlett Johansson falling in love with you, it’s probably going to be Llama-2 hallucinating 88% of the time.
Why it matters: We all knew that voice-driven AI assistants were coming. We just didn’t know that they’d be here so soon, or that they’d arrive in a series reboot of Google Glass.
Look, these are kind of cool (by AI standards, not by sunglasses standards). You can look at something and ask your glasses to identify it, unless it’s a meerkat or a lot of other things. Glasses are a sensible form factor in which to combine machine vision and generative AI, and talking to your glasses–and having them talk back–I guess is a sensible interface. The problem, of course, is that if these take off (or anything similar takes off), the streets will be filled with people muttering to themselves as they, I dunno, get directions to the bank, or something. And still no one will know what a meerkat is.
At last, Google finally has a way to make money
What happened: The Financial Times reported that Google is considering charging users for AI-assisted search. Google has been testing AI-assisted search, in which an LLM provides more complete synthesis of results with links to specific pages embedded or alongside the summary, but has yet to roll it out at scale. According to the report, the technology is almost ready, but the company hasn’t decided whether to launch it.
Why it matters: Generative AI presents a significant threat to Google’s advertising-supported search product. If LLMs can provide high-quality answers to users’ questions, why would they want a traditional page of partially-relevant links instead? And if the LLM provides a useful answer immediately, why would a user click on ads placed alongside the organic search links? AI-assisted search, which combines an LLM-style response with traditional links, might be the answer, but it also undermines Google’s existing model, and it’s expensive to run.
The people who run Google are smart, and they have lots of money. I suspect they don’t really care about the revenue potential here. Rather, I think they’re trying to lead the entire market towards charging for AI-assisted search (if only to cover operating costs), in the expectation that this will help their traditional search business stay relevant.
At last, evolution is a thing
What happened: Researchers from Sakana AI, a Japanese AI research lab, released a paper outlining a new approach to building foundation models in which different model structures and parameters are tested and refined through evolutionary approaches. In particular, the paper documents ways of combining different models, testing the results, and then refining the ensemble model with logic derived from the mechanics of variation and natural selection.
Why it matters: This might be a dead end, or an interesting idea that proves impractical. But if it works, it could point the way towards solving a significant challenge in the AI community, the challenge of finding more efficient ways of building specific-purpose ensemble models. We have strong automated testing and tuning regimes once a model structure is defined, but the first design steps are still highly manual. If the solution to that first step happens to be a nifty-sounding evolutionary algorithm, well, that’s a cool little bonus.
“Three remarkable things that happened in AI this week” is a more-or-less weekly roundup of the most noteworthy events that have transpired in the world of AI.
