Utilizing the otherwise quiet month of August gives Google a chance to show off new AI ideas on Pixel ahead of Apple’s next iPhones which, now, look to be packed with “Apple Intelligence.”
It's something similar like the dotcom crash where everyone and their mother is looking to capitalize on The Next Big Thing™ which will then crash spectacularly but will leave a host of survivors that'll be the next big companies.
Gemini Nano. It's in Pixel Recorder and Gboard but multimodal capabilities are coming soon. It's very limited to a couple phones and not many companies outside of Google have access, but it has been used in production since last year.
Apple does not have on device AI in production today. Microsoft's "Copilot+PCs" launched less than 2 weeks ago, but are significantly more powerful machines.
Writing it off as a fad is rather ignorant. Sure there’s a lot of hype and bullshit surrounding AI, but it already has strong use cases and earns a profit for the companies involved. It’s still the early years for the tech too, so it is reasonable to expect it to improve in the coming years, both in terms of accuracy and performance.
I don’t disagree with the idea that AI is being shoved into software without much purpose or thought, but that’s got little to do with whether it is here to stay or not. It’s here to stay for its many practical uses, be it new personal assistants like what Apple has shown off with Siri, text summarization like what is being added to browsers, rephrasing/tone checking like what is being added to office software, or code completion and debugging like what is being added to code editors. These applications have proved their worth, and even if some applications are just because hype, these applications are here to stay.
The thing is you don't really hear much about AI solutions which are actually disrupting multiple industries right now. You only hear about toys, this is done so you get used to AI over time. It's very important to prepare the ignorant and socially conservative public to avoid the next wave of luddites.
I’m sitting here really hoping that models hit a plateau in capabilities soon. Continuing to get smaller/more efficient would be great, but if the capabilities of our best models would plateau for a bit and give society time to adjust to the impact I would be very happy.
We’re already seeing a slight leveling off compared to what we had previously. Right now there is a strong focus on optimization, getting models that can run on-device without losing too much quality. This will both help make LLMs sustainable financially and energy-wise, as well as mitigate the privacy and security concerns inherent to the first wave of cloud-based LLMs.
It’s certainly not moving as fast as their promises (what ever does), and perhaps has slowed, but for me at least it’s too early to call a plateau. Perhaps someone who works in the field or follows more closely can provide a better characterization, though.