Maybe this is a little bit similar to VR and AR adoption. Arguably we have had stunning capabilities for years. But the headsets are uncomfortable and a bit expensive given that discomfort.
Likewise, the most capable models are huge, run on Anthropic or OpenAI's servers and are expensive to deploy on a large scale. That's uncomfortable from a privacy standpoint (although arguably essentially the same as any cloud service).
I believe that over the next few years we will finally start to see the transition from bulky VR headsets to comfortable AR glasses and goggles. That will lead to a huge increase in adoption.
As the LLM models, software, and hardware become more affordable and capable for running locally, production adoption of LLMs and multimodal models will continue to pick up.
I think that we are going to see huge excitement in the next few years as more and more very large and capable truly multimodal models come out. More specialized hardware like Groq etc. will become available. And within probably less than three years the truly radical new paradigms for neural network hardware such as those based on memristors, SNNs, etc. will likely show some very competitive prototypes.
As the new hardware paradigms are widely deployed, say within 5-10 years, human-level local AI will become ubiquitous.
What I take away from those charts is actually significant adoption that is steadily rising. But all of this is nested S-curves, as scientists and engineers work through small and large different bottlenecks and problems to get to the next capability or efficiency level.
Where are the numbers coming from? It's not only that chatgpt might have lost people but also that there are already a lot of options available now.
My company introduced helpful ai features so did Google and Microsoft.
I can now transcript and summarize my meetings in Ms teams for example.
In a conference I saw Bosch talking about a gen AI hub which they now internally use and saving time and money by using less external companies for these cases
GenAI and llms are already here.
And ai means a lot more. It's a Paradigma Shift.
The announcements of robots jumped and the quality of them too. Thanks to whisper and a lot of other advantages it's easier than ever to build a talking and listing robot.
Did we slow down a little bit after the original 'holy shit it's AI time let's restructure '? Sure but not that much.
And I'm definitely looking forward to the progress curve after all this capital has been moved and hardware gets cheaper and faster and more affordable.
>The problem is that most of them haven’t been back.
It's super task dependent.
We got a shiny AI at work (big finance place) that is both good from a technical PoV and is officially blessed for confidential info. Above is true - little use even by myself. I just can't figure out a good use...the dots just don't connect to what I need.
...go home & do some hobby coding / homelab tinkering...loads of AI use with great effect.
I do think this is somewhat temporary though, better integration into office products & OS etc will presumably still come, but a chatbot accessible easily isn't enough. Really needs something closer to the much feared MS's recall thingie...I know privacy issues, but I mean that sort of omnipresence would be needed to be useful.
Likewise, the most capable models are huge, run on Anthropic or OpenAI's servers and are expensive to deploy on a large scale. That's uncomfortable from a privacy standpoint (although arguably essentially the same as any cloud service).
I believe that over the next few years we will finally start to see the transition from bulky VR headsets to comfortable AR glasses and goggles. That will lead to a huge increase in adoption.
As the LLM models, software, and hardware become more affordable and capable for running locally, production adoption of LLMs and multimodal models will continue to pick up.
I think that we are going to see huge excitement in the next few years as more and more very large and capable truly multimodal models come out. More specialized hardware like Groq etc. will become available. And within probably less than three years the truly radical new paradigms for neural network hardware such as those based on memristors, SNNs, etc. will likely show some very competitive prototypes.
As the new hardware paradigms are widely deployed, say within 5-10 years, human-level local AI will become ubiquitous.
What I take away from those charts is actually significant adoption that is steadily rising. But all of this is nested S-curves, as scientists and engineers work through small and large different bottlenecks and problems to get to the next capability or efficiency level.