I appreciate this kind of summary, Ethan. Pausing for a second to look around is important, but putting what you see into context is even more valuable for folks reading this. It's a pleasure to read good writing, too!
Imbuing the edge of computer networks with intelligence is going to be a game changer, as you rightly state. Imagine billions of mobile devices, industrial sensors, robots, etc., all complemented with some kind of artificially intelligent autonomy. The implications are rather staggering.
Incidentally, is the title of your post an allusion to Carl Sagan's book A Demon-Haunted World?
I think the "army of AIs" absolutely where we're going - and I already find myself headed in that direction with GPTs. With a bit of prompting and custom tooling, we have custom AIs that act as specialized freelancers rather than general-purpose assistants: one for coding, one for graphic design, one for social media.
Understanding the implications of using open-source software in commercial products is crucial, as it involves navigating legal and licensing issues, ensuring security and stability, adapting to potential limitations in support and customization, and recognizing the impact on innovation and community collaboration. Additionally, it's essential to differentiate among various types of open-source licenses, such as permissive licenses like MIT and Apache, which allow broad usage and modification, and copyleft licenses like GPL, which require any derived works to also be open source, thereby influencing product development and distribution strategies. Understanding these distinctions helps in aligning with legal obligations and business objectives.
I wonder what effect this will have on human psychology and how we relate to our surroundings. In a very real sense, we are returning to an animist conception of the world, with everything around us possessed by a spirit with it's own personality and (eventually) motivations and desires.
In your Dec 19th article/newsletter, you said you said "In the last week, I managed to get an AI that is more powerful than the original ChatGPT to work on my home computer… and one slightly less powerful to run on my iPhone. The AIs ran entirely on my own equipment, no internet connection needed, and without anyone having access to my data".
I can't believe that YOU believe this is possible.
By definition, the entire essence of AI is that it has access to the accumulated millions of pages and billions of words and phrases that have been uploaded to "the cloud" and that this is the "database" from which any iteration of AI draws the answers for prompts that have been presented to it.
You're not suggesting that your phone or even your laptop would be able to "store" every law, every newspaper article, every university paper, every high school exam, every business contract.....etc, etc, etc that's ever been created......the way AI as we know it NOW functions.......are you?
Please tell me that I stopped reading your article too far "up the page", and that you went on to correct yourself.
LLMs handing off to each other reminds me of the 20-year-old trend toward parallel processors on devices and even more so in the cloud, as Moore’s Law started to break down, and progress was sustained that way.
If capabilities of individual CPUs double every 18 months (as was for example the case in the 1980s and 90s), you don’t really need “tricks” to improve performance, it increases at an almost insane rate, as you can see if you compare what a PC could do in 1990 to what it could do in 1980, or in 2000 compared to 1990.
Same thing with LLMs, as the difference between ChatGPT 2 to 3 and then to 4 was crazy. Now there are intimations of a plateau, and people are focusing on parallelizing LLMs. From my understanding, this is also what is happening with openAI’s rumored Q*, which uses two models to start to reason. For me, that’s the most important thing happening right now, not only in tech, but period, in the world. Cheap 3.5 models are all well and good, but they are just not quite good enough for the type of revolutionary changes people see coming. If we are honest, even ChatGPT 4 is not quite good enough, its limitations mean that after a year of hype, there is still not a killer use case and it functions mostly as a mind-blowing tech demo, an intimation of what is to come.
I’m most excited about Q*, if it is real, because the ability to plan and to reason is what will take us to agents, not just to chatbots. I don’t think even a Q* ChatGPT 5 will get there yet, but it could be again a tech demo of what agents, if they are nailed down, can do for us. For whatever reason google isn’t getting us there, it seems to be openAI making all the progress so far.
so I guess dumbbells will still be dumb
It all sounds incredibly tedious I have to say.
I appreciate this kind of summary, Ethan. Pausing for a second to look around is important, but putting what you see into context is even more valuable for folks reading this. It's a pleasure to read good writing, too!
Imbuing the edge of computer networks with intelligence is going to be a game changer, as you rightly state. Imagine billions of mobile devices, industrial sensors, robots, etc., all complemented with some kind of artificially intelligent autonomy. The implications are rather staggering.
Incidentally, is the title of your post an allusion to Carl Sagan's book A Demon-Haunted World?
What could possibly go wrong?
The genie is indeed out of the bottle. No amount of wishcasting would put it back.
I think the "army of AIs" absolutely where we're going - and I already find myself headed in that direction with GPTs. With a bit of prompting and custom tooling, we have custom AIs that act as specialized freelancers rather than general-purpose assistants: one for coding, one for graphic design, one for social media.
Understanding the implications of using open-source software in commercial products is crucial, as it involves navigating legal and licensing issues, ensuring security and stability, adapting to potential limitations in support and customization, and recognizing the impact on innovation and community collaboration. Additionally, it's essential to differentiate among various types of open-source licenses, such as permissive licenses like MIT and Apache, which allow broad usage and modification, and copyleft licenses like GPL, which require any derived works to also be open source, thereby influencing product development and distribution strategies. Understanding these distinctions helps in aligning with legal obligations and business objectives.
You're so good at helping us imagine what our future will be like. Thank you, Ethan.
I wonder what effect this will have on human psychology and how we relate to our surroundings. In a very real sense, we are returning to an animist conception of the world, with everything around us possessed by a spirit with it's own personality and (eventually) motivations and desires.
Quite interesting! You have shown in simple terms where AI is heading. It is a pleasure to read your postings. Thanks
In your Dec 19th article/newsletter, you said you said "In the last week, I managed to get an AI that is more powerful than the original ChatGPT to work on my home computer… and one slightly less powerful to run on my iPhone. The AIs ran entirely on my own equipment, no internet connection needed, and without anyone having access to my data".
I can't believe that YOU believe this is possible.
By definition, the entire essence of AI is that it has access to the accumulated millions of pages and billions of words and phrases that have been uploaded to "the cloud" and that this is the "database" from which any iteration of AI draws the answers for prompts that have been presented to it.
You're not suggesting that your phone or even your laptop would be able to "store" every law, every newspaper article, every university paper, every high school exam, every business contract.....etc, etc, etc that's ever been created......the way AI as we know it NOW functions.......are you?
Please tell me that I stopped reading your article too far "up the page", and that you went on to correct yourself.
I did a post you might find interesting on how we use animistic design for these types of agents and how they might coordinate in a smart home:
https://uxdesign.cc/a-smart-home-is-one-that-talks-to-itself-58bb9222d893
Overall, I think it helps a lot with creation of exceptions which the home is full of (and are constantly evolving).
Who knew!
You, apparently.
I obviously am in over my head... but appreciate being dragged along, rescued and educated, all in one.
Thanks very much for the enlightening response. Onward and upward!
LLMs handing off to each other reminds me of the 20-year-old trend toward parallel processors on devices and even more so in the cloud, as Moore’s Law started to break down, and progress was sustained that way.
If capabilities of individual CPUs double every 18 months (as was for example the case in the 1980s and 90s), you don’t really need “tricks” to improve performance, it increases at an almost insane rate, as you can see if you compare what a PC could do in 1990 to what it could do in 1980, or in 2000 compared to 1990.
Same thing with LLMs, as the difference between ChatGPT 2 to 3 and then to 4 was crazy. Now there are intimations of a plateau, and people are focusing on parallelizing LLMs. From my understanding, this is also what is happening with openAI’s rumored Q*, which uses two models to start to reason. For me, that’s the most important thing happening right now, not only in tech, but period, in the world. Cheap 3.5 models are all well and good, but they are just not quite good enough for the type of revolutionary changes people see coming. If we are honest, even ChatGPT 4 is not quite good enough, its limitations mean that after a year of hype, there is still not a killer use case and it functions mostly as a mind-blowing tech demo, an intimation of what is to come.
I’m most excited about Q*, if it is real, because the ability to plan and to reason is what will take us to agents, not just to chatbots. I don’t think even a Q* ChatGPT 5 will get there yet, but it could be again a tech demo of what agents, if they are nailed down, can do for us. For whatever reason google isn’t getting us there, it seems to be openAI making all the progress so far.
With AI helpers soon to be ubiquitous, who will they help the most? People who are good at asking for help?