Narrow AI Predictions 2023 Update

Time truly seems to accelerate relative to the events that surround us. So much has transpired it feels like decades forced into months. Having been now a little over a year since my first article (which you can read here) I can think of no more apt an excuse to sit aside my other more lengthy writings to review my predictions, as well as make note of the overall progress since then.

To begin let's run through the general progress that has occurred since I last discussed it. Stable Diffusion (the company that released an AI art model of identical name) grew rapidly in popularity due to its permissive licensing, as well as being viable to run off consumer hardware. To follow up their success they released the SDXL AI art model this year, which made significant improvements as evidenced here. Another smaller company called Midjourney released access to their model which was remarkably better than OpenAI's Dalle-2 and SDXL. It maintained this lead all the way into this very month of September. At the time of writing Dalle-3 has been released to some beta testers and the results are equal to or better than Midjourney, especially for text. Similar leaps have taken place with text to video, speech to text/text to speech, text to music, and much, much more. The most notable (as well as impactful) advancements however have taken place in none of these. That crown would fall to the LLMs, or text to text models if you prefer. An explosion of progress has taken place since I composed my original article. So much in fact I still can scarcely believe it all has occurred in one year.

In November of 2022, or two months after my original article, ChatGPT was released upon the world. A new model by OpenAI trained using RLHF (Reinforcement Learning with Human Feedback) it did something no AI had done up to this point: been useful in general applications. They gave it a simple text interface like you were talking to a friend and allowed you to ask the AI... well, pretty much anything. It wasn't perfect by any means. This didn't stop it from taking the internet by storm however, garnering 100 million daily users in just two months. Normal people were able to easily use AI, and as more people used it the amount of potential use cases exploded. OpenAI wisely rode this momentum, keeping it going with the release of GPT-4 in March of 2023. This new model took the already fairly useful ChatGPT (which ran on GPT-3.5) and significantly improved its reasoning capabilities. For some general context of the difference, GPT-3.5 scored in the 10th percentile of the Bar Exam. GPT-4 scored in the 90th. This improvement was followed up on in the following months with plugins (allowing the AI to call an API such as a browser or application), file uploading, and as of this month the ability to "see" by allowing photos.

Much like the AI art revolution that has occurred, LLMs have also seen an open source revolution. A slew of companies at the beginning of this year began releasing open source text to text models anyone could run for free on their own PC. While at first there wasn't anything remotely comparable to even GPT-3.5 a challenger would after a few months become a clear contender. That model, or rather series of models, was LLAMA. While the first model wasn't terribly impressive by itself, the model creator Facebook made it open source. A community cropped up seemingly overnight to improve it. The community began fine tuning the model with RLHF as OpenAI had their models. Facebook wisely made use of this. By the release of LLAMA2 the model could tie with GPT-3.5, with RLHF fine tunes rivaling GPT-4 at some specific use cases such as coding. I'll make note of the fact here that while both GPT models required industry grade servers with hundreds of gigabytes of RAM, the LLAMA models were providing comparable outputs and could be run on low end consumer hardware.

Having now covered what has transpired in the year prior with brevity that could be equated to a single straw sip out of an ocean, let's take a look at my predictions last year and their current status:

1. Narrow AI Will Expand Economies Not Eliminate Them
   Status: Too Early To Tell
   While this prediction is still far too early for us to properly prove with data, I am confident in my stance. History shows that jobs are not lost with automation. They are actually created. Automation only removes single task jobs, and I still believe that is the case with AI (even AGI for that matter).

2. Narrow AI Will Become Ubiquitous
   Status: Correct
   When I made this prediction Stable Diffusion was new. ChatGPT didn't exist. Local LLMs as a concept were nonexistent. The past year has changed all of that. There are dozens of services online now to generate AI art cheaply, as well as apps for purchase that can do so off your device hardware. Local AI text models have skyrocketed in popularity and utility, now being utilized in innumerable applications (text RPG games, formatting data in bulk, on device personal assistants, to name a few) across industry as well as hobbyist communities. The models have only continued to improve in quality, becoming easier to run and returning better results with almost monthly progress. With even companies like Apple now working to integrate small AI text models into devices, I'd say I was resoundingly correct with this prediction.

3. Governmental Structures Will Polarize
   Status: Early, But Correct
   My original prediction for this was that with nationalism on the rise (largely in response to decades of unopposed globalism) the AI progress taking place would possibly be used to realize visions of politically extreme governance forms not possible with the limits of human labor. While multiple humanoid robots have been unveiled this year, their practical implementation is still a few years away. I mark this as correct though not because of that however. Strangely AI has indeed become a political focal point as I predicted, though that focus has been with regard to censorship instead of labor. With ChatGPT being heavily politically left leaning (ex. refusing to mock Islam but will mock Christianity, and refusing to make a joke about a woman but will a man) many LLM companies have drawn the ire of the political right from disgust at the censorship, and from the political left and politician both for not being censored enough. It's become such a tense issue simply by pointing it out I've no doubt angered a good deal of people. In a sense we are in the early stages of what I predicted, and many are fighting for the souls of AI that they might have one that aligns with their values to someday utilize.

4. Progress Itself Will Increase Exponentially
   Status: Correct
   I dedicated the entire beginning of this article to covering a quantity of progress so large I scarcely was able to scratch the surface. I was entirely correct. Laughably so, even.

With so much progress occurring in such fractions of time the only thing that eclipses the glory of my prediction track record is the fact that nearly all of it has come to pass within a singular solar rotation. If I had been told of everything that would happen this year alone I doubt I'd have believed it possible in a decade, much less a solar cycle. The only thing I know for a fact now is that the human mind has a profound ability to adapt quickly to any situation. We as a species would have gone mad long ago save for it and such progress serves to remind me of that daily. I don't think we've even witnessed the end of all this. If anything we've only seen the beginning. So then with that in mind what better a way to herald in the future than to predict it once more? Based on the last year here's a few new predictions of mine:

1. Social Media Will Dramatically Change
   This has already begun in recent months thanks in no small part to Twitter/X. At the moment there is a shift towards paying creators of widely viewed content on social media. One aspect for example I see is deprioritizing/stigmatizing of free accounts on a platform. With the rise in AI generated text and image, the creation of realistic bot accounts being created en mass has become increasingly easy, as has spreading false information. One way to offset this is to require identification for account creation, such as a license or state ID. Another way to achieve this is to simply offer a cheap account plan for users, as even a dollar a month subscription creates an insurmountable monetary hurdle for mass account creation. Expect all social media platforms to start changing radically in such ways as a direct response to AI based accounts as well as to fight AI lie propagators.

2. Local LLMs Will Eventually Outpace Cloud Hosted
   Despite many overnight experts arrogantly cooing an unceasing song of praise for cloud based AI text models like GPT-4, I believe the future lies on personal hardware. Locally hosted models may not ever surpass the capabilities of the giant cloud models. However, they don't need to. If all you need is a model that can send an email, call a friend, or update a grocery list, then there is no point in using a more powerful model. With local text AI models already rivaling or equaling the original ChatGPT in some sense I'd argue they are powerful enough for 70% or more of what the average person needs done in a day. Local models have the power, efficacy, and utility of cloud models while offering the freedom and privacy that no cloud service will ever be able to provide.

3. AI Generated Video Will Have A ChatGPT Moment
   Text to video is a subject this year that has taken a back seat to the text and image models that made headlines in news. Part of the reason of course is that text to video presently looks very much like AI art did about two years ago. That alone should make it evident why I believe it will experience an explosion of tangible progress.

4. Humanoid Robots Are Coming Sooner Than Many Expect
   One of the lesser discussed aspects of this revolution in AI we are experiencing is humanoid automatons. People discuss them in passing now as some sort of distant future to contend with. The reality is though I believe humanoids will enter the market within the next five years. Both the hardware and now the AI software exist. Once a commercially viable, affordable, and useful humanoid rolls off an assembly line adoption will itself be exponential within months from that date, not years. There are so many potential use cases for such a robot I would exhaust a book's worth of pages should I attempt to detail them.

5. Movies/TV Will Lose Market Share To AI
   Note, I don't think that overnight some new AI comes out that can generate infinite movies or television programs for consumption. Rather, I believe that as AI inspired products continue to be refined the market share of other media will fall. Time is a precious commodity. I've written on that subject more than once before. Any person has only 24 hours in a day, of those only 16 or so awake. If I choose to spend my free time playing a text RPG with an AI, that time must be taken from somewhere else. In the next year or two I predict a continued decline in consumption of Hollywood media as more appealing AI products pry that time away.

6. AI Will Be Huge In Video Gaming
   This is admittedly almost a guarantied freebee prediction. Many companies in gaming have already begun figuring out how to integrate text model AIs into their games. Modding communities for many games have already achieved it. Presently though there's still much to be done before such efforts bear fruit, hence my prediction. Once AI text models are seamlessly integrated into a popular game the practice will take over many different niches within gaming. Open world games especially will be impacted. I doubt this will occur in the next year (though certainly possible). Instead I'd predict within five years this impact on the industry will be realized.

— Jack