In my last blog I wrote about the evolution of chip makers and how NPUs are now being put into various devices, making them more capable of executing on AI tasks.  Of course as soon as I wrote that piece the stock price of Nvidia jumped. Go figure. I take full credit.

 

But what is the newly capable hardware going to do?  

In this piece I will try to give some perspective on the other side of that equation, what these newly capable devices are going to do specifically. The use cases around A.I. type processes and models are endless, but there a few newsworthy examples worth noting that act as a microcosm for the larger picture. The first of which is the whole topic of deep fakes.  Without going too much into the weeds on this allow me to define some basic terms so we are all on the same page. In the “old” days fakes required manual dexterity and artistic skill. Think fake paintings, fake IDs, fake passports, fake money, fake jewelry, etc. These types of fakes were meant to look like a real item even though they were not the real item. Usually they had a physical element to them, aka, it was something you could see and touch but may not know it was fake. 

While those types of fakes will always be around, over the last 30 years the ‘digital fakes’ have also emerged. With the advent of digital images and videos and online content this newer category of fakes requires some software skills and a bit of creativity. But the idea is to create something that looks authentic online even though it has been edited and altered. Fake websites, videos and pictures with items or objects added in that weren’t in the original, you get the idea. Those types of fakes are pretty common to anyone that has used a computer in the last few decades.  Another similar off-shoot is “cheap fakes”:…..where an image or video maybe been cropped or altered so the context or understanding of the event is missing.  In those cases the events and objects are real and really did happen but the creator is trying to frame the message and narrative of what you are looking at in certain light. Again, those types of fakes have been around for a long time.

Which brings us to the topic of “deep fakes”. Deep fakes are a relatively newer category that involves A.I. Specifically a model is trained using inputs to get to a place where generative A.I. can produce desired outputs as if they were coming from an authentic source. The easiest way to think of this is in the context of a person. So imagine you wanted to deep fake a movie star or celebrity. The math around A.I. of course has not changed, but the availability of inputs to help train the model would be wide and plentiful in that context. Lots of images, sound bytes and video clips likely exist for the celebrity and these inputs would be fed into the model to, in essence, create a virtual version of the person. Assuming the training and model were good enough, that newly created virtual version would then be instructed to generate desired outputs…for example creating a video where it would look like and sound like the authentic celebrity was actually there speaking about whatever topic the user of the new model came up with.

Of course you can spot the obvious nefarious use case right away. Bad actors could create such virtual fake but compelling avatars to help sell a product or create misinformation and the original human being that was patterned would have no idea their likeness was being used in that manner.  Further complicating this is where the inputs to the model don’t even require authentic material or content of the person but rather some other person who looked and sounded very similar, think like training the model using a celebrity impersonator, not the actual celebrity.

If you are asking if this could happen to you, well maybe, but 95% of the population fall into the bucket where there is not enough actual content out there to train a model or there is no viable reason to do it.  The first part is easy to understand, if you haven’t really posted much on social media or starred in movies and television, then there is likely not enough available content for anyone to train a decent model of you, unless you specifically gave some one permission to have access to your personal images and videos and family archives and so on. The second part is even more relatable….if you have no celebrity status or are of little social importance then your ”selling power” is pretty limited so a digital avatar of you doesn’t have the type of commercial attraction for anyone to bother creating a model. Remember the training of models takes an investment in time and effort, especially to refine and get it right. Most of us aren’t important enough that a nefarious actor would even bother. 

There are use cases however where this type of approach might actually be desirable however strange it might seem. Consider for example using the same process for creating a virtual teacher or professor to teach an online course. Such an avatar would never age, never retire, never need to take vacations, never pause for a coffee, you get the idea.  Or to make it more personal, imagine creating a virtual version of a loved one before they die…thereby preserving the ability to “interact” with that person even when they are gone. Again, some may find that notion abhorrent while others might find it useful for the grieving process or to allow their grandchildren to get to know what grandpa was like when he was alive. You get the idea. 

 

BUT SHOW ME THE BIG MONEY

So services will pop up that in essence charge for the creation of virtual grandparents, virtual teachers, virtual companions of every sort. Its not hard to see these services charging some sort of fee to create and update these avatars and since most everyday consumers have neither the time nor money nor skills to train models and get into generative AI, they will gladly pay for it. And that is the exact dilemma on a grand commercial scale that is being played out right now in the larger economy.  Large corporations that might have a use for generative A.I. have to decide if they are going to invest in their own in-house resources to develop their own models or are they going to simply buy trained model products from companies that specialize in that space.  It is often the case that a device maker wants to use AI to enhance the user experience of their customers and they may not have experience with the AI side of that equation even if they are good at physically making the device itself. Decisions right? This is playing out across the economy right now and will do so for the next couple decades reaching into every sector.

What does it look like exactly? Insurance companies with call centers wanting to buy trained LLMs to help with the real time interpretation of conversations and communications with customers. Makers of cleaning robots wanting buy spatial awareness AI models to help their products autonomously clean buildings without running into people. Appliance makers that build ovens to accept downloaded recipes and proactively cook the food as needed rather than a human having to set timers and temperatures.  You likely have already seen the advent of cars that utilize AI routines for things like parking, highway cruising, and so on. The common theme here is that the AI model can make the devices more useful by becoming more autonomous, in essence, more in touch with the needs of the human users.

But the true nature of generative AI goes even further, not just becoming more useful to the human user, but actually becoming the human user itself. There is where things start to get a little weird but stay with me. The new smart phone equipped with modern NPUs is now going to become an extension of you. How? Imagine the phone maker has either developed (or bought) a trained model that studies you specifically, looks at your texts, your social media posts, listens to your conversations, records your locations, realizes your purchases, and so on. Putting aside the privacy issues and creepiness factor for a second, such a device could over time learn about you. How you communicate, your speech patterns, writing patterns, driving habits, etc. In essence, it would understand and become a virtual you. The generative part of this is where the value is. Once it has leaned all about you it becomes a virtual version of you to try to become ever more helpful.

Imagine this example. You are driving home from work, and your significant other sends a text asking when you think you will be home, aka what’s your eta and also what do you want for dinner. The phone, acting as a virtual you, automatically replies back based on current speed, traffic and many other inputs such as past food purchases etc. and without you actually doing anything sends a text back to that significant other saying you will be home in 20 minutes and maybe tonight would be great night for take out from one of their favorite restaurants. During the next 20 minutes while you are driving (which your car maybe doing for the most part automatically) the phone checks all of your social media accounts and appropriately creates a number of replies to the most common friend posts, so lots of likes, well done, and congrats. It then creates a couple of new posts from recent photos in your gallery of pictures and comes up with a catchy phrase and posting as if you had written it. In the mean time your partner’s phone has responded yes to take out, so your phone proceeds to place an order for pick up, and sends the updated instructions to your vehicle’s navigation system for the optimal route to the restaurant for pick up. Because this is going to take an extra ten minutes it has already sent a text to your significant other apologizing for the extra delay and advising of the new eta. It proceeds to then send out more messages to various friends asking what their plans are for the weekend. As you pull into the parking lot to get the take out it sends a message to your children’s tablet telling them to start their homework. It then scans your streaming services for recommendations to view something together with your significant other that evening and then sends a few of those suggestions to their phone.

I could on with that example literally forever but the point is the devices of the future could use generative AI to help you be, well, more of you. A super user version of you. There will be mixed reactions of course. Someone might be horrified by such an example. Do they really want a smart device in essence communicating with their loved ones or close friends in that manner? Doesn’t it all feel fake? There will be others who might see it as huge productivity boost or time efficiency play…look how much more I was able to do in the twenty minutes of driving time thanks to my device communicating for me.  Certainly social media influencers might see the value, now they can create even more posts and respond to ever more followers. A true influencer will likely have an army of such devices to push out record amounts of content and do what one human could never do in terms of scale.      

The implications of this are huge but allow me to throw out a couple of larger societal issues to consider.  If all the devices are acting as virtual people and are communicating with each other’s devices then of course it is much ado about nothing. In other words, if your phone and a friend’s phone simply sent texts back and forth (or even more scary left realistic sounding voice mails) but you and your friend did not actually create any of those communications then did you and your friend really communicate at all?  Or was it just devices being pleasant with each other trying to do their best impersonations of each of you. For that matter, in a world like that does certain communications even matter any more. For example if you created a social media post and got back a bunch of likes, were they real likes or just the devices automatically responding. Likes and similar forms of communications would become meaningless, unless there was some new designate marker for communication suggesting this communication was actually created by a real human. Imagine that, a “real like” versus a regular like. Social media companies, if you are reading this then you can thank me now for creating the next big addition….premium human communications versus bot communications.

The economic implications are huge as well. Regardless of the societal and ethical and privacy issues that will emerge, one thing is certain, the amount of data is about to explode like humanity has never seen before. The idea of bots talking to bots, generative AI in a never ending loop of creating content and reacting to that content, and human super users aided by a machine army of generative AI, all of it leads to more capacity requirements and the need to store more data, aka think cloud data services, server farms, energy requirements. Secondary services that provide logistical support to those sectors also will be needed, for example cooling supplies for the server farms, privacy and encryption services for the data and so on. If data is truly the new oil then we are about to hit the 1870s of the beginning of the oil boom, aka data is about to explode.

Now before we go too far down this path keep in mind that device makers will still allow you to be the boss, aka, there will be settings and options to shut off the “virtual you” or select levels so that you could choose a more harmless less verbose virtual you. This sort of grand settings selection will play out across many different devices and the economy as a whole. It means as a people we will all have our individual choices and preferences to make in terms of how much or how little of a role do we want this generative AI to play in our lives. It also means an increased skepticism of what is real and what is not. Usually what is rare is valuable so in the future it could be that human created communications, artwork, writing and just in general human creations will be highly valued in a sea of bot generated creations.  Time will tell.  

 

*          *          *

2067

“Hey Grandma, what’s that?”

“Oh, its an old painting your grandfather bought many years ago”

“You mean Grandpa, the guy on my phone who talks to me sometimes?”

“Well he bought it many years ago when he was alive, but yes that’s him,  your phone lets you talk to a memory of him now”

 “But’s its not the real him?”

“Not the real him, but kind of a special memory of him”

“Can I ask him about the painting?”

“Sure, but the memory of him might not know all the answers about it”

“Do you think he could tell me if it was painted by a real person?”

“Oh, I am sure it was painted by a real person. I was there when he bought it.”

“Wow, it must be worth a lot of money if it’s by a real human.”

“Hard to say, but yes, probably worth something”

“Grandma, do you think maybe we could sell it sometime?”

“Why would you want to do that?”

“Because I like you Grandma, and I want to have the money to buy a special memory of you too. That way you and Grandpa can both be on my phone together.”

“Well that’s a nice thought, but I am right here in person, we can talk anytime you want”

“Ya, but Mom says you won’t be around forever, and I should spend as much time with you as I can. Do you think when you become a special memory you will be able to talk to Grandpa’s special memory inside the phone?”

“I am not sure if it works that way. But happy to hear you will remember me once in a while”

“Grandma, do you think when I am old like you someone will pay for me to become a special memory?”