Can AI help those who suffer from memory loss?

Image
~ Pete Smith ~ 1 Comment

I was watching something on TV the other day and the characters were talking to someone with Alzheimer’s disease. Now this was NOT a documentary, and I am certainly not a doctor. I have had a few brief encounters with people suffering from cognitive decline and I found that experience unsettling, to say the least. Trying to tease from a lost elderly gentleman what his name was and where he lived when he had no clue…he was so frightened. My heart just about broke and since then losing my memory and sense of self has been one of my greatest worries, more so since I’ve officially become a senior citizen.

It occured to me that if I were ever to face cognitive decline that a specialized AI Chatbot might be really handy. I’m not talking about using ChatGPT but a smaller model that is trained specifically for the patient. Imagine that you or your loved ones could feed it a ton of facts about yourself, including pictures and names of people near and dear to you. It would have facts about you too.

Let’s be clear, this is all just my own personal speculation of what AI could do to help folk with memory issues. And it would not be a substitute for professional human care, just something extra for when the patient is alone.

The device would ‘live’ in something wearable like a locket, glasses, or a watch. It would act autonomously throughout the day, gently reminding you about things. If it was hooked up to a camera it could identify people you know and quietly whisper to you what their name is and what relationship they have to you.

When you’re alone it would keep you company, addressing you by your name, talking about what day it is, and where you are. What you need to do in that day, or what project you’ve been working on. Just little hints to help maintain some continuity in your life.

Basically it would be a patient companion as well as a back-up memory device. Something that is always there to help you and answer questions any time things started to get foggy.

Your chatbot could remind you to take medicines, remind you to eat. When to go to bed. Help you maintain your routine which, I’ve been told, makes life with memory loss a little easier to deal with.

I feel like this kind of device could be designed today, and within a few years it could be built into some kind of robot companion. There are AI toys demonstrating early versions of this idea already: imagine a hybrid between a Furby & a tamagotchi. You raise them and talk to them and they remember facts about you and are supposed to provide companionship. These are just toys, mind you, but they show at least a hint of the kinds of interactions that an AI with persistent long-term memory of you and your life can provide.

I’m imagining something like these toys only (of course) brought up to a much higher level of competency and ready to have adult conversations. There’s also going to have to be a lot of testing to ensure that appropriate guardrails are in place to keep them being helpful rather than distracting. Also of course privacy concerns, though I believe that most of the personal data would be stored locally.

Here in the US the population is aging. Anything that can help older people stay independent longer seems like a smart investment because it seems like there just isn’t going to be enough care to go around. This feels like one place that AI could really help improve a person’s quality of life.

Gemini 3 Builds Me An App

Image
~ Pete Smith ~ 4 Comments

Google’s Gemini 3 rolled out today and I got an email inviting me to try it. I clicked through and stared at the empty prompt box. What should I do? There’s an “I feel lucky” button and I clicked that and got a prompt to build some boring business app. But at least then I knew it wanted to build me something. I was at work so couldn’t spend a lot of time so just spewed out this prompt:

Build an app that I can use to track my gaming and that helps me to decide what to play next. It should also track what platform (PC, Xbox, Playstation) a game is on and should pick at least one Xbox and one PC app every day so I can earn Microsoft Rewards points. Once those conditions are met it should weight game choice based on how long it has been since I played that game (the longer it has been since I played, the more likely it gets picked). Also there should be a field for games with daily login rewards so I make sure to get those every day.

This is an interation on another (Python) app that I vibe-coded with ChatGPT. But I was just playing around, testing stuff.

I hit submit and then my boss pinged me on Teams so I switched over to my work desktop and, y’know, earned my salary. When I remembered the Gemini experience I went back and there it was, a fully fleshed out app that did pretty much what I wanted. I realized though that I need to differentiate between Xbox PC games and other PC games so I asked for a tweak:

I just realized I need TWO different PC categories. One is PC Xbox (games played through the Xbox App which qualify them for Microsoft Rewards) and one for all other PC games. Can we add that?

And off it went. Within a few minutes I had a pretty nice looking app:

Screenshot of the main dashboard of the NextPlay app
The Main Dashboard of NextPlay
Screenshot of the Library screen of Next Play
Note that we’re even tracking games with a daily login reward

The little blurb about the game is being pulled in automagically via AI; I didn’t ask for that, but I like it.

Now mind you I haven’t actually USED this yet but so far it looks like exactly what I wanted. I’ll tweak some of the colors (particularly the Playstation tag which needs to be a lighter blue) but it’s really close to EXACTLY what I want custom made for ME.

The downside is, where do I run this? For now I’m running it locally after failing to get it to run on Vercel, and during the troubleshooting for that I realized “before I put this anywhere public we’ll need to add an authentication layer and I’ll have to check the code for any security issues” both of which felt like more than I can take on right now. I’m actually glad I ran into Vercel issues because I hadn’t really thought through the ramifications of putting this online!

So to do: authentication, security check, find a place to host it (maybe just on this hosting account). Not finished yet, but it feels like a really good start!

When Can the AI Come Out & Play?

Image
~ Pete Smith ~ 4 Comments

I’m falling behind on the AI news, what with our upcoming move and the crazy-fast pace of innovation in the AI space.

So forgive me for just now getting around to talking about SIMA 2, Google’s newest game-playing AI model. My layman’s read is that SIMA 1 could follow learned, game-specific instructions (“turn,” “walk forward,” “use tool”), while SIMA 2 combines that motor-skill layer with Gemini-style reasoning. It can plan, decide what it wants to do, then hand those decisions off to its “action” module that knows how to execute them.

SIMA 2 can also learn new games a lot more easily — both by watching humans and through trial and error. It can transfer concepts across games, too. If it learns “mining ore” in Minecraft, it can apply that idea to Valheim or any other survivalcraft sandbox. To anthropomorphize it a bit, SIMA 1 was following instructions, while SIMA 2 now knows WHY it is hitting a chunk of ore with a tool.

Quick aside about why Google builds these things. The idea is that once they have a model that can learn to navigate and exist in any virtual world it is set loose in, that model can be used in robots to allow them to figure out how to navigate around a factory or a house. To Google this is not all fun and games and it explains why you generally see SIMA 2 playing ‘open world’ and procedurally generated games rather than corridor shooters. Real-world robotics needs improvisation, not scripted paths.

Anyway, if you want the true expert explanation, check the link above. As far as I know, SIMA 2 isn’t something we peasants can play with — it’s an internal Google project.

This is both a pity and a blessing.

Why I’m Glad SIMA 2 Isn’t Out in the World

If a model like this were loose in the wild, just imagine the impact on eSports or MMOs. Gold farmers would have an army of perfect workers. Competitive games would become impossible to police. How would you know if the player demolishing you is a human with insane skills or an AI with infinite patience? You wouldn’t!

Why I’m Sad SIMA 2 Isn’t Out in the World

But on the other hand — and why I selfishly think it’s a pity this isn’t an open model — what if we could have this AI as a companion in our games? I’m sure I’m not the only gamer out there that doesn’t have a crew of friends on-hand to play MP and Co-op games at any hour of day or night, but SIMA 2 could fill in.

It seems to me we’re like 9/10s of the way to this happening. In the video above the human asks the AI what it is doing in-game, and asks it to perform tasks. This seems to be happening via a side-channel, but if you could get the AI to take instruction through the game’s chat channel you’d have an AI sidekick to hang around with.

Of course it’d be tempting as heck to just have the AI do the boring stuff for you. While I’m at work my AI buddy can spend the whole day roaming around harvesting rare materials in an MMO, or grinding out dailies in my gacha games. Then when I’m free I just take over and do all the fun stuff myself.

That would feel pretty cheat-y to me, but man it would be tempting.

In any even, I doubt this genie will get stuffed back into the bottle and sooner or later someone will release an open source version of a model similar to SIMA 2. For now the cost of running an AI constantly while gaming will probably keep things in check, but eventually I do think we’ll have AI gaming companions. This even feels like a legitimate business opportunity to me. In the same way you can now sign up for a service that lets you create AI generated art, you could have services where you could ‘rent’ an AI gaming companion for a few hours. When you make your first million with that idea, don’t forget to comp me a subscription to your service!

Google’s Project Suncatcher is Blowing My Mind

Image
~ Pete Smith ~ 4 Comments

One of the big challenges of AI adoption is generating the compute required for all the systems. Not just the actual hardware, but the power to run the hardware. We keep reading about deals to build massive powerplants to run the data centers that AI is going to need.

It’s a thorny issue, particularly here in the US where we haven’t embraced renewable energy to the extent some other parts of the world have, meaning our data centers rely on burning fossil fuels.

Google is working on another potential solution. Project Suncatcher is an experimental initiative to explore the feasability of building a data center in space, using solar power.

The idea is to put a fleet of satellites in orbit, flying in formation and talking to each other via lasers. The satellites will be positioned such that they can ‘see’ the sun almost all the time so they don’t need a lot of batteries and such. No cloudy days in space, no nighttime either.

This sounds completely like science fiction, right? But there is speculation that Google could fly a test flight as early as 2027. The big challenge is to bring the cost of launching the hardware down enough to make the system viable, and if the rocket industry keeps going at the pace it is going now, Google projects we’ll hit that sweet spot around the mid-2030s.

I watched a long-ish video about this that goes into the project in more depth, and since I’m far from an expert let me just share this with you:

I find this really interesting not just in terms of solving an AI issue, but also in terms of moving real industry into space. I grew up during the Apollo era and wow have I been disappointed in the lack of progress in space since then. Glad to see private industry is starting to pick up the slack!

You Take it on Faith Until the Science Backs You Up

Image
~ Pete Smith ~ 2 Comments

I’ve spent the last few years deliberately trying to exercise my brain—Duolingo, Brilliant, Codecademy. Learning for learning’s sake. Why? Because I’m old and I was trying to keep my brain somewhat agile. I’m sure you’ve heard the sayings: “The brain is a muscle; use it or lose it.” or some variation thereof. I wasn’t sure I believed it but figured better safe than sorry, right?

Now there’s some science to back these sayings up, according to Mental exercise can reverse a brain change linked to aging, study finds at NPR.org. It’s nice to think we can do something to improve our brain, but it’s a little less so to find that playing Candy Crush had no impact but playing BrainHQ did. Honestly I’ve never played Candy Crush but I assume that it becomes more challenging the longer you play, in terms of either speed or strategy.

BrainHQ

I hadn’t tried BrainHQ either, so I decided to. You can do one exercise per day for free. The one I did had me visually tracking bubbles on your screen. You keep track of the ‘target bubbles’ in a field of identical copies, then after a few seconds you have to click on “your” bubbles (all bubbles look identical so its just watching and memory that tells you which are “yours.”) Initially there were two target bubbles and by the end, five. I’d tell you have many ‘distraction’ copies there were but I was too fixated to notice. Maybe 10?

What’d I’d love to see is the same study done and comparing people who played Dark Souls (or another game that is challenging in terms of either perception, quick thought, or good reflexes) to folks who played BrainHQ to see what the results there were. I think any semi-serious gamer would agree that at least SOME games count as mental exericise, right?

The article suggests 30 minutes a day with one of these learning apps and my experience with others is that 30 minutes of doing these things would feel like HOURS because they’re all pretty boring. That said, one of the notes that came up while testing BrainHQ stuck with me. BrainHQ says they make the games harder and harder until the user starts to fail, then the system backs down the difficulty a bit. They say that ‘threshold’ where a problem is difficult but solvable is where you get the good brain juice (I’m paraphrasing).

Reaching The Threshold

The reason that stuck with me is I find it hard to find that threshold in pre-made situations. I’m almost a year into doing Brilliant and most of it has been trivial except for the geometry course once we got to curves/circles. I’ve done something like 11 courses and that was the only one where I struggled at all. And what did I do? Switched to something easier. 🙁 That was dumb, but it’s because there’s a leaderboard and my gamer mind wants to be at the top of it in the most efficient way possible. So flying through 5 or 6 easy lessons rather than sticking with one that I’m struggling with just felt like the way to go. Probably not the intent of the leaderboard but, that’s gamer brain for you.

With CodeCademy, there’s learning there but also a lot of busy work and the “I learned something” dopamine hits just don’t outweigh the “this feels like a second job” (in that I’m writing code that someone else wants and that I have no interest in).

This “threshold” concept made me realize why most learning apps lose me: they either bore me or punish me for struggling. Real learning happens right at that line—hard, but still solvable.

The Actual Learning is Fun

When I really start to learn, and when I can kind of feel my brain waking up, is when I get excited about a project I want to do like (sorry to sound like a broken record here) this AI stuff. I had to learn about node.js and miniconda and pip installs (and git, but I already knew that) just to get started, then I was constantly running into issues that I’d have to research and troubleshoot in order to get all the parts to play happily together. Then I had to learn about checkpoints and loras and controlnets and VAEs and such; it was a whole new terminology. Once I got a pipeline running I had to start to learn camera terminology for prompting. It was initially overwhelming but I really wanted to learn it and pretty soon it started to make sense.

Weird thing, once I got it all working really nicely I kind of lost interest. LOL Maybe I’m addicted to the struggle—the problem-solving part lights up my brain, but once things run smoothly, the fun ends. Or at least, changes into something else.

So, still kind of groping around for the right system for me. Paying for Brilliant makes no sense since I’m not learning anything. CodeCademy is better but I still get SO bored with it. I’m actually thinking of making a custom ChatGPT that is an instructor who’ll create challenges for me, then nudge me when I’m stuck. I think I can do that; make it so it doesn’t just spit out the answer but gives me a clue. Then ideally it’ll check my response. I’ve never actually created a ChatGPT from scratch, so that alone could be an interesting learning experience.

Maybe the real trick isn’t finding the perfect app—it’s finding challenges that keep me right at the edge of what I can do. That’s where the brain stays alert and alive.