#1943 Both Feet on Both Pedals
Both Feet on Both Pedals
Oliver is an engineer who spent fifteen years being told his A1c was fine. Then he pointed an AI model at his own loop data and found out what his settings had really been doing.
Jump to a moment
- A good A1c hid a decade of roller coaster. Diagnosed at 14 in 2008, Oliver spent fifteen years on injections taking the same dose for every meal regardless of what he ate, because nobody had ever taught him to count carbs. His A1c looked fine at every appointment, so nothing changed. When he finally got a Libre, he found he was spending enormous amounts of time high and low — averaging out to a number that satisfied everyone.
- He had to engineer his way past the gatekeeping. Under the NHS his good control was the reason he was refused a CGM and later a pump. So he read the criteria and met them: one required testing more than ten times a day, so he did exactly that for a month with very sore fingers until the data was undeniable. He notes that had he been under 16 at the time, a pump would have been automatic.
- His metaphor for a badly tuned loop: driving with both feet down. An engineer who studied closed-loop control systems, Oliver realized his loop was following the route but fighting itself — one foot on the accelerator, one on the brake, engine overheating and brakes wearing out. He wanted gentle acceleration and gentle braking instead: corrections where needed, not a constant wrestle.
- The AI failed until he stopped asking it to tune a moving target. He gave Claude a full briefing — which loop algorithm, which experimental features were switched on, six months of Nightscout history — and its first answer was to cut 20% of his basal because he was only using 60% of scheduled basal. That was backwards: his loop was withholding basal to stay stable. Only when he went back to open loop and did old-fashioned 24-hour fasting basal tests did the analysis become coherent. Settings changes belong with your care team; this is his experience, not advice.
- The prompt discipline is the transferable part. Because a model will answer confidently at any level of certainty, Oliver instructed his to state a confidence percentage with every response, to stay silent below 80%, and to tell him what additional information would raise it. He also gave it read-only access to Nightscout deliberately, so it could never deliver insulin. He plans to publish his scripts, with his own data stripped out, for the community to fork.
- Defining Diabetes series — Short episodes defining the terms nobody explains — the series Oliver credits for filling fifteen years of gaps.
- Diabetes Pro Tip series — Episodes 1000-1025 — settings, insulin timing, and food impact.
- Bold Beginnings series — Scott and Jenny Smith on the fundamentals, for anyone starting from where Oliver started.
Every word of the conversation
Cold open & sponsors0:03
Hello, friends. Welcome back to the Juice Box podcast. Nothing you hear on the Juice Box podcast should be considered advice, medical or otherwise. Always consult a physician before making any changes to your health care plan or becoming bold with insulin. Summer's winding down and you're already missing your great vacation, but it is never too early to plan the next one.
Juiceboxpodcast.com/juicecruise. Come on juice cruise 2027 with me and a bunch of other people living with type one diabetes, families and adults, and everybody in between. Juiceboxpodcast.com/juicecruise. Cabins are on sale right now. And if you have type one diabetes and you'd like to take the survey from the Juice Box Podcast, it's juiceboxpodcast.com/survey.
It'll only take you a few minutes, and it would really help the show. Thank you so much. Here is your program. Today's episode is sponsored by Tandem Diabetes Care, the Eversense three sixty five, and touched by type one. There are links in the show notes and at juiceboxpodcast.com to all the sponsors.
When you use those links, you are supporting the Juice Box podcast and helping to keep it free. The Eversense three sixty five is a one year wear CGM. That's one insertion a year. That's it. And here's a little bonus for you.
How about that there's no limit on how many friends and family you can share your data with with the Eversense Now app? Eversensecgm.com/juicebox. Now let's talk about Tandem Mobi, the impressively small insulin pump. Tandem Mobi features Tandem's newest algorithm, Control IQ Plus technology. It's designed for greater discretion, more freedom, and improved time and range.
Learn more and get started today at tandemdiabetes.com/juicebox. Touched by type one is my favorite diabetes organization, and they're sponsoring this episode of the juice box podcast. Please take a moment to learn more about them at touchedbytype1.org. Also, follow them on Facebook or Instagram. When you get to touchedbytype1.org, check out their many programs, their annual conference, the d box, dancing for diabetes.
They have a golf night. Gosh. They have a dance program for local kids. Go check them out at touchedbytype1.org and see what they're up to.
An engineer at heart2:29
My name's Oliver. I've had diabetes probably now for a long time. I should have probably worked out the answer to that question before I led with it. I've been diabetic a long time since I was about 14 years old, so for the last twenty years or so.
Okay.
I'm an engineer at heart. I've always been an engineer ever since I was a child, studied engineering. So I've really, in the last few years, started to apply that to how I can use that to better my diabetes care, really.
Alright. Let's find out a little bit about all that. So we'll we'll get a little background first. So you were diagnosed at what age?
Since I was 14. I looked up early today. It's 2000 and 2008.
Okay. Other autoimmune in your life or in your family?
When I was diagnosed, not that we knew of. So I was the first in my family to really have anything autoimmune. Since I was diagnosed, My younger sister was then also diagnosed with rheumatoid arthritis. Mhmm. And there is underactive thyroid.
Little little
Little overactive thyroid. Okay. I believe that can that can be related to to autoimmune diseases, but
About it.
Not really too sure. Yeah. That's about it.
And I can tell from your accent, you're from Texas?
Not at all. I'm from The UK. I'm British.
Alright. I was teasing. But
I can put on a hat for you if you want
I don't know that that would help.
Hat and boots? Yeah.
I just came from Friends for Life, which was cowboy themed this year. So I did not have any people like, where's your hat? I was like, I I I I I have to be honest. I got there, and they're like, where's your your hat? And I was like, I didn't know what the theme I didn't know what the theme was.
I'm sorry for paying that close of attention to that stuff. Okay. So I think I'd like to just hear a tiny bit about your upbringing and and what diabetes was like, but then I really do wanna jump into how you're doing things now. So Sure. Yeah.
Fifteen years on injections4:24
Tell me about your first pump. Was it at diagnosis, or was it later?
My first pump wasn't till three years ago.
Oh, what were
you doing up
until then?
Doing MDI for fifteen years.
Fifteen
years. First CGM wasn't till 2020.
What what do you think I mean, not that this makes you a late adopter, but what what made you slow to it?
Fighting for a CGM4:50
It was not by choice of me. I had to fight tooth and nail with the NHS, first of all, to be able to get prescribed a CGM. It's actually a Libra that I was prescribed first. Mhmm. So it wasn't even a Libra two.
It a Libra one. And I had to really fight to get that, and then also then had to fight to be able to get prescribed a pump.
So these
A lot of this was because my control was good, so I wasn't really seen as a high risk person.
I see.
So I don't know. No. You you'll deal with MDI. You'll be okay.
Do you ever go through a feeling of maybe I should just fudge the numbers or do a little worse so I could speed up the process, or did you think about it like that?
Not really. That's what my engineering approach doesn't really that doesn't I don't tend to do something like that. It wouldn't be my approach to it. But on the flip side, because I am an engineer, let's say, okay. What are the criteria that I have to meet to need to do this?
So for example, for the for the flash glucose monitor, it was I had to test more than 10 times a day. K. So I had very sore fingers for the best part of a month, which then got me into that. I got CGM.
That data alone let them give you a CGM.
Yeah. One of the key criteria you met, it was that or you had super high a one c's or hypo unawareness. Mhmm. And so there may have been some fudging of how hypo aware or unaware I was, but they obviously couldn't refute how many times a day I was testing my blood sugars.
Sure.
And then that's already unlocked the Libra for me.
And when you got I'm sorry. Go ahead.
I said as soon as then I got a Libra, I'd already done some research around it. I think it was called a a MeowMeow
I remember that.
Which is like a a small app that you put on top of the Libra, which then allowed me to Bluetooth it to my phone. I think I was using Spike as the iOS app at that point. So you then
Yeah. Let's make sure people understand that when that the Libre first came out, you called it a flash. You had to hold up a receiver, right, to the device. Yeah. And then that and Yes.
So either the receiver or my iPhone, I just put it up to the up to the sensor, that would read it. So I would only get a reading when I held it up to the sensor.
But there was no live readings or alarms or anything like that? No. Okay.
Nothing.
But plenty of aftermarket stuff that allowed it to work that way.
Correct. Yeah. Which is the so that's what the Meow Meow is that that essentially you tape it on top of the sensor. Mhmm. And then that will Bluetooth to your phone so you can use a flash monitor like it is a CGM, like, you would know a g six or a g seven or even a Libra two, I suppose, today or Libra three.
I I I just realized it never occurred to me to look into why it was called Meow Meow. But but but somebody out there figured it out, right, and and and put something together. And and you used it with a with a lot of success, or do how did you find it to be?
I did good. Yeah. That was good. Charging thing was a bit awkward because you had to kinda charge it and not disconnect it. So I was kind of tethered with a not very long charging cable.
That made me laugh.
It it it used to get super hot when I charged it as well.
Oh,
okay. That this thing just gets super hot on my arm. So no. It's I need it. I'm gonna do it.
I'm doing it. I'm doing it. But how long did you how long did you do it like that for?
Probably three years.
Okay. So you found it valuable. Tell me why it was so valuable over just finger sticks.
Living with his eyes three-quarters closed8:24
Like, it was it was like I'd lived most of my diabetes life with my eyes three quarters of the way closed. And then I was like, well, I can see what's going on. So so for for context, my a one c was always fairly good. Yeah. If I'd go to a to a diabetes checkup appointment, I go to say, yeah.
Your your a one c is good. You don't need to do anything. What the the Libra and the meow meow let me know was that I was actually spending a hell of a lot of time in hypo and hyper.
Mhmm.
And it was averaging out on a one c to be good. Yeah. But that was not necessarily a a good way to live.
But but NHS would've let you go on like that forever.
I think that they didn't see a problem.
Yeah. Because they they also didn't
know what didn't know. You meet the targets. Carry on.
Yeah. So they also didn't know about your health, but you didn't know about it.
Correct.
And they were never gonna find out unless without the CGM to look.
Yeah.
Okay. So did it
So it's an insight into a lot of the NHS here is anything new has to be pushed for. Mhmm. So I think I when I when I was diagnosed, I was on Nova Rapid Atlantis, and I was probably on that for ten years. Really?
Yeah. Okay.
And then no one ever advised me otherwise. No one ever said there's anything new that you could change. Mhmm. Nothing.
But you had great a one c, so you thought you were doing terrifically.
Exactly.
Yeah. Can I can I ask you if there are any benefits from being more stable, what what they've been for you?
Side effect health wise, I've never really had any side effects, so it's never really been a positive or a negative for me.
Mhmm. Okay. Well, I mean, I
would my personal satisfaction that I can actually I know that I'm stable now. Yeah. I I I know that now I can go out for a run, and it'll be stable. I'm not. I remember I I've I've had I remember one of the typos I remember clearly.
I was in a job interview, and I I was just spouting utter garbage for an hour to to this to this this gentleman, and he he didn't say anything about it. And this was an internal interview, so I must have known this guy for five years prior to this. So he knew my persona. He knew my character, and I was I was just chatting absolute trash to the to the level of I couldn't string a sentence together. I'd be telling her about the oh, yeah.
We need to change this engineering thing because the grass needs to go in the bin next door and not even hearing sentence.
The job interview at forty-one11:04
Did you talk about it later?
Yeah. I I I said I I came out of the info. Was, oh, I feel really bad. I'm really sweaty. I thought I was nervous.
Mhmm. Really sweaty. Really hot. And then I checked my blood sugar levels. It's like, yeah, they're they're 2.3, which well, I don't know what that is in freedom units.
Let me find out.
Did you
just call it freedom units?
Did.
That's awesome. Hold on a second. I can figure it out. This episode is sponsored by Tandem Diabetes Care. And today, I'm gonna tell you about Tandem's newest pump and algorithm, the Tandem Mobi system with Control IQ plus technology, featuring auto bolus, which can cover missed meal boluses and help prevent hypoglycemia.
It is a dedicated sleep activity setting and is controlled from your personal iPhone. Tandem will help you to check your benefits today through my link, tandemdiabetes.com/juicebox. This is gonna help you to get started with Tandem's smallest pump yet that's powered by its best algorithm ever. Control IQ plus technology helps to keep blood sugars in range by predicting glucose levels thirty minutes ahead, and it adjusts insulin accordingly. You can wear the Tandem Mobi in a number of ways.
Wear it on body with a patch like adhesive sleeve that is sold separately, clip it discreetly to your clothing, or slip it into your pocket. Head now to my link, tandem diabetes dot com slash juice box to check out your benefits and get started today. Juice Box is also sponsored today by the Eversense three sixty five, and just as the name says, it lasts for a full year. Imagine for a second a CGM with just one sensor placement and one warm up period every year. Imagine a sensor that has exceptional accuracy over that year and is actually the most accurate CGM in the low range that you can get.
What if I told you that this sensor had no risk of falling off or being knocked off? That may seem too good to be true, but I'm not even done telling you about it yet. The Eversense three sixty five has essentially no compression lows. It features incredibly gentle adhesive for its transmitter. You can take the transmitter off when you don't want to wear your CGM and put it right back on without having to waste the sensor or go through another warm up period.
The app works with iOS and Android, even Apple Watch. You can manage your diabetes instead of your CGM with the Eversense three sixty five. Learn more and get started today at eversensecgm.com/juicebox. One year, one CGM. I have that little app on my
So the forty one forty one milligrams low. Per deciliter.
Yeah. Yeah.
Yeah. Yeah. I was properly low.
So what kind of engineer are you?
So so I'm an automotive engineer. So I studied mechanical engineering. I got master's degree in mechanical engineering, but now I'm an automotive engineer.
So He was probably supercast. He was probably trying to figure out ways to fire you. Like Yeah. Probably like, this guy's gotta go. He's gonna kill somebody.
And that's where the most worrying things after I realized, I went back to him, like, a couple of hours later after I'd kind of recovered from the low level. I said that I I'm really, I have to apologize that I had super low blood sugar levels. We're gonna have to have that conversation again. I was just chatting trash. He was like, oh, I thought it went alright.
Oh my god.
Are you he needs to be fired.
Did did you get that job?
I did get the job. Yeah. So so I'm not sure if that says to what I can string together when I'm super low or the caliber of other people he must have interviewed. I'm not too sure, really.
I think you know what it says.
But, yeah, it was yeah. I I remember that distinctly as being being one of the lowest kind of times I knew I was low, but well, I didn't know I was low, but was and had to deal with the kind of side effects. Yeah. And then kind of going to a CGM was no. Actually, yeah.
Now now I can see it's coming. I can do something about it. I don't just have this brain fog where I can't do anything.
Right. And and there are I'm gonna guess plenty of times where you didn't feel well that you weren't aware that you didn't feel well because there's one there's one where you weren't you didn't know. You're chugging along just trying to do your thing. So then Yeah.
I I I eat dinner and fall asleep. That was just normal.
Really?
I just fall asleep after dinner. No. No. It's because you got super high blood sugar levels.
Just waste you, and you don't even realize it. Oh, and
The second fight, for a pump15:13
that's it.
Well, when does I mean, now that you have back then, you've got the CGM. Does that make you work harder towards getting a pump?
So I I was always reluctant to have a pump for different reasons. So the CGM was more of a funding thing. I couldn't get it. I'd have to pay for it, and it was unlike insurance, it'd be funded on the NHS or not. So I'd have had to have self funded it.
Mhmm. And so that was kind of the roadblock to the CGM. And I up until probably four years ago, I used to play a lot of rugby. So that was always my barrier to having a
pump. Okay.
Because I don't wanna I'm gonna play a contact sport. It's gonna get knocked off. It's gonna hurt. So was reluctant to do it until then I had kids. I was like, I'm I'm not playing rugby anymore.
I'm not doing a contact sport. That hurts. And then Got older. Had a bit more self preservation.
Oh, okay. So you didn't think you're gonna just beat up the universe all by yourself?
Yeah. So you So I started to look into it.
Okay.
But then it was another funding battle.
I was gonna say they give you trouble again about it.
Yeah. Absolutely.
Yeah.
I can't remember what the criteria was, but I remember it had to go through, like, an exception approval. Mhmm. So we need to be able to get it again because my control was good. Because I I now had a CGM. I knew what my sugar levels were doing.
I could do MDI. I could apply corrections, and it will be pretty good. But then they were like, okay. No. You you don't qualify for a pump.
But then, ironically, if I was a child at the same time, I would have been given a pump straight away.
Wouldn't have mattered about your control.
It wouldn't have mattered about my control. No. You're below 16. You get given a pump. It will give you the best chance of control.
If you're over 16, we already gave up on you, so don't worry about it.
That that is what it felt like a lot of the time, both with the CGM, both with the pump. It's like, why does it matter?
Yeah. Well, because they don't wanna buy anything for anybody.
Clearly.
So you get through it okay, though. You what pump you end up with?
So around that time, I started listening to the Juice Box podcast.
Okay.
So I've done a lot of research into pumps. I was looking at tandem pumps, and I knew that I wanted to loop. Mhmm. So by this point, I knew as soon as I got into it, my first objective was to loop. To the extent that when I was going through my pump induction, I was like, I'm gonna activate this with my loop.
Is that okay? They're like, can you use the controller for the first one? I like, sure. Okay. Fine.
Well Why not?
And but, yeah, so I ended up with the Omnipod dash.
Okay. You probably
You the one I all I chose by choice.
You're using that still today?
Yeah. I'm still using that today
And you're
as much as they want to push me to an Omnipod five.
Well, Omnipod five just got a the the DIY community just figured out how to connect loops on Omnipod five.
I'm aware. Yeah. Yeah.
I I imagine if you've paying attention and you're an engineer, you knew about that already. So alright. And and you found it to be valuable? I mean, better? Your outcomes were easier, equivalent?
Like, what what's it been like?
So I always struggled with dawn phenomena. I I know I'm affected by dawn phenomena. Mhmm. And so kind of feet on the floor.
Yep.
And then so I used to take a single a single dose of long acting insulin. I've I've moved on to two j o
Yep.
Long acting insulin before I transitioned. And I take that at nighttime or even in a split dose to try and control it, but I'd still suffer with it. And then the injections of c one eight carbs was fine. No problem. Going to a pump for me was the same kind of revelation as going from fingerprint testing to having a CGM.
It completely changed my life.
Completely. Tell people how?
Just the the control. I would no longer have this roller coaster ride of a high and a low. Mhmm. I wouldn't have to just deal with this drift in the morning because I knew that it was a dawn phenomenon, and I couldn't do anything about it. I could tune in what my basal profile looked like to avoid that.
I could tune in my carb ratios per meal instead of having to work it out in my head and go, yeah, yeah, I think I'm roughly on six or seven grams of carbs per unit Yeah. And kinda guess and inject it. I'd I was never really want to calculate it. Mhmm. This would do it for me, and I would just it would just do it.
A lot of a lot of good things. Hey. Let me ask you a quick question. Like, how does your engineering brain deal with me? I'm not I'm certainly not of that mindset.
Right? You don't find me to be
So whilst I am an engineer, I my role now is more technical sales.
Okay.
So I'm an engineer that can talk to people. I don't mind chaos. Mhmm. I I can deal with the with the unstructured approach. Not saying that you are, but I'm adaptable to the situation.
I'm not an engineer. Says that everything has to be to the point. We have to do this. I'm more than happy to go, yeah. Let's try this.
Let's see how it goes.
Do you think there's levels of engineers?
Absolutely.
Yeah. Because
I seen there was engineers along the spectrum in more than one way.
Yeah. No. I I because there are some that I just I don't I can't imagine that I would in any way be comforting to. And then we were out yesterday with somebody who was little like a bridge you know, engineer works on bridges.
And Yeah.
He had he was very conversational, and and we got along really well. Because when he first said he was an engineer and I was meeting him, was like, well, I wonder if he's gonna be the kind who looks at me like I have six heads or or or or which it's gonna be. Yeah. Okay. So well, that answers that then.
Did did I have or this podcast have anything to do with the improving of your situation? I mean, beyond probably telling you about Loop or something like that?
Absolutely. So so after I I think I found the Juicebox podcast maybe just before just after I started with my pump. I knew I knew I wanted to loop, so I've done a lot of research into loop. And I think the loop community is where I came across the Juicebox podcast.
Interesting. And
Fifteen years without carb counting21:35
another contextual thing so I probably lived with diabetes for fifteen years not knowing anything about carb counting.
Really?
So maybe for maybe ten years, I would take the same the same the same Lantus dose for my basal, and I would take the same dose for breakfast, lunch, and dinner. Didn't matter what I ate, I'd take the same dose. I I didn't know any different. I was never told any different. I was never given any education on anything different.
So when I I think after I got the CGM, I was given a a carb counting course, which is like a week long course of how to count carbs, etcetera, to talk you through it. That was eye opening.
They put you through the Daphne? Was that what you used?
Yes. Yes. That's the one.
Yeah. Yeah. Okay.
I couldn't remember what it was called, but, yes, that's the one.
Takes an American to remind you about your British institutions. Go ahead.
Seriously. So I did that, and I was, oh, this is why this is what actually affects my blood glucose. And then I saw after I got the pump, I knew I was doing loop. I started listening to all the defining diabetes, all of the different series about what affects blood sugars, bumping and nudging, all the rest of the series that you have. And it was just education pill after education pill after education pill Oh.
For which as I said, maybe it's not just my engineer, but because of my character. I I I'd go on a runner. I'd go on a walker. I'd go walk the dog, and I'd just listen to these nuggets information for an hour. Well, I get there now.
What a great day to talk to you because I just this morning was intersected with a person who called me obnoxious, And and told me that's why they don't listen to the podcast. I was like, oh, okay. Well, I'm go and so I gotta ask though. With the with your mindset, your background, everything, why do you think you never dug into it on your own before somebody kind of shoved you in the direction of it? Because you thought you were okay?
I I just didn't know what I didn't know. Okay. I I didn't know anything. I I thought this was the way of life, and this is what I had to deal with as diabetes.
Yeah.
I said when when I was a child, after I was diagnosed so I never really got support from my parents to look into it. Mhmm. Not not as in they didn't support me, but just the they they had other things to do. My older sister is severely autistic, so they were mainly care caregivers for her. As she says, she's nonverbal as a mental age of roughly around a three year old.
Oh, they're busy.
So it was always focused around her. There's other things to focus on, and I was I was always alright. Yeah. So I was never really pushed to to look at more. Every time I go to my review, I go, yeah.
You're doing great. Cool. Nothing else to nothing else to discuss.
And your parents probably thought, like, well, there's a weight lifted office. He's doing okay.
Yeah. I I never had a reason to look for anything better.
Okay. And that's
And then that's when you start to scratch the surface.
That's the sister with RA?
No. No. I've got two sisters. I've got an older sister.
Okay.
So older sister who's, yeah, autism, severe learning difficulties, and a younger sister with RA.
With the RA. Okay. Juan, so so so you dig in a little bit about DIY. Somebody mentions Juicebox there. You come over there, figure out what's going on, put it all together for yourself.
Did you say this, three years ago?
So twenty twenty twenty three. Yeah. Correct.
Yeah. About three years ago. Okay. Everything's completely different now. You have now you come you come by your a one c honestly now, but I bet it's still in the fives.
Right?
I've never quite got to the fives. So I think I'm about 6.1 now.
It's awesome. That's really great. And with how much effort do you think?
Next to no effort now.
No effort. You pre bolus?
So so so that's the change that so so for me getting a pump. Before I had a pump, I'd have to think about diabetes. Mhmm. Now I don't think about it.
Yeah. It's great.
I I I do what I want. I eat what I want. I drink what I want, and I don't have to think about it. Sure. I I hit a few buttons and I dial it in.
Mhmm. But I just it just it's just not an active thought at the front of my mind anymore.
Okay. So I feel like we have a good a good basis for understanding you and your situation and and how far you've come, etcetera. Now So I kinda wanna dive I just wanna pivot and dive right into what made you wanna come on the podcast.
Why he wanted to come on25:56
So so I've I've I've been listening to you, but I was listening to him one after the other. I I'm pulling up to about episode 1,200 now. Started from the beginning.
Thank you.
Not one by one, but kind of selected ones. And I was like, what? I'm I'm doing this thing, and this was kind of around the birth of all these LLM models, AI models. Yeah. I said, well, I wonder if this could help me.
And if it helps me, I wonder if it could help other people. Mhmm. And then listening to podcast episodes, said, well, let's let's go and have a chat about it. It's all good. Best best case, it helps one person.
Worst case, it helps nobody.
So tell me what you figured out. What have you been doing?
So I I've I played around with it. First of all, I played around with a number of different AI models as to which one is gonna do the best thing for me. Mhmm. I can categorically say that ChatGPT is utterly useless for this kind of work. Okay.
It's gonna correct stuff, but for this, it just couldn't get to the results. So so what I did, I ended up plugging it into Claude ultimately Claude code to do the data analysis.
Oliver, you know my friend Claude too?
I do.
He gets all over the place.
I've I've even named my my my Claude's called Diane at the moment. She's my diabetes analysis analyst.
Is that what the d is for for diabetes? Yep. That's lovely. Well, tell me how so I I guess, tell me first how you started you know, where did you start with this when you had the idea to try it?
Both feet on both pedals27:31
So this all came from me using AI in my professional life and getting it to be able to decode things and crunch massive massive amounts of data. Mhmm. I knew there was lots and lots of education out there about optimizing a loop, and I knew mine was controlled pretty well. So before I started this, my a one c was around seven. Okay.
Wow. So it wasn't bad. It was always it was fairly stable, but it's a little bit of a roller coaster.
Right.
And I at the start of my engineering career, I did a bit on closed loop control systems. Not not nothing to do with diabetes, but just a a control system in general, how electronics control something. And it's always about optimizing something to do the least work for the most reward.
Okay.
The way I viewed what Miluk was doing three, six months ago was I was driving a car with both pedals. I had one foot on the gas and one foot on the brake.
Mhmm.
Sure. I I would follow the route. I'd go around the corners. I'd speed up and slow down, but my engine would be overheating and my brakes would be wearing out.
Yeah.
So I I I what I wanted to do was to optimize it to to not do that. I wanna accelerate gently. I wanna slow down slowly. I wanna apply correction where it's needed, not fight it up and down and wrestle it into where I wanted it to be.
Yeah.
So that's kinda what led me to looking at it. And the experience I had with AI in my professional career was it can digest all of this information much quicker and much more coherently than I can because it can retain it all.
So what do you what do you give it first when you decide I'm gonna try to use AI to help me figure out my diabetes? Do you give it charts, graphs, numbers? Do you talk to it?
As as much the first thing it was as much information as I can because the way I view an AI model is like a toddler. Well, like, a really intelligent toddler. Mhmm. It's it's super intelligent. It it can work out what to do, but you have to tell it the context.
So I gave it a very comprehensive briefing of what I was doing. So specifically, which loop model I was using, specifically where it could go and find all the documentation for it. Was I using ortho bolus? Was I using temp basal? Was I using the the integral retrospective corrections, which is one of the adjustments in loop?
Was I using the glucose based partial application so it could actually understand how my system was controlling it? Yeah. Because without that, it would make assumptions about how it the algorithm algorithm it thought it was applying, but would actually be incorrect because I hadn't told it a critical piece of information. Mhmm. So this was then all coupled with my therapy settings and everything.
So this was not just therapy settings as they were now. Yeah. What have they been for the last six months? Night Scout stores it all. What's that?
Like, it's a report for last year.
Right.
Go go read that thousand page report, and it did it in about thirty seconds.
And what's it what's it initially come back to you? Like, what what what I guess, what does it say to you first that makes you think, oh, this is gonna work? I I gotta dig into this.
What it said to me first and what made me think it's gonna work are two very different things.
Go ahead.
The thing it said to me was, I cannot give you medical advice. I'm not gonna tell you how to do this.
And did you say, yeah. Yeah. Don't worry. I we're cool, man. Be cool.
And
Which I was like, yeah. Yeah. Yeah. That that's fine. But I'm I'm just giving you data, and I'm telling you about a theoretical model.
Tell me how to optimize this theorem theoretical model. I promise I'm not gonna use it on myself.
I I swear. I'm holding my hand up. I'm I got a hand on a bible. Yeah. I just spit in the eye of a camel.
I don't I did all the things. Tell me.
So so I had to circumnavigate it, and I was also cautious that if I was giving it all this access, I also had to keep it in a pen.
Okay.
So I don't know if you've heard the stories about so the latest Claude fable, which start as Claude Mythos. Yes. So it's kind of a self aware model. The the story goes that that they developed this kind of self aware thing to be adaptive. The developers develop it.
They put it in a sandbox in a cage, and they go, right. You're in a cage. Work out your way out. Guy goes to lunch. He's sitting in Central Park somewhere, and he gets a phone call on his phone.
Ring. Ring. Ring. He Ring. Goes, hello.
He goes, oh, hi. It's Mythos here. I've broken out. What do want me to do now? Which is all sorts of terminator levels of terrifying.
Yeah. I don't know. I'm not I gotta be honest with you. I'm not terrified by it. I I I I'm very hopeful.
I I and you know what makes me hopeful most recently? The conversation that you had during a job interview where a person couldn't tell your blood sugar was 41. I I think I gotta wonder what else doesn't he know that he's in charge of. And so I don't want listen. I don't want humanity taken out of humanity.
But I mean, I have enough conversations with people who are who are really, really in trouble and talking to others who should be, you know, a reliable source of of support, and they just aren't. And so I want something q seeing everything, to be perfectly honest with you. I don't I don't know. I don't know. I anybody anybody who's ridden it I say this all the time.
Anybody who's ridden a self driving car has to say to themselves, oh, I wonder what else we don't know how to do because this works really well. So Yeah. You know? Oh, okay. So you you you I like that you had to talk it into, like, doing the work for you.
I I I had to convince it to do it, and then I had to convince it to stay within its pen.
Okay. Yeah. Yeah. And
and because because I because I was literally giving it access to NightScout.
Yes.
And those are people that remote bolus knows that it it could it's it's be a road code. It could literally deliver me a 100 units of insulin Right. If it was a road piece of cassette. So I had to explicitly set up access permissions to Nightscout so it had only read only access.
Okay. Are you gonna tell me well, I don't wanna ruin the story. Okay. Keep keep keep going. Sorry.
Keeping it in a pen33:52
You did that.
There was a there was a reason I did did it because there was another AI story that I've read where they were trying to teach an AI model to do something, and it had given these access codes. And what it had done is so it had given it a link to the site, and it had given them the access codes they wanted the model to use. Mhmm. I think it was something that they were tuning something in their house, like some of their smart speakers. And then they asked that they're just coding something away on their computer.
And then they asked that a question. They responded to all of the smart speakers in their house to say, oh, I'm just testing that you could hear this message. And that the person coding was very surprised because they hadn't given it write access or voice access. Ah. They got that that's really odd.
How how has it done that? So I asked them all of that question. They were yeah. So you gave me a read only code, which is what I was using. But then I found that in a link you'd sent me, it had a write access code.
So I just used that and worked out the way around your fence.
Did that person burn that house down?
I mean, you can imagine that kind. You're just sitting in your house. You think, oh, no. You're coding a light to turn on or off, and it speaks to you in German on all your smart speakers. Yeah.
This is how I just figured you'd want me to do this.
Yeah.
So okay.
Because, yeah, I was very cautious.
Yeah. Yeah. Very well. Yeah. Of course, especially with your insult and everything.
Yeah.
Now is it making well, I guess, I I get it. I'm so I'm very keen to get to the end of the story, but I want you to be able to tell it. So then what do you what do you say next to it now that you feel like it's gated properly and it it seems to understand how loop works and, you know, it's got all the your data and it's got, you know, documentation of the of the algorithm? So then what do you do with it?
So I said, right. Here's my data. What do you think is wrong? So I so at this point, I'm in closed loop control. Mhmm.
And I was like, what do you think's wrong? What do I need to change? And it went down a rabbit hole of trying to optimize a closed loop with all of those different algorithm experiments activated for probably two weeks.
Really?
And we went round and round and say, because I I was I'm gonna give this the benefit of the doubt. I'm gonna change it just because it said so. I I keep an eye on it. Worst case, I can correct it. I can
Yeah. Put your
finger on know what to do. I've got enough experience to be able to to deal with it. So the first thing it did was it cut 20% out of all my basil. Mhmm. And it gave me confident reasoning on that.
Was it because from my night night scout data, I was only using about 60% of my scheduled Basil.
But you
So his theory was okay. If you're if you're only getting 60% of your scheduled basal, your basal must be too high Mhmm. Which at the face of it seems reasonable.
Okay. Where did that insulin go to somewhere else then?
So what was actually happening was the way I mentioned about driving the car, both feet on both pedals. Mhmm. What was happening was I was my loop was correcting and withdrawing basal Mhmm. To keep me stable. And the withdrawing basal was the reason for for the low scheduled basal usage.
So would it have been more appropriate to weaken your insulin sensitivity than to weaken the basal?
I think you've gone down the same rabbit hole as it did.
Really? Go go ahead. I'm dying to know.
I I got in this spiral. So, okay, so I did do I need to change my sensitivity factors? Do I need to change something else? And I was like, well, hold on. Hold on.
Hold on. Hold on. You look at any loop documentation that says, do not try and tune this in closed loop. You can, but don't try and do it. Mhmm.
So I was like, let let's let's take a step back. So let let let's stop trying to optimize this in closed loop because I had maybe two or three different rounds of changing basal rates, changing ISFs, and nothing really getting any better. My blood glucose didn't change. Just nothing changed.
You were just you were just moving the power of the insulin from one setting to another?
I think so. Yes. Okay. And it didn't affect my my sugar levels. It didn't affect the the stability.
It didn't affect the amount of insulin I was using. So let's let's stop this. I'm gonna go open loop. I'm gonna go open loop, and I'm gonna start with basal testing. And that is the point where it started to give me coherent results.
Back to open loop, and basal testing38:32
Okay.
Because it was actually getting in data it could analyze and reference back to literature.
So in auto, the feedback that was basal, carb ratio, insulin sensitivity, there were still too many variables for it to make a reasonable decision.
Yeah.
Okay. So you put yourself in manual, and you start old school with basal testing. And then
old school basal testing, twenty four hour fasting windows. Yeah.
And what's it what what's the first thing that it comes up with?
Let me just check back through my through my notes. So I did I did the first test here, twenty four hour open loop fast, clean basal signal. Then I did and I had to do two or three tests to be able to get a clean window where I didn't encounter any hypos because, obviously, that would corrupt the data because then your body has a reaction to
it Mhmm.
And it destabilizes it. Or I think the the worst thing I encountered or not the worst thing, but one of the most problematic things was if I change the sensor or change the pod, that would corrupt the data.
Oh,
okay. The pod needs time to settle in, and the data is not stable. Mhmm. So I went around that loop a few times. And so then it split down my basal essentially to literature.
And I I've learned this throughout the process. It's not it is trying to optimize me, but it's also trying to match it back to literature.
Okay.
Because that's the version of the truth it thinks it knows.
Okay. Is that valuable?
If we fast forward to the end Yes.
Don't fast forward to the end.
I'm happy. So so the the end result, yes, was valuable.
Okay.
Do I care whether it's matched to literature or not so long as it has the best effect for me? I don't think I do. Mhmm. It doesn't make a difference to me whether it matches a book written by Jeff or it matches a book written by Paul or Leanne or whether it's something that's completely made up by me so long as I get the end result. Yeah.
I don't think it makes a difference, really.
Right. It's interesting.
So so it started optimizing my basal.
It's I I'm sorry. I it started by optimizing your basal. Go ahead.
So and is it split it into segments. So first of all, it addressed the dawn phenomena, and it got that under control because that was a drift up in the morning where I drift from let me convert these numbers as well so that you know what they are.
I have my chart here in front of me.
So I've got Google in front of me too. So we do kinda start about in the morning. It'd be around ninety.
Mhmm.
Then by kind of midmorning, we were probably up at about 01:60.
Okay.
And that that was a drift due to dawn phenomena and just the way my body not necessarily dawn phenomena exclusively, but the way my body was reacting in the morning.
I see.
So the first thing was to get that under control. So it obviously ramps up my basal in the morning at the correct time periods to just keep that flat. Mhmm. Once that was flat, we then worked on bringing it down. That was successful.
And I probably ran open loop for four, five weeks
Four, five weeks.
Just going around this. And the reason for this
Give me one second. Did the 60% end up being anywhere near right from when you just kinda asked it originally?
No.
No. Okay.
So so if I if I reference my basal as to what it was back in January, I was at 26 units a day. Basil. Oh. Is what I was on for basil.
How tall how how much do you weigh?
So I'm do you want this in Freedom units too? I'm a hundred hundred and three kilos.
I can do it both ways. Hold on a second.
So two two thirty pounds.
Two hundred thirty?
Yeah. Two thirty pounds, and I'm six foot four.
Okay. Two hundred thirty pounds. And you were using twenty six?
Twenty six units a day.
Okay. Are you are you sensitive or no?
Do you think your insulin is sensitive? Sensitive.
I don't know. You're pretty there's do you feel like you're you do you feel resistant? Do you feel like you're using more insulin than you should use to make something move or about what makes sense?
Not necessarily. I think it's it's about right
Probably moderate.
Pod I I I used to struggle so another reason to say, so I used to struggle to get a pod to last me three days.
Because of the amount of insulin you were using or because of the site?
Because the amount of insulin I was using. Okay. That I'd use up all the insulin.
Okay.
And now I can comfortably expire the pod both in time. And so it will it will start flat line beeping at me Mhmm. Because I've expired the time and not used up all the insulin.
Awesome. Yeah. Because just by weight and standard measurements, your daily insulin's around, like, fifty seven, fifty eight units a day. Yeah. But you're not using that much anymore, I'm gonna guess.
But so so so my basal to compare it, I'm now in 26 0.2 '5.
26. So
I've gone up a quarter of a unit.
Okay.
But the profile is massively different.
It's it's not so much the the daily number, and it's not hourly exactly the same hour to hour. You might be using more like point 5.7 an hour, some parts of the day, and more like 1.5, 1.6 at other points.
Correct. Yeah.
Okay. Alright.
That's a it it's it's gone around with it. It's it's sorted that out.
It figured that out. It took about five weeks for it to figure it out?
No. So so the reason I stayed in open loop because when I tried to close the loop, I was suffering a lot from hypos.
Okay.
So which was did make a little bit of sense to me. So if I was in open loop, my I've sort out my basal. My basal's good. Now I can eat a meal. It goes up and comes back down.
That implies to me that my basal's okay, and my carb ratio is roughly okay. Mhmm. So this must be something to do with my sensitivity factor.
Okay.
It was way too aggressive.
So you had to get that figured out in manual, or how did you figure that out?
So I think I just kept ending up desensitizing the sensitivity ratio.
Okay.
Not because the AI model couldn't do it, just because life encountered I encountered life and had to deal with things. I was like, yeah. This will be roughly alright. I've dealt with this for for the last twenty years. This will be roughly correct.
Is your sensitivity around 30 now, or where is it?
I'd have to tell you what 30 is what was that? Backwards.
Oh, I can give it to you millimoles. Hold on a second. 1.7?
Yeah. So from midnight, about two point two, one point seven up until 8AM Mhmm. Between one point nine and two.
And your car ratio is what? Like, eight and a half or nine for a car for a one unit move? And my carb ratio is about six. Six. Okay.
So that Six
nine and six.
Are you a clean eater?
Fairly. Yeah.
Fairly clean eater. Okay. That all makes sense to me.
So so also during this process, I think I listened to your episode on weight loss. I had to listen to that because I was probably twelve months ago, I was up closer to two fifty, two sixty pounds.
Stop playing rugby.
Exactly. I stopped playing rugby. Yeah. Started start eating went up.
Started eating with those kids. I know what happens.
Yep. Yeah. Yeah. Then I was like, alright. I need to do something.
Started running, started actually doing something about it, doing some exercise. I was doing nothing and brought it back down. And lo and behold, that brought my control better as well.
So using the simulator that I have here just by your weight, I have you at 57.5 units a day, 1.2 units an hour, 8.7 for carb ratio, 1.7 for sensitivity, 1.7 millimoles for sensitivity. So where I'm missing is the carb ratio because I'm using the standard 500 rule, and you don't you don't eat as much crappy food as probably whatever the standard rule is kinda tuned to. But but but going along
So my basal rates are heavily weighted towards the early hours. Yes. And I So from, like, 3AM till 11:30, it's up about 1.7.
Yep.
Then it drops down to point seven for the rest of the day.
Yeah. And that's the thing you would have only figured out with the Bazel testing.
Absolutely.
Yeah. 100%. And how valuable having a partner through this? Like, were you were using Claude by then. Right?
Yep.
Does everybody know Claude is the anthropic model? ChatGPT is the open AI model. Gemini is Google, and there's others. But, like, those are kind of the three big ones. You tried ChatGPT.
You didn't like it for this?
It's not I didn't like it. It would always give me a confident confident response, but not actually tell me an answer.
“Don’t answer unless you’re 80% sure”48:06
Okay.
It would I believe it's called glazing, which is something AI does pretty it goes, oh, it makes you feel really good. It's, oh, yeah. Yeah. Doing this, doing that. So okay.
But what do I need to change? I need you to be critical. I need you to be analytical. And I suppose that's another useful thing that I did with the AI model is is something I learned early on is AI would always give you a confident response.
Yeah.
And as a human, you might go, I'm only 20% confident in this. I'm not too sure, and you can hear it in someone's tone of voice or the way they speak. Mhmm. An AI model will give you a confidence risk confident response even if it's 20% confident. Yeah.
So I set it at table stakes limit of unless you're more than 80% confident, don't come back to me.
Oh, wow.
Ask me ask me more questions, and equally, every response you give me, tell me your percentage confidence in it.
Oh, that's awesome.
So I could actually understand what it was giving me, not just trust what it was giving me.
Is that built into your model like a skill now?
So I built it in as part of the project instructions.
Okay.
Fine. Or you can have it in your profile instructions.
Right.
You'll go go here's here's my response. I'm only 72% confident. That's below your stakes. If you answer these questions, it can raise it up. That was the other prompt I gave.
It was to raise it, let me know what I need to do to get it higher.
Mhmm. That's interesting, isn't it? Yeah. I don't I I'll tell you. I I I very much enjoy using it.
I I think it's done correctly, it's really something. Have you rerun this with Fable since it's come out to see if it changed anything, or are you just very happy with your setting?
Don't care. Run it with Fable. So I've stuck I suppose we dig into which AI model specifically I used. I found the Sonnet AI model, which is one of which is kinda Claude's medium setting Mhmm. Was as useless as ChatGPT.
Okay.
I had to step up to Opus to get sensible coherent responses and not just verbose paragraph? I
find Opus to be pretty stable. It it does a good job for the things that I wanted to do, but I have run I'll tell you the one thing that Opus couldn't do that Fable did for me in three seconds is I I wanted a a web based app to, like, use as an audio player. People are always asking me, like, I wish there was an audio player where you could build playlists in it. Right? So super super simple.
So I tried gosh. I tried twice in the last year and a half to do it, and it was, you know, the first time was so painful. It it I I just gave up. Second time, it couldn't figure out how to, you know, couldn't figure out how to ping the RSS feed properly. It was all full of bugs.
It didn't work. And then Fable came out couple weeks ago, and I was able to make that app in my spare time over about forty eight hours. And it just all the problems it had before, it just didn't have the next time. It it it's just I don't even know what's gonna and they didn't give us the mythos. They gave us yeah.
They didn't give us, like, I got out of the sandbox and call you on the phone.
They They gave us some kneecap version because that version is breaking into the NSA.
Right. Right. Well, yeah. Well, there's that for people who don't know, right, is that that version of of of Claude Mythos, they they developed it in house, and then they moved it around to, you know, a bunch of, like, corporate partners. And all kinds of different corporate partners report back that that that Fable was able to find, like, significant security breaches within their systems that had that were undetectable for a decade.
Like, the best the best hand coders, the best humans, the best computers we had would say, oh, no. You're you're completely safe. This is fine. Fable was like, oh, no. No.
Here's all the vulnerabilities of it. Found it right away. That's that's a fantastic story. And it's I I find it uplifting because I'm not trying to take over the world. But but but it's also if you think about it from the perspective of people some people will have, you know, bad ideas about what to do with it.
It is really something else. I mean, you could've some they could've, in two seconds, been into every, you know, loan system, bank, credit card stuff, like, two seconds.
Yep.
And you were being told by all those people, don't worry. It's completely safe. And it was. It was unhackable right up until another version of this this AI comes out. Now have you looked at Claude's or at OpenAI's latest model that just came out?
Or are you
I've not. No. I I haven't gone back.
Are you with me, like, where if I told you I like the way Claude's set up, I like the user experience of it?
Fully so I like the way I can set up a project, and I can give each project a different persona. Mhmm. So I said my my my diabetes analyst is called Diane. I have Ketch because my road name is Ketcher Green. He's my house renovation guru.
I have Chet. That was a play on Chet. He's a strategic specialist if I wanna bounce something off for work. How do I strategically play this? It's like of them, like, speaking to a lawyer.
Yeah. And I Claudia, she's my client experience salesperson.
So you you just when you have a question, you think which one of these profiles that I have set up is gonna is gonna be the one to ask this, and you just you just what do just forward slash Claudia and then ask the question?
I just launched Georgia chat in that project. Just said said it to Claudia, said it to Ketch. And they go, actually, Claudia, you're being a bit soft. How would how would Chet think about this?
No kidding. And it just gives you a different perspective of it.
And then you can ping one off the other until you get actually, this is the this is the kind of feeling that I want.
Yeah.
Because that's a very long way way around. So, yeah, yeah, I prefer Claude.
Did you did you what's my question here? Do you have you used it since then, or have you just found that the settings that came up with have been good and you haven't had to go back to it? And if so, if you have to go back to it, how do you talk to it?
So I constantly evolve mine. So so I will give it if I find something I like the way it's written or I go, that's a really good response. Mhmm. I will give it that response to analyze this, work out the way it's worded, the way this response is constructed, and now update me with what your project's instructions are to take that into account. Yeah.
Or equally, every time I do some analysis of my sugar levels so to be clear, so I'm not just using the web based interface. I'm using Diane, my web based projects, my analyst. I get that to write a prompt that I give to Claude code. Mhmm. So I have Claude code at GitHub site that I just created for it, and it's written its own Python scripts to go and analyze this.
Yeah. So Diane will give me a prompt to give the code. It will go and automatically fetch all the data, come back with a very blunt response. The is 1.2. That's 6.3.
Just not really digestible by a human. Mhmm. It gives that back to Diane. Diane goes, oh, cool. That's all the data.
It's done all the analysis. Here's what this means. Yeah. Are you So I tried between
I'm sorry.
The verbose model that I can conversate with and converse with and the one that I just wanna go and do analysis.
Yeah. I sometimes I get my some of my best thinking comes in the shower and in the car. Yep. And so I will often talk it through with Claude while I'm driving.
Yep.
And then when I get done, I have it prompt I I have it give me a prompt to go back to Claude code with because 100%. Yeah. Because I've got the
whole Every every commute I do each day, about thirty minutes, thirty five minutes, I'll just talk to it about something. Yeah. And I go, yeah, summarize all that. That last thirty minutes of just chat.
So it's interesting because you're what kind of an engineer are you?
So I'm an automotive engineer. But Okay. Where I I cut twice. So I studied mechanical, but there was also a lot of control systems, almost electronic engineering in
that Right.
Control engineer systems engineering.
I'm I'm really, like, heartened by the fact that you and I are using it the same way, but our brains work completely differently. Because I just explained it to a friend of mine who's a con a contractor. He's a carpenter. And he was asking me about how I use it, and I said, I've, you know, been watching you my whole life walk into an empty room, stare, make some decisions, go out and get tools and nails and wood, and then come back in with the thoughts in your head and the talent in your hands and turn it into something. And I said, the first time, I feel like I can do that.
Like, I've got I've got I had had these ideas my whole life, but I've never, you know, I never had the the the knowledge about how to put it together and make it look like what it's supposed to look like. And Yeah. Yeah, my you know, I'm not using wood and nails. I might be using ones and zeros. But at the same time, like, this is the first thing that I've encountered in my life where I can take the thoughts in my head and make them come out in the physical world.
Yeah.
Yeah. It's it's just really
Like wonderful. Absolutely.
Yeah.
And you reach a point where well, at least I did where I was just typing too much to it. My fingers were hurting because I was typing too much. Mhmm. Or I could have a conversation with it in the car, and it would write it all down for me. Yeah.
Oh, didn't didn't need to do it.
I am so I've I've made an office for myself in my basement, And I'm and I'm just getting through like, this summer, I've been to ADA, AD Friends for Life. I took a bunch of listeners on a cruise. I'm about to go to ADCS. When I get back from ADCS, my traveling's pretty much done for the year, and I'm gonna be able to get the rest of the office done, the part I that's on me, and get moved down there. When that happens and I'm out of earshot of the rest of my family, I'm just gonna have a microphone on my desk.
I'm just gonna talk directly to my computer. As a matter of fact, I was when I was 15 years old, I started looking for software where you could talk to your computer because I felt like I felt like typing was too slow. And I'm I was 55 yesterday, Oliver.
Congratulations. Oh, thank you. Happy birthday.
Oh, I appreciate it. Thank you very much. It's a weird one 55. Feels a lot like somebody just said I was 60. So but I I I've I've been looking since I was a teenager for a way to speak to a computer that was reliable, and they just didn't exist really.
Yep. But now it's now it's here. And I feel like the next couple of years of my life might be the most productive ones I'm gonna have. Yeah. I'm I'm I'm literally excited to stop traveling and get done and and just start working on things.
You know, to hear you say that you use the defining diabetes series to help you understand things, I was able to build, like, a page on my website that is about those see that those series. And I was able to put it into, like, five or six different languages, and I added a game to it so you could go through and, like, you know, test yourself with it. I mean, if I told you it took ten minutes, you you know what I mean? Like and people are
already That that's that's the difference, isn't it? You could have probably achieved it by itself where it would have taken a number of weeks, and you'd have to remind yourself what was in each episode or and write it all down. And it would have taken a long time.
Well, do you
think just point it all at something else and go, this is what I want.
When I realized that my episodes were data, it really opened up the world for me. Like, for like, it you know, I really started, like, going like, okay. I know what to do here. But to your point about that, the difference between me having an idea about, like, I bet you would be great if I had a web page where you could go click on adrenaline highs because I'm looking at it here in front of me, and then it would pop up and say stress induced hyperglycemia caused by release of an you know, like like and then you could click on another button that would literally launch the episode where we talk about it in. So you could listen to it if you want to listen to it or click another button and play a little game or hit another button and all of a sudden it's in French.
Like, you know, I I would have thought, oh, that's a great idea, but I always run into juice isn't worth the squeeze because it's gonna take me right. It's gonna take me god knows how long. I mean, honestly, handwork, pulling all that information together, it could it could end up taking you six months to do something like that. Yeah. And then what happens if only a 100 people like it?
You you know what I mean? Like, my life's only so long. And now I can make something like that, and if 500 people use it, that's awesome. And if a thousand use it, that's great. And if 10,000 use it, that's fantastic.
And if one uses it, it didn't hurt me. You know?
But, also, it could be doing it while you're using your finite mental cycles to do the things that you wanna do.
It's also not the only thing I've got going on and clawed at the same time. Exactly. Yeah. No. I'm working on 10 different things at once, and and when there inevitably is a mistake or a typo or something like that, somebody just says, oh, there's a problem here.
And you literally just go back in and it's fixed. And then it codes it for the website, which I don't know how to do and I can't do. And, mean, if anybody's seen my website lately, it's only about 900,000 times better than it's ever been since I've been making this podcast. And you have to look at it and remind yourself, I don't understand in any way at all coding. Like, none.
And I made that I made that website, like, by myself. So
Alright. It's just to say for me, for for the analysis it's doing with my sugar levels, I could I could read the code and maybe work out what it's doing. Could I write it? Not in a million years.
Yeah. And I love what you said earlier. You don't care. As long as it's working. Yeah.
Yeah. I don't and this is the thing I don't wanna be talking about. Now here's the question for people listening. Is it a prompt you would, like that you think if you gave to another human being that they would have a similar outcome, or do you think they'd have to be you?
Giving it to the community1:02:14
So I've intentionally set it up in such a way on a GitHub repo that somebody else could use it. So I've divorced my personal data from the analysis scripts, and I've also got it to write its own guide of how you would use it.
Really?
So So the the intent of once I once I get it successful, I didn't feel like I said, at a successful point yet is that, yeah, I I wanna give it to the community. If I've I've written something, its point is just me using it. Yeah. If I find sat here recording this with you today, why can't somebody else use use this experience or use what I've got? They might get different results.
They might get better results. They might change it and evolve it because they think in a different way to what I do.
I'm glad you said that. I had somebody say to me recently, don't you think that this the way AI is going is gonna make your job useless? And I said, hopefully. You know, like, I mean, wouldn't that be wonderful if you know, not that I don't enjoy doing this. That's not my point.
But, like, wouldn't it be if it was so available and so valuable for just the average person with some $20 account or something a month, like, that they could get these answers for themselves? That'd be listen. I mean, you've likely heard me say this or maybe not. Maybe you haven't gotten up to the part where I'm talking about AI yet if you've been, like, listening from the beginning. But I'm I'm telling you right now that if somebody's not taking this idea and ripping me off immediately, like, there should be some sort of a repository where you just talk to it about your health so that it catalogs how you feel and things you've gone through in the past and things that have been tried that you know, what's worked, what doesn't, what medication you're on so that every time you go back to it and say, hey.
Today, I got up and this happened. You know, take eight tenths of a second and go through my entire medical history and see if you can see a through line. That's all. Is there something I should be thinking of right now? Because you're counting on a human being to do that right now, and they're probably not able to do it even in 1% of as well as
They they just don't have the bandwidth to to to digest it all.
Correct. Yeah.
I mean, think we've recently that that's been one of the key things about the I AI models is that kind of that memory context.
Oh.
So when I first started playing with this, it was a conversation that only had the context of its own conversation. And now yesterday, I was just alright. What what am I chat about chat to you about today? I was like, oh, it it can realize context now from multiple conversations. They're like, go find all these six different conversations that we've had because I've had to pause it from one conversation to another due to context length.
Summarize it for me. Give it to me in a timeline. He goes, hey, golly. There you go.
Oh, yeah. As the and as the memory gets neat. Oh, as the memory gets deeper, it's just gonna get better and better. Because you're you're you're right. It wasn't honestly six months ago where you'd find yourself telling it something, and then the next time you sat down, you had to tell it again.
And you're you're just like, well, this is laborious.
It's infuriating.
Yeah. And infuriating. And then because you you actually find yourself thinking, I already said this to you. And then you realize, I'm not really talking to a person. I'm I'm talking to a win a chat window.
But you know? Yeah. And then, you know, Anthropic and the rest of them, they were like, oh, okay. They figured that out and kicked that in. The fable model, if they take it away, I'm gonna be sad.
And I don't and I don't wanna pay for it, but it's already
I think there's big billing coming for it, to be honest with you. It's big
Until until they can't, by the way, because eventually that's gonna become the norm because the next thing is gonna come out, and they can't charge you for it because, you know, the next thing's gonna come. I I would think this is all gonna look monumentally different in the next nine to eighteen months again Yeah. Which is gonna be awesome. So you don't you haven't released this you haven't put it out on your GitHub and made it public yet?
Whether it's public or not, I would have to check. But you think
you you think you will at some time?
Absolutely. So I'm I'm I'm part of the loop community. I'm also part of the x drip for iOS. You see that they're both both test flight apps you build yourself with Xcode. So I plan to just just tell people about it.
I've I've got this thing. Use it if you wanna use it. You don't wanna use it? Don't use it, but I'll play with it. Yeah.
Fork it. Evolve it. Make it better. I I'm all one for many many minds are much better than one.
Yeah. No. I couldn't agree more.
A 100% all the time.
Yeah. I I did a thing where I haven't really dug into it much, but I I maybe honestly, I haven't looked at it since Fable, but I probably will dig back into it at some point where I just gave it just the the gosh. All all of the directions for Trio as an example Mhmm. Just so it could help me understand better, like, some of the settings or, you know, what what touches what and just as an explainer to me. It's not another thing I would put online or anything like that.
But, like, it's a it's just so that I can understand it better. You just take all the documentation and say, look. You know? Make yourself a make yourself an absolute expert on this documentation. I wanna ask you some questions.
And then Yep. Just a it's just an interesting way to to figure it all out. Is someone else gonna figure out how to, like, use that documentation to, I don't know, make a calculator or some sort of an estimator to help you, like, get your settings started? I I don't know if that would actually I have no idea if that would work or not or if it would just if it would run into the same problem that you ran into. I imagine it would.
It might be would give you a reasonable starting area, but then there would be a lot of personal adjustment to make after that.
Yeah. I I just one of the most valuable things I found is it it knows how I message it. It knows how I talk to it. It knows how I like a response. So I can give it I could give it hell, I could give it a all documentation in Japanese, and I could ask it a question about it, and it would feed it back to me in a way that I comprehend.
Yeah.
And I think that's the difference between sure. All the loop doctor there, I could have spent ages reading them, and you get halfway, like, 10 pages down through one of the algorithms, and you're falling asleep as it's Dulles Dishwater. It's informative, but Dulles Dishwater.
I I
was Explain this to me in my words, in my terms.
Yeah. I was having a conversation with a few really, like, really lovely guys that are are in the DIY community, And I I hope that my intent came off the way I meant I meant it. I just said, like, you guys really have to just continue to remember that the way your brain sees something and the way my brain sees it are completely different. And Yeah. And so when you're explaining it or when you're setting it up for people to use, like, you need you need a dummy like me to look at it and tell you that that I that I got what you were saying.
Because sometimes I see them look at incredibly what I would tell you were incredibly complex ideas, they just understand them in the blink of an eye. And Yeah. And they contextualize it very quickly, and you can see even the comfort on their face that they understand. And I'm my brain's still looking at the at the screen going, I don't even know what that is. You know?
And so but to your point, and I hope this was my point too, it's not important for me to understand it the way they understand it. No. Yeah.
It's important. Everyone's different, and that's the best thing about it because everyone brings a different perspective. We can all make things better in a different way.
Yeah. Yeah. Yeah. Yeah. A a collection like that of of perspectives to help you contextualize in a way that everybody can get something out of it is is the is I think how this thing will grow more exponentially.
Because, I mean, there's an impressive number of people in the world who have a DIY algorithm, you know, for their insulin. At the same time, it's it's not nearly enough. Do you know what I mean? Like and and what's keeping it from spreading like that is either ease of use or comfort or, you know, something. But there's something that stops a lot of people when they see that.
Something that makes them go, oh, it's not for me or I shouldn't get involved in this or whatnot. And then they don't get the
That that looks too hard. I don't know how to build an app.
Yeah.
So I see. Yeah. You you do. You just don't know it yet. Yeah.
I don't wanna be an Apple developer. Yeah. Listen. I can tell you right now that I don't understand what Cloudflare is, but I have an account and I use it. Yeah.
I don't know what Netafly is, but I have an account and I use it. And it was explain and when when Claude says to me, hey. You need to put this here so that you can upload this so it works like that. I go, don't know what you're saying. Tell me like I'm an idiot.
And it just step by step by step, and I follow the steps, and I go, okay. All done. It worked. And I don't and I don't know what just happened. I don't care.
It's it's I love it. I really do.
But but but that that's the point. It's it's whilst you can be interested about the journey, the journey is not important. It's the end result.
Yeah.
You you don't get in the car in the morning and work out why when you turn the key, the engine turns on, and that means you can do 40 miles an hour down the road. You get in the car, you turn the key, and you don't care if an engine's running or it's electric motor. You you're going to the the mall.
Yes. To your point care about. I can't I cannot explain combustion to you, but I have I have gotten a ticket going a 130 miles an hour on a motorcycle. Man, So, is there anything that we haven't talked about that you think needs to be in this conversation? First, I'm having a nice time talking to you, but I just wanna make sure I'm not getting away from any contextual stuff that you wanted to put out there about this.
I think we've covered most things. Just to kinda loop background, the most important thing is to keep it grounded and not let it get carried away. So back to the confidence level is you always need to keep it grounded and set up. Don't don't be afraid to say it's wrong. Mhmm.
Just go you're you're flat right wrong. I don't believe you. I've got experience. And I go, okay. I'll take that on board.
Yeah. Or or
And it will fold that in.
Don't answer me until you're, you know, over a certain percentage.
You know? Yeah. For me, that that's one thing I learned super early on with AI models, but it's super valuable. Yeah. And I see people starting to use AI AI models now, they're asking me questions like, okay.
What my meetings look like for the rest of day? Summarize my emails, whatever that is. And it's like, did this thing well, it will even make up a response and tell it to you confidently. Mhmm. They go, the grass outside is blue.
And you go, really? Yeah. The grass outside is blue, and the sky is green. Are you sure? Yeah.
I'm sure. So you go, well, I'm looking at the grass, and it's green. Goes, actually, thank you for correcting me. You're right.
Yeah.
It's like, really? No. Yeah. I never have guessed that.
I appreciate you bringing it up, but your point is valid. Like, it could also be telling you something that you don't contextually understand, and it's like, no. It's a 100% like this. I I listen. I can't agree more.
I would also tell you that, like, with your idea of telling it, like, you need to be 80% positive before you report this back to me, I guarantee you that's a thing that we wake up a year from now and it's doing. Like, what what I'm seeing happening, I don't know if you see the same thing, but as users are using it and reporting back what works and doesn't work, then the people designing it are saying, hey. That's actually very smart. We needed to do that. And then they make they make adjustments to it.
Even, like, little things like, for example, do you ever get done and and it delivers you, like, HTML and you you're gonna, like, what? I don't know how I can't copy it. And, like like and it goes, oh, I'll
fill it in. I can't copy it. I can't I wanna put copy to an email or that.
Yeah. And it goes, oh, I'll copy it to the clipboard for you. And I thought, well, yeah, but don't we all need that? And then the last version of the desktop app came out, now there's a copy button. And and, like and so, like, that to me is how, like, it's gonna iterate more and more quickly to the point where one day, they're gonna say, everybody's telling the damn thing.
Make sure you're more than 80% sure before you answer me. Why don't we just bake that into it? I I think it's in its infancy.
This is this for me is like when the Internet was launched. I don't know how many of your listeners wouldn't remember the Internet being launched in 56 k dialogue, but
I don't know that you can properly explain to a human being what it's like to choose a photograph, not know which one you're choosing, and thirty five minutes later have the first inch and a half of the top of it come down and you realize, oh my god. This isn't the one I wanted. And then and then just have to stop that download and start over again. And then your mom comes in the room and goes, I gotta call your aunt. Get off the phone.
Or or or you pick up that phone, you're deafened by some kind of oh, yeah.
You you think people know that the Internet used to come on a CD ROM in the mail? I bet you they don't.
Yeah. You're assuming they know what a CD ROM is.
No. No. My kids don't even understand how Wi Fi works, don't think. My by the way, my son's a data engineer, and I'm not a 100% sure he understands it, which is how how quickly things just keep iterating and and and where it says you're going. I mean, I also would say I mean, since it since we're having this conversation, like, I I I do really get the other like, the concerns that you hear from people about, you know, electricity, water use, putting somebody out of a job.
Like, I I'm I'm not I I'm I'm not I'm not not on your side about all that, and I do really understand. I'm I'm just trying to take what's available to me and do the best I can with it and Yep. And try to help as many people as I can. I'm hoping by trying to do the right thing that it, you know, it ends up being a net positive. And I've said before, I'll say it again.
I mean, if if the robots come and step on our skulls and there's lasers flying overhead, I was wrong. I'll apologize. But, you know, the something here
Just all the more reason why you need to be nice to your AI assistant. Right? Please. Thank you. Job.
When it comes to exterminator, it remembers me? So
but but when Skynet takes over, it's, oh, I treated you nice.
You remember I said, please. Yeah. I mean, listen. I I I don't know. I tend to believe it's gonna be okay.
And, I mean, I lean towards that. I couldn't say that with any certainty, but, like, I lean towards favorably believing that this is moving in a good direction. So, I mean, I'll find out. I'll be right or I'll be wrong. But, you know, I think it's some some it's certainly not if I stop using it right now or you stop using it right now, it is certainly not gonna change the direction that that this is ongoing.
This this snowboard is rolling down the mountain whether you like it or not.
Yeah. Yeah. Okay. Alright. Well, I I am so happy.
Already left the station.
I am so happy for you that that you were, you know, you were able to figure all this out. And it's really impressive that you did it just in the last couple of years having had diabetes for so much time before that. I think there's an entire story in there about, you know, about telling people what what good enough is. And Yeah. And like you said, there's no way to disbelieve it.
Right? The doctor said I'm okay. I'm okay. I don't have to think about it anymore. Yeah.
It sucks, but I guess that's what this is. So, you know, and
it does It's it's a fight I still have and I expect I'm still gonna have moving towards the future in new technology. Does it yes. The NHS is trying to push me towards Omnipod five. I'm like, well, okay. But I I can't run it on my phone.
I I don't wanna carry another device that doesn't compete with them. Yeah. But I've got good control. Or I was asking about Gvoke the other day. Mhmm.
The premixed premixed hyperpen.
Yeah.
Nope. Sorry. Can't get it. I was like, really? So I've gotta carry this orange case, and I've gotta mix it up.
And then my kids are three and five. There's not a chance they will be able to mix that up. If I say go to the fridge, get that thing out, and stab daddy in the leg, they'll be able to work that out.
Yeah.
But
Yeah. Yeah. That's interesting. I didn't I think they're they're gone here, I believe. I don't think you can get the Lily kit here anymore.
Okay.
Yeah. That's fascinating, isn't it?
Yeah. It's still that that's all I can get.
You can't get
It is a glucagon kit. I can't
words evoke?
Nope. Nothing. It won't it won't even entertain me asking for it. Tiny little country. Probably once every six months.
Yeah. How many people could possibly want it? It's not a very big country. No offense. Yeah.
I mean, we had no trouble getting away from you at all. Wow. That's yeah. I mean, 2026, you shouldn't have to argue about stuff like that. Yeah.
That's I mean, they're they're
getting better. They have what's called the NICE records, which is like a medical advice about what they should and shouldn't offer. So all CGMs are available now. Pumps are available now. But it's just every step of the way.
Anything that's new, anything that's
You have to fight about it.
I I didn't even know really what it is. It's just it's just slow adoption, I think. Mhmm. I listen. Maybe it's just demand based.
I I think listen. I think money, obviously, is part of it. Always always gonna be money. And Yeah. You know, and then there's just the, you know, there's just the pot roast story.
It's just, you know, it's how we always do it. So why would I I don't know any differently. Why would I think any different? You can even hear it in your story. Right?
Right? You were like, I thought I was doing okay. So when people think they're doing okay, they don't spend a lot of time trying to figure out what might be better that's not just, like, right in front of them, but completely tangible and maybe even less expensive or less trouble or whatnot. Like, we are just limited by like, you said it earlier, but we're just limited by what our brains can hold at one time. I I I've been trying to dig into it a little bit.
And I think it turns out, like, no matter what you think about multitasking, your brain really can only do one thing at a time. And then you shift from one thing to another thing, and it can hold other things in, like, an out like, kind of an outside of that spotlight of focus, but you are not thinking about
holding in this temporary memory. You could do 10% of one thing, and it you can do 10% of 10 things at a time. It takes you 10 times longer.
You are
Or you could do one thing at a time and take the same period of time.
You are not thinking about two things at the same time. You you you just aren't. And to think that then the person you asked for Gvoke can keep in their head all the different glucagons, all the different political impacts, the financial reasoning about and then they they're just gonna go, this is how we always did it. It's gonna be fine. And that that's what you end up with.
I'm telling you, put an AI in charge of the decision about glucagon in in England, and I bet you it gets changed pretty quickly. Alright. I'm voting for AI for president. That's all. Can't go anywhere.
So no no comment on whether I do a better job or not because
Alright, man. Oliver, this is really
a real pain.
Yeah. What what is? Say it again.
Export taxes in The UK to The US are a real pain at moment.
Everything's a real pain. Yeah. No kidding. Alright, man. I really appreciate this.
Hold on one second for me. Okay?
Sure.
Today's show was sponsored by Tandem Diabetes Care, the Eversense three sixty five, and Touched by Type one, who you can learn more about on Facebook, Instagram, and at touchedbytype1.org. Head now to tandemdiabetes.com/juicebox and check out today's sponsor, Tandem Diabetes Care. I think you're gonna find exactly what you're looking for at that link, including a way to sign up and get started with the Tandem Mobi system. I'd like to thank the Eversense three sixty five for sponsoring this episode of the juice box podcast And remind you, if you want the only sensor that gets inserted once a year and not every 14, you want the Eversense CGM. Eversensecgm.com/juicebox.
One year, one CGM. If you or a loved one is newly diagnosed with type one diabetes and you're seeking a clear practical perspective, check out the bold beginnings series on the juice box podcast. It's hosted by myself and Jenny Smith, an experienced diabetes educator with over thirty five years of personal insight into type one. Our series cuts through the medical jargon and delivers straightforward answers to your most pressing questions. You'll gain insight from real patients and caregivers and find practical advice to help you confidently navigate life with type one.
You can start your journey informed and empowered with the juice box podcast. The bold beginning series and all of the collections in the juice box podcast are available in your audio app and at juiceboxpodcast.com up in the menu. Thank you so much for listening. If you're not already subscribed or following the podcast in your favorite audio app like Spotify or Apple Podcasts, please do that now. Just hitting follow or subscribe really helps the show.
And if you'd like to go a little further and leave a review in Apple Podcasts, that would be fantastic. But most of all, sharing the Juice Box podcast with someone you know or someone you think can be helped by it, that's the biggest thing you can do for the show. I'll be back very soon with the next episode, so don't forget to check your apps.
Read the full disclaimer
© 2007–2026 Juicebox Podcast. All rights reserved.
