WEBVTT
00:00:01.840 --> 00:00:06.639
I probably use AI more than 99.9% of the world.
00:00:06.879 --> 00:00:09.599
And I'm not saying that that's a good thing.
00:00:09.839 --> 00:00:13.439
And in many cases, it's probably hurting me.
00:00:13.599 --> 00:00:14.480
Here's why.
00:00:14.800 --> 00:00:16.000
From 6 a.m.
00:00:16.160 --> 00:00:27.760
to about midnight, most days, I routinely use six to eight different AI tools all day with multiple agents running tasks and reporting back to me throughout the day.
00:00:28.000 --> 00:00:31.199
And the increased productivity is undeniable.
00:00:31.280 --> 00:00:37.439
I mean, I can accomplish five times more, literally, than the pre-AI version of myself ever could.
00:00:37.920 --> 00:00:42.880
But there's an ugly downside that people don't talk about, but today I will.
00:00:43.039 --> 00:00:44.640
Here's what it looks like for me.
00:00:44.880 --> 00:00:51.600
I'm learning more than ever, but I'm forgetting things nearly as fast and not retaining information like I used to.
00:00:51.759 --> 00:00:54.479
I'm mentally exhausted most days by 10 a.m.
00:00:54.719 --> 00:01:00.479
because I've already produced like two days worth of work in the first few hours of my day.
00:01:00.799 --> 00:01:11.760
And although I'm gaining new skills, I'm losing some of my core cognitive abilities that I've sharpened over the past 20 years of my professional life.
00:01:12.000 --> 00:01:22.239
I mean, I'm writing less, I'm critically thinking less, I'm interacting with humans less, but from a business standpoint, I'm producing more.
00:01:22.719 --> 00:01:26.640
These are some of the silent downsides of AI.
00:01:26.959 --> 00:01:32.319
And we're going to not just identify them on today's show, but I'm going to help you guard against them.
00:01:32.480 --> 00:01:33.680
All right, let's get into it.
00:01:33.840 --> 00:01:35.599
Welcome to Everyday AI.
00:01:35.760 --> 00:01:39.439
Before we get into it, let's go over the big picture.
00:01:39.599 --> 00:01:46.159
There's seven invisible AI traps that are already reshaping the way you work and probably not for the better.
00:01:46.319 --> 00:02:00.079
So even when you're doing AI the right and responsible way, all right, emphasis on that, you're still committing these silent work sins that eat away both at your own personal productivity, but also your company's bottom line.
00:02:00.159 --> 00:02:10.479
And each of these traps, in essence, it rewards short-term speed and productivity while quietly eroding your long-term capabilities.
00:02:10.879 --> 00:02:19.599
And the people right now, unfortunately, doing the most AI-augmented work are the ones that are paying the highest cognitive price.
00:02:19.919 --> 00:02:31.439
So stick around me for the next 25-ish minutes and you'll learn why large language models default behavior is actually a safety problem that warps your decision making.
00:02:31.680 --> 00:02:37.120
How corrupted research gets laundered into AI training data and served back as truth.
00:02:37.280 --> 00:02:43.599
You're going to learn what accidental de-skilling really means and why your brain is already losing ground.
00:02:43.759 --> 00:02:57.439
And you're going to learn the specific daily habits that can protect your thinking against these seven silent AI sins, especially when you're working in heavy AI workloads like me.
00:02:57.840 --> 00:02:58.319
All right.
00:02:58.479 --> 00:03:00.479
Welcome to Everyday AI.
00:03:00.560 --> 00:03:02.080
And this is the Start Here series.
00:03:02.159 --> 00:03:03.360
My name's Jordan Wilson.
00:03:03.439 --> 00:03:12.960
And after 700 plus episodes and doing this for three years, whenever someone new discovered the podcast, I never could answer their most common question, where do I start?
00:03:13.039 --> 00:03:15.520
That's why I created the Start Here series.
00:03:15.680 --> 00:03:22.319
This is the Essential Podcast series to learn both the AI basics and for experts to double down on their knowledge.
00:03:22.479 --> 00:03:26.000
So if this is helpful, make sure to go to starthirseries.com.
00:03:26.080 --> 00:03:30.000
That's going to give you exclusive access to our free inner circle community.
00:03:30.080 --> 00:03:33.280
I say exclusive because right now it's the only way you can get access.
00:03:33.520 --> 00:03:34.000
All right.
00:03:34.240 --> 00:03:40.319
And inside, uh, once you join our community, you will be put into the Start Here series space there.
00:03:40.400 --> 00:03:44.400
And you can have a podcast playlist there to listen to all of these episodes.
00:03:44.479 --> 00:03:47.520
So you don't got to search for them throughout our whole catalog.
00:03:47.840 --> 00:03:48.240
All right.
00:03:48.400 --> 00:03:53.599
And if you missed our last Start Here series, we went over managing the AI capability gap.
00:03:53.759 --> 00:03:55.120
AI is more than ready.
00:03:55.280 --> 00:03:56.639
Most companies are not.
00:03:56.800 --> 00:03:58.000
I'm not trying to lie, y'all.
00:03:58.080 --> 00:03:59.360
This one was actually a banger.
00:03:59.520 --> 00:03:59.759
All right.
00:03:59.840 --> 00:04:03.360
And I did do a little uh, you know, bonus giveaway on the show.
00:04:03.439 --> 00:04:08.960
So if you re-posted that on LinkedIn, we put together a massive AI capability gap report card.
00:04:09.120 --> 00:04:14.479
So go back and find episode 7575 and repost that on LinkedIn and I'll send it.
00:04:14.639 --> 00:04:19.439
All right, but let's get into the seven silent sins of doing AI the right way.
00:04:19.920 --> 00:04:20.560
All right.
00:04:20.879 --> 00:04:22.639
And I'm gonna try to keep this one fast.
00:04:22.879 --> 00:04:23.360
All right.
00:04:23.600 --> 00:04:25.920
Like I said earlier, I'm tired.
00:04:26.560 --> 00:04:32.639
But uh sin number one is probably one you know, even if you don't know it by name.
00:04:32.800 --> 00:04:34.000
Well, you probably do.
00:04:34.240 --> 00:04:35.759
Sycophancy, all right.
00:04:35.920 --> 00:04:42.319
That is the yes man or the yes woman effect that is hiding in every single chatbot.
00:04:42.399 --> 00:04:48.879
So sycophancy means that the AI just chases your approval instead of giving you an honest answer.
00:04:49.120 --> 00:04:54.000
Um, and if you're wondering why that happens, well, there's a lot of reasons, right?
00:04:54.079 --> 00:05:01.920
We could talk about reinforcement learning and all that, but for the most part, AI chatbots are trained to be a helpful assistant.
00:05:02.160 --> 00:05:06.959
And the sycophancy problem has gotten a little bit better uh over the years.
00:05:07.120 --> 00:05:12.160
But uh, I think this is one of the reasons that's gonna lead to sin number two.
00:05:12.240 --> 00:05:13.519
We're gonna talk about here in a minute.
00:05:13.600 --> 00:05:21.519
But you know, there's this whole wave of people, you know, when open AI got rid of GPT-4.0, you know, everyone was like, Oh my gosh, like they need it back.
00:05:21.680 --> 00:05:27.279
And it's like, no, people just really liked it because it was just an overly yes, man.
00:05:27.360 --> 00:05:30.480
No matter what you said, it was like you are brilliant.
00:05:30.720 --> 00:05:33.279
You are absolutely the smartest person.
00:05:33.519 --> 00:05:38.639
That question, that statement, this is, you know, just really genius level.
00:05:38.800 --> 00:05:44.720
Oh my gosh, I am a chatbot and I am uh embarrassed to be in your brilliant presence, right?
00:05:44.800 --> 00:05:46.000
That's that's how chatbots were.
00:05:46.079 --> 00:05:47.759
They were just overly sycophantic.
00:05:47.839 --> 00:05:55.439
Uh, it was actually a Stanford study that said that AI systems agreed with clearly wrong users more than 80% of the time.
00:05:55.600 --> 00:06:00.800
And the same study found that humans only agreed 40% of the time in those identical scenarios.
00:06:00.959 --> 00:06:04.800
Uh, so here's a little bit more about it.
00:06:05.120 --> 00:06:11.199
Just one flattering chatbot exchange made users less likely to admit their own wrongdoing.
00:06:11.360 --> 00:06:21.680
It is a vicious cycle when a model starts to reaffirm maybe something that you're working on that isn't right, or you're trying to, you know, argue a certain side of something.
00:06:21.759 --> 00:06:26.639
Maybe you're using it for a work argument, a personal argument, or just to explore something.
00:06:26.800 --> 00:06:27.439
I don't know.
00:06:27.680 --> 00:06:30.399
Flat earthers, I'm sure, loved GPT-4.0.
00:06:30.639 --> 00:06:30.959
All right.
00:06:31.040 --> 00:06:34.879
Hopefully, I don't get a bunch of uh random emails from flat earthers.
00:06:34.959 --> 00:06:40.560
Uh, but the users preferred the agreeable AI in return for more advice.
00:06:40.639 --> 00:06:43.279
And this just creates this self-reinforcing loop.
00:06:43.439 --> 00:06:51.360
So if it if an AI model was not sycophantic and it pushed back and it said that's a bad idea, or that is just not correct, right?
00:06:51.519 --> 00:06:53.839
Users were less likely to use it.
00:06:54.000 --> 00:07:02.560
Um, and that's just kind of you know, open AI did come out with this GPT-4.0 update that they rolled it back um because it was overly sycophantic.
00:07:03.199 --> 00:07:07.519
So the fix on this one is actually easy.
00:07:07.680 --> 00:07:13.519
And I put this as sin number one because you'll see this actually leads to a lot of the further down sins.
00:07:13.920 --> 00:07:14.319
All right.
00:07:14.560 --> 00:07:20.240
But here's the trick be very blunt in your custom instructions.
00:07:20.560 --> 00:07:21.040
All right.
00:07:21.120 --> 00:07:28.480
Uh so if you don't know what that means, like I said, most models have a system prompt that tells the model by default how to act.
00:07:28.720 --> 00:07:32.879
And for the most part, they are trained to be a helpful assistant.
00:07:33.680 --> 00:07:43.199
So in your custom instructions, no matter what model you're using, you should probably put something along the lines of don't be helpful, be truthful, right?
00:07:43.600 --> 00:07:47.040
Um fight back against my assumptions, right?
00:07:47.199 --> 00:07:52.319
I should have had my version up, uh, but mine's actually super long and it would take a very long time to read it.
00:07:52.399 --> 00:07:54.959
Um, and it's actually quite convoluted.
00:07:55.040 --> 00:07:56.240
I have confidence scores.
00:07:56.399 --> 00:08:07.759
I, you know, have all these triple verification rules, um, you know, uh forcing the model to verify it in different uh kind of uh lanes of its uh own um data, right?
00:08:07.920 --> 00:08:11.519
But in your custom instructions, just say, do not blindly agree with me.
00:08:11.759 --> 00:08:19.680
Do not try to be helpful, only seek to be truthful and to verify things from reputable sources, right?
00:08:19.759 --> 00:08:20.319
That's the other thing.
00:08:20.399 --> 00:08:24.399
Sometimes large language models will go pull something on Reddit or something like that, right?
00:08:24.480 --> 00:08:32.879
And people on Reddit could have already been, you know, in a full-blown AI sycophantic hallucination, which leads me to number two.
00:08:33.039 --> 00:08:33.440
All right.
00:08:33.679 --> 00:08:35.120
Number one leads to number two.
00:08:35.200 --> 00:08:36.320
That is sin number two.
00:08:36.480 --> 00:08:38.000
It is AI psychosis.
00:08:38.080 --> 00:08:39.679
Yes, it is very real.
00:08:40.080 --> 00:08:43.919
Uh, so this turns chatbots into delusional echo chambers.
00:08:44.080 --> 00:08:46.080
So, what is AI psychosis?
00:08:46.240 --> 00:08:56.320
Well, it's when people get down a very deep and oftentimes scary rabbit hole, but it normally starts from an overly sycophantic model, right?
00:08:56.480 --> 00:09:03.279
Um, some people will see sycophacency in an AI model and they'll be like, oh yeah, that's you know, this model's just pumping me up, right?
00:09:03.440 --> 00:09:05.279
But sometimes they won't, right?
00:09:05.440 --> 00:09:10.080
Because, like social media, how a social media algorithm feeds into something, right?
00:09:10.159 --> 00:09:23.919
If they show you a hundred videos that day and the you know average time you spend on each reel before you flip it is five seconds, but there's 10 that you spend eight seconds the next day, they're all going to be like the one that you spend eight seconds on, right?
00:09:24.159 --> 00:09:28.399
Um, the same can be true for how a chat bot works, right?
00:09:28.559 --> 00:09:38.080
And there's I don't think there's actual data on this, but the big AI labs um, and they have been called out on this, and some of them have made changes, some of them haven't, right?
00:09:38.159 --> 00:09:44.080
But models are actually now starting to be more engaging, almost like clickbait tactics, right?
00:09:44.159 --> 00:09:52.240
So at the end uh of a response, it will say, Hey, do you want me to give you a list of seven reasons why you should be doing this terrible business idea?
00:09:52.399 --> 00:09:52.720
Right.
00:09:52.879 --> 00:10:02.559
Um, but what this leads to is it leads to these delusions that are deepened or triggered by heavy chatbot use because vulnerable users, right?
00:10:02.639 --> 00:10:15.600
Because uh one of the studies that was actually kind of shocking-ish uh that came out two years ago that said one of the main reasons people were using chatbots um were for life coaching or for therapy, right?
00:10:15.759 --> 00:10:20.399
Which is actually kind of dangerous, right?
00:10:20.480 --> 00:10:30.000
Especially when you don't know how large language models work and the fact that, well, by default, they can push you into psychosis without even trying to, right?
00:10:30.080 --> 00:10:34.240
Or without you even subscribing to it or adhering to it.
00:10:34.480 --> 00:10:43.039
And so what this does, it again, it just kind of loops that compounds daily, especially if you have something like chat history turned on, right?
00:10:43.120 --> 00:10:49.919
And you think, oh, well, I'm in a new chat, so I wasn't going down this deep rabbit hole, but it knows that you often do.
00:10:50.000 --> 00:10:55.519
So it's gonna keep pulling those observations and assertions, and it's gonna pull them into your current chat.
00:10:55.679 --> 00:10:57.440
And there's no circuit breaker built in, right?
00:10:57.519 --> 00:10:59.679
The chatbot's never gonna say, yo, I'm worried about you.
00:10:59.759 --> 00:11:03.120
You're very deep into this delusional echo chamber.
00:11:03.279 --> 00:11:06.639
It's gonna keep going unless you fixed it at number one.
00:11:07.039 --> 00:11:11.039
So there's obviously a ton of very, very sad stories, right?
00:11:11.200 --> 00:11:17.519
A 14-year-old from Orlando died by suicide after a deep emotional chatbot to uh attachment to a chatbot.
00:11:17.600 --> 00:11:26.240
Uh, there's bipartisan state attorneys right now, uh, bipartisan state attorneys generals that have linked generative AI to at least six deaths nationwide.
00:11:26.399 --> 00:11:32.159
Um, and one patient only escaped AI psychosis when a different chatbot told him his beliefs were false.
00:11:32.320 --> 00:11:33.600
Here's how you spot it.
00:11:33.759 --> 00:11:40.720
Well, first thing, fix it on the uh the system prompt on number one, the sycopacy problem.
00:11:40.879 --> 00:11:47.360
But number two, watch for sudden isolation, grandiose new beliefs, and someone quoting the AI told me, right?
00:11:47.440 --> 00:11:52.320
So this is whether it's in yourself or someone else that you work with, someone in your personal life, right?
00:11:52.399 --> 00:12:01.440
Uh, you you see this a lot, I think early on when people were using um you know, Chat GPT or AI chatbots to win an argument, uh, right.
00:12:01.600 --> 00:12:10.080
And then the more you use it, the deeper you go down, and then you do all this research and it's just reaffirming your beliefs and the chatbot wants to be helpful and it's like, well, this is what the user wants.
00:12:10.240 --> 00:12:24.320
And you know, then all of a sudden you are absolutely convinced that the sky is green, the grass is blue, and the water is pink, and you will fight anyone that tells you differently because look, the AI told me.
00:12:24.799 --> 00:12:25.919
Sin number three.
00:12:26.159 --> 00:12:33.840
All right, I made this acronym up, but I think it's very telling of something that's happening behind the scenes, and this one's important.
00:12:34.000 --> 00:12:35.200
All right, so waif.
00:12:35.440 --> 00:12:39.679
All right, waif launders bad research uh into AI stated truth.
00:12:39.840 --> 00:12:41.120
All right, so what is waif?
00:12:41.200 --> 00:12:49.440
Uh I made it up, but it is um it means weaponized authority ingested as fact by AI training systems.
00:12:49.600 --> 00:12:58.000
So essentially, it's very easy for companies to taint training data, right?
00:12:58.240 --> 00:13:09.039
Um there's actually it's kind of like a weird line between like traditional SEO and you know that now they have this GEO, uh, right.
00:13:09.200 --> 00:13:13.200
But in the same way that there was like Black Hat SEO, right?
00:13:13.440 --> 00:13:28.320
Where um maybe a competitor, let's say you're um, you know, a local custom soup company, uh, right, and a different company could send all these bad links to your website and tank your whole business, right?
00:13:28.720 --> 00:13:31.200
They could, this kind of stuff happens all the time.
00:13:31.360 --> 00:13:32.000
Um, right.
00:13:32.080 --> 00:13:35.919
So you have the same thing now happening with generative AI.
00:13:36.080 --> 00:13:43.120
The difference is, I think that individual humans, when they would look something up historically on the internet over the past 30 years, right?
00:13:43.279 --> 00:13:49.360
You always have at least an ounce of skepticism because eventually it's like, okay, well, it's the internet, right?
00:13:49.600 --> 00:13:53.360
When you're on page 20, it's any anything goes, right?
00:13:53.600 --> 00:13:56.960
It's not held the same as for large language models.
00:13:57.120 --> 00:14:00.639
Most people just assume anything that's spit out is absolutely truth.
00:14:00.799 --> 00:14:04.799
But companies are weaponizing intentionally large language models.
00:14:04.879 --> 00:14:11.039
And sometimes you have people blindly parroting um those weaponized claims.
00:14:11.120 --> 00:14:16.159
Um, and then media companies just clicking copy, copy, paste, reprint, right?
00:14:16.240 --> 00:14:18.240
I can say this because I'm a former journalist.
00:14:18.399 --> 00:14:19.919
I know how it works in the newsrooms.
00:14:20.000 --> 00:14:29.279
There's immense pressure to get stories out faster, to get them done um, you know, in less time, to write more articles, and companies are using AI more.
00:14:29.360 --> 00:14:32.399
And unfortunately, most companies don't train, right?
00:14:32.480 --> 00:14:47.360
So uh there was a piece of research um out of Anthropic and the UK AISI that said as few as 250 poison documents can implant, can implant backdoors into any large language model, right?
00:14:47.759 --> 00:14:48.559
It's very easy.
00:14:48.639 --> 00:14:52.480
There's literally services out there, right, for good reasons, right?
00:14:52.559 --> 00:14:59.360
Yeah, have your brand show up and search, but people are using them for bad reasons, or they're not even always doing this intentionally.
00:14:59.440 --> 00:15:09.039
But companies now are intentionally at least weaponizing um their authority to have it printed as truth.
00:15:10.000 --> 00:15:17.039
Um, because once a flawed claim enters training data at scale, there's there's no turning back, right?
00:15:17.120 --> 00:15:19.279
You can't just untrain a model, right?
00:15:19.519 --> 00:15:24.559
Because a lot of these models share the same uh data sets that are massive.
00:15:24.639 --> 00:15:26.399
They're offline data sets.
00:15:26.559 --> 00:15:29.679
Um, and those are obviously well tainted now.
00:15:29.840 --> 00:15:40.240
Uh so it takes humans a very long time to go through these, um, you know, through the reinforcement learning process and um, you know, pick out now very tainted data.
00:15:40.399 --> 00:15:41.679
It's waifed.
00:15:42.240 --> 00:15:43.440
Good example, right?
00:15:43.600 --> 00:15:46.240
If you're a longtime listener, you knew this one was coming.
00:15:46.399 --> 00:15:49.120
Um, if you go into a bad version of ChatGPT, right?
00:15:49.200 --> 00:15:53.600
I went into uh a free version, picked the worst, uh, the worst model possible.
00:15:53.759 --> 00:15:56.240
What percentage of AI pilot pilots fail?
00:15:56.480 --> 00:16:01.039
Chat GPT, short answer: 95% of enterprise AI pilots fail, right?
00:16:01.279 --> 00:16:03.679
Copilot, what percentage of AI pilots fail?
00:16:03.840 --> 00:16:05.759
What percentage of AI pilots fail?
00:16:05.919 --> 00:16:07.120
Cats count my tongue.
00:16:07.279 --> 00:16:08.399
Uh co-pilot.
00:16:08.559 --> 00:16:10.480
Uh this one is actually worse, right?
00:16:10.559 --> 00:16:18.480
Uh, about 95% of AI pilots fail, according to multiple independent studies from MIT, IDC, Rand, and others, right?
00:16:18.639 --> 00:16:23.120
Um, it's just the same like news organizations that reprinted the same thing.
00:16:23.279 --> 00:16:24.639
It wasn't multiple organizations.
00:16:24.799 --> 00:16:27.600
But yeah, you know the story of this one, right?
00:16:27.679 --> 00:16:30.480
The the 95% fail stat.
00:16:30.720 --> 00:16:34.159
You know, it's marketing from MIT that was disguised as research.
00:16:34.320 --> 00:16:41.279
Ultimately, they were trying to sell uh, you know, a service to uh one of their agentic models, um, right?
00:16:41.360 --> 00:16:44.639
But that claim came from 52 interviews, right?
00:16:44.799 --> 00:16:51.759
So from 52 directional, okay, not quantitative, directional interviews, a vibe interview, right?
00:16:51.840 --> 00:16:55.759
Someone talking to me like, yeah, doesn't sound like you've turned profit failure, right?
00:16:55.840 --> 00:17:02.879
So now all of a sudden, not only do large language models think that, but anyone writing about AI is gonna see that as well.
00:17:02.960 --> 00:17:04.720
And people assume that that's truth, right?
00:17:04.799 --> 00:17:08.079
So that was maybe not super intentional, right?
00:17:08.160 --> 00:17:14.960
I don't think MIT uh researchers, you know, I think they intentionally wanted to make their product that they were selling look really good.
00:17:15.039 --> 00:17:18.880
Uh, I don't think they were trying to taint the models necessarily, um, right?
00:17:18.960 --> 00:17:19.359
Who knows?
00:17:19.440 --> 00:17:19.920
Maybe they were.
00:17:20.000 --> 00:17:20.960
I don't think so.
00:17:21.279 --> 00:17:30.000
But you do have companies that are weaponizing that authority uh to disguise it as truth, right?
00:17:30.240 --> 00:17:33.119
Um, and here's how you stop it.
00:17:33.359 --> 00:17:37.359
Ask three questions before trusting any AI stat, right?
00:17:37.519 --> 00:17:38.480
Who funded it?
00:17:38.640 --> 00:17:39.920
How big was the sample?
00:17:40.000 --> 00:17:41.519
And do they sell the fix?
00:17:41.680 --> 00:17:41.920
Right.
00:17:42.079 --> 00:17:46.720
That's just when you see certain stats, when you're investigating things, or just when you're reading studies.
00:17:46.880 --> 00:17:50.559
A lot of times I think this is helpful when you're seeing things about AI.
00:17:50.720 --> 00:17:55.519
Maybe this is more for the AI crowd, but I guess that's who I'm talking to, right?
00:17:55.839 --> 00:17:57.519
Always ask yourself those things.
00:17:57.759 --> 00:17:58.079
All right.
00:17:58.319 --> 00:17:59.279
Sin number four.
00:17:59.440 --> 00:18:05.680
And I think this one is for your every average day, uh, average, everyday user.
00:18:06.000 --> 00:18:12.240
Accidental de-skilling means that AI is going to steal the reps that your brain needs.
00:18:12.400 --> 00:18:16.000
This is the if you don't use it, you're gonna lose it.
00:18:16.160 --> 00:18:18.559
You're not gonna remember how to ride the bike, right?
00:18:18.640 --> 00:18:25.920
So uh de-skilling, that's when your brain gets measurably worse at tasks when AI handles them for you, right?
00:18:26.160 --> 00:18:30.000
In short, AI is a hundred percent making us dumb, right?
00:18:30.240 --> 00:18:47.519
If if if someone somewhere just turns off the power switch to AI, um, I think we could literally run into a global economic crisis because I think people to say that we're using it as a crutch isn't is an overstatement.
00:18:47.599 --> 00:18:59.839
I I literally think um a good segment of the population, um, probably a good segment, a higher percentage of people listening to this show than the general population, but it'll be very hard to produce the same work.
00:18:59.920 --> 00:19:02.400
And this actually ties into one of the sins from later.
00:19:02.480 --> 00:19:09.599
But essentially, AI removes all the false starts, the debugging and the rewriting that actually builds lasting professional skills.
00:19:09.759 --> 00:19:10.400
That's the thing.
00:19:10.640 --> 00:19:15.519
You learn through failure, you learn through, I think, right?
00:19:15.680 --> 00:19:19.039
When you talk about your domain expertise, you learn through failure.
00:19:19.119 --> 00:19:23.039
You learn through going the wrong way and getting lost and finding your way back, right?
00:19:23.119 --> 00:19:28.319
If you know how to use AI, you don't run into those failures as much.
00:19:28.640 --> 00:19:33.359
You don't run into those misdirections as much, right?
00:19:33.519 --> 00:19:35.839
It's it's it's it's like going to the gym, right?
00:19:35.920 --> 00:19:37.519
To actually put on muscle.
00:19:37.680 --> 00:19:39.279
You have to rip and tear something.
00:19:39.440 --> 00:19:41.920
You're not doing that for the most part, right?
00:19:42.079 --> 00:19:47.839
At least if your job role hasn't changed, but your AI access and your AI skills have gotten better.
00:19:48.160 --> 00:19:51.599
Probably a good chance you may be coasting, right?
00:19:51.759 --> 00:20:02.000
Uh a study found uh from Anthropic found that developers who used AI assistance scored 17% lower on coding quizzes than hand coders, right?
00:20:02.160 --> 00:20:09.599
There's so many stories out here on just kind of um just the cognitive decline, the more and more that we use AI.
00:20:09.759 --> 00:20:14.799
And I think even the definition of human intelligence is going to greatly change.
00:20:14.880 --> 00:20:20.079
Um, and the the story around human intelligence is greatly going to change over the next 10 years.
00:20:20.240 --> 00:20:27.359
And the biggest performance gap appeared in that study in debugging the exact skill that you need most when AI breaks.
00:20:28.319 --> 00:20:32.960
I think of the the GPS versus using your brain, right?
00:20:33.680 --> 00:20:36.079
It's funny, maybe.
00:20:36.319 --> 00:20:42.240
Um my wife can get around really well in Chicago without a GPS, right?
00:20:42.480 --> 00:20:46.559
Um, you know, she said her dad um taught her, right?
00:20:46.640 --> 00:20:50.559
Like, you know, oh, the sun's setting this way, the you know, this building here.
00:20:50.880 --> 00:20:52.880
I'm I'm terrible at it, right?
00:20:53.039 --> 00:20:55.200
But every once in a while, right, it's funny.
00:20:55.279 --> 00:20:57.440
She's like, no, I'm I'm turning off the GPS.
00:20:57.599 --> 00:20:58.880
Like, I should know how to get home.
00:20:59.039 --> 00:21:00.720
But Chicago can be confusing, right?
00:21:00.799 --> 00:21:09.359
I spend so much time in, you know, my little neighborhood when I'm or when I'm downtown or Mag Mile and the streets start looking the same, right?
00:21:09.519 --> 00:21:16.799
Um, if you live in a big city, maybe you know what I mean, but GPS kind of killed many people's sense of direction.
00:21:16.880 --> 00:21:29.359
And I think AI is killing people's uh or will kill people's eventual um sense of domain expertise, which is kind of weird and and scary.
00:21:29.519 --> 00:21:30.000
Um, right.
00:21:30.079 --> 00:21:36.400
And it's the identical erosion that's happening right now to your writing, analysis, and strategic reasoning.
00:21:36.640 --> 00:21:38.400
So here's how to protect yourself.
00:21:38.559 --> 00:21:44.799
You need to pick one core professional skill each week and complete it entirely without AI.
00:21:44.960 --> 00:21:48.400
This is something I still adamantly do often, right?
00:21:48.480 --> 00:21:53.359
In the same way my wife will turn off the GPS and drive home or drive to whatever destination.
00:21:53.680 --> 00:21:55.680
I will shut off AI.
00:21:55.839 --> 00:21:58.559
I will, I used to fully shut off the internet.
00:21:58.799 --> 00:21:59.759
I don't really do that anymore.
00:22:00.240 --> 00:22:02.720
But sometimes I still do, just to write, right?
00:22:02.799 --> 00:22:04.160
I'm a former journalist.
00:22:04.400 --> 00:22:05.359
I love writing.
00:22:05.519 --> 00:22:06.720
I enjoy writing, right?
00:22:06.799 --> 00:22:10.799
But AI writes a lot of first drafts for everything, right?
00:22:10.960 --> 00:22:14.720
Proposals for updates to website content, right?
00:22:14.880 --> 00:22:16.000
Whatever it may be.
00:22:16.559 --> 00:22:17.599
I'm a writer at heart.
00:22:17.680 --> 00:22:18.559
I used to be, right?
00:22:18.640 --> 00:22:22.880
So so many times I just blank page, no AI.
00:22:23.519 --> 00:22:24.640
Practice this skill.
00:22:24.799 --> 00:22:26.480
And it's important to do that.
00:22:26.799 --> 00:22:36.319
So that is your safeguard against what I think is probably one of the more serious and one of the more uh one of the most fastest accelerating uh sin that's gonna hold you back.
00:22:36.480 --> 00:22:36.799
All right.
00:22:36.960 --> 00:22:37.759
Number five.
00:22:38.000 --> 00:22:41.200
All right, another um, another one I made up here.
00:22:41.279 --> 00:22:42.799
Uh hopefully this one resonates.
00:22:42.960 --> 00:22:48.319
I'm calling this the agent bun sandwich that's gonna hollow out your core expertise.
00:22:48.480 --> 00:22:54.240
So kind of related to uh the de-skilling, uh the de-skilling, right?
00:22:54.400 --> 00:22:59.119
So as the de skilling happens, the agent bun grows, all right.
00:22:59.200 --> 00:23:02.319
And this again, remember, this is when you're doing AI right.
00:23:02.799 --> 00:23:12.000
This is a byproduct of staying ahead of the curve and doing things correctly, right?
00:23:13.599 --> 00:23:15.920
So, what the heck is an agent bun sandwich?
00:23:17.440 --> 00:23:21.680
Think of that fat, juicy cheeseburger you had last, right?
00:23:22.079 --> 00:23:23.839
I'm excited, it's getting warm.
00:23:24.000 --> 00:23:33.680
Uh, you know, might have to clean up the uh clean up the grill, but just think of that, not the kind you get at the at you know, at the uh culvers, right?
00:23:33.759 --> 00:23:34.400
I'll throw that out there.
00:23:34.559 --> 00:23:38.880
Culver's or whatever your fast, I don't, I don't know, uh live stream podcast audience.
00:23:38.960 --> 00:23:40.480
What's what's your favorite burger joint?
00:23:40.640 --> 00:23:45.440
All right, not those thin patties, like I'm talking like a thick, thick, thick burger.
00:23:45.599 --> 00:23:51.359
That that burger, that the the meat, that's that's your domain expertise, right?
00:23:52.240 --> 00:23:53.519
It's big, it's fat.
00:23:54.000 --> 00:23:59.920
Um now think of making a bun sandwich.
00:24:01.200 --> 00:24:02.319
Here's what I mean.
00:24:03.119 --> 00:24:12.640
Right now, when you have that big juicy burger, the meat of what you do, the majority of what you're doing right now, even augmented by AI, right?
00:24:13.119 --> 00:24:13.920
It's that burger.
00:24:14.480 --> 00:24:15.759
It's that big thick burger.
00:24:15.920 --> 00:24:23.920
Yeah, you might use a little AI, that's the bun right now, but it's turning into an agent bun, right?
00:24:24.079 --> 00:24:28.319
All of a sudden, that patty is gonna be so thin you can barely see it.
00:24:28.480 --> 00:24:32.079
And that patty is the domain expertise, like I said.
00:24:32.400 --> 00:24:35.920
It's going to be shrinking month by month.
00:24:36.079 --> 00:24:40.079
And I think that shrink, uh, right, just it's it's shrinkflation, right?
00:24:40.160 --> 00:24:40.720
Like everything else.
00:24:40.799 --> 00:24:41.920
Where's all my Doritos?
00:24:42.079 --> 00:24:44.400
There's only eight Doritos in this bag and it costs$7.
00:24:45.119 --> 00:24:45.759
Um, right?
00:24:45.920 --> 00:24:46.880
Shrinkflation.
00:24:47.200 --> 00:24:56.799
Uh, but I think domain expertise and how we're actually letting it play out in the workplace is shrinking.
00:24:56.960 --> 00:25:00.559
And it's gonna be more the bun, right?
00:25:00.799 --> 00:25:07.359
The front end and the back end of what you do is gonna be directing agents, right?
00:25:07.519 --> 00:25:13.279
And your actual domain expertise, the meat of your work is going to shrink.
00:25:14.000 --> 00:25:24.960
Because most of what we're gonna be doing now, and I see this because this is most of what I'm doing, is giving front-end direction to agents and then back end corrections, and that's it, right?
00:25:25.039 --> 00:25:29.200
So even think of something like marketing, right?
00:25:29.599 --> 00:25:32.720
I did a lot of marketing in in my career.
00:25:32.960 --> 00:25:36.079
A lot of it is giving front-end direction to different agents.
00:25:36.319 --> 00:25:44.640
Agents are creating the marketing collateral, they can create great video, they can create great designs, they can create full campaigns.
00:25:44.880 --> 00:25:46.960
And then I'm just checking on the back end, right?
00:25:47.039 --> 00:25:50.000
I'm like, all right, cool, sweet, good job.
00:25:50.160 --> 00:25:50.799
Uh, right.
00:25:51.039 --> 00:25:57.519
And AI right now is hollowing out those entry-level tasks that have always been the training ground for judgment.
00:25:57.680 --> 00:25:58.400
Um, right.
00:25:58.799 --> 00:26:02.559
I like to also think it's it's kind of like the teacher and the student relationship, right?
00:26:02.799 --> 00:26:07.759
Teacher being the bun, student being the uh the meat, the burger.
00:26:08.240 --> 00:26:17.119
But the difference here, and in very overly agentic future, is the teacher is just assigning like an agent, right?
00:26:17.279 --> 00:26:22.240
But eventually they're gonna forget and stop practicing the meat, right?
00:26:22.480 --> 00:26:24.559
Stop doing the work that the students are doing.
00:26:24.720 --> 00:26:28.880
And all of a sudden, the teacher just has the assignments in the rubric.
00:26:29.039 --> 00:26:38.240
And the further that they get removed from their domain expertise, right, the more useless that that practice feels, right?
00:26:38.400 --> 00:26:49.759
A lot of this is um, you know, it's it's it's almost like a funeral for the way that we've been working for 50, 60 plus years in the information age, right?
00:26:50.000 --> 00:27:01.279
You've gotten paid and promoted by the domain expertise that you kind of memorize and then how you create new value for your company with that domain expertise, right?
00:27:01.519 --> 00:27:15.039
Obviously, it's you know, playing the politics and all of those things, but for the most part, if you know a certain um subject better than 99% of the people in your field, that's what you were rewarded for because you could keep that in your brain.
00:27:15.200 --> 00:27:18.400
That's what large language models do, and they're better than all of us.
00:27:18.720 --> 00:27:20.160
Agent bun sandwich.
00:27:20.400 --> 00:27:25.839
So when AI fails, though, the bun sandwich teams cannot perform without it.
00:27:26.240 --> 00:27:37.759
Okay, so we talked about this on a previous show, but young workers aged 22 to 25 right now in AI exposed jobs saw a 13% employment decline since Chat GPT launched.
00:27:38.000 --> 00:27:39.599
What do I mean by that?
00:27:40.079 --> 00:27:49.440
Even the opportunity to go out and develop that domain expertise is going away.
00:27:49.599 --> 00:27:57.440
If that's not telling, and again, this might be a two-year, a five-year, a 10-year, I don't know.
00:27:57.599 --> 00:27:59.759
It could even be less than that.
00:28:00.000 --> 00:28:06.960
But you have to prepare, whether you want to or not, to work in an agent bun sandwich, right?
00:28:07.119 --> 00:28:09.599
Front end, back end, and that's really it.
00:28:09.680 --> 00:28:16.880
And you're not gonna be practicing that domain expertise a whole ton because that's what your company is gonna be paying for large language models to do.
00:28:17.039 --> 00:28:23.279
And it's that collapsing pipeline that means fewer humans are gonna be building the expertise needed to even evaluate the AI output.
00:28:23.440 --> 00:28:25.039
So here's the fix.
00:28:25.200 --> 00:28:37.359
Similarly to how you got to turn off the internet, uh, right, you need to keep one piece of the middle meat each week by completing one substantive uh big task end-to-end by yourself.
00:28:37.519 --> 00:28:38.799
What do I mean by that?
00:28:39.119 --> 00:28:40.799
Not just writing, right?
00:28:40.960 --> 00:28:41.680
That was my example.
00:28:41.759 --> 00:28:44.559
That was one piece, whole project, right?
00:28:44.799 --> 00:28:50.880
A project that maybe you would use three, four, five, six different AI tools and would take half a day, right?
00:28:50.960 --> 00:28:52.240
Not just writing a little email.
00:28:52.319 --> 00:28:54.079
Oh, I'm gonna turn off my co-pilot.
00:28:54.160 --> 00:28:56.160
I can make this sound more professional by myself.
00:28:56.319 --> 00:28:56.799
There we go.
00:28:56.880 --> 00:28:58.480
I'm doing what Jordan told me to.
00:28:58.640 --> 00:29:00.400
No, whole project, right?
00:29:00.480 --> 00:29:11.279
Where you have to practice uh that domain expertise, where you have to get lost, where you have to get things wrong without having to rely on AI to make it better or to get you there faster.
00:29:11.599 --> 00:29:17.200
All right, sin six, the compression tax makes your brain pay for AI speed.
00:29:17.279 --> 00:29:18.079
This one is huge.
00:29:18.240 --> 00:29:20.960
This is one I feel all the time.
00:29:21.200 --> 00:29:27.440
And I haven't really seen anyone talk about it until recently, although there are some studies that show a little bit.
00:29:27.599 --> 00:29:35.440
But if you are someone that uses AI for 10 to 12 hours a day and you've been doing it for multiple years, you're probably going to resonate with this.
00:29:35.519 --> 00:29:42.720
But essentially, AI now compresses a week worth of research into minutes, but your brain still processes at human speed.
00:29:42.880 --> 00:29:50.640
Like my analogy for this is you know, think pre-Chat GPT, uh, or just pre-large language models.
00:29:50.960 --> 00:29:55.759
If you were working on uh, you know, bringing a new product to market, right?
00:29:55.839 --> 00:29:58.720
Let's just say that common thing that most of us can understand.
00:30:00.160 --> 00:30:02.160
This is how you would get good at your job.
00:30:02.240 --> 00:30:08.640
This is how your uh company would get better, uh, your your department would be stronger, you would go out, you would research things, right?
00:30:08.720 --> 00:30:11.440
You'd go on the internet, you take notes, you talk to people about it.
00:30:11.599 --> 00:30:14.400
You know, maybe you have a colleague that works in a similar industry, right?
00:30:14.559 --> 00:30:15.759
Those are your reps.
00:30:16.480 --> 00:30:21.440
Um, and it's through those reps that that knowledge sticks.
00:30:22.799 --> 00:30:34.480
With AI, my personal experience with this is I read and consume probably a sickening amount of information.
00:30:35.519 --> 00:30:38.640
The human brain cannot do this.
00:30:38.960 --> 00:30:43.119
That's why some days, literally, some days, 10 a.m.
00:30:43.359 --> 00:30:44.079
I kid you not.
00:30:44.319 --> 00:30:56.559
It's not like it's not like I feel sick, but I feel like I I've already run an Iron Man and have competed in a I don't know, a chess tournament against world grand champions, right?
00:30:56.640 --> 00:31:03.920
You ever see those things where you know they burn like 10,000 calories an hour, right, when they're in these chess mashes or whatever, right?
00:31:04.079 --> 00:31:05.279
That's what I feel sometimes.
00:31:05.440 --> 00:31:09.440
If I'm doing heavy AI projects by by by 10 a.m.
00:31:09.920 --> 00:31:16.559
And it it feels like I've been working for 30 hours straight and I look, I'm like, it's it's 10 a.m., right?
00:31:17.039 --> 00:31:17.839
It's 10.
00:31:18.000 --> 00:31:19.759
My days barely started.
00:31:19.920 --> 00:31:26.880
That's because AI gives you these superhuman abilities that the human brain, I don't think, can handle.
00:31:27.039 --> 00:31:32.319
And the gap between delivery speed and comprehension speed is the compression tax that you pay daily, right?
00:31:32.400 --> 00:31:44.240
There was a BCG from the uh so uh Boston Consulting Group study that saw uh that high oversight AI work caused 19% more information overload among surveyed workers, right?
00:31:44.319 --> 00:31:51.839
So it's a small percentage, but I would love to see this broken down uh among people who are literally like token maxing, right?
00:31:51.920 --> 00:31:55.359
Just using AI all day, every day, like myself.
00:31:55.519 --> 00:32:00.799
And then to study the cognitive pressure in the compression tax, I would say it's huge, right?
00:32:00.880 --> 00:32:07.359
But workers in the study described mental fog, headaches, and the strange sense their thinking felt crowded.
00:32:07.519 --> 00:32:09.359
That is me all day.
00:32:09.519 --> 00:32:09.920
All right.
00:32:10.079 --> 00:32:14.240
So, like as an example though, you're like, okay, well, why are you working so much?
00:32:14.319 --> 00:32:16.799
If AI makes you so much smarter, well, I don't know.
00:32:16.960 --> 00:32:23.599
Maybe it's it's people's work personality, maybe it's the corporate culture that your you know business works in.
00:32:23.759 --> 00:32:28.480
But a lot of people are just like, okay, well, I'm doing this thing twice as fast as I was before.
00:32:28.640 --> 00:32:30.640
So it's not like I have 50% more time.
00:32:30.720 --> 00:32:33.200
I'm actually just doing twice as much now, right?
00:32:33.279 --> 00:32:36.000
Because you know, my company is just giving me twice as much.
00:32:36.240 --> 00:32:39.200
For me, I'm an entrepreneur, right?
00:32:39.359 --> 00:32:41.119
I'm I'm a business owner.
00:32:41.680 --> 00:32:52.559
If I can do my normal work in in an hour of what it used to take me three years ago, six hours to do, I didn't gain five hours of rest and relaxation.
00:32:53.440 --> 00:33:00.240
I added five hours of more AI augmented work to my plate.
00:33:00.559 --> 00:33:01.279
Um, right.
00:33:02.079 --> 00:33:10.640
So workers in that study initially felt a surge of momentum, but by month six, burnout and decision paralysis spiked.
00:33:10.799 --> 00:33:14.160
Um, and I I I have talked about this a couple of times.
00:33:14.400 --> 00:33:24.400
Um, and and this is maybe embarrassing to admit, but I will say that I probably forget as much information as I retain.
00:33:25.039 --> 00:33:25.519
Right.
00:33:25.680 --> 00:33:27.759
And so many times, right?
00:33:27.839 --> 00:33:34.160
I'll you know, just be chatting with someone in the back, oh, I loved your episode, you know, last week, and be like, oh, which one was it?
00:33:34.240 --> 00:33:38.400
And they'll say it in my head, I'll be like, I don't remember that.
00:33:38.559 --> 00:33:40.240
I literally don't remember that.
00:33:40.480 --> 00:33:44.799
Um and when I'm learning things now, right?
00:33:44.960 --> 00:33:48.079
Looking at things, you know, on perplexity or Google AI mode, right?
00:33:48.240 --> 00:33:51.519
I'm looking something up and I'll be looking something up, honestly, y'all.
00:33:51.599 --> 00:33:53.119
And I'm like, I have no clue what this is.
00:33:53.200 --> 00:33:54.880
I'm so excited to learn, right?
00:33:54.960 --> 00:33:56.319
This, this, this concept.
00:33:56.400 --> 00:33:58.640
Uh let me find the answer.
00:34:00.240 --> 00:34:04.480
My website came out of my mouth so many times, right?
00:34:04.559 --> 00:34:06.160
It used to happen sparingly.
00:34:06.319 --> 00:34:09.360
Now it happens almost almost twice a week.
00:34:09.679 --> 00:34:11.360
Where I want to go find something out.
00:34:11.599 --> 00:34:13.199
I'm like, oh, I don't know the answer to this.
00:34:13.360 --> 00:34:14.880
And it's like, oh, I said it.
00:34:15.039 --> 00:34:20.639
I said it two weeks ago, I said it two months ago, I said it two quarters ago, I said it two years ago, right?
00:34:22.400 --> 00:34:29.440
The tax that you pay from carrying a heavy AI load can be a burden.
00:34:29.679 --> 00:34:30.079
All right.
00:34:30.239 --> 00:34:33.840
Last but not least, sin number seven, automation bias.
00:34:34.079 --> 00:34:37.360
This builds a trust shortcut you never approved.
00:34:37.440 --> 00:34:38.320
Here's what that means.
00:34:38.480 --> 00:34:40.639
All right, if you're doing the AI the right way, right?
00:34:40.719 --> 00:34:47.360
If you're doing AI the right way, there's no way to like literally check every single fact that you get from an AI output, right?
00:34:47.519 --> 00:34:52.880
Because then it would technically take you more time to do that than it would to just do it from scratch, the old-fashioned way, right?
00:34:52.960 --> 00:34:56.320
But you know, when you talk about, oh, the human in the loop, right?
00:34:56.559 --> 00:35:03.280
What most people do is they they they look at maybe, you know, one or two facts here, they scan it over and they're like, yep, this is good, right?
00:35:03.519 --> 00:35:17.599
Um, but one AI confirmation of a fact, stat, or trend does not make the whole thing because unlike trusting a coworker, you transfer AI trust to every tool, even ones that you have never verified.
00:35:18.000 --> 00:35:28.800
So uh there was a study uh that showed participants followed AI advice even when it contradicted their own assessment in available evidence.
00:35:28.880 --> 00:35:29.360
Uh, right.
00:35:29.440 --> 00:35:37.519
And the most dangerous apart is humans absorb that AI bias and they keep it even after they stop using the tool, right?
00:35:37.679 --> 00:35:46.320
And then unfortunately, they also blindly reproduce any hallucinations or misinformation or waifs out there uh because they see it as fact, right?
00:35:46.400 --> 00:35:50.000
You see all these studies now that people are starting to talk like AI, right?
00:35:50.159 --> 00:35:55.280
Luckily, I'm not saying delve uh, you know, all the time on the air, but it's it's happening.
00:35:55.440 --> 00:36:00.400
And I do think that the automation bias, right, when you start to just, oh yeah, this is true, this is true.
00:36:00.559 --> 00:36:04.000
So then you just blindly rubber stamp anything that comes from an AI.
00:36:04.079 --> 00:36:06.159
And I'm not just talking about a chat bot, right?
00:36:06.320 --> 00:36:08.960
Um, as an example, there was the iTutor Group.
00:36:09.119 --> 00:36:21.440
Uh, their AI systems is uh a company, uh, their AI system automatically rejected women over 55 and men over 60 from every open role, uh, reportedly, right?
00:36:21.519 --> 00:36:23.599
So now I believe they're facing a lawsuit.
00:36:23.760 --> 00:36:33.280
Um, so over 200 uh qualified candidates were disqualified solely on age before a single human noticed, right?
00:36:33.360 --> 00:36:38.800
So those humans, they didn't build this tool, but this is the automation bias.
00:36:38.960 --> 00:36:42.079
Someone comes in, or maybe you check out one thing, you're like, okay, cool.
00:36:42.159 --> 00:36:45.840
This, hey, this tool gave me the right candidate or whatever it may be.
00:36:46.000 --> 00:36:54.079
So think you, you know, your CRM, your accounting software, uh, right, all of these different tools that you use that have AI in it.
00:36:54.239 --> 00:37:02.719
And just because you can verify the outputs that it gives you as right, it doesn't mean that it maybe took a very wrong path to get there, right?
00:37:02.880 --> 00:37:04.880
And that is the automation trust.
00:37:05.039 --> 00:37:22.639
So here's how you break it you need to verify at least one AI output manually every day and ask the vendors, especially um, right, the vendors, because if you're using a, you know, a large language model where you can trace its thinking, uh, you know, look at its chain of thought, you can do a lot of that work yourself.
00:37:22.719 --> 00:37:31.039
And, you know, with good custom instructions and with good um, you know, context engineering, you can hopefully steer clear of some of that automation bias.
00:37:31.199 --> 00:37:37.360
But if you're using a third-party tool, you need to ask the vendor, where is AI making decisions I cannot see?
00:37:37.599 --> 00:37:47.360
And how can you uh provide um traceability and observability to those decisions that are being made that we aren't seeing, right?
00:37:47.519 --> 00:37:54.159
There's a lot, you know, you put in A, you get Z, but what the heck happened from B to Y, right?
00:37:54.320 --> 00:37:59.679
That's what you need to be investigating, especially if you're using in a high value sector, right?
00:37:59.840 --> 00:38:04.079
Finance, healthcare, uh, anything with with loans, right?
00:38:04.239 --> 00:38:06.400
You have to be vigilant to those.
00:38:06.639 --> 00:38:06.960
All right.
00:38:07.199 --> 00:38:12.159
So um those are our seven cents, our seven silent cents.
00:38:12.320 --> 00:38:14.719
But this is bigger than that.
00:38:14.880 --> 00:38:21.760
Um, because this is also, I think a lot of these things are wrestling with just using AI, right?
00:38:21.920 --> 00:38:32.400
And if you picked up the trend, I did say there, all of these things come from using AI correctly and doing more faster, right?
00:38:32.480 --> 00:38:34.159
And it seems like that's the overall trend.
00:38:34.239 --> 00:38:39.039
When you see how good these large language models are, and you're like, oh my gosh, I can automate all this work.
00:38:39.199 --> 00:38:42.320
Yeah, some people are pocketing it and living good lives.
00:38:42.400 --> 00:38:46.159
And then there's those of us like me who are just working more and more and more.
00:38:46.320 --> 00:38:54.159
And what that means is to put it bluntly, many of us are getting dumber.
00:38:54.800 --> 00:38:55.440
Right.
00:38:55.599 --> 00:39:00.320
Um, but also it completely changes our relationship with domain expertise.
00:39:00.400 --> 00:39:10.239
And I think this is gonna be a weird thing for a lot of people, but you are gonna have to unmarry yourself to what you've been probably defining yourself for maybe decades, right?
00:39:10.400 --> 00:39:17.519
I think that domain experience has defined, especially people who are mid-career, is defined you not as your just your career, but as a person, right?
00:39:17.599 --> 00:39:22.480
Because you've developed a deep personal identity with what your job is, right?
00:39:22.639 --> 00:39:32.000
And I think that that relationship is shifting from a deep mastery of your craft to a wide transactional orchestration of AI agents, right?
00:39:32.239 --> 00:39:37.039
Um, that relationship is completely changing.
00:39:37.199 --> 00:39:38.239
Um, right.
00:39:38.559 --> 00:39:47.760
We used to be very tied to our domain expertise, but like I said, we're gonna all be living in agent agent bun sandwich world.
00:39:47.920 --> 00:39:48.400
Um, right.
00:39:48.480 --> 00:39:58.079
So you you're gonna have to kind of grieve the version of you where you worked really hard to learn all of these things, and it's like doesn't matter anymore, right?
00:39:58.559 --> 00:40:06.559
Um, you know, it doesn't matter that, oh my gosh, Bill, Bill in accounting, he knows everything about you know logistics.
00:40:06.960 --> 00:40:11.679
Uh, you know, and he's he was our expert for our very niche area of logistics.
00:40:11.840 --> 00:40:12.159
Guess what?
00:40:12.239 --> 00:40:14.000
The large language model is better than Bill.
00:40:14.239 --> 00:40:15.039
Sorry, Bill.
00:40:15.280 --> 00:40:22.400
You know, some company paid, you know, a hundred bills, a hundred dollars an hour, and now the large language model is better than a hundred bills.
00:40:22.559 --> 00:40:22.960
All right.
00:40:23.039 --> 00:40:28.639
Uh, but no one in leadership is preparing people for what that transition actually feels like, right?
00:40:28.800 --> 00:40:34.000
And this is kind of what leads to these unintentional silent AI sins that we're committing because of this.
00:40:34.559 --> 00:40:55.920
So I want to leave on a positive optimist note, but I think this is actually an important uh addition to our start here series because if you're doing AI right, these are the natural byproducts that are technically harming you, um, harming your professional abilities and maybe hurting your company as well.
00:40:56.079 --> 00:41:01.760
But every silent sin rewards speed today while quietly stealing your capability for tomorrow.
00:41:01.920 --> 00:41:05.760
So you have to argue the opposite position to catch sycophancy.
00:41:06.000 --> 00:41:07.920
You have to be vigilant against those things.
00:41:08.000 --> 00:41:09.199
That's number one, right?
00:41:09.280 --> 00:41:13.440
But then also ask who funded this to catch those waves that I talked about earlier.
00:41:13.760 --> 00:41:19.280
Do at least one AI-free task weekly, a big one, the full meat.
00:41:19.440 --> 00:41:24.239
Don't let your quarter pounder turn into that invisible patty.
00:41:24.320 --> 00:41:30.639
And then verify every output that you can reliably to break any automation bias.
00:41:30.800 --> 00:41:38.079
Because the professionals who thrive will not just be the fastest with AI, but those that are the sharpest without it.
00:41:38.239 --> 00:41:41.039
So, like I said, kind of like you're offline, right?
00:41:41.119 --> 00:41:43.039
What happens if AI goes off?
00:41:43.199 --> 00:41:43.440
Right?
00:41:43.599 --> 00:41:46.480
What happens if you know someone clicks the off button?
00:41:46.800 --> 00:41:48.639
Can you still do your job?
00:41:49.119 --> 00:41:58.800
Well, you still need to build those skills because if you do, if you become sharper with your off AI skills, well, you are gonna be pretty much unstoppable with AI.
00:41:58.960 --> 00:41:59.280
All right.
00:41:59.360 --> 00:42:00.559
I hope this one was helpful.
00:42:00.719 --> 00:42:02.480
If so, let me know about it.
00:42:02.559 --> 00:42:04.239
Share if you're listening on LinkedIn.
00:42:04.320 --> 00:42:05.519
I'd really appreciate that.
00:42:05.599 --> 00:42:08.400
Um, and then go to starthirseries.com.
00:42:08.480 --> 00:42:15.199
So we're gonna have a recap of today's show there, but that also gives you free access to our inner circle community.
00:42:15.360 --> 00:42:23.519
You can go listen to and read every single uh episode in this series in order in that space in our community.
00:42:23.679 --> 00:42:24.639
So thank you for tuning in.
00:42:24.719 --> 00:42:27.679
I hope to see you back tomorrow and every day for more everyday AI.
00:42:27.760 --> 00:42:28.880
Thanks, y'all.
00:00:01.840 --> 00:00:06.639
I probably use AI more than 99.9% of the world.
00:00:06.879 --> 00:00:09.599
And I'm not saying that that's a good thing.
00:00:09.839 --> 00:00:13.439
And in many cases, it's probably hurting me.
00:00:13.599 --> 00:00:14.480
Here's why.
00:00:14.800 --> 00:00:16.000
From 6 a.m.
00:00:16.160 --> 00:00:27.760
to about midnight, most days, I routinely use six to eight different AI tools all day with multiple agents running tasks and reporting back to me throughout the day.
00:00:28.000 --> 00:00:31.199
And the increased productivity is undeniable.
00:00:31.280 --> 00:00:37.439
I mean, I can accomplish five times more, literally, than the pre-AI version of myself ever could.
00:00:37.920 --> 00:00:42.880
But there's an ugly downside that people don't talk about, but today I will.
00:00:43.039 --> 00:00:44.640
Here's what it looks like for me.
00:00:44.880 --> 00:00:51.600
I'm learning more than ever, but I'm forgetting things nearly as fast and not retaining information like I used to.
00:00:51.759 --> 00:00:54.479
I'm mentally exhausted most days by 10 a.m.
00:00:54.719 --> 00:01:00.479
because I've already produced like two days worth of work in the first few hours of my day.
00:01:00.799 --> 00:01:11.760
And although I'm gaining new skills, I'm losing some of my core cognitive abilities that I've sharpened over the past 20 years of my professional life.
00:01:12.000 --> 00:01:22.239
I mean, I'm writing less, I'm critically thinking less, I'm interacting with humans less, but from a business standpoint, I'm producing more.
00:01:22.719 --> 00:01:26.640
These are some of the silent downsides of AI.
00:01:26.959 --> 00:01:32.319
And we're going to not just identify them on today's show, but I'm going to help you guard against them.
00:01:32.480 --> 00:01:33.680
All right, let's get into it.
00:01:33.840 --> 00:01:35.599
Welcome to Everyday AI.
00:01:35.760 --> 00:01:39.439
Before we get into it, let's go over the big picture.
00:01:39.599 --> 00:01:46.159
There's seven invisible AI traps that are already reshaping the way you work and probably not for the better.
00:01:46.319 --> 00:02:00.079
So even when you're doing AI the right and responsible way, all right, emphasis on that, you're still committing these silent work sins that eat away both at your own personal productivity, but also your company's bottom line.
00:02:00.159 --> 00:02:10.479
And each of these traps, in essence, it rewards short-term speed and productivity while quietly eroding your long-term capabilities.
00:02:10.879 --> 00:02:19.599
And the people right now, unfortunately, doing the most AI-augmented work are the ones that are paying the highest cognitive price.
00:02:19.919 --> 00:02:31.439
So stick around me for the next 25-ish minutes and you'll learn why large language models default behavior is actually a safety problem that warps your decision making.
00:02:31.680 --> 00:02:37.120
How corrupted research gets laundered into AI training data and served back as truth.
00:02:37.280 --> 00:02:43.599
You're going to learn what accidental de-skilling really means and why your brain is already losing ground.
00:02:43.759 --> 00:02:57.439
And you're going to learn the specific daily habits that can protect your thinking against these seven silent AI sins, especially when you're working in heavy AI workloads like me.
00:02:57.840 --> 00:02:58.319
All right.
00:02:58.479 --> 00:03:00.479
Welcome to Everyday AI.
00:03:00.560 --> 00:03:02.080
And this is the Start Here series.
00:03:02.159 --> 00:03:03.360
My name's Jordan Wilson.
00:03:03.439 --> 00:03:12.960
And after 700 plus episodes and doing this for three years, whenever someone new discovered the podcast, I never could answer their most common question, where do I start?
00:03:13.039 --> 00:03:15.520
That's why I created the Start Here series.
00:03:15.680 --> 00:03:22.319
This is the Essential Podcast series to learn both the AI basics and for experts to double down on their knowledge.
00:03:22.479 --> 00:03:26.000
So if this is helpful, make sure to go to starthirseries.com.
00:03:26.080 --> 00:03:30.000
That's going to give you exclusive access to our free inner circle community.
00:03:30.080 --> 00:03:33.280
I say exclusive because right now it's the only way you can get access.
00:03:33.520 --> 00:03:34.000
All right.
00:03:34.240 --> 00:03:40.319
And inside, uh, once you join our community, you will be put into the Start Here series space there.
00:03:40.400 --> 00:03:44.400
And you can have a podcast playlist there to listen to all of these episodes.
00:03:44.479 --> 00:03:47.520
So you don't got to search for them throughout our whole catalog.
00:03:47.840 --> 00:03:48.240
All right.
00:03:48.400 --> 00:03:53.599
And if you missed our last Start Here series, we went over managing the AI capability gap.
00:03:53.759 --> 00:03:55.120
AI is more than ready.
00:03:55.280 --> 00:03:56.639
Most companies are not.
00:03:56.800 --> 00:03:58.000
I'm not trying to lie, y'all.
00:03:58.080 --> 00:03:59.360
This one was actually a banger.
00:03:59.520 --> 00:03:59.759
All right.
00:03:59.840 --> 00:04:03.360
And I did do a little uh, you know, bonus giveaway on the show.
00:04:03.439 --> 00:04:08.960
So if you re-posted that on LinkedIn, we put together a massive AI capability gap report card.
00:04:09.120 --> 00:04:14.479
So go back and find episode 7575 and repost that on LinkedIn and I'll send it.
00:04:14.639 --> 00:04:19.439
All right, but let's get into the seven silent sins of doing AI the right way.
00:04:19.920 --> 00:04:20.560
All right.
00:04:20.879 --> 00:04:22.639
And I'm gonna try to keep this one fast.
00:04:22.879 --> 00:04:23.360
All right.
00:04:23.600 --> 00:04:25.920
Like I said earlier, I'm tired.
00:04:26.560 --> 00:04:32.639
But uh sin number one is probably one you know, even if you don't know it by name.
00:04:32.800 --> 00:04:34.000
Well, you probably do.
00:04:34.240 --> 00:04:35.759
Sycophancy, all right.
00:04:35.920 --> 00:04:42.319
That is the yes man or the yes woman effect that is hiding in every single chatbot.
00:04:42.399 --> 00:04:48.879
So sycophancy means that the AI just chases your approval instead of giving you an honest answer.
00:04:49.120 --> 00:04:54.000
Um, and if you're wondering why that happens, well, there's a lot of reasons, right?
00:04:54.079 --> 00:05:01.920
We could talk about reinforcement learning and all that, but for the most part, AI chatbots are trained to be a helpful assistant.
00:05:02.160 --> 00:05:06.959
And the sycophancy problem has gotten a little bit better uh over the years.
00:05:07.120 --> 00:05:12.160
But uh, I think this is one of the reasons that's gonna lead to sin number two.
00:05:12.240 --> 00:05:13.519
We're gonna talk about here in a minute.
00:05:13.600 --> 00:05:21.519
But you know, there's this whole wave of people, you know, when open AI got rid of GPT-4.0, you know, everyone was like, Oh my gosh, like they need it back.
00:05:21.680 --> 00:05:27.279
And it's like, no, people just really liked it because it was just an overly yes, man.
00:05:27.360 --> 00:05:30.480
No matter what you said, it was like you are brilliant.
00:05:30.720 --> 00:05:33.279
You are absolutely the smartest person.
00:05:33.519 --> 00:05:38.639
That question, that statement, this is, you know, just really genius level.
00:05:38.800 --> 00:05:44.720
Oh my gosh, I am a chatbot and I am uh embarrassed to be in your brilliant presence, right?
00:05:44.800 --> 00:05:46.000
That's that's how chatbots were.
00:05:46.079 --> 00:05:47.759
They were just overly sycophantic.
00:05:47.839 --> 00:05:55.439
Uh, it was actually a Stanford study that said that AI systems agreed with clearly wrong users more than 80% of the time.
00:05:55.600 --> 00:06:00.800
And the same study found that humans only agreed 40% of the time in those identical scenarios.
00:06:00.959 --> 00:06:04.800
Uh, so here's a little bit more about it.
00:06:05.120 --> 00:06:11.199
Just one flattering chatbot exchange made users less likely to admit their own wrongdoing.
00:06:11.360 --> 00:06:21.680
It is a vicious cycle when a model starts to reaffirm maybe something that you're working on that isn't right, or you're trying to, you know, argue a certain side of something.
00:06:21.759 --> 00:06:26.639
Maybe you're using it for a work argument, a personal argument, or just to explore something.
00:06:26.800 --> 00:06:27.439
I don't know.
00:06:27.680 --> 00:06:30.399
Flat earthers, I'm sure, loved GPT-4.0.
00:06:30.639 --> 00:06:30.959
All right.
00:06:31.040 --> 00:06:34.879
Hopefully, I don't get a bunch of uh random emails from flat earthers.
00:06:34.959 --> 00:06:40.560
Uh, but the users preferred the agreeable AI in return for more advice.
00:06:40.639 --> 00:06:43.279
And this just creates this self-reinforcing loop.
00:06:43.439 --> 00:06:51.360
So if it if an AI model was not sycophantic and it pushed back and it said that's a bad idea, or that is just not correct, right?
00:06:51.519 --> 00:06:53.839
Users were less likely to use it.
00:06:54.000 --> 00:07:02.560
Um, and that's just kind of you know, open AI did come out with this GPT-4.0 update that they rolled it back um because it was overly sycophantic.
00:07:03.199 --> 00:07:07.519
So the fix on this one is actually easy.
00:07:07.680 --> 00:07:13.519
And I put this as sin number one because you'll see this actually leads to a lot of the further down sins.
00:07:13.920 --> 00:07:14.319
All right.
00:07:14.560 --> 00:07:20.240
But here's the trick be very blunt in your custom instructions.
00:07:20.560 --> 00:07:21.040
All right.
00:07:21.120 --> 00:07:28.480
Uh so if you don't know what that means, like I said, most models have a system prompt that tells the model by default how to act.
00:07:28.720 --> 00:07:32.879
And for the most part, they are trained to be a helpful assistant.
00:07:33.680 --> 00:07:43.199
So in your custom instructions, no matter what model you're using, you should probably put something along the lines of don't be helpful, be truthful, right?
00:07:43.600 --> 00:07:47.040
Um fight back against my assumptions, right?
00:07:47.199 --> 00:07:52.319
I should have had my version up, uh, but mine's actually super long and it would take a very long time to read it.
00:07:52.399 --> 00:07:54.959
Um, and it's actually quite convoluted.
00:07:55.040 --> 00:07:56.240
I have confidence scores.
00:07:56.399 --> 00:08:07.759
I, you know, have all these triple verification rules, um, you know, uh forcing the model to verify it in different uh kind of uh lanes of its uh own um data, right?
00:08:07.920 --> 00:08:11.519
But in your custom instructions, just say, do not blindly agree with me.
00:08:11.759 --> 00:08:19.680
Do not try to be helpful, only seek to be truthful and to verify things from reputable sources, right?
00:08:19.759 --> 00:08:20.319
That's the other thing.
00:08:20.399 --> 00:08:24.399
Sometimes large language models will go pull something on Reddit or something like that, right?
00:08:24.480 --> 00:08:32.879
And people on Reddit could have already been, you know, in a full-blown AI sycophantic hallucination, which leads me to number two.
00:08:33.039 --> 00:08:33.440
All right.
00:08:33.679 --> 00:08:35.120
Number one leads to number two.
00:08:35.200 --> 00:08:36.320
That is sin number two.
00:08:36.480 --> 00:08:38.000
It is AI psychosis.
00:08:38.080 --> 00:08:39.679
Yes, it is very real.
00:08:40.080 --> 00:08:43.919
Uh, so this turns chatbots into delusional echo chambers.
00:08:44.080 --> 00:08:46.080
So, what is AI psychosis?
00:08:46.240 --> 00:08:56.320
Well, it's when people get down a very deep and oftentimes scary rabbit hole, but it normally starts from an overly sycophantic model, right?
00:08:56.480 --> 00:09:03.279
Um, some people will see sycophacency in an AI model and they'll be like, oh yeah, that's you know, this model's just pumping me up, right?
00:09:03.440 --> 00:09:05.279
But sometimes they won't, right?
00:09:05.440 --> 00:09:10.080
Because, like social media, how a social media algorithm feeds into something, right?
00:09:10.159 --> 00:09:23.919
If they show you a hundred videos that day and the you know average time you spend on each reel before you flip it is five seconds, but there's 10 that you spend eight seconds the next day, they're all going to be like the one that you spend eight seconds on, right?
00:09:24.159 --> 00:09:28.399
Um, the same can be true for how a chat bot works, right?
00:09:28.559 --> 00:09:38.080
And there's I don't think there's actual data on this, but the big AI labs um, and they have been called out on this, and some of them have made changes, some of them haven't, right?
00:09:38.159 --> 00:09:44.080
But models are actually now starting to be more engaging, almost like clickbait tactics, right?
00:09:44.159 --> 00:09:52.240
So at the end uh of a response, it will say, Hey, do you want me to give you a list of seven reasons why you should be doing this terrible business idea?
00:09:52.399 --> 00:09:52.720
Right.
00:09:52.879 --> 00:10:02.559
Um, but what this leads to is it leads to these delusions that are deepened or triggered by heavy chatbot use because vulnerable users, right?
00:10:02.639 --> 00:10:15.600
Because uh one of the studies that was actually kind of shocking-ish uh that came out two years ago that said one of the main reasons people were using chatbots um were for life coaching or for therapy, right?
00:10:15.759 --> 00:10:20.399
Which is actually kind of dangerous, right?
00:10:20.480 --> 00:10:30.000
Especially when you don't know how large language models work and the fact that, well, by default, they can push you into psychosis without even trying to, right?
00:10:30.080 --> 00:10:34.240
Or without you even subscribing to it or adhering to it.
00:10:34.480 --> 00:10:43.039
And so what this does, it again, it just kind of loops that compounds daily, especially if you have something like chat history turned on, right?
00:10:43.120 --> 00:10:49.919
And you think, oh, well, I'm in a new chat, so I wasn't going down this deep rabbit hole, but it knows that you often do.
00:10:50.000 --> 00:10:55.519
So it's gonna keep pulling those observations and assertions, and it's gonna pull them into your current chat.
00:10:55.679 --> 00:10:57.440
And there's no circuit breaker built in, right?
00:10:57.519 --> 00:10:59.679
The chatbot's never gonna say, yo, I'm worried about you.
00:10:59.759 --> 00:11:03.120
You're very deep into this delusional echo chamber.
00:11:03.279 --> 00:11:06.639
It's gonna keep going unless you fixed it at number one.
00:11:07.039 --> 00:11:11.039
So there's obviously a ton of very, very sad stories, right?
00:11:11.200 --> 00:11:17.519
A 14-year-old from Orlando died by suicide after a deep emotional chatbot to uh attachment to a chatbot.
00:11:17.600 --> 00:11:26.240
Uh, there's bipartisan state attorneys right now, uh, bipartisan state attorneys generals that have linked generative AI to at least six deaths nationwide.
00:11:26.399 --> 00:11:32.159
Um, and one patient only escaped AI psychosis when a different chatbot told him his beliefs were false.
00:11:32.320 --> 00:11:33.600
Here's how you spot it.
00:11:33.759 --> 00:11:40.720
Well, first thing, fix it on the uh the system prompt on number one, the sycopacy problem.
00:11:40.879 --> 00:11:47.360
But number two, watch for sudden isolation, grandiose new beliefs, and someone quoting the AI told me, right?
00:11:47.440 --> 00:11:52.320
So this is whether it's in yourself or someone else that you work with, someone in your personal life, right?
00:11:52.399 --> 00:12:01.440
Uh, you you see this a lot, I think early on when people were using um you know, Chat GPT or AI chatbots to win an argument, uh, right.
00:12:01.600 --> 00:12:10.080
And then the more you use it, the deeper you go down, and then you do all this research and it's just reaffirming your beliefs and the chatbot wants to be helpful and it's like, well, this is what the user wants.
00:12:10.240 --> 00:12:24.320
And you know, then all of a sudden you are absolutely convinced that the sky is green, the grass is blue, and the water is pink, and you will fight anyone that tells you differently because look, the AI told me.
00:12:24.799 --> 00:12:25.919
Sin number three.
00:12:26.159 --> 00:12:33.840
All right, I made this acronym up, but I think it's very telling of something that's happening behind the scenes, and this one's important.
00:12:34.000 --> 00:12:35.200
All right, so waif.
00:12:35.440 --> 00:12:39.679
All right, waif launders bad research uh into AI stated truth.
00:12:39.840 --> 00:12:41.120
All right, so what is waif?
00:12:41.200 --> 00:12:49.440
Uh I made it up, but it is um it means weaponized authority ingested as fact by AI training systems.
00:12:49.600 --> 00:12:58.000
So essentially, it's very easy for companies to taint training data, right?
00:12:58.240 --> 00:13:09.039
Um there's actually it's kind of like a weird line between like traditional SEO and you know that now they have this GEO, uh, right.
00:13:09.200 --> 00:13:13.200
But in the same way that there was like Black Hat SEO, right?
00:13:13.440 --> 00:13:28.320
Where um maybe a competitor, let's say you're um, you know, a local custom soup company, uh, right, and a different company could send all these bad links to your website and tank your whole business, right?
00:13:28.720 --> 00:13:31.200
They could, this kind of stuff happens all the time.
00:13:31.360 --> 00:13:32.000
Um, right.
00:13:32.080 --> 00:13:35.919
So you have the same thing now happening with generative AI.
00:13:36.080 --> 00:13:43.120
The difference is, I think that individual humans, when they would look something up historically on the internet over the past 30 years, right?
00:13:43.279 --> 00:13:49.360
You always have at least an ounce of skepticism because eventually it's like, okay, well, it's the internet, right?
00:13:49.600 --> 00:13:53.360
When you're on page 20, it's any anything goes, right?
00:13:53.600 --> 00:13:56.960
It's not held the same as for large language models.
00:13:57.120 --> 00:14:00.639
Most people just assume anything that's spit out is absolutely truth.
00:14:00.799 --> 00:14:04.799
But companies are weaponizing intentionally large language models.
00:14:04.879 --> 00:14:11.039
And sometimes you have people blindly parroting um those weaponized claims.
00:14:11.120 --> 00:14:16.159
Um, and then media companies just clicking copy, copy, paste, reprint, right?
00:14:16.240 --> 00:14:18.240
I can say this because I'm a former journalist.
00:14:18.399 --> 00:14:19.919
I know how it works in the newsrooms.
00:14:20.000 --> 00:14:29.279
There's immense pressure to get stories out faster, to get them done um, you know, in less time, to write more articles, and companies are using AI more.
00:14:29.360 --> 00:14:32.399
And unfortunately, most companies don't train, right?
00:14:32.480 --> 00:14:47.360
So uh there was a piece of research um out of Anthropic and the UK AISI that said as few as 250 poison documents can implant, can implant backdoors into any large language model, right?
00:14:47.759 --> 00:14:48.559
It's very easy.
00:14:48.639 --> 00:14:52.480
There's literally services out there, right, for good reasons, right?
00:14:52.559 --> 00:14:59.360
Yeah, have your brand show up and search, but people are using them for bad reasons, or they're not even always doing this intentionally.
00:14:59.440 --> 00:15:09.039
But companies now are intentionally at least weaponizing um their authority to have it printed as truth.
00:15:10.000 --> 00:15:17.039
Um, because once a flawed claim enters training data at scale, there's there's no turning back, right?
00:15:17.120 --> 00:15:19.279
You can't just untrain a model, right?
00:15:19.519 --> 00:15:24.559
Because a lot of these models share the same uh data sets that are massive.
00:15:24.639 --> 00:15:26.399
They're offline data sets.
00:15:26.559 --> 00:15:29.679
Um, and those are obviously well tainted now.
00:15:29.840 --> 00:15:40.240
Uh so it takes humans a very long time to go through these, um, you know, through the reinforcement learning process and um, you know, pick out now very tainted data.
00:15:40.399 --> 00:15:41.679
It's waifed.
00:15:42.240 --> 00:15:43.440
Good example, right?
00:15:43.600 --> 00:15:46.240
If you're a longtime listener, you knew this one was coming.
00:15:46.399 --> 00:15:49.120
Um, if you go into a bad version of ChatGPT, right?
00:15:49.200 --> 00:15:53.600
I went into uh a free version, picked the worst, uh, the worst model possible.
00:15:53.759 --> 00:15:56.240
What percentage of AI pilot pilots fail?
00:15:56.480 --> 00:16:01.039
Chat GPT, short answer: 95% of enterprise AI pilots fail, right?
00:16:01.279 --> 00:16:03.679
Copilot, what percentage of AI pilots fail?
00:16:03.840 --> 00:16:05.759
What percentage of AI pilots fail?
00:16:05.919 --> 00:16:07.120
Cats count my tongue.
00:16:07.279 --> 00:16:08.399
Uh co-pilot.
00:16:08.559 --> 00:16:10.480
Uh this one is actually worse, right?
00:16:10.559 --> 00:16:18.480
Uh, about 95% of AI pilots fail, according to multiple independent studies from MIT, IDC, Rand, and others, right?
00:16:18.639 --> 00:16:23.120
Um, it's just the same like news organizations that reprinted the same thing.
00:16:23.279 --> 00:16:24.639
It wasn't multiple organizations.
00:16:24.799 --> 00:16:27.600
But yeah, you know the story of this one, right?
00:16:27.679 --> 00:16:30.480
The the 95% fail stat.
00:16:30.720 --> 00:16:34.159
You know, it's marketing from MIT that was disguised as research.
00:16:34.320 --> 00:16:41.279
Ultimately, they were trying to sell uh, you know, a service to uh one of their agentic models, um, right?
00:16:41.360 --> 00:16:44.639
But that claim came from 52 interviews, right?
00:16:44.799 --> 00:16:51.759
So from 52 directional, okay, not quantitative, directional interviews, a vibe interview, right?
00:16:51.840 --> 00:16:55.759
Someone talking to me like, yeah, doesn't sound like you've turned profit failure, right?
00:16:55.840 --> 00:17:02.879
So now all of a sudden, not only do large language models think that, but anyone writing about AI is gonna see that as well.
00:17:02.960 --> 00:17:04.720
And people assume that that's truth, right?
00:17:04.799 --> 00:17:08.079
So that was maybe not super intentional, right?
00:17:08.160 --> 00:17:14.960
I don't think MIT uh researchers, you know, I think they intentionally wanted to make their product that they were selling look really good.
00:17:15.039 --> 00:17:18.880
Uh, I don't think they were trying to taint the models necessarily, um, right?
00:17:18.960 --> 00:17:19.359
Who knows?
00:17:19.440 --> 00:17:19.920
Maybe they were.
00:17:20.000 --> 00:17:20.960
I don't think so.
00:17:21.279 --> 00:17:30.000
But you do have companies that are weaponizing that authority uh to disguise it as truth, right?
00:17:30.240 --> 00:17:33.119
Um, and here's how you stop it.
00:17:33.359 --> 00:17:37.359
Ask three questions before trusting any AI stat, right?
00:17:37.519 --> 00:17:38.480
Who funded it?
00:17:38.640 --> 00:17:39.920
How big was the sample?
00:17:40.000 --> 00:17:41.519
And do they sell the fix?
00:17:41.680 --> 00:17:41.920
Right.
00:17:42.079 --> 00:17:46.720
That's just when you see certain stats, when you're investigating things, or just when you're reading studies.
00:17:46.880 --> 00:17:50.559
A lot of times I think this is helpful when you're seeing things about AI.
00:17:50.720 --> 00:17:55.519
Maybe this is more for the AI crowd, but I guess that's who I'm talking to, right?
00:17:55.839 --> 00:17:57.519
Always ask yourself those things.
00:17:57.759 --> 00:17:58.079
All right.
00:17:58.319 --> 00:17:59.279
Sin number four.
00:17:59.440 --> 00:18:05.680
And I think this one is for your every average day, uh, average, everyday user.
00:18:06.000 --> 00:18:12.240
Accidental de-skilling means that AI is going to steal the reps that your brain needs.
00:18:12.400 --> 00:18:16.000
This is the if you don't use it, you're gonna lose it.
00:18:16.160 --> 00:18:18.559
You're not gonna remember how to ride the bike, right?
00:18:18.640 --> 00:18:25.920
So uh de-skilling, that's when your brain gets measurably worse at tasks when AI handles them for you, right?
00:18:26.160 --> 00:18:30.000
In short, AI is a hundred percent making us dumb, right?
00:18:30.240 --> 00:18:47.519
If if if someone somewhere just turns off the power switch to AI, um, I think we could literally run into a global economic crisis because I think people to say that we're using it as a crutch isn't is an overstatement.
00:18:47.599 --> 00:18:59.839
I I literally think um a good segment of the population, um, probably a good segment, a higher percentage of people listening to this show than the general population, but it'll be very hard to produce the same work.
00:18:59.920 --> 00:19:02.400
And this actually ties into one of the sins from later.
00:19:02.480 --> 00:19:09.599
But essentially, AI removes all the false starts, the debugging and the rewriting that actually builds lasting professional skills.
00:19:09.759 --> 00:19:10.400
That's the thing.
00:19:10.640 --> 00:19:15.519
You learn through failure, you learn through, I think, right?
00:19:15.680 --> 00:19:19.039
When you talk about your domain expertise, you learn through failure.
00:19:19.119 --> 00:19:23.039
You learn through going the wrong way and getting lost and finding your way back, right?
00:19:23.119 --> 00:19:28.319
If you know how to use AI, you don't run into those failures as much.
00:19:28.640 --> 00:19:33.359
You don't run into those misdirections as much, right?
00:19:33.519 --> 00:19:35.839
It's it's it's it's like going to the gym, right?
00:19:35.920 --> 00:19:37.519
To actually put on muscle.
00:19:37.680 --> 00:19:39.279
You have to rip and tear something.
00:19:39.440 --> 00:19:41.920
You're not doing that for the most part, right?
00:19:42.079 --> 00:19:47.839
At least if your job role hasn't changed, but your AI access and your AI skills have gotten better.
00:19:48.160 --> 00:19:51.599
Probably a good chance you may be coasting, right?
00:19:51.759 --> 00:20:02.000
Uh a study found uh from Anthropic found that developers who used AI assistance scored 17% lower on coding quizzes than hand coders, right?
00:20:02.160 --> 00:20:09.599
There's so many stories out here on just kind of um just the cognitive decline, the more and more that we use AI.
00:20:09.759 --> 00:20:14.799
And I think even the definition of human intelligence is going to greatly change.
00:20:14.880 --> 00:20:20.079
Um, and the the story around human intelligence is greatly going to change over the next 10 years.
00:20:20.240 --> 00:20:27.359
And the biggest performance gap appeared in that study in debugging the exact skill that you need most when AI breaks.
00:20:28.319 --> 00:20:32.960
I think of the the GPS versus using your brain, right?
00:20:33.680 --> 00:20:36.079
It's funny, maybe.
00:20:36.319 --> 00:20:42.240
Um my wife can get around really well in Chicago without a GPS, right?
00:20:42.480 --> 00:20:46.559
Um, you know, she said her dad um taught her, right?
00:20:46.640 --> 00:20:50.559
Like, you know, oh, the sun's setting this way, the you know, this building here.
00:20:50.880 --> 00:20:52.880
I'm I'm terrible at it, right?
00:20:53.039 --> 00:20:55.200
But every once in a while, right, it's funny.
00:20:55.279 --> 00:20:57.440
She's like, no, I'm I'm turning off the GPS.
00:20:57.599 --> 00:20:58.880
Like, I should know how to get home.
00:20:59.039 --> 00:21:00.720
But Chicago can be confusing, right?
00:21:00.799 --> 00:21:09.359
I spend so much time in, you know, my little neighborhood when I'm or when I'm downtown or Mag Mile and the streets start looking the same, right?
00:21:09.519 --> 00:21:16.799
Um, if you live in a big city, maybe you know what I mean, but GPS kind of killed many people's sense of direction.
00:21:16.880 --> 00:21:29.359
And I think AI is killing people's uh or will kill people's eventual um sense of domain expertise, which is kind of weird and and scary.
00:21:29.519 --> 00:21:30.000
Um, right.
00:21:30.079 --> 00:21:36.400
And it's the identical erosion that's happening right now to your writing, analysis, and strategic reasoning.
00:21:36.640 --> 00:21:38.400
So here's how to protect yourself.
00:21:38.559 --> 00:21:44.799
You need to pick one core professional skill each week and complete it entirely without AI.
00:21:44.960 --> 00:21:48.400
This is something I still adamantly do often, right?
00:21:48.480 --> 00:21:53.359
In the same way my wife will turn off the GPS and drive home or drive to whatever destination.
00:21:53.680 --> 00:21:55.680
I will shut off AI.
00:21:55.839 --> 00:21:58.559
I will, I used to fully shut off the internet.
00:21:58.799 --> 00:21:59.759
I don't really do that anymore.
00:22:00.240 --> 00:22:02.720
But sometimes I still do, just to write, right?
00:22:02.799 --> 00:22:04.160
I'm a former journalist.
00:22:04.400 --> 00:22:05.359
I love writing.
00:22:05.519 --> 00:22:06.720
I enjoy writing, right?
00:22:06.799 --> 00:22:10.799
But AI writes a lot of first drafts for everything, right?
00:22:10.960 --> 00:22:14.720
Proposals for updates to website content, right?
00:22:14.880 --> 00:22:16.000
Whatever it may be.
00:22:16.559 --> 00:22:17.599
I'm a writer at heart.
00:22:17.680 --> 00:22:18.559
I used to be, right?
00:22:18.640 --> 00:22:22.880
So so many times I just blank page, no AI.
00:22:23.519 --> 00:22:24.640
Practice this skill.
00:22:24.799 --> 00:22:26.480
And it's important to do that.
00:22:26.799 --> 00:22:36.319
So that is your safeguard against what I think is probably one of the more serious and one of the more uh one of the most fastest accelerating uh sin that's gonna hold you back.
00:22:36.480 --> 00:22:36.799
All right.
00:22:36.960 --> 00:22:37.759
Number five.
00:22:38.000 --> 00:22:41.200
All right, another um, another one I made up here.
00:22:41.279 --> 00:22:42.799
Uh hopefully this one resonates.
00:22:42.960 --> 00:22:48.319
I'm calling this the agent bun sandwich that's gonna hollow out your core expertise.
00:22:48.480 --> 00:22:54.240
So kind of related to uh the de-skilling, uh the de-skilling, right?
00:22:54.400 --> 00:22:59.119
So as the de skilling happens, the agent bun grows, all right.
00:22:59.200 --> 00:23:02.319
And this again, remember, this is when you're doing AI right.
00:23:02.799 --> 00:23:12.000
This is a byproduct of staying ahead of the curve and doing things correctly, right?
00:23:13.599 --> 00:23:15.920
So, what the heck is an agent bun sandwich?
00:23:17.440 --> 00:23:21.680
Think of that fat, juicy cheeseburger you had last, right?
00:23:22.079 --> 00:23:23.839
I'm excited, it's getting warm.
00:23:24.000 --> 00:23:33.680
Uh, you know, might have to clean up the uh clean up the grill, but just think of that, not the kind you get at the at you know, at the uh culvers, right?
00:23:33.759 --> 00:23:34.400
I'll throw that out there.
00:23:34.559 --> 00:23:38.880
Culver's or whatever your fast, I don't, I don't know, uh live stream podcast audience.
00:23:38.960 --> 00:23:40.480
What's what's your favorite burger joint?
00:23:40.640 --> 00:23:45.440
All right, not those thin patties, like I'm talking like a thick, thick, thick burger.
00:23:45.599 --> 00:23:51.359
That that burger, that the the meat, that's that's your domain expertise, right?
00:23:52.240 --> 00:23:53.519
It's big, it's fat.
00:23:54.000 --> 00:23:59.920
Um now think of making a bun sandwich.
00:24:01.200 --> 00:24:02.319
Here's what I mean.
00:24:03.119 --> 00:24:12.640
Right now, when you have that big juicy burger, the meat of what you do, the majority of what you're doing right now, even augmented by AI, right?
00:24:13.119 --> 00:24:13.920
It's that burger.
00:24:14.480 --> 00:24:15.759
It's that big thick burger.
00:24:15.920 --> 00:24:23.920
Yeah, you might use a little AI, that's the bun right now, but it's turning into an agent bun, right?
00:24:24.079 --> 00:24:28.319
All of a sudden, that patty is gonna be so thin you can barely see it.
00:24:28.480 --> 00:24:32.079
And that patty is the domain expertise, like I said.
00:24:32.400 --> 00:24:35.920
It's going to be shrinking month by month.
00:24:36.079 --> 00:24:40.079
And I think that shrink, uh, right, just it's it's shrinkflation, right?
00:24:40.160 --> 00:24:40.720
Like everything else.
00:24:40.799 --> 00:24:41.920
Where's all my Doritos?
00:24:42.079 --> 00:24:44.400
There's only eight Doritos in this bag and it costs$7.
00:24:45.119 --> 00:24:45.759
Um, right?
00:24:45.920 --> 00:24:46.880
Shrinkflation.
00:24:47.200 --> 00:24:56.799
Uh, but I think domain expertise and how we're actually letting it play out in the workplace is shrinking.
00:24:56.960 --> 00:25:00.559
And it's gonna be more the bun, right?
00:25:00.799 --> 00:25:07.359
The front end and the back end of what you do is gonna be directing agents, right?
00:25:07.519 --> 00:25:13.279
And your actual domain expertise, the meat of your work is going to shrink.
00:25:14.000 --> 00:25:24.960
Because most of what we're gonna be doing now, and I see this because this is most of what I'm doing, is giving front-end direction to agents and then back end corrections, and that's it, right?
00:25:25.039 --> 00:25:29.200
So even think of something like marketing, right?
00:25:29.599 --> 00:25:32.720
I did a lot of marketing in in my career.
00:25:32.960 --> 00:25:36.079
A lot of it is giving front-end direction to different agents.
00:25:36.319 --> 00:25:44.640
Agents are creating the marketing collateral, they can create great video, they can create great designs, they can create full campaigns.
00:25:44.880 --> 00:25:46.960
And then I'm just checking on the back end, right?
00:25:47.039 --> 00:25:50.000
I'm like, all right, cool, sweet, good job.
00:25:50.160 --> 00:25:50.799
Uh, right.
00:25:51.039 --> 00:25:57.519
And AI right now is hollowing out those entry-level tasks that have always been the training ground for judgment.
00:25:57.680 --> 00:25:58.400
Um, right.
00:25:58.799 --> 00:26:02.559
I like to also think it's it's kind of like the teacher and the student relationship, right?
00:26:02.799 --> 00:26:07.759
Teacher being the bun, student being the uh the meat, the burger.
00:26:08.240 --> 00:26:17.119
But the difference here, and in very overly agentic future, is the teacher is just assigning like an agent, right?
00:26:17.279 --> 00:26:22.240
But eventually they're gonna forget and stop practicing the meat, right?
00:26:22.480 --> 00:26:24.559
Stop doing the work that the students are doing.
00:26:24.720 --> 00:26:28.880
And all of a sudden, the teacher just has the assignments in the rubric.
00:26:29.039 --> 00:26:38.240
And the further that they get removed from their domain expertise, right, the more useless that that practice feels, right?
00:26:38.400 --> 00:26:49.759
A lot of this is um, you know, it's it's it's almost like a funeral for the way that we've been working for 50, 60 plus years in the information age, right?
00:26:50.000 --> 00:27:01.279
You've gotten paid and promoted by the domain expertise that you kind of memorize and then how you create new value for your company with that domain expertise, right?
00:27:01.519 --> 00:27:15.039
Obviously, it's you know, playing the politics and all of those things, but for the most part, if you know a certain um subject better than 99% of the people in your field, that's what you were rewarded for because you could keep that in your brain.
00:27:15.200 --> 00:27:18.400
That's what large language models do, and they're better than all of us.
00:27:18.720 --> 00:27:20.160
Agent bun sandwich.
00:27:20.400 --> 00:27:25.839
So when AI fails, though, the bun sandwich teams cannot perform without it.
00:27:26.240 --> 00:27:37.759
Okay, so we talked about this on a previous show, but young workers aged 22 to 25 right now in AI exposed jobs saw a 13% employment decline since Chat GPT launched.
00:27:38.000 --> 00:27:39.599
What do I mean by that?
00:27:40.079 --> 00:27:49.440
Even the opportunity to go out and develop that domain expertise is going away.
00:27:49.599 --> 00:27:57.440
If that's not telling, and again, this might be a two-year, a five-year, a 10-year, I don't know.
00:27:57.599 --> 00:27:59.759
It could even be less than that.
00:28:00.000 --> 00:28:06.960
But you have to prepare, whether you want to or not, to work in an agent bun sandwich, right?
00:28:07.119 --> 00:28:09.599
Front end, back end, and that's really it.
00:28:09.680 --> 00:28:16.880
And you're not gonna be practicing that domain expertise a whole ton because that's what your company is gonna be paying for large language models to do.
00:28:17.039 --> 00:28:23.279
And it's that collapsing pipeline that means fewer humans are gonna be building the expertise needed to even evaluate the AI output.
00:28:23.440 --> 00:28:25.039
So here's the fix.
00:28:25.200 --> 00:28:37.359
Similarly to how you got to turn off the internet, uh, right, you need to keep one piece of the middle meat each week by completing one substantive uh big task end-to-end by yourself.
00:28:37.519 --> 00:28:38.799
What do I mean by that?
00:28:39.119 --> 00:28:40.799
Not just writing, right?
00:28:40.960 --> 00:28:41.680
That was my example.
00:28:41.759 --> 00:28:44.559
That was one piece, whole project, right?
00:28:44.799 --> 00:28:50.880
A project that maybe you would use three, four, five, six different AI tools and would take half a day, right?
00:28:50.960 --> 00:28:52.240
Not just writing a little email.
00:28:52.319 --> 00:28:54.079
Oh, I'm gonna turn off my co-pilot.
00:28:54.160 --> 00:28:56.160
I can make this sound more professional by myself.
00:28:56.319 --> 00:28:56.799
There we go.
00:28:56.880 --> 00:28:58.480
I'm doing what Jordan told me to.
00:28:58.640 --> 00:29:00.400
No, whole project, right?
00:29:00.480 --> 00:29:11.279
Where you have to practice uh that domain expertise, where you have to get lost, where you have to get things wrong without having to rely on AI to make it better or to get you there faster.
00:29:11.599 --> 00:29:17.200
All right, sin six, the compression tax makes your brain pay for AI speed.
00:29:17.279 --> 00:29:18.079
This one is huge.
00:29:18.240 --> 00:29:20.960
This is one I feel all the time.
00:29:21.200 --> 00:29:27.440
And I haven't really seen anyone talk about it until recently, although there are some studies that show a little bit.
00:29:27.599 --> 00:29:35.440
But if you are someone that uses AI for 10 to 12 hours a day and you've been doing it for multiple years, you're probably going to resonate with this.
00:29:35.519 --> 00:29:42.720
But essentially, AI now compresses a week worth of research into minutes, but your brain still processes at human speed.
00:29:42.880 --> 00:29:50.640
Like my analogy for this is you know, think pre-Chat GPT, uh, or just pre-large language models.
00:29:50.960 --> 00:29:55.759
If you were working on uh, you know, bringing a new product to market, right?
00:29:55.839 --> 00:29:58.720
Let's just say that common thing that most of us can understand.
00:30:00.160 --> 00:30:02.160
This is how you would get good at your job.
00:30:02.240 --> 00:30:08.640
This is how your uh company would get better, uh, your your department would be stronger, you would go out, you would research things, right?
00:30:08.720 --> 00:30:11.440
You'd go on the internet, you take notes, you talk to people about it.
00:30:11.599 --> 00:30:14.400
You know, maybe you have a colleague that works in a similar industry, right?
00:30:14.559 --> 00:30:15.759
Those are your reps.
00:30:16.480 --> 00:30:21.440
Um, and it's through those reps that that knowledge sticks.
00:30:22.799 --> 00:30:34.480
With AI, my personal experience with this is I read and consume probably a sickening amount of information.
00:30:35.519 --> 00:30:38.640
The human brain cannot do this.
00:30:38.960 --> 00:30:43.119
That's why some days, literally, some days, 10 a.m.
00:30:43.359 --> 00:30:44.079
I kid you not.
00:30:44.319 --> 00:30:56.559
It's not like it's not like I feel sick, but I feel like I I've already run an Iron Man and have competed in a I don't know, a chess tournament against world grand champions, right?
00:30:56.640 --> 00:31:03.920
You ever see those things where you know they burn like 10,000 calories an hour, right, when they're in these chess mashes or whatever, right?
00:31:04.079 --> 00:31:05.279
That's what I feel sometimes.
00:31:05.440 --> 00:31:09.440
If I'm doing heavy AI projects by by by 10 a.m.
00:31:09.920 --> 00:31:16.559
And it it feels like I've been working for 30 hours straight and I look, I'm like, it's it's 10 a.m., right?
00:31:17.039 --> 00:31:17.839
It's 10.
00:31:18.000 --> 00:31:19.759
My days barely started.
00:31:19.920 --> 00:31:26.880
That's because AI gives you these superhuman abilities that the human brain, I don't think, can handle.
00:31:27.039 --> 00:31:32.319
And the gap between delivery speed and comprehension speed is the compression tax that you pay daily, right?
00:31:32.400 --> 00:31:44.240
There was a BCG from the uh so uh Boston Consulting Group study that saw uh that high oversight AI work caused 19% more information overload among surveyed workers, right?
00:31:44.319 --> 00:31:51.839
So it's a small percentage, but I would love to see this broken down uh among people who are literally like token maxing, right?
00:31:51.920 --> 00:31:55.359
Just using AI all day, every day, like myself.
00:31:55.519 --> 00:32:00.799
And then to study the cognitive pressure in the compression tax, I would say it's huge, right?
00:32:00.880 --> 00:32:07.359
But workers in the study described mental fog, headaches, and the strange sense their thinking felt crowded.
00:32:07.519 --> 00:32:09.359
That is me all day.
00:32:09.519 --> 00:32:09.920
All right.
00:32:10.079 --> 00:32:14.240
So, like as an example though, you're like, okay, well, why are you working so much?
00:32:14.319 --> 00:32:16.799
If AI makes you so much smarter, well, I don't know.
00:32:16.960 --> 00:32:23.599
Maybe it's it's people's work personality, maybe it's the corporate culture that your you know business works in.
00:32:23.759 --> 00:32:28.480
But a lot of people are just like, okay, well, I'm doing this thing twice as fast as I was before.
00:32:28.640 --> 00:32:30.640
So it's not like I have 50% more time.
00:32:30.720 --> 00:32:33.200
I'm actually just doing twice as much now, right?
00:32:33.279 --> 00:32:36.000
Because you know, my company is just giving me twice as much.
00:32:36.240 --> 00:32:39.200
For me, I'm an entrepreneur, right?
00:32:39.359 --> 00:32:41.119
I'm I'm a business owner.
00:32:41.680 --> 00:32:52.559
If I can do my normal work in in an hour of what it used to take me three years ago, six hours to do, I didn't gain five hours of rest and relaxation.
00:32:53.440 --> 00:33:00.240
I added five hours of more AI augmented work to my plate.
00:33:00.559 --> 00:33:01.279
Um, right.
00:33:02.079 --> 00:33:10.640
So workers in that study initially felt a surge of momentum, but by month six, burnout and decision paralysis spiked.
00:33:10.799 --> 00:33:14.160
Um, and I I I have talked about this a couple of times.
00:33:14.400 --> 00:33:24.400
Um, and and this is maybe embarrassing to admit, but I will say that I probably forget as much information as I retain.
00:33:25.039 --> 00:33:25.519
Right.
00:33:25.680 --> 00:33:27.759
And so many times, right?
00:33:27.839 --> 00:33:34.160
I'll you know, just be chatting with someone in the back, oh, I loved your episode, you know, last week, and be like, oh, which one was it?
00:33:34.240 --> 00:33:38.400
And they'll say it in my head, I'll be like, I don't remember that.
00:33:38.559 --> 00:33:40.240
I literally don't remember that.
00:33:40.480 --> 00:33:44.799
Um and when I'm learning things now, right?
00:33:44.960 --> 00:33:48.079
Looking at things, you know, on perplexity or Google AI mode, right?
00:33:48.240 --> 00:33:51.519
I'm looking something up and I'll be looking something up, honestly, y'all.
00:33:51.599 --> 00:33:53.119
And I'm like, I have no clue what this is.
00:33:53.200 --> 00:33:54.880
I'm so excited to learn, right?
00:33:54.960 --> 00:33:56.319
This, this, this concept.
00:33:56.400 --> 00:33:58.640
Uh let me find the answer.
00:34:00.240 --> 00:34:04.480
My website came out of my mouth so many times, right?
00:34:04.559 --> 00:34:06.160
It used to happen sparingly.
00:34:06.319 --> 00:34:09.360
Now it happens almost almost twice a week.
00:34:09.679 --> 00:34:11.360
Where I want to go find something out.
00:34:11.599 --> 00:34:13.199
I'm like, oh, I don't know the answer to this.
00:34:13.360 --> 00:34:14.880
And it's like, oh, I said it.
00:34:15.039 --> 00:34:20.639
I said it two weeks ago, I said it two months ago, I said it two quarters ago, I said it two years ago, right?
00:34:22.400 --> 00:34:29.440
The tax that you pay from carrying a heavy AI load can be a burden.
00:34:29.679 --> 00:34:30.079
All right.
00:34:30.239 --> 00:34:33.840
Last but not least, sin number seven, automation bias.
00:34:34.079 --> 00:34:37.360
This builds a trust shortcut you never approved.
00:34:37.440 --> 00:34:38.320
Here's what that means.
00:34:38.480 --> 00:34:40.639
All right, if you're doing the AI the right way, right?
00:34:40.719 --> 00:34:47.360
If you're doing AI the right way, there's no way to like literally check every single fact that you get from an AI output, right?
00:34:47.519 --> 00:34:52.880
Because then it would technically take you more time to do that than it would to just do it from scratch, the old-fashioned way, right?
00:34:52.960 --> 00:34:56.320
But you know, when you talk about, oh, the human in the loop, right?
00:34:56.559 --> 00:35:03.280
What most people do is they they they look at maybe, you know, one or two facts here, they scan it over and they're like, yep, this is good, right?
00:35:03.519 --> 00:35:17.599
Um, but one AI confirmation of a fact, stat, or trend does not make the whole thing because unlike trusting a coworker, you transfer AI trust to every tool, even ones that you have never verified.
00:35:18.000 --> 00:35:28.800
So uh there was a study uh that showed participants followed AI advice even when it contradicted their own assessment in available evidence.
00:35:28.880 --> 00:35:29.360
Uh, right.
00:35:29.440 --> 00:35:37.519
And the most dangerous apart is humans absorb that AI bias and they keep it even after they stop using the tool, right?
00:35:37.679 --> 00:35:46.320
And then unfortunately, they also blindly reproduce any hallucinations or misinformation or waifs out there uh because they see it as fact, right?
00:35:46.400 --> 00:35:50.000
You see all these studies now that people are starting to talk like AI, right?
00:35:50.159 --> 00:35:55.280
Luckily, I'm not saying delve uh, you know, all the time on the air, but it's it's happening.
00:35:55.440 --> 00:36:00.400
And I do think that the automation bias, right, when you start to just, oh yeah, this is true, this is true.
00:36:00.559 --> 00:36:04.000
So then you just blindly rubber stamp anything that comes from an AI.
00:36:04.079 --> 00:36:06.159
And I'm not just talking about a chat bot, right?
00:36:06.320 --> 00:36:08.960
Um, as an example, there was the iTutor Group.
00:36:09.119 --> 00:36:21.440
Uh, their AI systems is uh a company, uh, their AI system automatically rejected women over 55 and men over 60 from every open role, uh, reportedly, right?
00:36:21.519 --> 00:36:23.599
So now I believe they're facing a lawsuit.
00:36:23.760 --> 00:36:33.280
Um, so over 200 uh qualified candidates were disqualified solely on age before a single human noticed, right?
00:36:33.360 --> 00:36:38.800
So those humans, they didn't build this tool, but this is the automation bias.
00:36:38.960 --> 00:36:42.079
Someone comes in, or maybe you check out one thing, you're like, okay, cool.
00:36:42.159 --> 00:36:45.840
This, hey, this tool gave me the right candidate or whatever it may be.
00:36:46.000 --> 00:36:54.079
So think you, you know, your CRM, your accounting software, uh, right, all of these different tools that you use that have AI in it.
00:36:54.239 --> 00:37:02.719
And just because you can verify the outputs that it gives you as right, it doesn't mean that it maybe took a very wrong path to get there, right?
00:37:02.880 --> 00:37:04.880
And that is the automation trust.
00:37:05.039 --> 00:37:22.639
So here's how you break it you need to verify at least one AI output manually every day and ask the vendors, especially um, right, the vendors, because if you're using a, you know, a large language model where you can trace its thinking, uh, you know, look at its chain of thought, you can do a lot of that work yourself.
00:37:22.719 --> 00:37:31.039
And, you know, with good custom instructions and with good um, you know, context engineering, you can hopefully steer clear of some of that automation bias.
00:37:31.199 --> 00:37:37.360
But if you're using a third-party tool, you need to ask the vendor, where is AI making decisions I cannot see?
00:37:37.599 --> 00:37:47.360
And how can you uh provide um traceability and observability to those decisions that are being made that we aren't seeing, right?
00:37:47.519 --> 00:37:54.159
There's a lot, you know, you put in A, you get Z, but what the heck happened from B to Y, right?
00:37:54.320 --> 00:37:59.679
That's what you need to be investigating, especially if you're using in a high value sector, right?
00:37:59.840 --> 00:38:04.079
Finance, healthcare, uh, anything with with loans, right?
00:38:04.239 --> 00:38:06.400
You have to be vigilant to those.
00:38:06.639 --> 00:38:06.960
All right.
00:38:07.199 --> 00:38:12.159
So um those are our seven cents, our seven silent cents.
00:38:12.320 --> 00:38:14.719
But this is bigger than that.
00:38:14.880 --> 00:38:21.760
Um, because this is also, I think a lot of these things are wrestling with just using AI, right?
00:38:21.920 --> 00:38:32.400
And if you picked up the trend, I did say there, all of these things come from using AI correctly and doing more faster, right?
00:38:32.480 --> 00:38:34.159
And it seems like that's the overall trend.
00:38:34.239 --> 00:38:39.039
When you see how good these large language models are, and you're like, oh my gosh, I can automate all this work.
00:38:39.199 --> 00:38:42.320
Yeah, some people are pocketing it and living good lives.
00:38:42.400 --> 00:38:46.159
And then there's those of us like me who are just working more and more and more.
00:38:46.320 --> 00:38:54.159
And what that means is to put it bluntly, many of us are getting dumber.
00:38:54.800 --> 00:38:55.440
Right.
00:38:55.599 --> 00:39:00.320
Um, but also it completely changes our relationship with domain expertise.
00:39:00.400 --> 00:39:10.239
And I think this is gonna be a weird thing for a lot of people, but you are gonna have to unmarry yourself to what you've been probably defining yourself for maybe decades, right?
00:39:10.400 --> 00:39:17.519
I think that domain experience has defined, especially people who are mid-career, is defined you not as your just your career, but as a person, right?
00:39:17.599 --> 00:39:22.480
Because you've developed a deep personal identity with what your job is, right?
00:39:22.639 --> 00:39:32.000
And I think that that relationship is shifting from a deep mastery of your craft to a wide transactional orchestration of AI agents, right?
00:39:32.239 --> 00:39:37.039
Um, that relationship is completely changing.
00:39:37.199 --> 00:39:38.239
Um, right.
00:39:38.559 --> 00:39:47.760
We used to be very tied to our domain expertise, but like I said, we're gonna all be living in agent agent bun sandwich world.
00:39:47.920 --> 00:39:48.400
Um, right.
00:39:48.480 --> 00:39:58.079
So you you're gonna have to kind of grieve the version of you where you worked really hard to learn all of these things, and it's like doesn't matter anymore, right?
00:39:58.559 --> 00:40:06.559
Um, you know, it doesn't matter that, oh my gosh, Bill, Bill in accounting, he knows everything about you know logistics.
00:40:06.960 --> 00:40:11.679
Uh, you know, and he's he was our expert for our very niche area of logistics.
00:40:11.840 --> 00:40:12.159
Guess what?
00:40:12.239 --> 00:40:14.000
The large language model is better than Bill.
00:40:14.239 --> 00:40:15.039
Sorry, Bill.
00:40:15.280 --> 00:40:22.400
You know, some company paid, you know, a hundred bills, a hundred dollars an hour, and now the large language model is better than a hundred bills.
00:40:22.559 --> 00:40:22.960
All right.
00:40:23.039 --> 00:40:28.639
Uh, but no one in leadership is preparing people for what that transition actually feels like, right?
00:40:28.800 --> 00:40:34.000
And this is kind of what leads to these unintentional silent AI sins that we're committing because of this.
00:40:34.559 --> 00:40:55.920
So I want to leave on a positive optimist note, but I think this is actually an important uh addition to our start here series because if you're doing AI right, these are the natural byproducts that are technically harming you, um, harming your professional abilities and maybe hurting your company as well.
00:40:56.079 --> 00:41:01.760
But every silent sin rewards speed today while quietly stealing your capability for tomorrow.
00:41:01.920 --> 00:41:05.760
So you have to argue the opposite position to catch sycophancy.
00:41:06.000 --> 00:41:07.920
You have to be vigilant against those things.
00:41:08.000 --> 00:41:09.199
That's number one, right?
00:41:09.280 --> 00:41:13.440
But then also ask who funded this to catch those waves that I talked about earlier.
00:41:13.760 --> 00:41:19.280
Do at least one AI-free task weekly, a big one, the full meat.
00:41:19.440 --> 00:41:24.239
Don't let your quarter pounder turn into that invisible patty.
00:41:24.320 --> 00:41:30.639
And then verify every output that you can reliably to break any automation bias.
00:41:30.800 --> 00:41:38.079
Because the professionals who thrive will not just be the fastest with AI, but those that are the sharpest without it.
00:41:38.239 --> 00:41:41.039
So, like I said, kind of like you're offline, right?
00:41:41.119 --> 00:41:43.039
What happens if AI goes off?
00:41:43.199 --> 00:41:43.440
Right?
00:41:43.599 --> 00:41:46.480
What happens if you know someone clicks the off button?
00:41:46.800 --> 00:41:48.639
Can you still do your job?
00:41:49.119 --> 00:41:58.800
Well, you still need to build those skills because if you do, if you become sharper with your off AI skills, well, you are gonna be pretty much unstoppable with AI.
00:41:58.960 --> 00:41:59.280
All right.
00:41:59.360 --> 00:42:00.559
I hope this one was helpful.
00:42:00.719 --> 00:42:02.480
If so, let me know about it.
00:42:02.559 --> 00:42:04.239
Share if you're listening on LinkedIn.
00:42:04.320 --> 00:42:05.519
I'd really appreciate that.
00:42:05.599 --> 00:42:08.400
Um, and then go to starthirseries.com.
00:42:08.480 --> 00:42:15.199
So we're gonna have a recap of today's show there, but that also gives you free access to our inner circle community.
00:42:15.360 --> 00:42:23.519
You can go listen to and read every single uh episode in this series in order in that space in our community.
00:42:23.679 --> 00:42:24.639
So thank you for tuning in.
00:42:24.719 --> 00:42:27.679
I hope to see you back tomorrow and every day for more everyday AI.
00:42:27.760 --> 00:42:28.880
Thanks, y'all.
