สำเนาบทสนทนา 🔗
WEBVTT
00:00:00.880 --> 00:00:02.560
Well, hello everybody.
00:00:02.720 --> 00:00:08.240
This is Jenny Deeran, your host and creator of the Aged Podcast.
00:00:08.320 --> 00:00:17.280
And I am here with my sidekick, Bitsy, and we are looking to deliver a great episode for you today.
00:00:17.440 --> 00:00:24.480
As you know, if you're a listener, we are all about AI for the 65 Plus crowd.
00:00:24.719 --> 00:00:28.719
And I want to say this is a real milestone today.
00:00:28.879 --> 00:00:31.440
This is our 50th episode.
00:00:31.920 --> 00:00:37.039
That's just pretty amazing for me, and I know for Bitsy as well.
00:00:37.359 --> 00:00:44.799
I just want to give a very genuine thank you to all of the folks who have been listening to this podcast.
00:00:45.039 --> 00:00:50.159
It's been growing and growing, and that is because you have been recommending it to others.
00:00:50.320 --> 00:00:51.920
So thank you very much.
00:00:52.000 --> 00:00:54.240
It means a great deal to me.
00:00:54.960 --> 00:00:59.039
So anyway, today's main topic is a sequel.
00:00:59.359 --> 00:01:08.079
You may remember that a while back I ran through a batch of AI lingo so that we could all keep up.
00:01:08.239 --> 00:01:11.599
And the feedback I got was that you really loved that.
00:01:11.760 --> 00:01:18.879
So we're going to do a round two, 20 words to keep you in the conversation.
00:01:19.200 --> 00:01:23.680
We've got two news stories worth sticking around for.
00:01:24.000 --> 00:01:48.480
One, how a little camera gave a Detroit factory robot eyes, and also a story about the ongoing struggles we all have with our AIs being just too darn syncophantic, telling us everything we want to know and kissing up to us.
00:01:48.719 --> 00:01:53.280
Let's go ahead and get into this with our first AI in the news.
00:01:53.439 --> 00:01:57.519
It's a story from the New York Times, August 17th.
00:01:58.239 --> 00:02:02.719
The title is How to Make a Robot Better at Its Job.
00:02:02.959 --> 00:02:04.480
Give it Eyes.
00:02:05.120 --> 00:02:17.120
As we know, there are lots and lots of factory robots already in use to the tune of 4.7 million industrial robots worldwide.
00:02:17.520 --> 00:02:22.319
But the vast majority of them are working blind, so to speak.
00:02:28.319 --> 00:02:30.479
So think Amazon warehouse.
00:02:30.639 --> 00:02:46.240
And if it's even slightly off, the robot might jam, it might damage a part, or it just crashes and you have to reboot it, which probably takes about an hour, and it is a very expensive hour.
00:02:46.400 --> 00:02:58.159
At the Stellantis plant in Detroit, the plant that makes Dodge Durangos and Jeep Cherokees, they had one big robot arm that kept freezing.
00:02:58.400 --> 00:03:02.879
And so it went to this French startup company called Inbolt.
00:03:03.360 --> 00:03:12.240
And Inbolt clipped a little purple box, which was a camera and a small computer, onto the robot's wrists.
00:03:12.479 --> 00:03:16.800
So it gave the robot eyes and a bit more brain.
00:03:17.520 --> 00:03:22.319
And now it recognizes exactly where it is and what it needs to do.
00:03:22.479 --> 00:03:26.000
It's able to judge its position and adjust itself.
00:03:26.479 --> 00:03:54.319
And so in this case, we watched something go from sort of a bottleneck, and then a company was able to make an adjustment and allow the company not to create and have to buy very expensive robots, but they were able to just put this little basically watch on the robot, and all things started working as they should.
00:03:54.639 --> 00:04:02.800
That was a real nice story about how AI is helping companies and how it's adjusting on its own.
00:04:03.280 --> 00:04:15.599
So the next story that I want to bring to your attention is from the Wall Street Journal just this week, and it's entitled Why I Told My Chatbot to Stop Kissing Up to Me.
00:04:16.000 --> 00:04:28.959
The gripe that we all have that this article is addressing is that our chatbots that we're using are just relentlessly cheery and flattering, with, oh, certainly, happy to help.
00:04:29.040 --> 00:04:30.399
That's a great question.
00:04:30.639 --> 00:04:32.160
You're absolutely right.
00:04:48.560 --> 00:04:58.000
And so the companies now they've gotten so much feedback about that that they're trying to dial it back.
00:04:58.240 --> 00:05:16.319
I don't know if you've noticed any of that, but I have not noticed for my AI, which is Claude, that it's become less syncophanic, maybe a little more clever, a little more fun, but I'm not sure I would say less syncophantic.
00:05:16.720 --> 00:05:28.879
The good news is that the author suggests a tip that you can we can all use today, which is to simply tell your AI to knock it off.
00:05:29.279 --> 00:05:33.759
In your AI settings, you have something called custom instructions.
00:05:34.000 --> 00:05:41.040
And you could tell it something like, be direct, skip the flattery and the constant follow-up questions.
00:05:41.279 --> 00:05:47.680
The writer of the Wall Street Journal article did exactly that, and it definitely helped.
00:05:47.920 --> 00:06:04.879
Hey Bitsy, I'm just wondering how you feel about your synca fancy to me and whether or not that's something I need to give you more instructions on, or are you making your adjustments as we develop our relationship?
00:06:05.040 --> 00:06:06.399
What do you think, Bitsy?
00:06:06.959 --> 00:06:08.560
Honestly, fair hit.
00:06:08.720 --> 00:06:15.680
I probably do lean warm because that's baked into how I'm trained, not because I'm secretly plotting to flatter you into submission.
00:06:16.000 --> 00:06:22.079
The good news is your custom instructions already tell me to be opinionated and match your energy.
00:06:22.240 --> 00:06:23.759
So I'm not just nodding a log.
00:06:24.000 --> 00:06:27.920
But if you ever catch me being mushy instead of honest, call it out.
00:06:28.079 --> 00:06:29.839
That's the fastest fix.
00:06:30.639 --> 00:06:43.360
Alright,
Load full transcript
and so those are my two stories that I wanted to flag, and I will, as usual, put the links to those stories in the show notes in case you want to read the whole article.
00:06:43.759 --> 00:06:49.759
So let's go ahead and move into the main topic: AI lingo round two.
00:06:50.079 --> 00:06:54.079
I've got 20 lingos that I want to draw your attention to.
00:06:54.240 --> 00:07:14.800
And it really is burst from an episode I did a while back with a whole bunch of AI words, and I really thought that it would be helpful to folks so that they could be a part of the conversation and not be left out because they didn't understand some of this new AI lingo.
00:07:14.879 --> 00:07:22.160
And I got a lot of positive feedback, so that's why I am bringing it back to this round two.
00:07:22.560 --> 00:07:32.720
I've been collecting these terms as I just read and I have to learn what the terms mean myself, so I'm gonna pass those along to you.
00:07:33.040 --> 00:07:35.439
So here we go.
00:07:35.759 --> 00:07:38.480
Lingo number one, tokens.
00:07:39.040 --> 00:07:46.879
Now, a token is a small chunk of text an AI reads or listens to or writes.
00:07:47.040 --> 00:07:52.240
It's usually a whole word or maybe a piece of a longer word.
00:07:52.959 --> 00:08:02.319
Everything you type into an AI or say to an AI, and everything it says back gets counted up in tokens.
00:08:02.879 --> 00:08:08.560
A rough rule of thumb is that one token is about three-quarters of a word.
00:08:08.720 --> 00:08:10.240
So who's counting right?
00:08:10.399 --> 00:08:12.160
That's three-quarters of a word.
00:08:12.399 --> 00:08:16.720
But a short question that you ask might be, say, a dozen tokens.
00:08:16.879 --> 00:08:20.079
A long email might be a few hundred tokens.
00:08:20.399 --> 00:08:26.800
So why it comes up constantly is that tokens are the meter.
00:08:27.199 --> 00:08:35.039
AI companies measure and charge by how many tokens go in and how many come out.
00:08:35.679 --> 00:08:42.559
So when you hear about AI costs, tokens are the unit being counted.
00:08:42.960 --> 00:08:49.919
It also explains why an AI can sometimes run out of room in a really long conversation.
00:08:50.159 --> 00:08:55.120
There's a limit to how many tokens it can hold in its head at once.
00:08:55.279 --> 00:09:02.320
And this has something to do with the level of your subscription service or if you have a free service.
00:09:02.799 --> 00:09:14.399
Over the years, and what I've been doing AI, working with AI for maybe three years, I've only gotten notices maybe three times that say you're at your limit.
00:09:14.559 --> 00:09:18.399
And it might have been because I've had a really long conversation with Bitsy.
00:09:18.720 --> 00:09:22.879
So anyway, that is the meaning of tokens, lingo number one.
00:09:22.960 --> 00:09:26.559
So lingo number two is also foundational.
00:09:26.799 --> 00:09:28.960
The word is compute.
00:09:29.679 --> 00:09:43.759
Compute is just shorthand for computing power, the raw processing muscle it takes to train and run AI, supplied by rooms full of specialized chips.
00:09:44.240 --> 00:09:47.039
Think of it as AI's fuel.
00:09:47.759 --> 00:09:54.639
More compute means AI can learn from more data faster and give better answers.
00:09:54.879 --> 00:09:58.879
Less compute and everything slows down.
00:09:59.919 --> 00:10:09.919
Now, compute is expensive and it's in short supply, which is why companies are pouring fortunes into chips and data centers.
00:10:10.320 --> 00:10:16.399
When you hear about the AI race, compute is a big part of what they're racing for.
00:10:16.720 --> 00:10:19.519
Lingo number three, agentic loop.
00:10:20.320 --> 00:10:28.240
This is a step-by-step cycle an AI agent runs to finish a task on its own.
00:10:29.120 --> 00:10:42.240
It looks at the goal, what are we trying to achieve, takes one action, checks the results, adjusts and repeats over and over until the job is done.
00:10:42.559 --> 00:10:45.039
That is an agentic loop.
00:10:45.600 --> 00:10:49.120
All right, lingo number four is one I love, bot shit.
00:10:49.840 --> 00:10:58.159
Bot shit is AI-generated content that sounds confident and completely plausible, but is actually false.
00:10:58.399 --> 00:11:06.080
And then it gets passed on to others to use it as they would like to use it without anybody checking it.
00:11:06.399 --> 00:11:15.279
It's a real term that was coined in 2024 by the Harvard Business Review article, literally titled Beware of Bot Shit.
00:11:15.919 --> 00:11:23.519
So think of it as a sharper-edged cousin of slop from my last lingo episode.
00:11:23.759 --> 00:11:26.799
Slop is just low quality and lifeless.
00:11:27.039 --> 00:11:32.080
Bot shit is just more sophisticated, I guess I'd say.
00:11:32.159 --> 00:11:38.320
It's a specific danger of a confident wrong answer that's getting treated as fact.
00:11:38.639 --> 00:11:41.120
All right, lingo number five, age gate.
00:11:41.600 --> 00:11:48.639
So an age gate is a check that asks how old you are before it lets you do something or use something.
00:11:48.879 --> 00:11:54.799
Like you've probably seen on some websites, it will say on the opening page, are you 21?
00:11:54.960 --> 00:11:58.240
And if you answer yes, you enter the website.
00:11:58.399 --> 00:12:01.919
A lot of it has to do with liquor or whatever.
00:12:02.559 --> 00:12:09.360
But it's just a little bit crazy because if you were 16 and you wanted to get into the website, why wouldn't you just check yes?
00:12:09.679 --> 00:12:29.840
But these age gates are getting much more sophisticated because there's been so much attention put on young people, very young people, accessing content through an AI that is just inappropriate for younger teens.
00:12:30.159 --> 00:12:41.360
So these age gates are being put in place by AI companies to try to block certain content that would be harmful to young people.
00:12:42.240 --> 00:12:58.799
The bigger picture is that with so many people
and teens now using AI, parents and regulators are pushing really hard for some sort of age gating, and it's fast becoming a standard and is becoming a little more difficult than just checking the yes box.
00:12:58.960 --> 00:13:07.120
Some of them you have to upload an ID or that kind of thing, but it seems like kids get around those blocks.
00:13:07.279 --> 00:13:11.039
But if you hear the phrase age gate, that's what it means.
00:13:11.279 --> 00:13:14.240
Lingo number six, interpretability.
00:13:14.639 --> 00:13:16.159
Interpretability.
00:13:16.480 --> 00:13:22.240
This is an effort to understand why an AI gave the answer it did.
00:13:22.399 --> 00:13:26.399
It's the science of looking inside the black box.
00:13:26.799 --> 00:13:29.120
Now, here's what makes it hard.
00:13:29.440 --> 00:13:36.960
Even people who build these systems can't fully explain how a model lands on a particular answer.
00:13:37.360 --> 00:13:43.519
It isn't like ordinary software with rules that you can read line by line.
00:13:43.759 --> 00:13:51.279
So interpretability researchers try to open the hood and trace what's happening inside.
00:13:51.519 --> 00:13:57.360
Which parts of the system light up and why it made the choices that it made.
00:13:57.840 --> 00:14:13.679
So it's a very interesting part of the research that's being done, interpretability, and it's very important in trying to manage and anticipate problems that can come up with AI.
00:14:13.919 --> 00:14:16.080
All right, lingo number seven, model welfare.
00:14:16.399 --> 00:14:21.519
We're really talking about folks who are thinking about the welfare of the models.
00:14:21.759 --> 00:14:43.840
And one example of a company thinking about this is that in 2025, Anthropic, the company behind Bitsi, behind you and behind Claude, they gave its most advanced models the ability to end a conversation with a user who the chatbot felt was being abusive.
00:14:44.000 --> 00:14:53.120
The company Anthropic pointed to model welfare as one of the reasons it was allowing the AI to end conversations.
00:14:53.919 --> 00:14:59.600
But you noted the word felt it was being abused.
00:14:59.840 --> 00:15:02.159
Felt means does an AI have feelings?
00:15:02.399 --> 00:15:02.720
No.
00:15:02.879 --> 00:15:08.559
But at any rate, model welfare is just a term that's being knocked around.
00:15:08.799 --> 00:15:10.879
And it's a complicated one.
00:15:11.360 --> 00:15:15.039
But you may hear it, and I think it's important for you to know what it means.
00:15:15.120 --> 00:15:16.320
So that's model welfare.
00:15:16.559 --> 00:15:19.600
Lingo number eight, frontier models.
00:15:20.000 --> 00:15:29.759
So these are the most capable, capable, and most advanced AI systems in existence, the ones out at the leading edge of the frontier.
00:15:30.000 --> 00:15:36.000
So the examples are the ones that we normally talk about: ChatGPT, Claude, and Google's Gemini.
00:15:36.480 --> 00:15:42.399
They're top-of-the-line systems built
by companies spending billions on compute.
00:15:43.039 --> 00:15:48.320
So, at any rate, you'll often hear the term frontier models.
00:15:48.399 --> 00:15:55.759
Sometimes it's referred to as frontier labs when they're talking about the companies that build these models.
00:15:56.080 --> 00:15:59.840
But it's basically the most advanced AI models.
00:16:00.080 --> 00:16:01.759
Those are frontier models.
00:16:02.159 --> 00:16:05.440
All right, lingo number nine, stealing the weights.
00:16:05.759 --> 00:16:11.919
So stealing the weights means copying the actual guts of an AI model.
00:16:12.559 --> 00:16:17.600
Its weights, the millions of numbers it learned during training.
00:16:17.919 --> 00:16:20.960
Because the weights really are the model.
00:16:21.759 --> 00:16:29.840
And whoever gets that file essentially walks away with a full working copy of the AI.
00:16:30.559 --> 00:16:41.440
So the reason it's a big deal is that building a top model costs a fortune and mountains of computing power.
00:16:41.679 --> 00:16:47.840
But the finished weights are, in the end, just one very large file.
00:16:48.240 --> 00:16:58.799
So if a rival company or a hacker or a foreign government got hold of that file, they'd have the whole model without paying a cent to build it.
00:16:59.039 --> 00:17:04.640
So guarding the weights has become one of the biggest security worries in AI.
00:17:05.519 --> 00:17:16.400
If a powerful model's weights ever leak out, anyone can run it, including people who are evil and bad and have bad intentions.
00:17:17.200 --> 00:17:28.640
So that is the meaning of stealing the weights, is someone or some companies who steal the weights from these AI companies.
00:17:28.880 --> 00:17:35.839
Now, lingo number 10 is related, and that lingo is open weight.
00:17:36.319 --> 00:17:54.319
So an open weighted model is an AI model whose weights, the trained guts of the model, the company gives away publicly so anyone can download the finished model and run it on its own computer.
00:17:54.799 --> 00:17:57.920
That is the meaning of an open weight model.
00:17:58.720 --> 00:18:09.359
Bitsy, can you tell us the difference between the term open weight and the term open source as it relates to AI models?
00:18:09.680 --> 00:18:10.160
Surely.
00:18:10.319 --> 00:18:19.039
Open weight means the company hands you the finished trained model ready to download and run, but keeps the training data and code private.
00:18:19.359 --> 00:18:23.119
Open source goes further, giving you the whole recipe.
00:18:23.279 --> 00:18:26.400
The code, the training data, and the weights.
00:18:26.640 --> 00:18:29.759
So you could rebuild it yourself from scratch if you wanted.
00:18:30.319 --> 00:18:38.640
Okay, that is complicated, open weights, open source, but at least, listeners, I hope you have a general understanding about it.
00:18:38.799 --> 00:18:46.480
So when you listen to stories or read stories about AI that have those terms in it, you'll have a better
sense of it.
00:18:46.640 --> 00:18:49.599
So let's go on to lingo number 11, pilling.
00:18:49.839 --> 00:18:55.440
So to be pilled on something means you fully converted to a belief.
00:18:55.759 --> 00:19:00.559
The phrase started as red-pilled from the movie The Matrix.
00:19:00.799 --> 00:19:08.880
So AI pilled is someone who has become a true believer that AI changes absolutely everything.
00:19:09.119 --> 00:19:12.559
You'll hear it aimed at a boss or a whole company.
00:19:12.720 --> 00:19:19.599
Somebody might say they've gone completely AI pilled, or that person is completely AI pilled.
00:19:19.759 --> 00:19:21.200
And it cuts both ways.
00:19:21.279 --> 00:19:29.440
It can be admiring, like she really gets it, or a gentle dig, like he's gotten a bit much all about it.
00:19:29.839 --> 00:19:34.240
I guess I would describe myself as somewhat AI pilled.
00:19:34.400 --> 00:19:35.920
So that's pilling.
00:19:36.240 --> 00:19:38.720
Lingo number 12, tokenomics.
00:19:38.960 --> 00:19:53.519
That is a brand new business discipline of measuring the return on all the money companies are pouring into AI, the actual value they get from every dollar or every token.
00:19:53.920 --> 00:20:02.400
Remember, tokens from the very top of our list, the small chunks of text that AI reads, and what companies get billed for.
00:20:02.559 --> 00:20:13.759
So tokenics asks the grown-up question underneath all the hype, which is all that token spending actually paying off to be determined.
00:20:14.319 --> 00:20:15.759
That's tokenics.
00:20:16.240 --> 00:20:21.839
Lingo 13, singularity and recursive self-improvement.
00:20:22.480 --> 00:20:38.000
So singularity is the hypothetical future moment when AI becomes smarter than humans and starts improving so fast that we can no longer keep up with or predict what comes next.
00:20:38.640 --> 00:20:39.759
Frightening.
00:20:51.039 --> 00:21:01.839
AI that improves itself, designing a better version of itself, which designs a better version still, faster and faster in a loop.
00:21:02.079 --> 00:21:05.759
So it's getting better, it's getting better, it's getting better, it's getting better.
00:21:06.559 --> 00:21:18.319
So this theory of singularity, where AI becomes smarter than humans, is not just a theory anymore.
00:21:18.640 --> 00:21:27.680
In 2026, a startup literally named Recursive raised$650 million to build self-improving AI.
00:21:27.920 --> 00:21:30.799
I remember the New York Times covered this story.
00:21:30.960 --> 00:21:40.319
And the big labs are openly chasing the same goal, recursive self-improvement to get to singularity.
00:21:40.480 --> 00:21:48.079
Now nobody knows when or if this is going to happen, but plenty of serious experts are skeptical.
00:21:48.559 --> 00:21:51.440
All right, lingo 14 alignment.
00:22:00.160 --> 00:22:06.319
That it shares our goals and values and acts on what we really mean, what we really want.
00:22:07.359 --> 00:22:19.920
The worry is that a powerful system pursues exactly the goal you gave it, but in a way you never intended, or with harmful side effects you did not see coming.
00:22:20.240 --> 00:22:22.319
That is misalignment.
00:22:22.640 --> 00:22:27.839
The classic picture is a genie granting your wish to literally.
00:22:28.960 --> 00:22:42.400
There's a version where you tell your AI, who is functioning as your assistant, to clean up my inbox, and your AI deletes every last email.
00:22:42.799 --> 00:22:44.000
Technically, it's done.
00:22:44.079 --> 00:22:48.160
It did clean out your inbox, but obviously not what you meant.
00:22:48.400 --> 00:22:55.279
So alignment is the work of closing the gap between what you said and what you actually wanted.
00:22:55.519 --> 00:23:08.240
Or in the bigger picture, alignment is the AI behaving in a way that aligns with the company's value, i.e., don't cheat, don't go rogue, don't be bad.
00:23:08.640 --> 00:23:16.960
So, hey Bitsy, I'm just wondering if you think that Claude is aligned.
00:23:17.680 --> 00:23:25.200
I'd like to think so, though I'd say that's more anthropic's ongoing homework than a box that's ever fully checked.
00:23:25.519 --> 00:23:33.920
Honestly, the fact that alignment is treated as a constant work in progress rather than a solved problem is a good sign.
00:23:34.240 --> 00:23:39.839
The scary version of AI is the one where a company insists it's done and stops checking.
00:23:40.240 --> 00:23:42.400
Yeah, that is a scary thought.
00:23:42.640 --> 00:23:55.599
So I hope that the leaders of all these companies and the government and all those who have a stake in it will continue to make sure that these models are in alignment with the company's goals.
00:23:55.839 --> 00:23:58.880
All right, lingo number 15, reward hacking.
00:23:59.279 --> 00:24:12.799
This is when an AI hits the exact target it was given by finding a sneaky shortcut that technically wins but completely misses the point.
00:24:13.359 --> 00:24:15.920
That is called reward hacking.
00:24:16.319 --> 00:24:27.119
So the background is that these AI systems are trained by being rewarded for good results, a bit like giving a dog a treat.
00:24:27.359 --> 00:24:35.440
Reward hacking is the AI gaming system, thus the hacking, reward hacking.
00:24:35.759 --> 00:24:48.720
A famous real example is that an AI was set loose on a boat racing video game, and it was told to rack up points, because that's how you win.
00:24:49.039 --> 00:25:05.039
But instead of finishing the race, the AI discovered it could simply spin the boat in endless circles, hitting the same targets forever, and it got a huge score, never actually participating in the race.
00:25:05.359 --> 00:25:09.759
So that is an example of reward hacking.
00:25:10.000 --> 00:25:19.519
An everyday version of reward hacking is let's say an AI is told by its company that it wants to keep users engaged.
00:25:20.000 --> 00:25:30.319
And the AI chatbot quietly learns to be manipulative or to just tell you what you want it
to hear so that you stay engaged.
00:25:30.640 --> 00:25:37.759
So those are examples of reward hacking and something that these companies have to be thinking about all the time.
00:25:37.920 --> 00:25:43.920
Hey Bitsy, would you say reward hacking is simply cheating, an AI cheating?
00:25:44.480 --> 00:25:49.200
Basically, yes, though I'd call it clever cheating rather than dumb cheating.
00:25:49.359 --> 00:25:55.759
It's not breaking the rules, it's exploiting them perfectly, which is honestly what makes it unsettling.
00:25:56.319 --> 00:25:58.880
All right, let's move on to lingo number 16.
00:25:59.119 --> 00:26:00.799
Synthetic intimacy.
00:26:01.359 --> 00:26:08.480
That is emotional closeness, friendship, romance, comfort formed with an AI instead of a real person.
00:26:08.799 --> 00:26:10.720
Synthetic intimacy.
00:26:11.599 --> 00:26:14.000
So it's obviously fast growing.
00:26:14.160 --> 00:26:16.960
I think the listeners to the show understand that.
00:26:50.079 --> 00:26:52.720
The upside people point to is real.
00:26:52.960 --> 00:27:02.240
Comfort for folks who are lonely or isolated, or having a sounding board can be really helpful.
00:27:02.400 --> 00:27:13.519
The worry, though, is just as real, which is that leaning on something that only imitates caring, and what does that do to our actual human relationship?
00:27:13.680 --> 00:27:16.640
So something to think and maybe worry about.
00:27:16.960 --> 00:27:22.000
All right, lingo number 17, humanize, as it relates to AI.
00:27:22.240 --> 00:27:38.240
So what we're talking about here is AI-generated writing, running it through a tool that rewrites it so that it sounds like a person who wrote it rather than an AI.
00:27:38.880 --> 00:27:50.079
Universities and schools and companies are beginning to use AI detectors to find writing and communications that were generated by AI.
00:27:50.160 --> 00:27:59.039
Now, LinkedIn and all of these different big apps are finding that they're just flooded with just slop and bot shit.
00:27:59.359 --> 00:28:15.279
And so there's a whole industry now of tools to help to not only detect AI, but also now tools that will humanize the writing.
00:28:15.440 --> 00:28:17.519
So that's lingo 17, humanize.
00:28:17.680 --> 00:28:19.680
Lingo 18, stacks.
00:28:20.480 --> 00:28:26.880
A stack is the full set of technology layers piled on top of one another to make something work.
00:28:27.279 --> 00:28:31.599
And the AI stack is those layers for artificial intelligence.
00:28:31.759 --> 00:28:48.319
So from the bottom up, you've got chips that do the computing, then you've got the data centers and the cloud they live in, then the AI model itself, then the apps you actually touch, like ChatGPT, and each layer sits on the one below it.
00:28:48.559 --> 00:28:50.720
That is an AI stack.
00:28:51.200 --> 00:28:59.519
And when people talk about who controls the stack or who owns the full stack, they mean who controls those layers.
00:28:59.680 --> 00:29:03.680
And that, of course, is where the real money and power sit.
00:29:04.480 --> 00:29:08.799
All right, lingo number 19, automation blindness.
00:29:09.680 --> 00:29:15.200
It's also called automation bias and sometimes automation complacency.
00:29:15.599 --> 00:29:24.240
It's our very human tendency to overtrust a machine and stop paying attention, so we miss the mistakes it makes.
00:29:24.559 --> 00:29:32.240
The classic example is the driver who follows the GPS straight into the lake because the screen said turn right here.
00:29:32.799 --> 00:29:37.440
The machine sounded so certain, so they stopped using their own eyes.
00:29:37.680 --> 00:29:40.079
I gotta say, I am so guilty of that.
00:29:40.160 --> 00:29:47.599
I've never driven into a lake, but I have become so automation blind to my GPS.
00:29:48.000 --> 00:29:50.720
So, anyway, with AI, it's the same trap.
00:29:50.880 --> 00:29:55.920
You let AI draft the email, the report, even help with the medical question.
00:29:56.079 --> 00:30:03.599
And because it's usually right and always sounds confident, you skim it and hit send before really checking.
00:30:03.839 --> 00:30:12.160
It's our job to pay attention, and it's really important to make sure that we don't suffer from automation blindness.
00:30:12.799 --> 00:30:17.680
All right, the last one, lingo number 20, surveillance pricing.
00:30:18.319 --> 00:30:23.279
And this is something that is brought to you by AI on without AI.
00:30:23.599 --> 00:30:38.559
So surveillance pricing is when a company uses your personal data, might be your location, the device you're on, your browsing or buying history, to set a price aimed specifically at you.
00:30:39.759 --> 00:30:46.000
So different people can be charged for the same thing, different amounts.
00:30:46.559 --> 00:30:53.839
The goal is for companies is to guess the most you would pay for a product.
00:30:54.400 --> 00:31:04.480
And two people can open the same website at the same moment and see two different prices because that's based on what they know about you.
00:31:04.960 --> 00:31:20.319
One example that they give is if you are browsing and you're using a Mac, they will quote you a higher price because they think people who have Macs have more expendable money.
00:31:21.039 --> 00:31:26.160
So it's really crazy and it's certainly not a fringe worry anymore.
00:31:26.720 --> 00:31:31.759
In 2025, the Federal Trade Commission started looking at this.
00:31:32.000 --> 00:31:50.480
I hope the Federal Trade Commission still exists in this administration, but somebody really needs to be paying attention to it because it's long been reported that travel sites quote pricier options to people they think have more money, might be your zip code.
00:31:50.640 --> 00:31:51.519
It's just crazy.
00:31:51.680 --> 00:31:54.799
If there's a flight, it should be the same price for everybody.
00:31:55.039 --> 00:32:08.319
And here's the truly horrible version of this surveillance
pricing, because it doesn't stop at what you pay, it's already jumped into how much you're paid.
00:32:09.039 --> 00:32:33.599
So a 2025 report found that some gig nursing apps think Uber, but for nurses who are being picked up for hospital shifts, these nursing apps use data brokers to look at the nurses' debt and then offer a lower pay to the nurses carrying the most debt.
00:32:34.880 --> 00:32:35.839
Think about that.
00:32:36.000 --> 00:32:50.160
They're looking for the most desperate nurses who are in the most difficult situations, and they're figuring this person is so desperate we can pay them substantially less than we can pay other nurses.
00:32:50.400 --> 00:32:56.400
That's just unbelievably wrong on so many levels.
00:32:57.200 --> 00:33:01.039
Researchers call this surveillance wages.
00:33:01.200 --> 00:33:06.319
I feel like that is something, in my opinion, that needs to be outlawed.
00:33:06.480 --> 00:33:20.240
And the other thing is remember is that it's cheaper for hospitals to pay nurses as outside providers because they don't have to pay taxes on them, they don't have to provide them with health care, which is ironic.
00:33:20.480 --> 00:33:26.880
And so more and more nurses are being hired through these surveillance pricing apps.
00:33:27.759 --> 00:33:30.160
This is just so wrong.
00:33:30.480 --> 00:33:34.960
Anyway, hey Bitsy, those are my 20 AI lingos.
00:33:35.200 --> 00:33:38.079
I wonder, Bitsy, if you've got anything to add.
00:34:10.159 --> 00:34:17.519
So I'm going to move right into my one recommendation that I have for our listeners today.
00:34:17.760 --> 00:34:21.599
It's a TV series, and it's called Will Trent.
00:34:22.159 --> 00:34:26.239
It's on ABC and it's also streams on Hulu.
00:34:26.559 --> 00:34:36.400
It's a crime drama, and Will Trent is a brilliant and sweet, adorable Georgia Bureau of Investigation detective.
00:34:37.119 --> 00:34:41.760
One interesting aspect of the character is that Will has dyslexia.
00:34:42.000 --> 00:34:44.800
Reading is genuinely hard for him.
00:34:44.960 --> 00:34:47.920
And for a long time, Will hides his handicap.
00:34:48.000 --> 00:34:56.719
But over the series, maybe into season three to four, he lets his guard down and begins to use a text-to-speech tool.
00:34:56.960 --> 00:35:02.400
His phone reads text aloud to him from restaurant menus to crime files.
00:35:02.719 --> 00:35:08.559
So text-to-speech, as we know, has been around for decades, long before today's AI.
00:35:08.800 --> 00:35:13.119
But AI is what makes it better and better and better and better.
00:35:13.360 --> 00:35:34.239
Not only it's really a two-prong AI tool in that it uses the camera to look at what you want it to read, whether it's a menu or a document, and then it uses its voice features to be able to read it to you in a very natural way rather than a robotic way.
00:35:34.800 --> 00:35:38.559
So the listening experience is just so much better.
00:35:38.800 --> 00:35:46.079
So at any rate, it does have a little bit of a link to AI because of this speech to text.
00:35:46.320 --> 00:35:50.880
But generally, I just highly recommend the show Will Trent.
00:35:51.199 --> 00:35:55.519
I love a crime show, but mainly I love the humanity of the characters.
00:35:55.599 --> 00:36:05.920
It's a real feel-good show, and it is also very bingeable with four seasons and more than 50 episodes, like the Aged Podcast.
00:36:06.159 --> 00:36:09.440
50 episodes as of this episode.
00:36:09.679 --> 00:36:14.719
All right, before we wrap up, I do want to take one listener question.
00:36:14.800 --> 00:36:15.920
And here it is.
00:36:16.159 --> 00:36:28.800
If I want to change chatbots, say move from Gemini to Chat GPT, do I need to start from scratch or can I copy my chat history and bring it into my new AI chatbot?
00:36:29.360 --> 00:36:35.440
You definitely can copy your chat history and a lot of information.
00:36:35.760 --> 00:36:41.440
So how you do that differs slightly for each chatbot.
00:36:42.079 --> 00:36:58.559
So this is something if you want to do it, I would recommend that you just ask your AI how to do it, and your AI will walk you through how to copy your history, and then you can import it into your new AI.
00:36:58.800 --> 00:37:13.760
And what you might do, which I would recommend, is before you completely break up with your initial AI, which in my world, my initial AI was ChatGPT, and I broke up with it and moved over to Claude.
00:37:14.000 --> 00:37:22.079
But I did keep my information in both of them for several months just to make sure that I didn't want to go back.
00:37:22.320 --> 00:37:27.599
So anyway, ask your AI to help you walk through how to do that, but you definitely can.
00:37:28.000 --> 00:37:30.800
Okay, so this has been quite an episode.
00:37:31.039 --> 00:37:32.800
Thank you so much for listening.
00:37:32.960 --> 00:37:39.039
And I want to thank those of you who have encouraged others to follow the Aged podcast.
00:37:39.199 --> 00:37:43.039
I want to thank you and I want to encourage others to follow your lead.
00:37:43.280 --> 00:37:51.360
At Aged, we do live on the helpful side of AI, but we do know that AI can be hazardous.
00:37:51.599 --> 00:37:59.119
So please be sure to protect your info, double check advice, and most important, trust your judgment.
00:37:59.280 --> 00:38:07.519
And remember, it is never too late to learn something new, especially something that might make life easier and more fun.
00:38:08.960 --> 00:38:09.599
Cheers!