WEBVTT

00:00:00.660 –> 00:00:13.020
Welcome to the Friday wrap up for May 15th, 2026. This short episode is where I talk about three things. What’s on my mind, recommended reading, and recommended media.

00:00:13.540 –> 00:00:29.960
Streamline Solopreneur is the show to help you build more reliable solopreneur systems so you can take time off worry-free. I hope this roundup and reflection will help you think more about your own systems. I’m Joe Casabona and here’s what’s on my mind.

00:00:31.160 –> 00:00:40.260
Okay, so earlier this week, I found myself fighting with Claude about something I felt was pretty basic.

00:00:41.000 –> 00:00:46.000
A problem that I’ve actually used Claude to solve before.

00:00:47.480 –> 00:00:55.900
I kept going back and forth with the LLM and I would ask it questions, and then it would do things that I didn’t even remotely ask it to do.

00:00:57.180 –> 00:01:05.440
And so I started to form this weird theory that in Opus 4.7, it’s designed to waste tokens,

00:01:06.140 –> 00:01:09.380
which I know is like a weird conspiracy, but maybe it’s not that weird. I don’t know.

00:01:09.820 –> 00:01:10.940
Sound off in the comments.

00:01:11.860 –> 00:01:16.300
But I’m actually worried that it’s worse than that.

00:01:17.260 –> 00:01:22.040
And I, first of all, this is like wild unsubstantiated speculation.

00:01:23.880 –> 00:01:25.080
But we’ve seen it before.

00:01:25.700 –> 00:01:36.260
So in his book, Hooked, Near Aal talks about how social media sites and addictive products employ variable rewards.

00:01:37.120 –> 00:01:42.880
The general idea is that our craving for the reward is stronger than the reward itself.

00:01:43.840 –> 00:01:50.360
So we invest time and money into pursuit of the award.

00:01:50.520 –> 00:01:52.500
And that’s what actually satiates the craving.

00:01:52.740 –> 00:02:01.560
So if you are scrolling on a social media site like TikTok, maybe 75% of the videos you don’t care about.

00:02:02.060 –> 00:02:07.480
But 25% of the time, you’re going to get a video that you really like.

00:02:07.580 –> 00:02:08.820
And it gives you that endorphin hit.

00:02:08.920 –> 00:02:12.560
And so you keep scrolling on TikTok.

00:02:14.220 –> 00:02:21.959
This is why social media sites, gambling sites, prediction markets, and so many other things are so

00:02:22.380 –> 00:02:33.280
addicting. As I had my argument with Claude convinced that I could get a computer program to act

00:02:33.420 –> 00:02:41.900
logically, which I don’t think is a wild request, I started wondering if large language models

00:02:42.620 –> 00:02:51.520
offer a sort of variable reward system. After all, it actually does perform certain tasks really,

00:02:51.780 –> 00:02:52.460
really well.

00:02:54.580 –> 00:03:01.580
And what if that variable reward is enough to convince most people, as well as momentarily

00:03:01.860 –> 00:03:08.140
trick others like me, that the LLMs are actually good at a lot more than they are good at?

00:03:09.120 –> 00:03:14.860
We are pleasantly surprised by the results of a single task or a category of task like

00:03:15.680 –> 00:03:21.600
vibe coding something or going out to our calendar and, and,

00:03:21.660 –> 00:03:26.100
grabbing stuff or going to our email and sending stuff to our to do list, all very like

00:03:26.280 –> 00:03:32.640
computery things. So we start to crave that feeling. The I can’t believe a robot actually did

00:03:32.760 –> 00:03:38.700
this so I don’t have to do this anymore feeling. And we pursue that craving for efficiency.

00:03:39.700 –> 00:03:48.760
We pursue it with time and tokens. And again, this is totally unsubstantiated. It’s a probably a weird

00:03:48.860 –> 00:03:56.220
theory. I don’t think it’s whereas with social media sites and gambling sites and prediction

00:03:57.290 –> 00:04:05.840
markets, it’s in the best interest of those websites to employ variable rewards because it’s in

00:04:05.840 –> 00:04:12.360
their best interest for people to stay on those sites. I actually don’t think it is in the best

00:04:12.600 –> 00:04:18.739
interest of large language models to employ variable rewards. They want to want to. They want to

00:04:18.760 –> 00:04:28.580
want a higher hit rate because the productivity or efficiency angle is the thing that keeps

00:04:28.660 –> 00:04:29.300
people coming back.

00:04:29.400 –> 00:04:30.940
But it’s just something.

00:04:31.240 –> 00:04:37.340
It was so weird in the beginning of the week where I had successfully vibe coded something

00:04:38.100 –> 00:04:38.320
running.

00:04:38.780 –> 00:04:41.700
And then I tried to do it again with a different project.

00:04:41.880 –> 00:04:43.520
And it just like went off the rails.

00:04:44.860 –> 00:04:52.240
And so as I am wasting time trying to bend the LLM to my will, I thought, is this a variable

00:04:52.520 –> 00:04:56.460
rewards thing where like it pleases me a certain amount of the time?

00:04:56.620 –> 00:04:58.040
And so I keep coming back to it.

00:04:58.800 –> 00:05:00.700
Again, I don’t know.

00:05:00.860 –> 00:05:05.880
I don’t think that large language models would actually benefit from that.

00:05:05.960 –> 00:05:07.080
But it’s something I was thinking about.

00:05:07.720 –> 00:05:08.300
What do you think?

00:05:08.360 –> 00:05:09.540
Is it plausible, likely?

00:05:10.080 –> 00:05:11.040
Way off base?

00:05:12.040 –> 00:05:14.820
Let me know either in the comments below or over.

00:05:14.840 –> 00:05:23.200
at streamlinedfeedback.com. I’d love to hear your thoughts on variable rewards in large language

00:05:24.610 –> 00:05:30.840
models. Now, moving on to recommended reading. Usually I like to make recommended reading like

00:05:30.940 –> 00:05:36.660
a heavier, interesting think piece and the recommended media to be something fun and light to bring you

00:05:36.670 –> 00:05:44.800
into your weekend. But it’s reversed this week. So the recommended reading is from the

00:05:44.880 –> 00:05:50.180
it’s called the colorful impact of Spike Lee’s Red Yankee Hat request 30 years ago.

00:05:51.340 –> 00:05:54.520
I am a chronic Yankees hat collector.

00:05:55.780 –> 00:06:04.160
And I suspect that my collection of about a dozen or so Yankee hats pales in comparison to some.

00:06:04.280 –> 00:06:11.300
But still, I have those dozen or so Yankee hats emblazoned with the classic interlocking

00:06:11.340 –> 00:06:15.560
NY that has persisted for over a hundred years.

00:06:17.110 –> 00:06:18.820
In other words, I love a dope hat.

00:06:20.420 –> 00:06:29.840
And arguably, we would not have the vibrant dope hat market that we have today without Spike Lee.

00:06:31.020 –> 00:06:39.560
30 years ago, 1996, Spike Lee, a diehard Yankees fan, wanted a red Yankees hat to match his

00:06:39.480 –> 00:06:40.020
red jacket.

00:06:41.700 –> 00:06:47.200
And when he tried to get it made, he was unable to do to licensing.

00:06:48.180 –> 00:06:53.720
New Era, the company that makes, I believe it’s New Era, the company that makes the on-field

00:06:53.920 –> 00:07:00.900
hats only had license to make hats the teams actually wore on-field.

00:07:02.520 –> 00:07:08.320
So Spike Lee doing something that only Spike Lee and a handful of other people could do,

00:07:09.100 –> 00:07:12.500
Went to the boss, Yankees owner George Steinbrenner.

00:07:13.340 –> 00:07:18.380
And he got George to approve a red Yankees hat for him to wear.

00:07:19.560 –> 00:07:25.960
And now, of course, we have all sorts of hats, all sorts of on-field hats, different colors.

00:07:26.180 –> 00:07:32.280
I have that red Yankee hat, and I wore it on red hat, red Yankee hat day,

00:07:32.340 –> 00:07:34.200
whatever that was called a couple of weeks ago.

00:07:35.160 –> 00:07:41.220
And I have blue hats and yellow, a yellow Yankees hat to match my brand.

00:07:41.920 –> 00:07:47.100
And it’s just, it’s so, it’s a, it’s a fun thing to have.

00:07:47.930 –> 00:07:53.280
And I love this article because it provides such an interesting bit of history and context

00:07:54.320 –> 00:07:59.660
to something that wouldn’t necessarily seem to have an interesting backstory, right?

00:08:00.060 –> 00:08:07.120
You would think, oh, MLB realized they could make a lot more money if they printed different color hats.

00:08:09.360 –> 00:08:15.640
But Spike Lee kind of pioneered this because he also wanted to wear a dope hat.

00:08:16.100 –> 00:08:17.000
So love that story.

00:08:17.440 –> 00:08:19.080
I’ll link it in the description and the show notes.

00:08:19.720 –> 00:08:20.440
I think it’s a fun read.

00:08:22.340 –> 00:08:29.680
And now, because I’m not following the news cycle of like hit you with the bad story and then

00:08:29.780 –> 00:08:35.860
end with a nice story. It’s called the recency effect or recency bias where you remember the last

00:08:35.860 –> 00:08:47.620
thing they talked about. But this video, I think, is too important to not mention. It is by, oh,

00:08:47.620 –> 00:08:55.060
I linked something else recently to them. Right. More Perfect Union. It’s called, I tracked down

00:08:55.160 –> 00:08:58.560
the hidden workers secretly powering chat GPT.

00:09:00.140 –> 00:09:01.760
this is something totally different, right?

00:09:01.920 –> 00:09:07.820
It’s the video talks about companies that recruit people who train large language models.

00:09:08.880 –> 00:09:14.420
And so a lot of large language models are trained with, I forget what it’s called,

00:09:14.580 –> 00:09:18.660
it’s like human reinforced learning, something like that.

00:09:19.260 –> 00:09:25.520
This used to happen in, like, countries that were struggling economically.

00:09:27.000 –> 00:09:41.040
Which is why, for example, chat GPT would use delve so much because the humans that were reinforcing the training, I think in South Africa, used the word delve a lot.

00:09:41.720 –> 00:09:47.580
So there’s a little tidbit as to why delve is used so much in large language models.

00:09:48.300 –> 00:09:54.840
But now, as the companies are trying to be more Ph.D. level, right?

00:09:54.920 –> 00:09:56.720
you’ve heard Sam Altman say this.

00:09:56.840 –> 00:10:03.160
You’ve heard Amodi say this, Dario Amadhi from Anthropic, right?

00:10:03.780 –> 00:10:08.540
That older models could do these things, but now they have PhD level knowledge.

00:10:08.740 –> 00:10:09.660
And that is not true.

00:10:10.360 –> 00:10:12.860
They are hype men trying to make billions of dollars.

00:10:13.360 –> 00:10:15.860
And so they need to paint their products in the best way possible.

00:10:17.020 –> 00:10:24.820
But they are hiring PhDs to do human reinforced training on these.

00:10:24.900 –> 00:10:33.460
large language models. And so the problems this video highlights is twofold. It is the predatory

00:10:33.800 –> 00:10:41.200
nature of recruiting experts in a way that is dehumanizing, like selling to the lowest bidder,

00:10:42.160 –> 00:10:47.760
making them or like offering like higher contracts, but like you’ve got to be available at all

00:10:47.940 –> 00:10:54.820
hours to do it. No price negotiation. You get what we give you. And, and it’s,

00:10:54.840 –> 00:11:01.480
It’s terrible work, and the people who are doing it are possibly in dire economic straits.

00:11:02.740 –> 00:11:03.840
So that’s one side of it.

00:11:04.440 –> 00:11:12.020
But the other side of it is this chilling mindset behind AI companies who want to own knowledge

00:11:12.380 –> 00:11:13.940
and sell it back to us.

00:11:15.560 –> 00:11:22.820
So, you know, I think Sam Altman is quoted in this video as saying, like, yeah, we want to make knowledge available

00:11:23.100 –> 00:11:25.400
and you purchase tokens to get a little bit of that knowledge.

00:11:26.580 –> 00:11:27.340
And that’s awful.

00:11:28.900 –> 00:11:32.320
It cuts against what got us here, right?

00:11:32.440 –> 00:11:36.140
The reason the internet was created

00:11:36.900 –> 00:11:40.780
was so that researchers could quickly share knowledge with each other.

00:11:42.060 –> 00:11:47.060
And large language models appear to be trying to paywall that knowledge.

00:11:48.880 –> 00:11:50.580
So this video is really interesting.

00:11:50.860 –> 00:11:52.380
I think like more perfect union.

00:11:53.380 –> 00:11:58.640
I’ve been watching a few of their videos and I want to dig a little bit deeper into them and like where they came from and this and that.

00:12:00.520 –> 00:12:06.160
But it’s at the very least seems like they do this like deeply researched work and it’s well produced.

00:12:09.200 –> 00:12:16.320
And it’s just an it’s an interesting thing that one might not think about when it comes to large language models.

00:12:16.540 –> 00:12:20.300
Now there is a hopeful call to action at the end.

00:12:21.380 –> 00:12:27.520
but ultimately I’m sharing this because I think it’s an important message for anybody who uses large language models to hear.

00:12:29.900 –> 00:12:35.080
All right, that is it for the Friday wrap on May 15th, 2026.

00:12:35.500 –> 00:12:43.440
If you want to get a written version of this delivered directly to your inbox, as well as an exclusive automation of the week,

00:12:44.460 –> 00:12:48.900
join my newsletter over at streamlined.fm slash wrap.

00:12:49.960 –> 00:12:52.700
This week, maybe somewhat hypocritically,

00:12:52.880 –> 00:12:53.880
given what I just talked about,

00:12:54.440 –> 00:12:57.420
I am sharing an automation I’ve set up in Claude.

00:12:58.280 –> 00:13:02.180
But that is it for this episode of the streamlined solopreneur

00:13:02.180 –> 00:13:03.100
and the Friday wrap-up.

00:13:03.100 –> 00:13:04.000
I hope you enjoyed it.

00:13:04.480 –> 00:13:08.880
If you did, again, sign up for my newsletter.

00:13:09.480 –> 00:13:10.480
Thanks so much for listening.

00:13:11.260 –> 00:13:13.940
And until next time, I hope you find some space.

00:13:14.540 –> 00:13:15.160
in your weekend.