Ilan (00:00)
Made Larry Ellison the richest man in the world.

David (00:00)
boy, yeah, that's...

Yeah, which I'm sure he's going to do lots of great philanthropical stuff with that money. All

Hey, welcome to Prompt and Circumstance My name's David.

Ilan (00:14)
and I'm Ilan

David (00:15)
and let's look into some mid-September news.

Ilan (00:31)
You know, David, I sometimes think that we should just turn this into a news podcast. There's just so much coming out at all times about AI and vibe coding and how it's impacting the world and productivity and the workplace. I feel like it's impossible to keep up. We did a news episode two weeks ago and there's a whole flood of new stuff going on.

David (00:55)
Yeah, there is a lot. ⁓ and, you know, when I was looking at some interesting things to bring up, ⁓ I've got something from, I think two months ago and I was thinking, that's, that's ancient news. That's, that's old time stuff. Yeah.

Ilan (01:09)
Right

Why don't you start off?

David (01:14)
Yeah, sure. ⁓ So OpenAI and Microsoft, you know, they've been really close for a long time. And I think things are starting to cool down for one reason or another. I'm not going to speculate on that. But I think there's some news that really ⁓ shows how they are maybe diversifying their relationships. ⁓ So for example,

Ilan (01:25)
Mm-hmm.

David (01:43)
OpenAI just signed a contract with Oracle to handle some cloud computing, some cloud capabilities. ⁓

Ilan (01:52)
Yeah,

$300 billion contract.

David (01:55)
Yeah, yeah. Imagine being the sales guy for that, right? Just inking that deal. like, yeah, yeah, I'll take $3 billion. That's fine. And then just retire. Yeah, high fives all around, I'm sure, at Oracle. I think that's sent their stock up by double digit percentages, right?

Ilan (01:59)
Got my 1 % commission

% at one point, yeah.

Made Larry Ellison the richest man in the world.

David (02:20)
boy, yeah, that's...

Yeah, which I'm sure he's going to do lots of great philanthropical stuff with that money. All

right, and no more speculation.

so so that is a really big deal because OpenAI used to use Azure Microsoft's cloud platform for all for all of their compute, I believe. So now they're diversifying. Yep.

Ilan (02:47)
the first time that they've diversified. In May, they signed a deal with Google, funnily enough, because Google one of their main competitors in the LLM space with Gemini, of course. But they need that compute, man. So they signed a deal in May. I don't think it was reported for how much and how much compute and what were the terms of the deal.

yeah, they're definitely diversifying away from Azure as their only compute provider.

David (03:15)
What sounds like it's a great time to be ⁓ an outside sales rep for cloud computing services. Just work the AI market. ⁓

Ilan (03:28)
That's right.

David (03:28)
right, so it goes both ways in this relationship. So Microsoft is diversifying away from OpenAI. So they are going to be using Anthropic.

for, it looks like going to be the office suite. So as you know, there's co-pilot that's embedded into office. And based off of some studies that ⁓ the Microsoft team had done, it looks like Anthropic actually outperformed OpenAI.

Some quotes from the article saying things like, they found that Anthropics latest models, such as Claude Sonnet 4 performed better than OpenAI in automating tasks, such as financial functions in said that, you know, although GPT-5 is a step up in 4, it actually performs better in creating

PowerPoint presentations that were more aesthetically pleasing.

Ilan (04:28)
You know, not having tried out these tools for creating PowerPoint presentations, I will say that, ⁓ the vibe feels right there. ⁓ that's generally the, the impression that I've gotten when comparing chat GPT and Claude. They are obviously both super powerful models. Chat GPT is incredibly competent at a lot of tasks.

But Claude has a

taste just doesn't seem to have.

David (05:03)
Yeah, you know, people have been saying that they often go to Claude to get something that to create copy that sounds a lot better than what GPT has This also ties into some recent announcements from Anthropic where Claude can now create and edit files.

Ilan (05:14)
Mm-hmm.

David (05:26)
They have this really impressive demo video where somebody provides information on their food truck business and it generates an entire report with charts and analyses and so forth, which is very impressive. Now I've yet to try it out,

Ilan (05:35)
Mm-hmm.

He

Actually, at the same time, Anthropic also announced that Claude now has memory and team memory, which is pretty cool. I'm excited to dig into that more, but the way they've done it is really cool. It's memory.

across your teams at a business. So it can maintain context across your entire business, but also inside of projects, it maintains the memory specific to that project so that you're not kind of like leaking context from one closed item across other items. ⁓

David (06:21)
Hmm.

yeah, that's interesting. I do know that with GPT projects, it ⁓ can have access to the general memory ⁓ of GPT for the user. I like the idea that it can have a Teams memory because I haven't heard much in terms of enterprise adoption of Claude.

Ilan (06:38)
Mm-hmm.

separately, Google released an AI report recently talking about cloud AI trends and it's a 40 odd page report. We'll link it in the show notes. and it had a lot of really interesting insights from them, of course, pretty self-serving to, Google case studies, talking about how people have used Gemini. It makes sense.

An interesting fact from this report.

came out of a study from Stanford, MIT, and NBER found that access to AI assistance increases worker productivity by about 15 % on average. But interestingly, the study also found that workers with less experience and skills improved both the speed and quality of their output, while the most experienced and highest skilled workers saw few gains in speed.

and a decline in quality,

David (07:53)
Yeah. And, ⁓ you know, I wonder whether this has to do with how, there simply is more research and work being put in to use cases that are at, at the front lines versus that, which would be at the executive level. if you just, just search online for, help with prompts and so forth, you, you'll see that they are all frontline oriented.

suspect that might be the gap that's manifesting here. ⁓

Ilan (08:20)
Yeah, absolutely. It's very hard to say, Hey Claude, what should my business strategy be?

David (08:25)
Yeah. ⁓

Ilan (08:27)
tell you about one of our favorite is an AI agent which sits on top of your data stack and allows you to ask natural language questions to pull out data insights.

It's the perfect tool if you're looking to understand market segments and user data, as well as trends your product.

Click on the link in our show notes to try it today and let them know that David and Ilan sent you for two months free.

Ilan (08:54)
on a little bit of a darker note ⁓ with regulators are looking into AI for child safety A report came out last month about the types of prompts that ⁓ Meta

was comfortable with their ⁓ AI character models.

responding to to minors and it was Recommend that people read the report themselves it went into a very uncomfortable place both in terms of ⁓ racism and in terms of ⁓ sort of sexual activities for kids

David (09:34)
Mm-hmm.

Ilan (09:35)
That's triggered regulators to start looking more deeply into child safety for AI. And I think that point that you made about optimizing for the here and now or the short term is very much what Meta was going for, right? Let's optimize for getting users on board and like whatever, we'll figure out the future when we get to the future. Exactly.

David (09:54)
engagement. Yeah,

whatever it gets them to continue to talk, right? Even if that means, like, like romantic conversations inappropriately. Yeah.

Ilan (10:04)
That's right.

David (10:06)
All right. Well, ⁓ in even older news, going back to March this year, remember back then, back in the GPT 4.5 days. some of us might know about the Turing test. And for those who don't know, yeah. So, so here's, here's how it works. ⁓ this goes back to Alan Turing, where he was speculating about AI. ⁓

Ilan (10:09)
You

Mm-hmm.

What's the Turing test?

David (10:32)
so the Turing test is one where you have a person who is going to evaluate whether or not something talking to that person is a human or a machine.

the idea is that if, if a machine can convince a human that it is a human more than half the time, ⁓ then that it passes the Turing test. And this is something that two researchers at, ⁓ UC San Diego, ⁓ had been investigating for some time.

They've been doing a lot of research on the Turing test and this paper from March shows that with a decent sized prompt GPT-45 passes the Turing test at guess the rate.

Ilan (11:19)
Well, it has to be above 50 % to pass the Turing test, is it, David?

David (11:25)
It was 73%. So I mean, think about the implications of this, right? So you have a, and here's how they ran their test, by the way. You have a human on one side and you actually have two responders on either side. Okay. So as a human, ⁓ I'm the interrogator as an, I'm the one who's deciding whether, like which one is human, which one is machine, right?

So I send a message and I get a response from A and I get a response from B and I'm going to say, ⁓ B sounds like that's the human. Right. you know, think of the implications of this, that, that a machine sounds more human than a human does, right. And in terms of like writing marketing copy, in terms of being a customer support agent, in terms of there's all the implications in terms of being a podcast host. ⁓ you know, the, are we real?

Ilan (11:57)
Mm-hmm.

Are we real?

David (12:17)
Or are we just

Now, related to ⁓ how AI models might be becoming more human, ⁓ there's some other news. That's a bit more recent.

So the August issue of Vogue had an ad for Guess starring an AI model. Not starting GPT, but starring an image of a person that was entirely synthetic. And ⁓ so this has meaningful implications, right? So does this mean that

there is a threat to human models when it comes to finding work

Ilan (12:59)
Mm-hmm.

David (13:03)
All right, so here is the article on the matter and ⁓ this is what those synthetic images look like. So you can imagine that, well, there might be a little bit of a ⁓ this image is that of a blonde haired blue eyed woman.

And the interesting thing is that, so the company that had generated the advert, ⁓ it was founded by these two people, Andrea and Valentina And they had mentioned ⁓ that ⁓ they had tested more diverse models on Instagram. So here's some examples of, ⁓ I guess, some of the didn't engage nearly as much with those models. So...

Ilan (13:40)
Mm-hmm.

David (13:49)
What does this mean? What does this for people as a whole? I some audiences remain biased, unfortunately, ⁓ and that ⁓ AI Models might be a reflection of that.

Ilan (14:01)
Yeah, I agree. I would also say that there's a little bit of

selection bias or I don't know maybe reversion to the mean is the right term here because of the things pointed the article in that BBC article is in the previous 10 years up until about the point of chat GPT the fashion industry had been making moves to become more inclusive

more diverse, and in the recent years, there's been a trend away from do think that society as a whole tends to comfortable in stasis, right? Like it's uncomfortable to change the types of images that you're seeing. It's uncomfortable to change

standards of But if we agree as a society that that is there's a certain amount of discomfort that people will just have to face and you know leaders like maybe Vogue magazine would have to just accept that Maybe engagement won't be quite as high across the board, but little by little they will pull the mean

of what people think is the standard of beauty towards something that they feel is more inclusive, more diverse, know, better reflects just society as a So I think saying, yeah, well, we tested this on Instagram and people really engage with the blonde haired blue eyed it might be a good test of what's gonna earn you the most money or get the most clicks right now.

But it's not a good test of, we want, where do we want the fashion industry to be? Do you know what I'm saying?

David (15:45)
Mm-hmm.

Yeah, yeah, absolutely. It's, it's, ⁓ it's optimizing really hard for the short term rather than long term.

Ilan (15:53)
Mm-hmm.

Exactly. Which sounds like LLM's overall.

interesting bit of AI news, this is ancient history from two weeks ago, Claude found that a hacker had extensively used their products in order to extort 17 companies to the tune of millions of dollars. And basically, this guy went end to end. He used Claude to figure out what companies he should target. He used it

David (16:15)
Wow.

Ilan (16:24)
to write malware, then used Claude to write the copy of the emails that he should send to the companies to let them know what information he had stolen from them and ⁓ suggest the amount that ⁓ he should request from them. All of this happened within the Claude ecosystem. And of course, the anthropic folks said, well, we're trying our best to maintain security and...

and not allow bad actors to do things like this with our platform.

David (16:56)
Well, the rising tide lifts all boats, including pirate ships.

Ilan (17:03)
mean this was inevitable and if this is the guy who got caught, I wonder how many are out there who are maybe a little bit more careful ⁓ and a little bit more savvy and are not getting caught.

David (17:17)
Yeah, it's like ⁓ when you see one cockroach, it means that there's a whole lot of others.

Ilan (17:24)
That's right.

So we'll link to the article, curious what your thoughts are on this topic and anything that we discussed this week. But with that, we're gonna wrap up. You can follow us on the socials @pandcpodcast You can like and subscribe on Spotify, on Apple Podcasts, on YouTube, or wherever you got your podcasts. And we'll see you next week.

David (17:50)
See you at the next one.