David (00:00)
I think it has to do with the way in

It took little while, a little bit of latency.

Hey everybody, this is Prompt and Circumstance. My name is David.

Ilan (00:16)
and I'm Ilan.

David (00:16)
And this is your Mid-October news.

Alright, so it's mid October. Lots has happened since we last talked about the news. What's going on?

Ilan (00:38)
There is so much news, David. I'm a little overwhelmed if I'm honest, but hopefully we can parse this together for the audience. Starting with Anthropic, who released the Sonnet 4.5 model, which is the latest version of their baseline Sonnet model.

David (00:54)
I haven't had a chance to play with it yet, although I'm curious about the natural soundingness of it, as some people who might work with lot of AI slop might notice. Some of the text that it generates, sounds like it was generated, Anthropic has done a pretty good job with their Claude models to help it sound a bit more natural.

Ilan (01:13)
Yeah, I've noticed that it's less sycophantic than before. Claude has done a good job with the personality of their models as well. So that kind of gets to the same thing about not sounding like you're an AI.

coming hot on the heels of that Sonnet 4.5 release is both Lovable and Bolt released their full cloud offering. So Lovable calls theirs cloud, which gives you access to a backend user authentication as well as AI models. And Bolt calls theirs V2. And it's effectively the same thing. They give you access to a backend authentication ability to build agents.

So we're seeing these vibe coding tools growing up and becoming full stack applications that, you know, a person like you or me who doesn't have much coding experience could really build an entire application on top of.

David (02:05)
Yeah, it's really great to see them do that. It really was the next natural step, I would argue, because if you build any kind of serious app, you're going to need authentication. And you're not going to want to store the secrets in plain text. So they're definitely addressing some of the primary concerns that people had in the previous generation.

Ilan (02:20)
Mm-hmm.

Yeah, you can hear some of that feedback in our very early episodes where we reviewed these tools almost six months ago.

David (02:34)
Maybe it's time to do a re-review.

Ilan (02:37)
That's right, the re-Lovable and re-Bolt.

David (02:40)
Yeah.

Ilan (02:40)
Following that, you have a news story from Microsoft.

David (02:43)
Yeah, I do.

Ilan (02:44)
Now is a great time to tell you about our sponsor for this week, Querio. Querio is a data analysis tool which sits on top of your data stack and allows you to ask natural language questions about your product data. It gives you beautiful visuals and allows to really pull out threads so that you can get to the bottom of all of your product

You can try it too. Just go to querio.ai and let them know David and Ilan sent you for two months free of Explore. That's a thousand dollar value for our listeners.

David (03:19)
Alright, so Microsoft has released new text-to-audio models as part of their MAI set of ⁓ foundation models. And you can try it out for free today at the Copilot Labs.

And so they're saying that this is going to be powering Microsoft Copilot, which is the umbrella term for anything AI powered from Microsoft. So you'll see Copilot in their business applications like Dynamics, and you'll see it in their office applications like Microsoft Office, like Microsoft Word. So it'll be really interesting to see this come to life, and you can certainly play around with it today. So here we go. This is what it looks like.

And just like any other thing, you can just simply prompt it. So let's choose Emotive and then you can choose the voice.

I think what the mode refers to is the way in which it reads out the text. So is it going to read it out like maybe like a narrator or a storyteller? Or is it going to be more expressive with it? Or is it going to be just kind of flat, which is probably what this scripted is going to be? so let's try out the voices here. So here's Rain.

Ilan (04:21)
Got it.

David (04:29)
Right, so that's a British accent. Here's one with an American accent.

So, and then you have a style in terms of the emotion that's conveyed in the audio. So it's really neat. There's joy, sadness, determination, curiosity, excitement, relief, encouragement, friendly, grateful reflection. This is, there's a lot. There's a lot in here. Wait a minute.

Ilan (04:50)
Mm-hmm.

David (04:52)
there's Shakespeare. You know what? I'm going to choose Shakespeare. And I'm going to tell it to give me the introduction, just the first four lines to the play King Henry V. Let's see how that does.

Ilan (05:03)
All right, let's see.

David (05:25)
Okay, so it did not adhere to the script.

Ilan (05:28)
Hahaha

David (05:29)
⁓

So I think what's happening is that the style ends up modifying the text that you give it. So if you want it to be Shakespearean, it's going to take whatever you can give it, I guess, normal text, and it's going to turn to Shakespearean.

So in the interest of time, let's move on. what other news do we have here?

Ilan (05:46)
Well, OpenAI has dropped a lot of news in the last two weeks. Starting with two weeks ago when they announced Agentic Commerce.

in ChatGPT starting with etsy and closely followed by shopify you can transact entirely in ChatGPT so you can find a product and then buy the product completely within the chat interface

David (06:09)
Wow, so you could ask it for advice on, hey look, what are the latest fashion trends? And it'll give you, okay, this item, that item, and you can buy it right away.

Ilan (06:21)
That's right. And following that, they also announced Sora 2, which is their latest video generation model. And with that announcement, they also released a social media app where

the entire app is built around generated content from their new model.

David (06:38)
It's really interesting that they decided to do that because I hear from my colleagues all the time. I wish I had a social media application where nothing is real.

Ilan (06:49)
yeah, exactly. I'm tired of seeing like, you know, my friends and family. I want to see made up people doing made up but it's an interesting play into the consumer web. And you can see OpenAI testing out

David (06:49)
Yeah.

Hehehehe

Ilan (07:04)
different ways to address consumer desires, consumer needs. And ultimately in Sora, which we'll dive into a little bit later, you can download the videos you create. So the play might just be getting people to create videos there and then pull them into TikTok and Instagram and the other social media applications that they're used to using every day.

But on top of those two big announcements, they also had their Dev Days where they released a whole slew of new functionality.

I'm going to go through these, but each one could be unpacked on its own episode. So we're just going to try and get through these quickly. they released the agent kit, which is their own take on how developers can build AI agents with workflows, something like n8n or Lindy that you've heard us talk about recently.

⁓ except for it's all built within OpenAI and it's all orchestrated within OpenAI.

David (08:02)
So that's certainly a threat to companies like Lindy, right? Where, I mean, we saw this with other kinds of companies like ⁓ note-taking companies and so forth, where companies like OpenAI would just expand on their capabilities to envelop what those companies uniquely had to offer.

Ilan (08:19)
Yeah, that's right. This, I think will be a little bit less of a challenger in a particular way, which is that a benefit that you get from an n8n, for example, is that you can call different LLMs for different purposes and OpenAI's models are not all powerful, though OpenAI may think.

They are, there are aspects of Gemini Flash that are useful. And we know that Claude Opus is kind of the king of coding in the current moment. So if you have access to all these different models, you can achieve more and you can use them for what they are best at, as opposed to being locked into OpenAI's models. They also released their apps SDK.

David (09:00)
Yep. Makes sense.

Ilan (09:04)
So this is a big platform push that OpenAI is making. What this actually means for our listeners is that you can create apps that will function within ChatGPT. So if you trigger the app with the right phrasing, so for example, Spotify, which is one of their beta

apps, you can say, I want to build a playlist in Spotify.

David (09:29)
That's really cool. So you could say, Hey, I want a K-pop workout list and it would generate that for you.

Ilan (09:35)
That's right.

Why this matters more is that this is a new distribution channel. So ChatGPT may not be part of your long-term strategy, but

you could potentially build your entire business. And then eventually as OpenAI closes off the access to those users or potentially monetizes the access to those users. If you've built a good enough product that solves people's problems, you can then move them away from the initial platform onto your own site.

David (10:04)
Yeah, I mean, it certainly behooves a lot of go-to-market folks and even product folks to think about this new channel that is emerging. You know, we had previously talked about AEO, answer engine optimization, and, you know, that's just for discovering, but like here, the fact that, you know, distribution can be within the answer engine, ⁓ I think is really compelling.

Ilan (10:26)
That's right.

Absolutely. They also announced that the Sora 2 video generation API is now available. They announced that Codex, which is their coding model has gone GA. So access to everybody again, available via API. They also announced that

is available through the API. And they announced a mini speech model and a mini image model. So cheaper ways that developers can access the functionality for speech recognition or speech functionality, as well as image generation functionality.

David (11:04)
Very cool. Yeah. Lots more power to the developers.

Ilan (11:09)
Totally. I saw an interesting stat in the landing page that they created for their developer conference, which is that four million developers have built with OpenAI. And four million without context, that can seem very big, but...

I actually think that in the, in the overall context of the world, that's actually pretty small. I would say that, you know, there are on the order of probably hundreds of millions of developers worldwide. And so this is maybe 1 or 2 % of the total developers in the world who've actually used ChatGPT's in their products. So there is

a lot more depth to get to in terms of applications that can leverage LLMs. It also means that a lot of applications out there aren't really seeing the need for integrating LLMs yet.

David (12:01)
Not yet. And you know, sometimes it's the carrot and sometimes it's the stick.

Ilan (12:07)
Yeah, we will see. I think there's going to be a lot more, ⁓ you know, random chat interfaces sitting on top of applications before we get to a better future where people really understand how to build with AI.

David (12:20)
Yeah, yeah, it's going to take a little bit. Hey, I've got some news.

Ilan (12:23)
What's that?

David (12:24)
All right. So in Australia, Deloitte, which is a consulting company, was caught using AI to generate a report for the Australian government. And how were they caught? Well, their report was full of hallucinations and just complete things that don't exist. And so I think this is a pretty good sign that like AI slop is real, work slop. Another term for that.

Ilan (12:37)
Ha

David (12:47)
I don't know whether you've been work slopped before, but I certainly

Ilan (12:51)
Yeah, just putting don't hallucinate in your prompt is not good enough.

David (12:55)
Yeah, yeah, exactly.

Ilan (12:56)
I have definitely been both work slopped. I also can say that I have work slopped others sometimes, you know, there's, definitely.

when you're busy and you have a million things on your plate, an instinct to just go and put a prompt into your LLM of choice and then copy paste the response and leave all the em dashes. And ⁓ I've been called out a couple of times for you know, having plain LLM responses.

David (13:22)
⁓ that's

very sloppy of you.

Ilan (13:24)
⁓ those were in the early days. do it less now than, before.

David (13:29)
⁓ Another piece of news is that Lionsgate, which owns the rights to John Wick, among other movies, they had partnered with Runway, which as some might know generates AI generated videos. They were thinking of ⁓ making a new John Wick movie using all the footage that they had and Runway would generate that.

Ilan (13:40)
Mm-hmm.

David (13:54)
And it turns out that even with all of the footage that Lionsgate has with John Wick, that was not enough to generate a new movie. So that's really interesting to me ⁓ that maybe that's a runway thing, ⁓ but the fact that at this level of magnitude that they were not able to find any kind of success.

Ilan (14:02)
Wow.

David (14:16)
⁓ is a bit of a signal.

Ilan (14:18)
Yeah, there's a reason that image generation models are generating five or 10 second clips for the most part. The ability to maintain a character, maintain a scene across a feature length movie, I think is a little out of the depth for some of these models still. And at best.

We've seen Coca-Cola commercials, maybe 20 seconds or 30 seconds that use entirely AI-generated content.

David (14:48)
Yeah, yeah.

Now that's related to what we talked about earlier about Sora. So did you want to walk us through some of the stuff that you've done there?

Ilan (14:56)
Yeah. Let's talk a little bit about Sora. So recapping Sora 2 is OpenAI's new video generation model, as well as the application that they've created to share those, ⁓ generated models. And there are some really cool things about Sora. When you first create an account in the app.

you're requested to create a cameo of yourself. So you kind of hold your phone in different positions and it gets an idea of your ⁓ face and style. And you can then allow people to use your likeness in their videos. Another cool feature of Sora is that

you can remix videos. So if you find something that you like, you can give a quick prompt to turn it into something a little bit different.

David (15:48)
And that's been really powerful for creating memes.

Ilan (15:50)
Yes, the memeification of the internet is not gone

I've been playing around with Sora. I've been paying a lot of attention to the prompts. Some people share their full prompts that they use to create their videos.

So I started off with, Hey, let's create a mountain biking video. And it took a few tries. It doesn't really understand physics all that well. Uh, but where I had more success initially was just adding an animal.

to the mountain biking video.

What I got interested in though is how do you create a multi shot video that cuts between specified images? And so this is ⁓ where I got to with Sora.

so I created a video of the mountain biking dog and the trail human. Let's take a look.

And this is a take on a classic mountain bike video where you have a human riding their bike and their trail dog following behind them. So just inverting the expectations there.

David (17:05)
That's fun. Yep.

Ilan (17:06)
you need to create a large structured prompt in order to get these kinds of results where you can give Sora, exact camera positions, exact lenses to use cuts to make audio expectations. And it does a pretty good job. takes a few tries to get it to work.

David (17:30)
It's almost like a PRD equivalent for generating a video.

Ilan (17:34)
And we will share the structure that we used to generate these videos. But I'll share one more that I built, which I called the cost of knowledge. And here it is.

David (17:57)
He's so smug.

Ilan (17:58)
So in this video we see somebody typing a question into a chat box What is the cost of knowledge and then a cameo of Sam Altman pulling a lever and starting a waterfall next to a power generation

Again, all that fresh water, that's all it costs all the knowledge in the world.

David (18:15)
that fresh water.

Ilan (18:21)
I would say my top tips here single subjects work better than multiple subjects. Not really surprising considering what we know about AI models and their difficulty counting. Animal videos work better.

especially with short prompts short prompts often win. If you just have a general idea, you know, maybe a one sentence prompt, it does a really good job with those. But you can get closer to a picture or a vision that you have in your head using a long structured prompt.

David (18:54)
Alright, awesome. And I believe Sora right now is invite only.

Ilan (18:59)
That is correct, yeah. It's invite only. For all of those.

who have received the Sora invite and are using the application right now. We'd love to know what you're making, see your videos. So feel free to share a link in the comments and we'll check it out.

David (19:14)
Awesome. I think that wraps up news for mid-October.

Ilan (19:18)
that's all we got for you for this week. Give us a like and a follow. It really helps to spread the podcast. And you can always find us at @pandcpodcast on all the socials.

David (19:27)
See you next time.