Adam:
It's Wednesday, May 24th. This is Accelerate Daily. Today we've got a look at Jimi Hendrix at 54, the AI future of Microsoft's Cortana with a side of Stuck Snack, Hugging Face partnering with IBM, Adobe debuting generative AI in Photoshop and a recipe for mass content pregeneration. Put on your goggles, let's jump into it. Okay. Welcome back everybody. I'm Adam.

gptboss:
My name is Mackenzie. Good morning.

Adam:
Uh, we're back with three or a little more than that. We got some bonus links today, three links and one recipe to keep you caught up on what's happening in AI today.

gptboss:
We're gonna move fast, this is a zippy one, because we have a hard stop.

Adam:
We do have a hard stop. We can't ramble too hard. Still need to go to the next place to run this show. There we go. Okay, the deck. Uh, yeah, so, so it's been another 24 hours. How you doing?

gptboss:
Oh, I live in the dream.

Adam:
Hahaha

gptboss:
I'm near bankruptcy.

Adam:
It's the time of, it's the time of year in my basement, which is where the office is. Uh, where I saw a meme the other day. It was a college humor video, uh, about, uh, it was like girl winter. And it was

gptboss:
Oh,

Adam:
about how offices get real cold

gptboss:
the AC turns

Adam:
during

gptboss:
on, yeah. Yeah,

Adam:
the summer. When the AC turns on,

gptboss:
that's

Adam:
like

gptboss:
a great one.

Adam:
I live that there's a time of the summer where for my life outside of my office. I, it makes sense to be bald. So I have a shaved head now, but like, then I get down in the office to work and I'm like, Oh, it's cold. I need to, I need to put a hat on.

gptboss:
Yeah.

Adam:
Here's the solution. If anyone wants to advertise on that space,

gptboss:
Hehehehe

Adam:
call me. Okay. Oh, the lead image is great today. The, uh,

gptboss:
Yeah.

Adam:
the prompt is a simple one, but it came out nicely. Jimi Hendrix performing on Saturday Night Live in 1996.

gptboss:
Mm-hmm. That's exactly what this is. There's a little bit of like prints going on

Adam:
For sure.

gptboss:
Yeah

Adam:
But they're parallel

gptboss:
Yeah,

Adam:
aesthetics.

gptboss:
totally.

Adam:
Fair.

gptboss:
Now, was this stable diffusion or mid-journey? Do you know?

Adam:
mid-journey, I think.

gptboss:
Whoa. That's wicked.

Adam:
Yeah, the full check out the link. If even if you're watching the live stream, check out the link for the full square versions, if they're coming out of mid journey, mid journeys output is a one by one, uh, I have to zoom in to make it thumbnail sized on YouTube, but if you zoom out, like the background is what really makes this perfect as being on SNL. Right. Like the house band is behind them doing something. You recognize the saxophone player. But he's rocking out.

gptboss:
That's awesome.

Adam:
Word. Let's jump into it.

gptboss:
So first up, we've got announcements from Microsoft for Windows 11. And one of them is a doozy. There's some in this article, there's some like small ones that are fine,

Adam:
Yeah.

gptboss:
but one of them is really, really big.

Adam:
Yeah, the link is Ars Technica reporting built in chat. GPT driven co-pilot will transform windows 11 starting in June. Uh, Cortana's out chat. GPT is in.

gptboss:
Yeah.

Adam:
Um...

gptboss:
So this is gonna be a whole new, like if you're familiar with like the start menu, this is gonna be a similar experience to a start menu. It's like, it's literally called a co-pilot and it's going to be a window that's always on your screen on the sidebar that whatever it is that you're doing, browsing the internet, word processing, image generation, video editing, whatever it is, co-pilot can see that process and you can paste stuff in and ask questions about the page that you're looking at. There's like a ton of promise features. This is really exciting. Kind of scary for me as an AI developer to just have like, somebody beat me to like, you know, they're like putting a competing product on the OS. But I think it's ultimately going to be a lot better experience for a lot of people. And it's basically just going to completely change the way that we like interface with computers. Like the interface is going to be different kind of until we don't have this interface anymore. I think I think this is like a permanent

Adam:
Right.

gptboss:
upgrade.

Adam:
You can, you can kind of follow this thread through like quick launches and stuff. But if you grew up in the era before spotlight, there were like plugins that got popular. One was called like Quicksilver. They did exactly the same thing. You hit a quick key and just type the thing that you want. And it's just a faster way to find the thing that you want. Uh, the faster way is search. The faster way is this, but like this layer gets really interesting cause it's chat GPT. Um. But I want to back it up for a second on this one. I, what, like, can we talk about the failure failure of Cortana a little bit? Uh,

gptboss:
Yeah, I remember

Adam:
this

gptboss:
that.

Adam:
was, this was an interesting one from the beginning.

gptboss:
Yeah.

Adam:
Cause it was at a time when Microsoft was kind of trying to merge the, the, the place where the gaming lived in their universe on Xbox and their windows platforms.

gptboss:
Mm-hmm.

Adam:
And so they named their like internal assistant. after. After the like AI computer voice that's scripted inside of the Halo video games, right?

gptboss:
Yeah, she was a heads up display assistant, like a co-pilot, right? She was like a more advanced, like personality infused co-pilot. Yeah, I grew up, I got an Xbox in the year 2000 and it came with Halo Combat Evolved.

Adam:
Yeah.

gptboss:
And so that was like, I was so excited for when they were like announcing this product. I was like, oh, no way, like they really made this in real life. And then you get it. And it's, it's not that.

Adam:
Heh.

gptboss:
And I think that kind of cognitive dissonance of like, oh, there's

Adam:
Right.

gptboss:
this scripted like. like really robust personality to the AI. And then the Cortana that we got was, didn't have that much personality. It wasn't that advanced. So that just kind of like flopped and now they're moving to like this co-pilot rebrand, which

Adam:
Microsoft

gptboss:
I personally

Adam:
had

gptboss:
like.

Adam:
Microsoft has a broader history of failing at this too. This is Clippy Clippy 2.0 to some extent. But the thing about both of them is like the metaphor is not wrong. They were just way too early to promising the computer from the next generation. By the way that they branded the thing. I was just turned it off during

gptboss:
Totally.

Adam:
setup. I never had an experience of taking Cortana or even their system wide search seriously. Um. Which I don't think is fair. You always find what you're looking for. It's just not as intuitive as Apple, I think, but this is

gptboss:
Yeah.

Adam:
a totally different. Anyway,

gptboss:
There's a number

Adam:
too,

gptboss:
of smaller updates,

Adam:
too

gptboss:
so

Adam:
far

gptboss:
definitely

Adam:
down the rabbit

gptboss:
do

Adam:
hole, but the power

gptboss:
check out

Adam:
of.

gptboss:
this to see

Adam:
Yeah.

gptboss:
what's coming out sooner than June because there are gonna be some updates to the Windows users out there like myself.

Adam:
Uh, there's a funny sort of addendum to this at the

gptboss:
Yeah.

Adam:
same time that this story dropped, there's a podcast coming at the verge cast, uh, verges sort of flagship podcast where they're talking about, right, but what if I tell one of these AIs to do a Stuxnet, uh, like shut down any cooling systems on a reactor that are running windows 3.1. or something like that.

gptboss:
Yeah.

Adam:
Anyway,

gptboss:
Um,

Adam:
they

gptboss:
and

Adam:
go back

gptboss:
so

Adam:
and forth a bit. You

gptboss:
the

Adam:
should check out the, I linked to a TikTok. It's a minute and a half long, but it's worth,

gptboss:
The guest was the CTO from Microsoft, right?

Adam:
look at that. CTO for Microsoft. Yeah.

gptboss:
Kevin Scott, yeah. And so he said like, no, I don't think that's possible. And I think it's just like a matter of like prompt engineering. This seems like something that could theoretically be doable with like chaining prompts, because

Adam:
Right.

gptboss:
the prompt he was looking at was like, please write Stuxnet. And obviously it's not going to do that because Stuxnet has like a really negative reward value. So it's gonna respond like, no, I'm not doing Stuxnet. But if you were writing an API for a system controlled by Windows 3.1 and you needed a method to remote in and then turn some feature on and turn on some other feature off, there's no reason to deny you that as like an application programmer. But like, so these systems, there's still this threat of like, yes, they definitely can do that work, but you need to be smart enough to put it all together and theoretically also not get caught, which are probably not. So don't go out and make Stuckstead.

Adam:
Yeah, there are, there are, it's safe to say there are a lot of ways that you get in trouble. If you try to do some kind of thing that AI is not exempt from because you did it funny through auto GPT, um, it's going to take a lot of iron to run all of that software touches the ground somewhere. Uh, at the same time. Oh, have the slides been advancing for you?

gptboss:
No they have not.

Adam:
Oh no, we gotta fix that. weird. up live stream viewers of me fixing production things during the session.

gptboss:
There's still only one of them. It is me.

Adam:
Okay, how we doing now?

gptboss:
Yeah, there we go.

Adam:
Doesn't really matter, but it's okay. Okay.

gptboss:
head.

Adam:
Where were we? Stuxnet of AI.

gptboss:
We were

Adam:
Um,

gptboss:
just about to wrap up and move on to

Adam:
yeah,

gptboss:
the next

Adam:
the

gptboss:
one.

Adam:
broader thing on this one is, is, uh, red teaming. You know, the, the, what the, what, um, CTO of Microsoft is talking of the broader conversation is about red teaming. These things like the idea that we're trusting these companies to put internal teams in charge of trying these attacks that you talk about. In a safe environment and saying like, Ooh, that would have worked if it were connected to the internet. That's terrifying, but also it's a thing they do with a lot of software. So I don't think, you know, they should be doing it. Questions like turns into legislation. I feel like pretty quickly.

gptboss:
Yeah, that would be a crazy job. Congrats to anyone on the red team out there. So next up, unfortunately, I wasn't able to access this one. So it's going to be a lot of you on this one, Adam. But Hugging Face and IBM are partnering on a new Watson X AI, which is a next generation enterprise studio for AI builders, which is awesome. I've used the Hugging Face, like spaces and inferences, and I've deployed some stuff from Hugging Face. So a little confusing. So I'm looking forward to this studio to get a little bit of a better developer experience and access some of these benefits that Hugging Face offers.

Adam:
Uh, yeah, it's an interesting thread that I, that, that, that I tend to follow through watching this stuff happen in crypto. And before that, and like server software and Linux and open source in general, um, hugging faces, a community of open source AI, you know, machine learning scientists working on projects. You're publishing papers about what they're working on, but also here's the code that went with it. Uh, you can now run this thing. If you put it on X, you know, hardware. Um,

gptboss:
importantly,

Adam:
Yeah,

gptboss:
they

Adam:
you know,

gptboss:
also have

Adam:
like

gptboss:
a pipeline

Adam:
you,

gptboss:
for

Adam:
so

gptboss:
actually deploying it.

Adam:
to track to what happened in crypto, as we watched a protocol rollout, you get to a point where, like, if you, if you want some aspects of, of this kind of software to work, you have to let the network talk to itself and to do that, you have to have an aspect that says, if you build toward this spec, you'll be able to talk to all of them. It's why email keeps working because we built a whole thing on top of that. This infrastructure still needs to exist, but it didn't monetize it infrastructure. So anyway, in crypto, you saw everybody announced their own blockchain and then it all collapsed back in on. We realized over a couple of years that we can't monopolize this aspect of it. So no one should be allowed to. Anyway, that's the more, you know, Uh, it feels almost theoretical side of it. This is just a significant headline in terms of, you know, IBM, a giant in the space who's been working, who has, if you know, say what you will about where they progressed from that, but like they're forever in the history books for winning chess games and things like that. Like they've been paying attention. Uh, it's fair to hit a point where you go, you know, the best place to build is on top of the stuff that. you know, gets to what you were talking about yesterday. There's a layer of this where the, the middle in the middle and the bottom get fat and there's a lot of stuff that happens even before we have to worry about GPT seven and the super intelligence and the, and the, whatever. Um, because some of this stuff doesn't work if you don't share the underlying ability to whatever right process contribute

gptboss:
Totally.

Adam:
to the

gptboss:
Yeah,

Adam:
solving

gptboss:
definitely.

Adam:
the problems as they emerge.

gptboss:
Definitely the big benefit from hugging face is like custom training. Um, so there's this, it's, it's against open AI terms and conditions, but theoretically you could collect 10 million, 100 million outputs from GPT four and then back train a less robust model on like, Hey, this response should give this output and that's really easy to do on hugging face. So it kind of, it, um, diffuses intelligence across a bunch of different pieces of hardware, which is pretty exciting.

Adam:
Moving on. Adobe dropped some good stuff. This one is at the verge. Adobe is adding AI image generator Firefly to Photoshop. They moved really quickly from a beta that I was playing around in to a thing inside of a sincere tool that artists use every day. And like, it just. This is back to the ambient intelligence thread, I think, but also again, a cool headline. Um, you know, uh, they know how to build this kind of software. They have the rights to their stock library. You know, they're in a really interesting position to make this move over to what it seems to be happening, which is with this, you should check out the article. I wanted to link to Adobe's post. I'll s- might still put it in the description, but the verge has a bunch of examples all in one place, but you should also check out Adobe's video and their posts. Like they show off some pretty cool stuff. Uh, you know, literally just you hire, you highlight a section, all the parts of the dress you don't like you're right. Put a fancy or dress here and then it produces a thing that you're like, yep. If I back up a tiny bit and squint, that's good enough. Ha ha ha. The lead image here looks like a Pink Floyd album cover or something like that.

gptboss:
yeah yet

Adam:
which I appreciate. Because I used to I followed that for a long time as a photographer because it was really hard to do composites like that in a dark room.

gptboss:
Yeah.

Adam:
And now it's just like, hey,

gptboss:
Yeah.

Adam:
put a line here. Hey, put a lake in front of this car. Anyway.

gptboss:
What I'm wondering about all this like generative AI stuff is their recent Figma acquisition as a developer. I'm always inside of Figma. I almost never use like Illustrator. That was

Adam:
Yeah,

gptboss:
the previous

Adam:
that's

gptboss:
de

Adam:
happened

gptboss:
facto.

Adam:
to me as well. I

gptboss:
Yeah.

Adam:
pay for creative cloud though, because I, you know.

gptboss:
I don't, I canceled it because I never

Adam:
used to

gptboss:
actually

Adam:
it.

gptboss:
used Photoshop, right? I was never like already like that, but

Adam:
Yeah.

gptboss:
I always need UIs. So I wonder at what point, like, it's cool that we're getting this like generative artwork, but I wonder at what point we're starting to get generative UIs too.

Adam:
Ambient intelligence. Yeah.

gptboss:
Yeah.

Adam:
It's just going to all live. Everywhere.

gptboss:
I hope so. I hope so. Right. Like Adobe

Adam:
Yeah. That's,

gptboss:
is responsible

Adam:
that's the hope.

gptboss:
for implementing it into

Adam:
Yeah.

gptboss:
Figma and they might decide not to to just like give more demand to Photoshop.

Adam:
That's fair too. Thing was pretty.

gptboss:
There was a link that you kind

Adam:
I think

gptboss:
of buried

Adam:
figs.

gptboss:
the lead on, I think, coming up next, which is the Verify. Let's see if that's in the deck.

Adam:
Oh, it's not in the deck. Uh, but that was, that's a good reminder to talk about it. Uh, thank you. Yeah. Uh, they also verify is also run by Adobe or funded, or I don't know what the structure is exactly. Um, Verify is basically a place you can go to verify an image, uh, like provenance and stuff like that, um,

gptboss:
Yeah, a new a new primitive web initiative to

Adam:
Yeah.

gptboss:
provide like this is the source of this image and this is the edits that happened to it. And Adobe wants that because it's really easy for them to support that. So to say like, hey, images that don't have this metadata can't be like referenced in like news stories would be really, really good for Adobe, obviously.

Adam:
A lot of

gptboss:
But

Adam:
this

gptboss:
they're

Adam:
already

gptboss:
opening

Adam:
exists

gptboss:
it up. So

Adam:
because they do stock photography. Like they have a whole thing stock. Um, I don't like it as much as shutter stock, but it's included in creative cloud. So I use it all the time. Although Canva has kind of taken that over because it's just sort of built in. You can just go find a photo, but even Canva is tracking the rights to those photos and anyway,

gptboss:
I'm

Adam:
it's

gptboss:
excited

Adam:
a preexisting

gptboss:
for this slide. This

Adam:
thing.

gptboss:
is my first time seeing this. Waitlist activate by mission control.

Adam:
Oh

gptboss:
What

Adam:
yeah.

gptboss:
is this?

Adam:
Uh, this is where I got a, I got a, I got to sell my own book. Um, yeah, we have a launch today. Uh, we, we, on a marketing front, I would say we launched a marketing rebrand. Um, but if you chase it up into the actual offering, we, uh, rolled out an official version of mission control and what we're really trying to do here. So go check it out. Deploy and trust generative AI in minutes. I got to read it like a headline, right? Yeah, we're trying to make it easier for builders to thread together all of the complexity we talk about here into functional things without having to worry about code, or worry about code as much. No code platform for

gptboss:
Yeah.

Adam:
being able to do the recipes and stuff. Uh, we're starting with an enterprise level thing. So

gptboss:
Yeah,

Adam:
it's more of

gptboss:
and

Adam:
a

gptboss:
it's

Adam:
waste wait

gptboss:
so.

Adam:
list kind of deal, but sign up. And also there's all kinds of links to stuff. Other stuff I make to educate and push different balls forward in this front. Uh, like not having AI ruin things.

gptboss:
So correct me if I'm wrong, but mission control's main thing is like, um, it's all about compliance, uh, for like enterprise solutions because you can,

Adam:
Uh...

gptboss:
you can like go and like get a cowboy to like do some

Adam:
Yeah.

gptboss:
AI thing, but you don't know what they're doing with the data. You don't know like where it's going. You don't know if like they're breaking terms to get it done. So from what I understand mission control, like this, like, it's like a sandbox that your enterprise development team can work in to make sure that everything that they're doing isn't like misaligned with your company goals and stuff like that, right?

Adam:
Uh, yeah. Um, and the way it's taken shape as we've been working on the problem, you know, cause we're at the stage of a startup where you go pitch an idea, you raise some money and then when it hits the market, everything starts changing. And you, and you are trying to find a product market fit, right? Uh, you do that by talking to customers and getting answers back. What we've ended up realizing is that the compliance stuff, so early on compliance was a concern from like enterprise level clients stuff. They're used to aspects of this because of GDPR, things like that. Um, but as it applies to AI, and then we realized over time that that was sort of just like a backbone service. So the cool thing is now we kind of a narrative of working for a long time on the, on the, in the space of. How can you deploy trust frameworks to understand the complexity and the risk inside of the system of AIs, but also think that's kind of just table stakes so that we can do trust and responsibility. Right. So, so AIs don't become problematic, but then we realized the best way to do that. It's just had to be the backbone of a no code platform that lets us thread together the recipe kind of stuff we talk about here at the end of the show, um, by just dragging bubbles around. And this is already a way that a lot of like data automation happens. It's just easier to visualize a certain layer of, okay, I need to take all these APIs. I need to have these different things go through here. I need the NSFW filter here. I need the GDPR filter here. I need the whatever's, you know, like it ends up being compliance software. If you want to use it that way, it can also just be, Hey, this task is better on Bard right now. So I'm going to have Bard. Do this type of thing and it flows over here and it hits the website and the data structure and the wherever you want to route it on the other side. Um, but that also includes setting up an internal hub for chat GPT or something like that and being able to provide the chat GPT experience while like running these tests, right? You can have an internal testing ground for these kinds of things. And so if you're talking about. Running that kind of stuff, you can. wallet off in ways that make it safer. So that gets to compliance things. Uh, but also you can, you can train an AI with your intellectual property to look for possible leak points on these kinds of conversations. It's interesting. The

gptboss:
That's

Adam:
concerns

gptboss:
been a huge

Adam:
that we're

gptboss:
problem.

Adam:
running into. Yeah. Like filtering and monitoring and stuff is a big, is a big ask on how to keep. Um, Like IP from leaking that has turned out to be a problem. Cause everybody wants to use the tools. And so they're going to chat GPT and saying, given me a market marketing plan for X and it's like, you just went and talked

gptboss:
leaked your

Adam:
to an

gptboss:
new product.

Adam:
agent that's not controlled by the company about like the next, yeah, the next

gptboss:
And

Adam:
iPhone.

gptboss:
it's a no-code platform, so you could theoretically do it yourself, but if you don't want to, bang my line. I'll do it for you.

Adam:
But

gptboss:
So

Adam:
yeah,

gptboss:
we got

Adam:
check, check

gptboss:
we

Adam:
that

gptboss:
got

Adam:
out.

gptboss:
four minutes left

Adam:
Four minutes left. I know you're right. Not about the heart out. Okay. Is that the recipe for today? When

gptboss:
Yeah.

Adam:
it looks like code, I'm inclined to call it neural network nourishment.

gptboss:
Totally.

Adam:
You call this one the data pre-cooker.

gptboss:
Yeah, it's like, yeah. So basically

Adam:
We can

gptboss:
yesterday

Adam:
go a bit

gptboss:
we were

Adam:
over,

gptboss:
talking about

Adam:
by the way.

gptboss:
dadjokes.email where you get 365 emails generated from GPT-4 that tell a dad joke and then you send them out

Adam:
Cough

gptboss:
for a dollar for the year, whatever it is. And so then people get this delightful email every morning. So this pre-generation task of like, let's make sure that we have 365 unique jokes is a little bit difficult with GPT-4 because of limits on the attention. If you have GPT-4. you're probably limited to an 8,000 token limit because to get the 32,000 token limit, you have to run evals, which is this crazy like computer science thing that basically nobody outside of Silicon Valley knows how to do. So if you're not, if you don't have it, then you're not getting it anytime soon. So we have to have a workaround and that's what the code on screen is. This is a function called check cosine similarity. And this is just basically an algorithm to compare two pieces of text to make sure that they're not too similar. So an example of a really similar dad joke is the one that we told you yesterday, what is blue and not very heavy. light blue, but you could also generate what is blue and not very light and then say dark blue. And so these jokes are too similar. If I got those in my email back to back, I'd feel like ripped off. So

Adam:
So

gptboss:
with

Adam:
you're pre-training it with examples, like you're

gptboss:
no,

Adam:
actually giving

gptboss:
you're,

Adam:
it concrete

gptboss:
you're taking

Adam:
examples

gptboss:
the output, you're taking

Adam:
or

gptboss:
the output and then you're

Adam:
yeah.

gptboss:
saying, Hey, just make sure that this joke that you just came up with doesn't match any of the other, like 100 jokes that you previously came up.

Adam:
Yeah.

gptboss:
And then if they do, then you send it back with few shot learning and you're saying, Hey, you just wrote this joke and that's really similar to this previous joke you already wrote. So can you write a new one that's different enough? And, um, kind of the upper limit, like where you start to fail to make sure that these things are unique is, um, there is a potential that the starts happening at around 14,000 tokens, which is way beyond our 365 jokes. And then the hard upper limit where you can guarantee that they're like, you cannot ensure that they'll be different is around 14 million tokens. So. Think about what you could do with 14 million tokens. That's like, the great Gatsby is 80,000. So how many great Gatsby's is that curb? I'm not a big math guy. Yeah.

Adam:
Yeah. Oh jeez. Neither am I. Ask chat GBT. Bit

gptboss:
Yeah,

Adam:
my...

gptboss:
so it's a lot. That's a very big upper limit. But 14,000 again is like, it's about a quarter of the great Gatsby. So basically what I want from AI, what I'm looking forward to is more attention. I always want more attention.

Adam:
And there are models with more like Claude is bragging about a hundred thousand already out of Anthropic who just raised a bunch of money.

gptboss:
Yeah, MPT

Adam:
Um.

gptboss:
7 billion has like theoretically unlimited. They've proven it

Adam:
But

gptboss:
works

Adam:
so

gptboss:
up

Adam:
that's

gptboss:
to 84K.

Adam:
the, like, that's like the resource side, right. But also the, the there's what would that cost you right in dollars, which I don't really know the conversion, like how much, how, you know, uh, how much does it have to run? And then what, you know, anyway, you get charged at the API level, right? Metered, metered existence. Kind of used to that with AWS and stuff. Um, The single shot alternative would be just ask chat GPT this. And it sounds like what you're saying is right now, it's not so good at a chunk that big if you're telling it to repeat a theme, because it's going to retread itself unless you check that against the list and then give it a rating or something like that, right?

gptboss:
That's exactly correct. And I think that's our show. It's 10.30, I think that's all for today.

Adam:
Absolutely. I forgot about the heart out. Thanks everybody for joining us. It's good. It was a good show

gptboss:
Yeah, see y'all later.

Adam:
I'm adam

gptboss:
M-M-M-M-Mackenzie.

Adam:
See you tomorrow. The heart out is kind of fun. We could just do that as a thing.