Subscribe
Share
Share
Embed
Trying to keep up as AI accelerates toward escape velocity? Let the Mission Control crew help. Tune in daily for a quick take on the last 24 hours in AI news.
Adam:
It's Wednesday, May 30th. This is accelerate daily. Today we've got Nvidia making moves in the market and on the ground. What happens when lawyers don't check their chat, GPT, mind reading AI and a primer on the practice of AI role playing. Get your hand away from that abort button. It's time for accelerate daily.
gptboss:
Hahaha!
Adam:
trying to follow the rocket metaphor.
gptboss:
It'll be fine, probably.
Adam:
Okay, music break.
gptboss:
Probably.
Adam:
Whatever. It's just, I'm just going to try dad jokes until at some point that devolves to just trying dad jokes. Something that's working. Okay. Welcome back everybody. I'm Adam.
gptboss:
My name is Mackenzie. Good morning.
Adam:
And we're back with three headlines and one how to, to keep you caught up on what's happening in AI together. Together today, together also though. That's what I get for like trying to
gptboss:
There
Adam:
side
gptboss:
are two of us
Adam:
side
gptboss:
and we're both
Adam:
eye the
gptboss:
doing
Adam:
script.
gptboss:
it.
Adam:
I didn't put the script in the wrong place. Anyway, thanks for joining us. Okay. Let me go back to the slides. This is a great one for our image today, but There there's a funny interaction as well that you can find in the chat if you click through to the Yeah the reddit thread On that thread people ask what their prompts are because everybody's trying to learn mid-journey The title for this one is clothing line inspired by tectonic plates They asked what the prompt was And the reply was, I don't want to give all of my secrets away, but I find the trick is really to think of words that have good reference for visuals based on the internet culture, but a great start is to use baseline subject and type quote re-imagined as, or inspired by and follow up with your reference. And they, they give an example.
gptboss:
This
Adam:
Um,
gptboss:
is what I'm always harping on, right? Like it's, you need to access the latent space as a proud prompt engineer. And then also it's better at interpolation than generation.
Adam:
Right.
gptboss:
So it knows what clothes looks like. It knows what tectonic plates looks like. And then you could just plug those things. You can get your pen pineapple, apple pen on the go. Um,
Adam:
Yeah.
gptboss:
and that's what this is for.
Adam:
But good stuff. Check it out. There are a bunch of sort of it's, you know, it looks like high fashion photography, uh, but they're all, uh, outfits based on sort of geological, uh, geological stuff. The one I picked is a volcano dress. It's
gptboss:
Yeah,
Adam:
pretty,
gptboss:
this is sweet.
Adam:
pretty epic, but it also has all the other characteristics of fashion. Like the like the scowling, skinny,
gptboss:
Odd looking
Adam:
you
gptboss:
high
Adam:
know.
gptboss:
fashion model.
Adam:
Yeah, exactly. It's a striking image, to be sure. There's
gptboss:
Totally.
Adam:
a bunch more of them. So check that out. Okay, let's jump into the topics. I might've switched the order. I did from the links. Okay. In video making moves. Um, this headline is from the financial times and it's, it's actually thinking that's a little bit secondary to, I think the hot, the hot take right now on Nvidia would be about the, them hitting a trillion in valuation, which is certainly a signal in terms of demand and things like that happening in the AI space, but I'm more interested in this. Uh, from the financial times WPP teams up with Nvidia to use generative AI and advertising, and then there's quotes in the piece from people at WPP about how they're going to be able to do 10,000 things where they used to do 10 and they're going to be more customized than ever.
gptboss:
Who is WPP?
Adam:
Uh, they're a big, um, advertising agency in the UK. Like huge there in the U
gptboss:
Well,
Adam:
I guess.
gptboss:
they should have called me because I'm a big advertiser. I'm six foot six and two 50.
Adam:
I'm
gptboss:
It makes
Adam:
sorry.
gptboss:
me bigger than half the NFL. If they want, if they want to go big, they should hit, give me a call.
Adam:
Uh... Ha ha ha ha! Um... This is, this is. a part of what seems to be a broader expansion by Nvidia, which is like, it feels like five years ago, they were just the GPU company that the
gptboss:
Mm-hmm.
Adam:
gamers knew because they could get the best. you know, chips, the best best graphics cards. Then those cards turned out to be good for other stuff. And now you're seeing this expansion where it's turning into this is a chip company partnering with a giant ad agency saying this stuff that our chips allow you to do. We can partner with you to optimize your whatever engine.
gptboss:
Yeah, because
Adam:
You know.
gptboss:
they have an AI lab and there's stuff there's like algorithms that they have under the hood at Nvidia that I really, really desperately want to get my hands on, like the research coming out of their lab is at the forefront, but there's like very limited access, so I'm turning green here. I want this, give it.
Adam:
Hahaha.
gptboss:
I've been an Nvidia customer literally my entire
Adam:
Uh,
gptboss:
life. Um,
Adam:
yeah.
gptboss:
yeah, it's really exciting. A trillion dollars in market cap. The other thing about that, I think this is like kind of an obvious, maybe Luke Worm take, but that feels like a speculative price tag and there's always like a worry that when people start speculating on something, it doesn't ever really live up to the hype necessarily, even though it is, it is literally a trillion dollar company, but will it be able to like meet? that great expectations placed on it. You know what I mean?
Adam:
Yeah, the reel in the weeds, I gotta check the clock before I jump this far in the weeds on the first slide. I'm excited the extent to which it's a victory for like arm chips. Uh, cause that's the more sort of open platform compared to the other dominant chip architecture, like chiplets, uh, not like chiplets. I mean, the art anyway, it's in the weeds of that stuff, which is cool to the extent of pushing in the direction of, of. Swifter growth, right? So that thing of like, okay, the speculative, the speculative part is people are saying, okay, this is a good bet because the demand for the processing power to run AI is about to be great and people are buying servers for it and their, their numbers are off the charts in terms of demand, but you're right, it could just be. Could be a bubble.
gptboss:
I hope that they use their new capital responsibly. There's a lot of stuff to do. And like mostly the bottleneck in my opinion isn't really like the compute power, it is the power, the energy, right? There's an ongoing energy issue. It's just, it's always been scarce and Sam Almond's making promises that it won't be somehow with AI. We'll see what happens. But this is all kind of dependent on energy as well.
Adam:
Okay. Next up, this one is so delightful. All it got was an emoji face palm
gptboss:
Hehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehe hehehe hehehe hehehe he he he he he he
Adam:
as a headline in the deck here. The headline from the New York Times. Here's what's happened. Here's Take two. Here's what happens when your lawyer uses chat GPT a lawyer representing a man who sued an airline relied on artificial intelligence to help prepare a court filing. It did not go well. Uh, my first beef is with the headline, which is, uh, this isn't what happens when your lawyer uses chat GPT. This is what happens when a bad ill prepared to use AI lawyer uses chat GPT. Um, the failure here anyway,
gptboss:
Lawyers,
Adam:
a lawyer.
gptboss:
if you want GPT lessons, follow me on TikTok.
Adam:
Yeah.
gptboss:
Because this one came up, I did, I did a tutorial recently explaining the architecture of something called a generative adversarial network, which uses a step called a discriminator. LLMs are designed to have a human being be the discriminator to say this output is good and correct. This output is bad and incorrect. So what this guy did was he took a discriminator out of the loop. He didn't read the output. And so when he walked into court, he was underprepared because he took away a foundational fundamental tool that AI systems need to work properly, which is something monitoring the output. Okay, so a lot of people have this expectation that oh this does the work for me, right? I could get more done in less time because this does all of this work for me and that's somewhat true. But you can't unload everything. You can't replace yourself. You can't replace a human being with this, right? It's just it lets you get the work done faster. So you still got to you have to discriminate. You still got to check this stuff.
Adam:
There's, uh, if the production time exists after we're done recording this, I'll go dig it up, but there's a, there's a Billy Corrigan. There's a tick tock of Billy Corrigan from smashing pumpkins, talking about this that I end up sharing with people. But like, he talks about exactly that. The discriminator thing. He's like, you experience it. If you're in the studio, there's somebody kind of sitting on the couch just going, nah, that's not cool. And look. Rick Rubin seems to be right more than most people. So he gets to do whatever he wants. But like, the actual story. So he, so the, I'm saying he, I don't even necessarily know. Anyway, let me back that up. The lawyer in the story. Uh. asked chat GPT to write a filing and the, and chat GPT made up cases. And then I think he got in a bit deeper because when it came up, he, they went back to chat GPT and told it to spit out details about those cases, which it'll also do
gptboss:
I'm sorry.
Adam:
if you start from a, from, from a nonsense request, you're going to get it. You get a nonsense thing. Uh, And then it just got worse from there, but follow the link, read the, read the article for the specifics.
gptboss:
The way that you actually get around this is you use an embedding system to convert all of case law into a token representation. And then you use a tool former system to allow GPT-4 to research real cases against your DB. And then it can actually source things effectively. There is a way to do this. You can do
Adam:
Right.
gptboss:
this, but you're not getting it done through the chat GPT GUI, right? It's going to make stuff up and
Adam:
Right.
gptboss:
lead you on. My big thing is like, don't be lazy. It's like the quickest way to shoot yourself in the foot, right? This is possible. There are ways to do this and, and allow you to kind of be lazy, but it takes a prep step. So don't, don't walk through life being lazy because it's, you're going to get hurt. I don't know. Simple ads. Like a lot of
Adam:
Yeah.
gptboss:
people already knew this.
Adam:
You got to keep watching. You got to keep watching the thing, but also the, your description felt way technical, right? Like the tech, you're talking about a machine learning problem. There's also just a workflow process problem, which is understanding you have to review the output after you do things. And you have to set limiters upfront. In this case, this, this could have been easily solved if The research went into the right cases and then a limiter upfront is set in terms of saying, these are the cases you're allowed to reference. Here's the case text. That's it solved, right? If you have the context, if you, if the model you're using can handle that much context, you can put all great Gatsby in there. anyway, it's not AI, it's not AI's fault people, but also You know, be careful.
gptboss:
Please do be careful and please read your outputs.
Adam:
Face palm emoji. We just do emojis for all of them. That would save me some head writing, headline writing time.
gptboss:
This one I would put shocked emoji. I hate this this is
Adam:
Uhhh...
gptboss:
There's there's cool medical applications, but this is like I've seen so many like dystopian science fictions where this is like a core plot point
Adam:
Uh, this, yeah, this is from the department of weird stuff. The X files, uh, the link is to a tweet. And then I think I also linked to the paper. Um, it's, uh, well, I'll just read the pertinent part of the tweet. Mind eye is a novel FMRI to image approach to retrieve and reconstruct viewed images from brain activity. There's there. Few different things like this have been popping up. Um, my understanding is this is the first one where they just say, think about a thing or think about a, uh, a memory and then it reconstructs it. The, the most recent one I saw otherwise was people looking at a specific picture and trying to memorize it and then being asked to recall it later and they produce sort of fuzzy images, but. Check out the tweet for more on what's in the paper, but also verify, read
gptboss:
Yeah.
Adam:
the paper if you're really freaked out.
gptboss:
So this is like a core, we're seeing this more and more in models like Facebook just released a new, it's called ImageBind, I think, or something. That's like a true multimodal because all that you need to do for these systems like transformers, GPT of course is a transformer to work, is tokenize some input. It doesn't really matter what the input is. It could be an audio file, it could be a video file, it could be an fMRI, it could be an x-ray, it could be text, right? It doesn't matter. As long as it's tokenizable, it's transformable. So we're gonna start seeing a lot more of these. kind of multimodal applications and what tool forms we can plug them into, right? Like we could take like image to text and then text to a tool. And then that tool produces response that can recreate an image. And that's exactly what's going on here, right? We're like checking, tokenizing, FMRI brain scans, and using that with another tool to just reconstruct what you're imagining inside of your head. And so this is like, it's crazy, but it's awesome because there's a lot of people that have like a walked in kind of syndrome that can't communicate. And this goes away to like start solving that kind of issue, which is, I think, a really important issue just to return like dignity to people that definitely deserve it. But we just didn't have a way to give it to them until now.
Adam:
It's super interesting through that lens. Also, just the extent of like our obsession with reproducing reality, like even just for video games, reproducing this aspect of it. So we can re-experience it is how we got the GPUs we talked about earlier in the episode. Video games and what not. The convergence of these these. Areas of compute, super interesting, so get in the weeds, people. Okay. Oh yeah, this one's a mission control shout out. Um, Ramsey Brown, who has occasionally hosted this show, but the, uh, one of the co-founders of mission control has accepted a position at Jesus college. Cambridge working on AI things in the academic says senior research associate is the title. So
gptboss:
So that's in Cambridge,
Adam:
shout
gptboss:
Massachusetts.
Adam:
out for that one. Um, No, it's in England.
gptboss:
Oh. Ha
Adam:
Yeah.
gptboss:
ha.
Adam:
It's the older than that Cambridge.
gptboss:
Cool.
Adam:
Uh, no, we partnered with them, um, last year to do an event, a leadership summit that was worked out well. So that's evolved into ongoing work, which is exciting for everybody, but Ramsey
gptboss:
Yeah,
Adam:
in particular.
gptboss:
totally.
Adam:
So good work, sir. Okay. onto the how to I'm calling it this week because it's a little easier, but also I think as I've gotten feedback from people watching show, listening to the podcast, uh, they were like even broader concepts, like where we're onboarding to this weird new way of trying to do whatever it is, uh, which actually gets to a thing I didn't mention today's title. Was brought to you by right clicking on a thing and click up and asking it to summarize it into a title for a podcast. I just pasted it in. If it's 10 X better than me, I'm not as good at writing headlines as I thought I was.
gptboss:
We're gonna have to take you behind the barn.
Adam:
Anyway. Yeah. So lay out the, the idea of role playing.
gptboss:
sure, this
Adam:
in
gptboss:
is
Adam:
terms
gptboss:
how
Adam:
of
gptboss:
my company
Adam:
prompt
gptboss:
works actually.
Adam:
and prompt engineering. Yeah.
gptboss:
Yeah, my whole platform is literally just this. So this is, I mentioned it right at the start, right? When we were talking about the slide, right? So these AI models, they have a large, very, very, very large training data set. And it's too much to think about, like a metaphor, right? We're going into the world of metaphor to explain this. There's too much data for them to reference, to think about everything in their data set at one time. So what your job is when you're prompting, when you're interacting with these systems, kind of like your first step, is you want to refocus the attention of the model on different parts of the training data. So if it has a lot of training data on copywriting, it might not always reference that when you ask it for details on a writing task, right? So if you just say, hey, can you please write me an email for my company, which is like a siding company, I'm doing a 20% off sale for this week only, and I just, need to get a job, right? So just write an email for me. That result is going to focus on siding and writing, and it's going to refer to other siding company emails that it's seen before. However, if you, before you get to the task, you say, this is who you are, this is your identity. That refocuses the attention in the latent space inside of the training set to reference copywriting tips. And so when it's remembering these copywriting tips, it'll use them in that email, and so the email ends up being better. This is from it's on open AI's website. Open AI explains you should do this every single time you're interacting with these systems.
Adam:
care.
gptboss:
And something, an hilarious example of this is I have a customer who runs a company called three B's digital marketing, like the animal B B E E.
Adam:
Yeah.
gptboss:
And so when it's writing copy for her, because that's the name of her company, it makes B puns, right? It's like, we know all of the correct buzzwords for your industry.
Adam:
Plus words.
gptboss:
We're reaching out to the hive mind and stuff like this, right? So,
Adam:
Yeah.
gptboss:
this latent space identity crafting is a really, really important part of prompt engineering. So you can do this with Chat GPT, just Chat GPT blank, because Chat GPT pays a lot of attention to what's going on inside of the chats. So the very first chat that you send it, say, for the rest of this conversation, you're acting as an expert copywriter. And it'll kind of keep that in mind. It'll start to, like, reference tokens that have to do with professional copywriting. And so the rest of the chat, it is going to work like this. It's not as good as what I do with GPT-BOSS, which is to use, there's a special identity field inside of the programming that developers have access to that works specifically for GPT-4. So, um, if you want to do these role-playing things, use GPT-BOSS, especially for business tasks. But if you just want to try it out with GPT to see what other kinds of ways you can take this, for example, you're a country music composer, or you are a fashion marketing director, or whatever it is. Like if you want to try alternative things, you can just go and... and do this exact prompt inside of ChatGBT and it's gonna work really well for you.
Adam:
I love the idea of refocusing attention. But it's. It's always funny how much we, we, as we start talking about how to work with the algorithm, it turns into a, like we end up using human words, right? Governance attention, stuff like that. Right. But if, if the metaphors are right and they help you understand, Hey, whatever. Uh, but it's, it's, um, right way of thinking about it, I think. Uh, or, or it's a way to think about it that helps you get to understanding what's happening with the emergence of AI and working with the tools and stuff like that, the, the broader thing is it's generalized, right? You're talking about focusing the attention, but like, this is the part that, that caused us to stick in my mind in terms of how to interact with things like, like GBT, uh, Was like these, these it's the agentic part, basically, like it maps to human society in a sense where we would plug a person as the discriminator, like we talked about before. And so we're so used to going to lawyers that when this magical, you know, like answer portal appears, we don't remember that if you're talking to a generalized thing, we achieve the. Attention focusing in this sort of like query dynamic by hunting for a lawyer and then talking to a person on the phone in your district, blah, blah, blah. You have to remember that and do the same thing with the AIs or with the, with the, you know, with the, with the algorithmic generative, whatever we want to call these things. I don't know if the AIs feels weird, but yeah,
gptboss:
Yeah.
Adam:
you were the best copywriter in the world, right? Uh, you, you, and you can put in stuff like that, right? Like, like it's talking about a corpus where it's just trying to organize this stuff at the same time that it surprises provides answers. So there are places where as you have more experience, you'll still have an advantage, right? If you have a framework memorized for deploying marketing projects, you can tell it that framework because thousands of blog posts have been written about any of these frameworks, right? Write me a plan for an agile based blah, blah, blah. Really good at that. It's because it's interpolation, like you were saying before.
gptboss:
Yeah, so
Adam:
It's
gptboss:
it's
Adam:
also
gptboss:
already been
Adam:
really
gptboss:
trained
Adam:
good
gptboss:
in
Adam:
at...
gptboss:
Agile, and it's interpolating it to your specific request.
Adam:
And it's also really good at falling back on credential-based merit structures. So you can tell it grade levels and things like that. Which sometimes is a hack, right? I'm not targeting fifth graders, but by saying to it, fifth grade. I can calibrate for a level of simplicity in explanation, right? Sometimes they say reading levels or what?
gptboss:
I have two tools like this on GPTboss. There's one at the fifth grade and one at the third grade level. And so the fifth grade one is for like, when you're explaining what it is that you're marketing, you have a curse of knowledge, you know your product really well, so you're going to use like jargon and stuff that you're familiar with to explain it to people. So rewriting it into the fifth grade level makes it digestible for people. That's the point of the fifth grade one. The third grade one points out logic errors. The language is so simple that if you made a mistake of like attribution or something doesn't follow logically. it's going to be extremely obvious. And I think that is such a remarkable skill to have, right? Because if you know so many things, if you're so smart, explain it to a third grader. And a lot of people can't. And being able to see that happen can start poking holes in your theories and help you do better research.
Adam:
That's role-playing. Past that, where we're way too deep in prompt engineering things. But back it up. Remember that part. You got to tell the bot who it's supposed to be.
gptboss:
a sign and identity, yeah.
Adam:
Yeah.
gptboss:
And I think that's our show.
Adam:
Absolutely. Thanks, everybody, for joining us. Thanks to anyone checking out the livestream. Let's go to script. I didn't write anything crazy. Okay, that's Accelerate Daily for... Whoops! It says Tuesday, May 30th. Tuxen... I'm not smooth enough to just fix that in my head on the fly yet. Okay, that's Accelerate Daily for Wednesday, May 31. We put a question mark on the teleprompter. Okay. That's accelerate daily for Wednesday, May 31st. And this show has grown up. So it's time for bake sale. Uh, at some point I'll move on from this talking about bake sales angle, but it's only been a couple of days. Uh, yeah, if you're watching and listening, um, Jump in with a like, subscribe, comment, whatever it's, it's. The part where we ask for it and talk about it and stuff like this, we call it bake sale sometimes because. like, you know, your kid's selling cookies to pay for their... football jerseys. Um, but our version is attention and engagement. So we can make the product better so we can rise in the algorithmic rankings. If those gods see fit, uh, but anyway, those metrics really do make a difference when it comes to reaching more people, uh, working on the future of AI. So, uh, and if you can make schedule work, jump in the live stream. Uh, we're always watching chat. Otherwise, that's our show. Thanks again, everybody.
gptboss:
See y'all tomorrow.