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:
Okay, we're on. Let's do this. It's Tuesday, May 23rd. This is Accelerate Daily. Today we've got the weaponization of medical AI, open AI, talking super intelligence, micros warning to white collar workers, and a request for feature. We're calling it now for written style cloning. We'll talk through what we mean by that toward the end of the episode. Anyway, put on your goggles. Let's jump into it.
gptboss:
Bye bye.
Adam:
Welcome back everybody. I'm Adam.
gptboss:
My name is Mackenzie, good morning.
Adam:
How you doing?
gptboss:
Oh it is what it is, living the dream. Hehehehehehehehe
Adam:
I realized that like, and I, and I've said it to our editor on this recording, which I don't know if he'll cut out or not, but we should leave some banter in here, right? Like it's, it's a, it's worth acknowledging that the reason that we do a project like this is yeah, marketing, but also because we're both doing weird startup things in the AI space. And so
gptboss:
Mm-hmm.
Adam:
like almost every day there is a thing where I'm like, yeah, but did you have you thought about anyway.
gptboss:
Well,
Adam:
How you doing?
gptboss:
I got a one. I got a one to edit in. What's blue and not very heavy? light blue.
Adam:
Oh
gptboss:
Come
Adam:
Lord.
gptboss:
on. Come
Adam:
Hey,
gptboss:
on, Adam.
Adam:
I can get on board for daily dad joke.
gptboss:
Ha ha ha.
Adam:
Uh, I actually own daily dad joke dot email. And my
gptboss:
I'll
Adam:
plan
gptboss:
stop
Adam:
was
gptboss:
all these
Adam:
just
gptboss:
smokes.
Adam:
like, I'm just going to use GBTA to generate 365 dad jokes and have them automatically emailed out every
gptboss:
Yeah,
Adam:
day
gptboss:
we'll do that request
Adam:
for
gptboss:
for
Adam:
a
gptboss:
future
Adam:
dollar
gptboss:
tomorrow.
Adam:
per subscriber. There's there's there's good luck.
gptboss:
Hehehe
Adam:
Okay. Uh, lead image. So we decided we're going to start putting the prompts for these images in the description. So if you're on a podcast, wherever you are, check out the description. We're going to start pulling actual prompts for the lead images that we talk about here. Um, this one though is, uh, Edgar Allen Poe on the red carpet.
gptboss:
Indeed.
Adam:
Um.
gptboss:
I thought maybe it was like a sick Charlie Chaplin, but it was definitely
Adam:
Like,
gptboss:
Edgar Allen Poe.
Adam:
like, ill Charlie Chaplin. Yeah. Um, the, the prompt that's this is, it's funny pulling these from Reddit because then you have the comments. So check out the Reddit link to see people talking about this stuff as they work on mid journey images and stuff. Uh, they say, Hey, what was the prompt? And the guy says, basically, uh, it was just Edgar Allen Poe on the red carpet. And then
gptboss:
Yeah.
Adam:
you try it a bunch of times and it's like, there's five images that are all kind of like, yeah, okay.
gptboss:
Yeah, totally.
Adam:
That's definitely what I just said. And then there are also people there with very complicated prompts. So sometimes we'll also have prompts that are a beefier, if you kind of want to understand how prompting. Otherwise, let's jump into it.
gptboss:
Sorry, I'm having a technical difficulty. Looks
Adam:
I'll
gptboss:
like
Adam:
get
gptboss:
my camera's
Adam:
it.
gptboss:
a little blurry.
Adam:
Might just be the backlighting. You can workshop that after, or keep doing it. First
gptboss:
Mm-hmm.
Adam:
up,
gptboss:
Hehehehe.
Adam:
we got medical AI weaponization. Uh, this is from Axios. It says medical AI's weaponization.
gptboss:
Yeah,
Adam:
The graphic
gptboss:
this
Adam:
is
gptboss:
one's
Adam:
great. Uh,
gptboss:
awesome.
Adam:
yeah, you threw this one out. It's, it's a little heavy. So I thought we'd lead with it so we can taper off before we get to the silly stuff at the end.
gptboss:
Yeah. So, well, one of the things that they're talking about, this is not the way that I interpreted the headline. The headline, I think, is way scarier than the actual thing that is being talked about here, which is that the World Health Organization is warning about the risk of bias, misinformation and privacy breaches and the deployment of large language models into health care. So there's this is the ongoing problem of AI of like, who watches the watchman, right? Like people just refer to AI and they're like, Oh, this thing is a super intelligence. It knows all the facts. So whatever it says must be a fact. And so that it can be like a really big issue, especially with healthcare information, like it might misdiagnose you. And then the big problem with that is who is actually culpable for that misdiagnosis. We don't like turn the AI off. We don't kill the AI when it's wrong about one diagnosis. We don't take away its medical license. We just have to keep using it. So it's something that we need to kind of bake in at a model level or institutional level to make sure that these things have checks and balances.
Adam:
There's a portion of it though that talks about like using AI to sequence. Um, like you mentioned, you mentioned something about, uh, causing yourselves to produce THC.
gptboss:
Yeah, I can't remember the TikTok user, but this guy developed an invasion plan for Lichenstein. You can create a virus that causes cells to produce a specific protein and that protein can have magnetic activation requirements. So you could figure out the geomagnetic profile of certain regions on earth that cause a protein to kind of become activated. So this virus could spread all over the world and do nothing in any population until it gets to the region where it needs to be in. and then it activates. And okay, Kurt, if you want to go one more,
Adam:
I...
gptboss:
the same restriction on the geolocation is not the be all end all. You could also build a protein that only activates when in contact with a specific DNA profile. So you can have targeted individual viral loads.
Adam:
Okay. So the idea of how this mat, how AI matters though,
gptboss:
Yeah.
Adam:
is, uh, the massive complexity and trying to do that and maybe actually build it into a protein is the thing that AI possibly allows. Um, but that's a little bit more on the like, okay, that's terrifying, but also you gotta, you gotta look at the counter sort of hype side of it, which is, um, yes, but also because we have M R M R and a vaccines. We can do all kinds of crazy stuff that's going to be great for us. So you don't want
gptboss:
Yeah.
Adam:
to kill this thing, right?
gptboss:
Yeah.
Adam:
Even if you look at it through the doom scenario. Uh, but at the same time, like it, uh, it, it, the reason I throw on the goggles, right, uh, it's more in the space of like. forces us to look at the extent to which like a lot of medicine is just the coin flip. Right. Like like people can be better educated at how to recognize what's happening. But you have to like let's let's take it away from medicine which is like emotional and visceral. driving. Humans are not good at driving. People die in accidents constantly and we just go, well, it was an accident, but we hold machines to a different standard. And so a thing that I've talked about long before AI was in the picture about this kind of stuff, it's just like... not only are my kids probably not gonna drive around, they're not gonna be allowed to drive around.
gptboss:
Hehehehehehe
Adam:
Like, I don't know what the time scale is exactly, but like once the AI and the machines are better at driving than humans, we're gonna suddenly go like, wow, the motor vehicle death rate has dropped a remarkable amount. Like, and
gptboss:
Yeah, and this used to be science fiction as recently as like two years
Adam:
yeah.
gptboss:
ago. Have you seen Upload?
Adam:
Uh, uh, uh, no. Is that the one with the, like the, the paralyzed guy that has a like. Insert thing where the AI can control his body or something.
gptboss:
No, no. The guy
Adam:
Oh
gptboss:
he's
Adam:
no,
gptboss:
like
Adam:
upload
gptboss:
a rich.
Adam:
is the, yeah, the intelligence one, the afterlife, the AI,
gptboss:
The afterlife,
Adam:
the VR
gptboss:
AI afterlife.
Adam:
afterlife
gptboss:
Yeah,
Adam:
one. Yeah.
gptboss:
yeah, exactly. Yeah. So the reason that he dies is because he was in a car that had like an AI driver that refused. It was like hacked to refuse his ability to like override the car.
Adam:
Yeah.
gptboss:
So it crashed him because somebody changed his like targeting parameters. So there's yeah, this is another like having machines drive cars
Adam:
Yeah,
gptboss:
is another doom scenario.
Adam:
right. It is. But also we're going to hit a point where they're probably still better at it than humans.
gptboss:
Yeah,
Adam:
And
gptboss:
totally.
Adam:
and then it gets weird once you get a generation far enough away from how it used to be to to to realize like they might look at us driving cars around it like it's savagery.
gptboss:
Totally. One last thing on this Axios article, which is inverting rewards in AI systems. So there's a paper that they mentioned where some medical researchers took an AI that was designed to reduce the toxicity in the development of medicine and just inverted the reward. So the more toxic the chemical was, the more the AI was rewarded to create that. So in six hours, they created 40,000 new chemical weapons that all need to new kinds of detection algorithms in six hours. Okay, moving on. And
Adam:
It's goggles, for sure.
gptboss:
that was last year.
Adam:
Well, you know, it's Tuesday, so it feels like this one stays almost as heavy. Uh, super intelligence, open AI dropped a blog post where they really are actively introducing the idea of super intelligence as part of this conversation. And this is like, file this under, this is why we call it accelerate daily. Um, how quickly we went from. conversations about like, Hey, Siri sucks to the Overton window has shifted into like, okay, we have to have a sincere conversation about what happens once there are like a gentic actors who are a hundred, a thousand times smarter than any human has ever been. Um, like Elon Musk has been talking about it for awhile and they kind of look at him like a cook, uh, it's, it's. Here we are, right? Sam Altman just testified before United States Congress last week, and now they're dropping papers saying, Hey, we have to take this idea very seriously and how we govern these things and puts forward some basic ideas.
gptboss:
Totally. Um, I see it. Yeah. Like, I think that there's like some whiplash from like chat GPT a little bit where people were like, oh, this thing is crazy. Like I can like, oh, I could do all this stuff. And then now that we're like getting into like the request for features, um, or the neural network nourishment formerly known as, uh, we start talking about ways that this like, like new startups that can exist to like start to like automate these things. And so one, one of the like questions or like concerns that I have about super intelligence is like, is that a single model? Is there a single point of like, you turn this off or is super intelligence? the work that I do, where I take a generally intelligent model and plug it into so many places that it achieves this effect. What they talk about in the paper is that these AMP models would produce as much productivity as today's large corporations. Of course, referring to things like the largest corporations that we have today would be like Apple, Google, Microsoft, IBM, Raytheon. So imagine a single model producing that much kind of activity. And that's what they're looking for. And I think that like GPT-7 might pass that check, but also just GPT-4 plugged into enough plugins might do it too.
Adam:
The right it's, it's, it feels closer to the, like a Greg who type conversation where you imagine like. Is the ultimate answer for a lot of things more like fungal than it is, than it is, uh, you know, sort of. corporate. I don't know if that's the right way to say it, right? But like you're talking about sort of, we have these entities, we give them rights, like people, they're just groups of people currently. Um, when you look at it that way, you go, okay, so if we already have something that's up to humans, aggregate all of that, you know, um, The weird part is like, it's all crazy and interconnected because of the internet, which is the thing that like, when we invented how corporations work, it was like, we're going to put a little silo over here. We're going to have rules. You can do blah, blah, blah. You can go public. You have to have shareholder meetings, all that kind of stuff. Like it's possible that we just end up with a singular super intelligence. That's that, that is made up of the entire globe, which gets you to like weird Zen. like Gaia things where
gptboss:
Yeah.
Adam:
you're just sort of like, yeah, that's how nature works. This is just nature, nature, even though it looks to us like gadgets and technology, but another way to look at this is yeah, you're going to have a gradient, right? But it does, it sure does feel like at the top, you're going to have this thing that's built on an assemblage of decentralized capacitive engines that,
gptboss:
Yeah,
Adam:
that,
gptboss:
this is a good question. What is the body plan of a super intelligent organism?
Adam:
right.
gptboss:
Slime mold, I think is a good contender. Crab
Adam:
Yeah.
gptboss:
is another popular.
Adam:
Oh,
gptboss:
Anyways.
Adam:
anyway, that's where we are in the over to the window.
gptboss:
Yeah.
Adam:
Let's
gptboss:
Um,
Adam:
keep let's
gptboss:
so
Adam:
keep moving.
gptboss:
next story is Mike. What'd you call a micro? I always said this row.
Adam:
Mike Rowe. Yeah, Mike Rowe is
gptboss:
But
Adam:
how
gptboss:
there's
Adam:
you say
gptboss:
the
Adam:
it.
gptboss:
E. Like, what are they doing?
Adam:
I'm, you know,
gptboss:
That must just be my Canadian
Adam:
we're Americaning.
gptboss:
showing. Yeah.
Adam:
Not, that's not fair to say, cause America includes North America.
gptboss:
Yeah.
Adam:
We're United States thing.
gptboss:
Well, regardless, Mr. Michael, warning
Adam:
Anyway.
gptboss:
to white collar workers, the robots are coming for your white collar job. Yeah, kinda, yeah.
Adam:
Yeah, this
gptboss:
We've
Adam:
is
gptboss:
noticed this.
Adam:
Fox Business. It says Mike Rose warning to white collar workers quote, the robots are coming for you for your
gptboss:
But
Adam:
quotes, your white collar job.
gptboss:
this has been his position his entire career. And then just a quick aside, if anyone is at home listening and trying to figure out how to make money, go look up Mike Rowe. There's a, what's the word, an uneven advantage to doing these dirty jobs, especially local service jobs, which is that you don't have as much competition as these white collar jobs. White collar jobs, the results can be transmitted anywhere in the world. So your competition is everyone on earth, everyone with an internet
Adam:
Right.
gptboss:
connection,
Adam:
Right.
gptboss:
which is literally everybody. So because of that high degree of competition, even if the work is valuable, you're not likely to get paid that much for it. But installing a floor, there's only like 10 people in your city that could do that properly. So you become the 11th and then you join this elite mysterious cabal of floor installers that engage in like price fixing, right? It's just, it's
Adam:
Hahaha
gptboss:
way, if you want to make some money, like learn a blue collar skill, totally.
Adam:
Yeah. So, so, so yeah, for context that anyone that, that is not familiar, right? Mike Roe is the, the host, former host of a very popular discovery show called dirty jobs, where he just went and like, you know, worked as a pig farmer for a day or whatever. Um, but he's always been an advocate for the space and for learning the trades and, and all that kind of stuff. And you should watch the clip here because he, you know, his take on it is, is. is a relevant part of the conversation, which is like, there is, there is an opportunity here, right? We talk about retraining and re-skilling and people are always sort of like, uh, but that doesn't work. Okay. But there's another way to look at it, which is there's an opportunity here to grow into the space, like orthogonal to what's going to happen here and, and, and find the opportunity that's created. Uh, it's, it's just usually been the the, the, you know, it's been the realm of, of, Hey, I live in a town and they need people to dig ditches. Uh, I can charge money for that. And like, I think we've, we've, we, with our phase of the industrial revolution, we've gone through a phase of not looking at those jobs as entrepreneurial like opportunities, but somebody in a town who's just like, Yeah, I can deliver your hay and puts together a business. It's not getting replaced by AI in the same way.
gptboss:
Those
Adam:
Um,
gptboss:
are the best entrepreneurial place, right? So American viewers, you might not know this, but the average home price in Canada is about double what it is in the United States, mostly due to restricted inventory. So there's a lot of houses that would be $800,000 in Texas here in Canada go for two, three million dollars. So you need to be a certain kind of person to live in a place like that. And the people that are that are not, they're not working at Microsoft, right? They're not working at Amazon here, not even in the tech department. They are running like an electrical installation company by themselves. And that's it. That's all they have to do is just like install some wires. It's like, do like definitely investigate these blue, especially if you're a young person, definitely, definitely, definitely do this stuff because it's going to take a lot longer to automate this kind of thing. And I can tell you that with a lot of confidence because I try to automate this kind of thing and it's a mess.
Adam:
And, and he's spot on in terms of the white collar, like just, it's interesting watching technology eat that space. Um, and it's, and it's right where I live. Certainly. Cause you know, marketing was, was safe behind this wall of, you know, uh, oh, you need people to do that. Turns out you don't adjust or die. And that's kind of part of what he talks about. Anyway, check it out. It's only like six minutes. Moving on. Uh, yes. Neural network nourishment, as we've been calling it. Uh, we realized like coming up with actual recipes that we're working on, probably not a sustainable project for the 550, 250 days a year that we need to come up with one. But what I do think about constantly is, Hey, here's something that if somebody built, I would use it. Um, we could also call this segment like, you know, giving away the good stuff for free, but like I can only build so many companies and there's so many widgets and there's plenty of design space here. So I think it'd be fun to talk about if other people have ideas they want to talk about or things like this that we can jam on how to actually build it. Uh, that's what this chunk of the show will be for, uh, this one request for feature written voice clone. So what I want to do here is, uh, build a system that would scrape a profiles, uh, social profiles, uh, like existing output, and then use it to establish a written sort of voice print situation so that I could then say, right. LinkedIn updates for my CEO that would be on voice, have them quickly approved, get them out the door so that we can better leverage the influencers inside of our organization or network. Um, How hard would that be to do?
gptboss:
So you want the voice output?
Adam:
No, no, I just want, I want voice from like a marketing standpoint. Right.
gptboss:
Right, like
Adam:
So
gptboss:
tone
Adam:
it's like, okay,
gptboss:
and like,
Adam:
brand
gptboss:
yeah,
Adam:
voice.
gptboss:
yeah,
Adam:
Yeah.
gptboss:
brown
Adam:
Tone
gptboss:
voice.
Adam:
and like, you know, do you use contractions or not, stuff
gptboss:
Yeah.
Adam:
like that. And a lot of times that's sort of an authentic thing that somebody will have built up over 10 years of social media. If they're
gptboss:
Mm-hmm.
Adam:
my age and have been running a company right now.
gptboss:
People change
Adam:
Um,
gptboss:
over time too. So that's kind
Adam:
yeah.
gptboss:
of like something that's questionable about this. And this is getting into some other things that we talked about, which is like AIs feeding themselves. The way that I would solve this is few-shot learning, which is a little bit expensive, but because it's for like a corporate thing, it like, you know what I'm, for you at home, the cost on this might be like 80 cents, a dollar a post, but for a corporation paying a dollar for a post, they like, they don't care because they were willing to pay. for like the CEO to sit down and do that, to take 20 minutes straight of post, they've spent $700 or whatever it is, right? So like the dollar starts to make sense for them, but if you're like grinding this out on your own at home and you're taking my advice and starting a flooring company, that might be a little bit too expensive. So the way to do it is to do the scrape, develop a database of all these posts and then sample, take some random indices of posts and then put them into a few shot training sequence together. So this rapidly expands the input that's going to these models like GPT-4. Um, but then at the end, the final prompt is like, can you write a new blog post on this topic based on that? Sounds like these previous examples of writing that I've shared with you. And generally, yes, it can do that. So this is, yeah, it's like kind of an expensive way to do it. The ideal way would be to take this stuff and fine tune an existing model to say that like, like we need to use a little bit of AI here to say like, the prompt write a topic about X, which is what the post is about produces that post output. So then in this way, we're fine tuning something to have this brand voice, but fine tuning is quite expensive. Continuing to use fine tuning is quite expensive. And it's difficult for you as the feature vendor to manage all of these different models because there's gonna be a lot of people that wanna just like try this out and then maybe not come back. So you've eaten the cost on the training that you can't recoup all the time. So that's why I recommend like few shot learning. That's the way that I would do it. But once you have an established business, then you would want to go and identify your power customers and build them custom models.
Adam:
That's from the context of you providing that, like, like your context, GPT boss, right? Is it worth building this as a tool? It's how you're thinking about it. Um,
gptboss:
Yeah, totally.
Adam:
The distinction. So I think maybe one thing to hit here that's interesting is the distinction between few shot and one shot learning. So when you say few shot, you mean you want to send a couple of successive, a few successive messages to the. Model to get responses back or to say, here's new context on this, like sort of chat thread, basically
gptboss:
Exactly.
Adam:
in chat,
gptboss:
Yeah.
Adam:
G J P GPT, um, versus one shot, which would be tried up. put a workable data load into the initial prompt as context, which hits upper thresholds on like,
gptboss:
attention.
Adam:
yeah, attention, like
gptboss:
Yeah.
Adam:
possible, like how much you can get in context
gptboss:
I, yeah, so
Adam:
and
gptboss:
one argument
Adam:
how
gptboss:
I
Adam:
much
gptboss:
have
Adam:
it
gptboss:
here.
Adam:
costs, right? Cause it's also going to be more tokens
gptboss:
One argument
Adam:
to process
gptboss:
I have here
Adam:
the bigger
gptboss:
is that you
Adam:
load.
gptboss:
do not need 10 years of history to get this working. You need like 10 posts, which could be the last week.
Adam:
That was going to be my next question. Right. So, so, so there is a one shot way to do this that might be good enough for some contexts, which is you don't even have to, you can do this manually, like just pull those posts, put them in a, in a spreadsheet or whatever, right. Do you know, you made a spreadsheet, put them in a single document. You can literally label them post one post two posts three, like the agent will understand and then say, you know, match the style of the above posts. Um, and just do it, but, but I present it as requests for feature because you could also just build in a thing where I give you a LinkedIn, uh, URL and then you, you know, prep prep the virtual CEO bot. So build that, I'll use it. Ha ha ha ha.
gptboss:
Yeah. I guess I got to spend some time. Sorry. Fox is like sending me an ad that I didn't click on. Um, okay. So, uh, I got to build this for, for Gary Vaynerchuk. Gary would love that if I made like a Gary bot. Yeah.
Adam:
Okay. Uh, well that kind of actual, and that brings up an interesting aspect of it to get into some of the stuff we've talked about. If we for sure don't shy away from on this show, I'll say, um, you could also use this maliciously, right? Which is where this is going to get interesting because if people can clone Gary V and start making videos that come out of that voice, um, we're going to have to figure out how to deal with that. Gary V says, I didn't say that. But anyway, stick around if you want to keep jamming on these kinds of things at the end of the show every day. Uh, this has been another accelerate daily. Thanks for joining us. Like subscribe, follow us wherever we are. If you catch, if you catch us on YouTube, we're we're live here. And I'm, I got, I'm always looking off to the left awkwardly. It's cause I'm watching the chat. So getting the chat.
gptboss:
Where's the damn chatters? Okay,
Adam:
Thanks.
gptboss:
I'll see you tomorrow.
Adam:
Thanks for joining us. See you tomorrow.