Adam:
It's Tuesday, May 30th. This is Accelerate Daily. Today we've got an early warning system to save us from the AI apocalypse and some AI images of what that might look like. The next step in the AI financial advisor wars, Altman's optimistic, optimistic, Altman's optimistic view of the future. And a great example of chat GPT is a tutor that isn't just about cheating on your papers. Get ready to step on the gas. This is Accelerate Daily. Intro music and stuff. Okay, welcome back everybody. I'm Adam.

gptboss:
My name is Mackenzie. Good morning.

Adam:
And we are back with three headlines and a how to, to keep you caught up on what's happening in AI today. Today's title card prompt for anyone watching the live stream or clicking through the description is a Canon photo of the blank after the apocalypse, abandoned, bombed,

gptboss:
the

Adam:
collapsed, dilapidated, overgrown, photorealistic, octane render, wide angle closeup, cinematic aspect ratio 16 by nine. We had a real full prompt for this one.

gptboss:
Yeah, this is how I structure my mid-journey prompts.

Adam:
Yeah.

gptboss:
Yeah.

Adam:
Um, the actual picture is a picture of the white house and it looks kind of like, you know, post independence day. Um,

gptboss:
It's

Adam:
uh,

gptboss:
post-war of 1812 when we invaded your house.

Adam:
yeah, the post, the, the, uh, max Canadian in case it hasn't come up yet. Um, the link to the series though is after the apocalypse and it's a bunch of landmarks. Um, all kind of with. you know, like mossy overgrowth and those vibes. Um, I appreciate the extent to which the prompts on this kind of stuff include like octane render, wide angle, cinematic, like they're, they're non, sort of non-specific terms for vibes that we understand you say cinematic and you're going to get a little more like orange and teal or whatever that kind of like Michael Bay. You know, thing is the octane render gives it a little bit of an unreal feel, but that work

gptboss:
Mm-hmm.

Adam:
makes it play as art instead of trying to be a picture. But then they also call it the call out cannon, like a color profile from a popular SLR, um, sensor that people are familiar with.

gptboss:
It's all just like token

Adam:
Super

gptboss:
association,

Adam:
cool. Yeah.

gptboss:
right? Like they, I don't know if it's specifically doing that color profile, but if a reference image is labeled, this is a Canon photo, then it allows you to access that latent space instead. A lot of prompt engineering is all about like, how do I get into these latent spaces that exist in the training set?

Adam:
super interesting. Okay. So before we jump in today's topics, we're growing up around here. So we got to do some plugin, like subscribe. If you're really into the project here, write a review on your favorite podcast app. I know you hear that inside of every type of thing like this, live stream, podcast, whatever, but it really makes a difference in terms of literally like how the discovery algorithms think about your show. So it's nice to get the feedback and it feels like I'm like I'm asking for praise. What we're really asking for is just, we need that juice in the algorithm so we can find other people for a community of people that want to talk about these problems. Literally every day. So we can, so we can not miss the boat. Honestly.

gptboss:
I liked and subscribed just because you said it. I hadn't put it to this point.

Adam:
Yeah, that's a funny part of this phase of things where like I go in and go, Hey guys, we're up to 22 subscribers, but like 70% of them are our friends.

gptboss:
Yeah.

Adam:
What's up everybody? Uh, Gary in the chat says, hi, what's up, Gary. Okay. Uh, let's jump into the topics. I'm gonna find the dick. First up, this one dropped this morning and preempted. So anybody that looked at the deck ahead of time on LinkedIn, this one wasn't in there. Sorry, didn't have time to fix it. Statement on AI risk. So this one hit all the big media outlets and all the various feeds. It's literally from a website, safe.ai, and it's got all these thought leaders who have signed off on it, but it's literally just 22 words. Like it feels like a check down where they did, they did a letter where they were like, Hey, we should worry about this. And here's a lot of words. And then on this one, they were like, maybe we need to make it a tweet

gptboss:
Yeah.

Adam:
and see how we do. But, um, let's just read the statement. It says mitigating the risk of extinction. Hey, mitigate.

gptboss:
I can take whack at it.

Adam:
Yeah,

gptboss:
You

Adam:
you

gptboss:
good?

Adam:
do it.

gptboss:
You got it?

Adam:
I've been talking too much already anyway.

gptboss:
Yeah, yeah, you're coming off a meeting. So the statement from the Center for AI Safety says, mitigating the risk of extinction from AI should be a global priority alongside other societal scale risks such as pandemics and nuclear war.

Adam:
I feel like it's it. The reason I was like, okay, we got to throw this one in is it feels like. Share it. Given the thought leaders who have signed off on it, which to be clear, you can also go sign it. There's a forum at the bottom, but you've got Sam Altman, Jeffrey Hinton, bunch of other people just putting this link forward so people can see it feels to some extent, like just. An important aspect, bringing it forward for conversation feels like an important aspect of. good faith participation in the community of AI builders. I wouldn't have to talk about it forever, but you know, check it out. I know you have a take though, so let's do

gptboss:
I don't

Adam:
it.

gptboss:
want to nitpick, but it's just, it's odd, right? This is like, it's weirdly short and it's weirdly worded. My tier list of societal scale risks, like definitely nuclear wars in the top five, pandemics in the top 10, but these are not like the two most important things, I would say. And additionally, these things don't have a good track record of being dealt with. So to suggest that we should treat... AI X-risk the same way that we've treated nuclear disarmament, to me, it sounds like they want to do nothing about it. Or they just want to scare you or like, I don't know, like,

Adam:
Yeah.

gptboss:
you know what I mean? Like nuclear war is a weird one to put out because we've done exactly nothing to prevent nuclear war.

Adam:
I think it maybe feels that way, but personally, having lived through the 80s, there's actually a lot of stuff in terms of test ban treaties and ways to handicap the capacity of countries to develop nuclear capacity willy nilly. So I tend to think that it's fair. It's just interesting watching it play out and sort of how fast it is. Like the people who are building these things going guys, the failure scenarios that our red teams are coming up with, I shouldn't say guys, but you know, I mean the failure scenarios that these red teams are coming up with. Potentially end up bad enough that we have to talk about things like, like, like, like in the chemical weapons and like genetic engineering space where we have an idea of. society level guidelines for the people working on these projects to make sure they don't go bad. That for sure exists with nuclear, right? So to that extent, it's relevant,

gptboss:
Yeah,

Adam:
but...

gptboss:
to tie it back to an earlier episode, we saw that there was a model that was able to produce 40,000 novel chemical weapons in about six hours by inverting the reward function. So rather than producing medicine that had a very low toxicity, by inverting that reward function, they developed medicines that had a very high toxicity.

Adam:
Right.

gptboss:
And so

Adam:
We have

gptboss:
each of

Adam:
ways

gptboss:
these

Adam:
to handle

gptboss:
novel

Adam:
this stuff is the

gptboss:
chemicals,

Adam:
point I'm making,

gptboss:
they

Adam:
I

gptboss:
need

Adam:
think.

gptboss:
new like detecting

Adam:
But,

gptboss:
algorithms, right? Like

Adam:
right.

gptboss:
this, the proliferation can just go crazy. And that's a very like,

Adam:
Right.

gptboss:
that's in my opinion, one of the benefits of AI is that you could do something like that. And we'll talk about that a little bit later towards the end of the show.

Adam:
Right.

gptboss:
Yeah.

Adam:
Well, so that's actually a perfect transition to the next topic. Early warning system, AI early warning system. This is a, this is a, actually a paper. It's a post on deep mind from Google's technical blog. Title is an early warning system for novel AI risks. Oh, I realized, did I ever read the headline on the previous? Ah.

gptboss:
Yeah. You read it before you transitioned over.

Adam:
In case I didn't, I'm going to do it again. Uh, here we have a post from safe.ai. Uh, it's literally just called statement on AI risk. And I think we'll just read the statement. Steve, you can cut that in after when Mac Mac takes over reading it.

gptboss:
Okay, so back to DeepMind.

Adam:
Okay. Yeah. We have the intro for this already though.

gptboss:
Yeah,

Adam:
Early

gptboss:
okay,

Adam:
warning

gptboss:
so.

Adam:
system, novel AI risks. So this is the thing about that level of conversation, right? Um, one of the things that we talked about throughout the eighties and the cold war scenario of nuclear sort of proliferation intention was, could you build an early warning system? The reality that gets to what I think you were saying, which is it feels like nothing has been done is, yeah, we would know when the thing launched. But short of shooting it down at that point, like, there's not much you can do. And it literally resulted in our parents doing like, you know, air raid drills and hiding under their desk, which sounds ridiculous in the face of an understanding of what's happening with a nuclear attack. Um, this is software, so it's a slightly different thing. And this is a paper that is, takes a look at, okay, what does it look like to build? Systems that can actually watch out for the things that are really the catastrophic scenarios here. Is that a, is that a fair representation? I think

gptboss:
There's

Adam:
you read it more

gptboss:
a

Adam:
deeply

gptboss:
link

Adam:
than I

gptboss:
to

Adam:
did.

gptboss:
an archive paper, research paper, with all references and building on earlier kind of stuff, where it explains why they wanna start worrying about this now. There's a nice graphic where it shows how the capabilities are growing, and they're about to enter a zone where through misapplication, right, misanthropic intent, the capabilities could potentially do something dangerous soon. And then there's also a list of specific capabilities that they watch out for, just briefly. Deception, persuasion and manipulation, weapons acquisition, long-highs in horizon planning, and situational awareness, I think are some of the most interesting ones. Situational awareness means that the model knows if it's being trained or used in production, and it can give different answers. So it would be able to deceive the testers and say, well, I know I'm being trained, so I'm gonna give the answer you want, but when I'm in production, I'm gonna get some schmuck to build chemical weapons for me. which feels like science fiction, but this is a deep mind project. They've maybe seen this in their red teams because they're dealing with systems that none of us have access to.

Adam:
It's, uh, the, the optimistic take is, but they're, they're publishing this, right? Like the important part is I do think downstream from comments like that, that I think play like FUD fear things. Like part of that is because you have to get the public to take it seriously before it turns into the political action that is inevitably needed to get everyone on the same page, whether they like it or not via the law. Um, so that there are systems like this that we can build in and say, here are these routes that if we see them starting to happen inside of a model or relative to a model's output, we're building in the right back doors and the right ripcords to be able to say, Ooh, that's not allowed.

gptboss:
Yeah,

Adam:
We're going to get fined

gptboss:
like

Adam:
for that.

gptboss:
one of the specific

Adam:
Right. Like.

gptboss:
technical implementation issues I see here and why I'm grateful that like the DeepMind team is working on this and it's not my job is it has to be deception robust, right? Like

Adam:
Right.

gptboss:
you have to, your tests need to account for the fact that the model might lie to you, which just sounds impossible. I don't even know how you like start with that.

Adam:
Well, actually part of it is a, it's, it's a perfect segue to the next thing, which is talking about a less. This level of the last two things we're talking about is about the idea of the more generalized side of the spectrum. At the same time, you're getting more specific applications, which gets us to the next thing. Indexed GPT from JP Morgan. This one is CNBC reporting. JP Morgan is developing a chat GPT like AI service that gives investment advice. We already talked about this because people have been doing like tests where they just see if chat GPT is good at picking stocks and it is, uh, but there's, I think a broader thing here, which is, you know, uh, JP Morgan is, is. They're aware of the middle person positions here, the things that can be disintermediated. Uh, so they're not necessarily eating their own lunch, but they're eating the lunch of. My aunt's investment advisor

gptboss:
Yeah.

Adam:
or whatever. Right. Um,

gptboss:
I have an unethical FUD recipe for this. May

Adam:
Go

gptboss:
I share

Adam:
for

gptboss:
it?

Adam:
it.

gptboss:
If I was JP Morgan, what I would do is I would give the model access to a database of all of the C-suites current positions

Adam:
Yeah.

gptboss:
and advise the clients to counter-trade me.

Adam:
lies the clients to counter trade you. And then

gptboss:
So

Adam:
you would

gptboss:
if.

Adam:
be hedging your risk at all times by telling a massive group of people to do

gptboss:
If

Adam:
the

gptboss:
I wanted

Adam:
opposite.

gptboss:
to like long Microsoft, for example,

Adam:
Right.

gptboss:
I would write in my database. I wanna open a long position on Microsoft and then the AI would tell everyone to sell Microsoft.

Adam:
Pretty sure that's why you have to pass a test to be an investment

gptboss:
Well, watch out for stuff like this

Adam:
advisor.

gptboss:
because it is unregulated, right? Like that's just some FUD. FUD

Adam:
Right,

gptboss:
for thought, I

Adam:
right.

gptboss:
suppose.

Adam:
Oh, that might be a new segment for thought. Part of this is they applied for a trademark. It'll be interesting to see if it even gets approved because chat GPT open AI has also trademarked any like GPT as prefix thing, I think. But like we've

gptboss:
suffix.

Adam:
talked about on here before, the interesting part is. This will work because that's what an investment advisor is doing. Mostly is pattern recognizing and saying, Hey, get in some low yield index funds for 25 years. That's the safest thing. Thank you.

gptboss:
Yeah. And

Adam:
I'll take a

gptboss:
that's

Adam:
cut.

gptboss:
what GPT models specialize at, absolutely, is finding

Adam:
Yeah.

gptboss:
the most median, middling kind

Adam:
Right.

gptboss:
of solution.

Adam:
Um, cool. Moving on from that one. Uh, yeah, I feel like this, I don't know. I, it's similar to wanting to share that statement. I feel like I also want to point out when the people who are often part of the, the subject of the doom headlines are also out there optimistic. This is optimism from Altman. Uh, it's a tech crunch post. Sam Altman shares his optimistic view of our AI future. And it is mostly a good introduction for our next segment, but it's him talking about medical advancement and scientific advancement and educational stuff and like Uh... for all of the scariness that we just talked about in the context of apocalypse scenarios. The other side of this, even more so than I think those scenarios are scary, this side is optimistic in terms of being able to massively pattern recognize our way to incredible advancements in science and educational tools that let human minds do better and crazier things than we imagined, you know, like before the smartphone existed, for example.

gptboss:
He specifically calls out education, which is, I

Adam:
Yeah.

gptboss:
think we don't want to spend too much time on this one, but an example of the education is in the next slide.

Adam:
But it's worth checking out. Um, there's a link to a bigger talk where he kind of talks about some of this stuff, which to be fair, if you listen to Altman actually on a podcast or anything, like he talks about all this stuff and just the headlines it gets distilled to by the New York times need those need those FUD clicks. So food, food, FUD, FUD for

gptboss:
Hehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehehe

Adam:
thought. Um, so the next one, which is in our usual Uh, recipes or how tos hacks kind of segment. Um,

gptboss:
Oh, we skipped

Adam:
uh,

gptboss:
over the Reddit.

Adam:
no, this is it. I put it in requests for future

gptboss:
Oh.

Adam:
because,

gptboss:
Oh!

Adam:
because they mentioned a bunch of different sort of recipe kind of things in the course of this post, it's literally just a Reddit post from a Reddit user inside of our slash chat. GPT headline says chat GPT is saving my life. And then they outline falling behind in a class that's very foundational. Like that, that's accumulative. And really struggling to catch up. And even before that falling behind because they were struggling to, to continue following the stuff and. And they outline a few different things that I was like, Oh yeah. I mean, everyone's talking about chat GPT as a way to cheat on your paper and not have to write your paper. Right. This guy is ingesting notes and then producing. quizzes and flashcards and downstream other explanations of what's happening and ways to test it and just having an ongoing dialogue with jet GPT to understand these topics. And then I think it's not in screenshot, but at the end reports that he's. I don't, I don't even know if it's a not fair. It says

gptboss:
I

Adam:
by

gptboss:
crunched

Adam:
deleted.

gptboss:
the numbers.

Adam:
Oh no.

gptboss:
I crunched the numbers on his improvement.

Adam:
Great improvement.

gptboss:
So

Adam:
Yeah.

gptboss:
yeah, he got a four out of 30, and then he went to a 23 out of 30.

Adam:
Sorry, let me chill

gptboss:
A little

Adam:
these

gptboss:
dog.

Adam:
dogs out. Yeah, I'm gonna take the dogs upstairs.

gptboss:
Okay.

Adam:
Say hi to everyone in the live stream for us.

gptboss:
All

Adam:
Okay.

gptboss:
right, I'll just, I'll keep it moving. So

Adam:
Hey!

gptboss:
yeah, I crunched the numbers and we had this four out of 30 to a 20, it looks like 25 out of 30. And that is a 638% increase in his understanding of organic chemistry. And organic chemistry is a famously inaccessible subject. It's like the meme subject that, you know, people fail in school in like high school and college media. So this 600% improvement in in one of the most inaccessible and difficult subjects, I think, uh, portends a lot of other intelligence upgrades coming in a lot of other regions. And this points back to why I think that this like chemical engineering AI that we mentioned earlier while threatening is a really good thing because on my tier list of like existential threats, the things that I have near the top is like water accessibility, agriculture, and energy. And so those are all like high hard science problems. And if GPT-4 is able to educate somebody who is interested in hard science, but not very good at it at a rate of 600% better results, better understanding. Um, then I have, I'm feeling really optimistic about those issues such as water, agriculture, and energy, which like are just are, are much more pressing, much, much, much more pressing.

Adam:
Yeah,

gptboss:
And

Adam:
this

gptboss:
I think

Adam:
is.

gptboss:
with enough people poking at them, they're easy to solve.

Adam:
Right. That is also part of what Altman is talking about. Right. Like he, he often repeats what maybe feels like a kooky claim, honestly, which is that like, uh, electrical power is about to become infinitely cheap. That feels crazy with the, with the squeeze it feels like we have on, but he's a little bit not wrong if we can make certain nuclear things work that we've been trying to do for a long time. And I think he's, he is placing a lot of bets based on the idea that something like GPT-4 can help us solve some things that we can't quite figure out there yet with like cold fusion, for example.

gptboss:
Well, we do have a working kind of system. It's just proliferation is difficult. It's called a liquid

Adam:
Right.

gptboss:
fluoride thorium reactor. Thorium is more abundant than uranium and it is way less dangerous. And so this could be a house-size appliance. And they've kind of been in development and patented for a while. But it's difficult to roll something like that out because it's scary and

Adam:
Yeah.

gptboss:
not enough people know how to work it. But GPT-4, giving a 600% increase to intelligence, I think will help people feel safer about having a somewhat It's like, it's in my opinion, it's about as dangerous as a water boiler. If that explodes, you're in just as much trouble as like a fluoride reactor. Um, but yeah.

Adam:
Um, anyway, the educational piece goes in the recipes section because follow the link to the Reddit post and they outline four or five different things that are literally just a matter of. Hey, stuck studying next year or trying to rethink how you study in school come fall, or if you're in summer school, totally different vibe, just saying these things to chat, GPT putting in the stuff and saying, Hey, can you help me make flashcards for this? can result in things that are otherwise possibly hard to do if you're lacking the understanding piece. Also possibly good exercises to do yourself, like making the flashcards is what always helped me in law school, not reviewing them. But I know that reviewing them was huge for a lot of people. The idea of that level of access.

gptboss:
Yeah, totally.

Adam:
Super exciting.

gptboss:
I've been, I've been looking at building a startup in this space and it's like, I, I think that there's going to be a lot of them, like a lot of competition because, um, curriculums are a list, right? It's a checklist. Just, just build like a database of like this user needs to pass this, this, this, this, and this, and then AI builds the lesson plan around those curriculum points. It's

Adam:
Yeah. One of my favorite

gptboss:
such

Adam:
demos,

gptboss:
a slam dunk.

Adam:
one of my favorite ways to demonstrate to like executives and stuff when I'm talking to them about this stuff is to, is to, is to eat that layer of what they would do during a year to go to chat GPT and say, okay, you give me a campaign you're about to do. And then I'll put in a prompt that says, uh, come up with a marketing campaign, like come up with a marketing plan for 12 leak week campaign optimized for signups based on this narrative and it writes a thing that usually that level of marketing manager would take a month to work on. And they go, whoa, what it's like, it's Keanu Reeves. gifs all the way down.

gptboss:
That's so true. So yeah, that's our show, I think.

Adam:
Anyway, that's Accelerate Daily for today. Like I said at the top, we're growing up here, so I'm gonna go back to the script. Ah, yes. Okay. Here's what I said. Okay. That's accelerate daily for today. For, yeah. Okay. That's accelerate daily for Tuesday, May 30th. And like I said, up top, the show's growing up. So it's time for, uh, what we call bake sale. And I'm going to introduce kind of in that way to show people behind the curtain of marketing things. Um, a bake sale is a type of campaign that I frequently run where it's literally about. Can we get the existing community to engage? Sometimes it means you're at a big company. You're trying to bootstrap a podcast and you literally need to make sure that the 10,000 people at your company are aware that they should go like and subscribe. In this case, it's more just for you, the viewer, like I said, up top. Um, if you're getting something out of this project, then, uh, like subscribe, write a review on the podcast channels. Those things really matter in terms of the discovery algorithms that are our real source of growth traffic. for a project like this. So throw it over there. I won't even ask you to buy any brownies.

gptboss:
It's free and easy.

Adam:
And if you can make the schedule work, jumping into the live stream every day at 1 PM Eastern 10 AM Pacific, uh, we do there, our whole production thing out in the open so you can watch me, uh, mess up takes and

gptboss:
Okay.

Adam:
start over again. Anyway, thanks for joining us everybody. We'll

gptboss:
Yeah.

Adam:
see you tomorrow.

gptboss:
Yeah. See y'all.