OpenAI’s latest GPT-4 upgrade has Monkey and Badger both amazed and a little concerned. We dive into the mind-blowing new features: human-like voices, lightning-fast responses, and conversations that flow like never before.
If you’re a math student or a science student, great news! Your life has just gotten way easier and potentially more optimized if you learn how to use this tool correctly!
If you’re an investor in a company like Duolingo $DUOL, caution is advised! Badger notes that learning languages through gamification and a fancy-looking app will no longer be sufficient, while Monkey wonders whether Adobe $ADBE is in a similar quandary; memories of Kodak $KODK linger.
Is the future of therapy and coaching … robotic?
Join us as we explore the potential and pitfalls of this groundbreaking technology. Is it time to rewatch ‘Her’ with a newfound sense of unease, or excitement?
We discuss reasons to be both sceptical and optimistic.
#GPT4o #GPT4omni #ChatGPT4o
—
WSW E27 – OpenAI
===
[00:00:02] Luke: Hi, and welcome to the latest episode of Wall Street Wildlife. Krzysztof and I just watched the OpenAI demo, their spring release, and we have thoughts about their new product, ChatGPT 4o.
[00:00:18] Krzysztof: Many thoughts, but before we get to the open the AI and all our hot takes. I’m shocked to come back from my week long retreat, find that you did not burn down the entire Wall Street Wildlife podcast and brand in my absence. I thought I was going to come home to, I don’t know, digital springs hanging out from everywhere.
how did it go?
[00:00:42] Luke: Uh, yeah,
took a week off from Wall Street Wildlife and I, uh, managed to badger my buddy Albert from the Telescope Investing Days. He joined me on the pod, was an excellent stand in for yourself, and we did a Telescope Investing Reunion episode as a one off.
[00:00:59] Krzysztof: Okay, well, I was also worried that when I tried to log back into the Wall Street Wildlife podcast admin stuff, I would get the, sorry, your password no longer works. But I still have a job, right? I still
[00:01:13] Luke: You, yeah, you absolutely do, my friend. It’s good to have you back. It was great to chat to Albert, but it’s good to have you back on the show.
[00:01:20] Krzysztof: you say the sweetest things.
[00:01:22] Luke: Did you enjoy your, uh, week off? You were off doing fun stuff at a zen retreat
[00:01:27] Krzysztof: Yeah, it wasn’t so much fun as it was profound and meaningful. I was in a completely beautiful part of the world in Hawaii on Malachi, uh, cleansing my spirit and soul. And we have recorded an episode all about that. To launch in some not yet determined time, right?
[00:01:51] Luke: unexpected vacation and he left me with like 24 hours notice going, ah, we have no, we have no episode. He’s like, I’m going to go and do some Zen for a week. I’m turning my phone off now. so yes, so we have recorded an episode just about, uh, the emotional journey of unplugging and doing Zen and yoga and stuff, and we will release that at some future date.
But today is all about chatGPT4o.
[00:02:17] Krzysztof: So we just watched the live demo, so you didn’t have to, and what do we discover? So the first thing let’s, let’s maybe go play by play in, in right? So the first thing they demoed was, uh, an ability to talk in real time to the platform and they, the hosts were attempting to get the GPT voice just right in terms of emotional valence and in terms of tone and in terms of I guess getting it as Uh, fine tuned to their specifications as, as possible.
What did you think of that?
[00:02:59] Luke: You should, you should go away and watch the demo. If you’re interested in this topic, it’s very interesting. And yeah. It’s, I mean, it’s clearly it’s not having emotions. It’s not alive. Right. But it was faking emotions. So accurately. And as you say, that part of the demo, they were getting it to tell a story.
And, and then the host was like, no more emotional, no more emotional. And it was really responding and becoming this, you know, you can imagine this kind of creature behind the screen being very animated. Uh, and really doing a bang on job of sounding like an emotional human being.
[00:03:33] Krzysztof: Right, so Hokie in one respect, because it’s a cool toy, but you know, how often are you going to play with an avatar like that? But, I think impressive on another level because of what we experienced as almost no latency. So they were talking, the hosts were talking to the interface and in this case it was a female voice, female sounding voice, trying to be dramatic, more dramatic than they turned her into robotic, dramatic voice.
But the changes were nearly instantaneous. So it wasn’t like you had to write down, read, download stuff, close apps. It was like, Oh, let me make this adjustment and tweaking whichever way wanted one
[00:04:14] Luke: And, um, and yeah, so that, like, as you say, that latency was, Almost nothing. And I guess they are like direct, they almost have no latency, right? They’re probably sat in some part of their offices directly plugged into their servers, you know, you can play online games with like super low latency.
So I’m assuming that when they release this to the public, it will be almost as quick. But the thing that also impressed me was like the model is continually listening. So they were just interrupting, a bit like human flow, you know, you suddenly start waving and talking over each other and you just kind of navigate a conversation.
And I’ve used chat GPT for voice. Uh, and I’ve done a couple of demos for like my parents, my dad was really impressed with it. But yeah, you had to like, wait, you know, you said something and then the model has its turn and you kind of wait for it to finish speaking, then you speak again. And this was just like a back and forth conversation.
[00:05:03] Krzysztof: use case. I see it in my future, immediate future. Once this is released to the public is I’m going to take a wall street wildlife podcast episode or a link. I’m going to feed it to GPT and I’m going to ask it to give me a sexy badger voice. I’m going to, I’m going to give me, Give me the badger, but make him a lot sexier, uh, for, for nighttime lullabies.
[00:05:35] Luke: What do you mean my voice is not sexy, but two,
[00:05:40] Krzysztof: Hard to imagine, right?
[00:05:42] Luke: but two, uh, so assuming they, they follow a similar pattern they did with chat GPT four, you’ll have like a bunch of preselected voices you can pick from. So it won’t be able to simulate my voice. Um, like at the moment, the state of technology, I gather I read a few days ago is now there are AI models that can reproduce your voice almost pitch perfectly accurately with just five seconds example of you talking.
Um, and I, that stuck in my mind because I’m thinking, well, you know, you know, that famous scam where it’s pretty low tech, but someone just phones up and says, Oh, you know, you Uh, you know, pretend to be like a grandchild and, you know, granny, send me some money. I’m stuck somewhere. Well, you know, it’s all done.
It’s done with like text messages or on like a bad line. Now you could literally just grab anyone off a podcast or anyone off a phone call, you know, call someone, have a conversation with them, record it. Uh, even if it’s like, Oh, sorry, it was the wrong number. You’ve got five seconds of them chatting, that’s enough for you to then simulate their voice and call their grandparent and try and, uh, extract some money.
So these, this technology is now so good. I’m about to establish a family password. Um, so whenever I talk to my folks or they talk to me, if there’s something kind of financial going on or something serious, we just know to check in with each other and say, what’s like the family code word.
[00:07:04] Krzysztof: And this is one of the benefits of having a criminal mastermind as your coast co host, you’re one step ahead of, of the dark side. But yeah, uh, We’ve been talking about this stuff for over a year now, but the reality of, of its potency is getting more and more real and sophisticated and absolutely you had better bet that the dark forces will use this from their nefarious end.
So any listeners you, you know, with regard to your own passwords and security, you got to really start getting on top of this stuff. It will be, it will be taken advantage of in the near term.
[00:07:44] Luke: For real, but let’s not go back into cybersecurity. That’s a whole can of worms. Let’s come back to chat GPT 4. 0. So yeah, it was like. Interrupt it, super low latency, back and forth conversation, simulating emotions. Something I noted, like, uh, the guy, so it had vision embedded as well, which is in the existing chat GPT 4, but it’s all kind of tied together in a really nice way.
The guy was like, he pointed the camera at himself and he was like, like, and he made a big kind of cheesy grin and looked excited and chat GPT was able to tell him, uh, what his emotional state was based on what it was seeing.
[00:08:18] Krzysztof: Yeah, that’s a little bit of a neat trick. I’m sure there’ll be interesting applications, including some for therapy. that you, you mentioned. Let’s return to that if we can, uh, the therapy angle, because I think that’s a nice way to end the episode. But in the meantime, sequentially, the next thing they demoed was one of the guys writing a linear algebra equation, three X plus one equals four on a piece of paper and asking GPT, interestingly, not just to solve it because that’s old hat news now, but to help him.
Think through the process by which he could solve it and my immediate reaction to that was oh man, like being a high school student in 2024 is a whole nother world than when we were in school back in the 15th century like, you know, I mean really name, name a, any kind of problem, physics, math, whatever, where, you know, being stuck in the cubicle, you’re scratching your head.
Like, damn, I don’t get this. You, you, you, you show your textbook to the camera and you have this mastermind that will help you think through it. It’s amazing.
[00:09:33] Luke: So like, if you’re a math or really a tutor in anything, like, is your job suddenly on the line?
[00:09:40] Krzysztof: I, it seems like, Anything that is, right, anything that is knowledge based that can be, you know, written, like, facts, factual things, so I guess math and, right, those kinds of tutors, sure, I think are in great danger. I guess we could talk about language tutors, right, because that was another demo.
[00:10:03] Luke: I, uh, let me jump in there. Cause I came out, uh, on the 7investing discord maybe like a year ago. And I said, um, Oh, I’m going to short Duolingo. And I bought some puts on Duolingo at the time. Um, cause I said like this large language model technology, and this was back in the days of probably chat GPT 3 then, maybe, maybe four hadn’t quite come out.
But I was saying Duolingo and other language learning tools are going to get massively disintermediated because suddenly. Like I don’t have to get, I’m trying to learn Spanish. I don’t have to get like a Spanish teacher at so many dollars an hour. I could just chat to the LLM and I could tell her I want to learn Spanish and it could kind of help me along the way.
And. Maybe that, maybe that seemed like it was a bit further away, but the demo we just saw was exactly that. Some canny audience member posed that question and ChatGPT acted as a, uh, translator between an Italian and an American host and did a, a quick, but pretty effective job of helping them speak to each other in their native languages.
[00:11:05] Krzysztof: Right, so your immediate reaction was, and remind me, are we still a family friendly show?
[00:11:11] Luke: Uh, yeah, here we are.
[00:11:13] Krzysztof: We are. So, uh, Badger’s reaction was, Duolingo are, and it rhymes with a duck. So,
[00:11:22] Luke: Yeah.
[00:11:23] Krzysztof: I Now, let me just, let me just say, because, uh, Duolingo, I actually sold my puts at a loss. Uh, cause I came around to the idea that actually Duolingo have a partnership with, with open AI specifically, and they’ve embedded chat GPT into their interface. But this is interesting, right? Because maybe I’m going to go back into buying some more puts mode, because like the whole, the whole thing with Duolingo is it’s a very pretty interface.
[00:11:49] Luke: And you know, I’m, I’m like on a 520 day streak now on Spanish. Uh, and it’s like the engagement and the gamification. I know, I know Duolingo is like a bridge to get me to the point that I’m good enough, I have enough vocabulary to talk to a real native Spanish speaker and then, and then go into like immersion mode and try and improve my language.
So I was gonna get like a thousand words of vocab and do that because I don’t want the cost. And also I don’t want the embarrassment of just being like a inept person in this like one-to-one conversation with a human being. I would be happily to chat to chat GBT right now ’cause I know that’s not being heard by any other people.
And I could be as inept as I like. So yeah, maybe, maybe the, maybe the LLM has taken the technology out of the way. You don’t need the gamified fancy. You literally just talking to a person,
[00:12:40] Krzysztof: right, so a couple of things. I’m trying to learn Italian now. So it was a nice coinkydinky that they used Italian on the demo. My understanding of something like language is an interesting use case because like you said there are levels of
Legitimate mastery versus something like just learning vocabulary that is nowhere enough to fluency and It makes sense to me that Duolingo, like you described it, is a small step that builds vocabulary, right? But it is nowhere near what is required to spontaneously speak to a human. However, if a language model, like ChatGPT, could obviously handle the sophisticated conversations, then obviously you could also Work backwards and tell it to teach you 20 vocabulary words a day, right?
Using say, give me spaced repetitions in terms of, you know, memory. Like here are words I memorized last week. Help me remember those right. And building a roadmap like that. So,
[00:13:47] Luke: I mean, just, just on that, if it’s got a big enough context window, it will remember the conversation you have with it. As long as you reopen like that chat, I don’t know how they’re going to implement it, but like, I often go back to, I’ve got to have like multiple long running chats with an LLM and I’ll go back to an old one just to continue the conversation like weeks later.
So you could do that. You know, you have your Italian learning context and it knows where you got to, you just pick up where you left off.
[00:14:14] Krzysztof: so I’m with you on, on this Duolingo’s doomed call. It was spontaneous. Uh, when, when you said it, however, I don’t know, I don’t understand this bit. And maybe you could help. Help me help clarify things. I made a similar call with regard to Adobe and once we saw the image generators popping out mid journey and Dali and all that.
And yet, as far as I could tell, Adobe as a company has continued to perform on, on. Well, and they’re sort of also integrating AI into their products. So I don’t quite understand. I don’t know if it’s as easy a call, whether it’s this black and white, that you have a leader in a certain niche and that, that, that competition will eat it up.
[00:15:06] Luke: I think they’re fundamentally different things. What, um, cause I agree with what you said in terms of what happened. Adobe is, let’s say the products are all about editing photos and maybe also to some extent editing video if you’re using like Premiere and After Effects and other parts of their suite.
And although ChatGPT can create photos and create video, but that’s not public domain yet, um, uh, and create text. Like the, they focused in the demo today. They really focused on human conversation, talking back and forth. So one of the big challenges say in replacing something like Adobe is it’s quite hard to talk to an LLM because it’s a large language model and use language to describe, okay, you gave me this photo, but I want you to make like the sky bluer and like, Make the guy wear glasses and put some like seagulls in the background.
I’ve just, I’ve tried doing that. You know, your ability to edit a photo using language, it sort of works, but it doesn’t quite work. You go around the houses a little bit. You never quite get what you want. because that’s the complexity of trying to use language to describe the changes you want in a picture.
But if you’re just talking to something and it’s talking back to you, well, you’re just talking, right? It’s not like you’re trying to edit something, you’re having a conversation. So in some ways, obviously it’s incredibly complicated, but in some ways it’s much easier than mixing these modalities.
[00:16:28] Krzysztof: okay. So is the takeaway that Duolingo, because it’s accomplishing a more linguistic task, more block like approach to language that could be disrupted way more easily than something that requires a lot of sophistication that Adobe provides. And even if both companies integrate AI into their products, the one will be more likely disrupted than the other.
[00:16:59] Luke: Yeah. Because let’s not get stuck on Duolingo because that was just one little example. And I think I’ve got a bigger. position here about anything that involves talking, but, um, but yeah, you know, fundamentally Duolingo is trying to teach you to speak and listen. And so that’s best achieved through a conversation, right?
Writing stuff down is maybe an intermediate step to help things stick in your head. But I just, I mean, ultimately my goal is to go into, you know, into a Spanish language. environment and chitchat. So you’ll get there by chitchatting.
[00:17:31] Krzysztof: Right. Okay. So we’re making a call here that if you’re a stock owner of Duolingo, be careful.
[00:17:38] Luke: Well, let me broaden it because I don’t wanna get stuck on Duolingo. I think if you’re a stockholder now, having just seen that demo and I’m quite animated and excited, hey Luke, turn the emotions down. Uh, uh, I think potentially you’ve just been disintermediated if your mode of engagement With a person is conversational because suddenly this thing can, it’s obviously got the intelligence and the deep intelligence of a large language model without really any true intelligence or comprehension.
And it can fake emotions and be, you know, very engaging. If you hadn’t told me that that was them talking to a piece of technology, I would, I would a hundred percent believe that it’s just two humans chatting away. And with the latency and with the talking over each other and the human’s nature of the engagement, I think almost anything one human can do for another over a telephone call, let’s say, uh, within reason, obviously hallucinations and all the other problems we know about, which haven’t gone away, I assume, um, I reckon there’s technology potentially dissing to me.
It’s all of those. So language learning. Yep. Um, tutoring. Yep. Yep. Yep. Um, well. Coaching, mentoring, uh, life coaching, psychotherapy, business
mentorship.
[00:18:54] Krzysztof: right. So this is where I push back just a little bit because. Of my own, , experience and expertise I think categorically makes sense to me that there will be levels of disruption, but to complicate things, there are multiple kinds of safe therapists and doctors and, uh, artists, right?
It’s not one category does not define all of them. I agree with you that perhaps the lowest tier of let’s say coaches or therapists will get disrupted if all they’re doing is they’re performing like reciting facts back at you, that kind of level. Uh, so give me some good advice on how I could. deal with my anxiety, and that level could be replicated pretty easily. But once you get into maybe the middle or higher tiers, I know that human beings, because we’re biological creatures, a lot depends on knowing you’re being seen and heard by another human. So even if, let’s say, the words themselves are exactly the same, the impact of knowing you’re talking to a robot versus another human is not the same.
And there’s still some embodied stuff that comes back to this in a, say, live session. There’s a big difference between like putting your hand on somebody’s shoulder has a somatic effect that can rewire someone’s emotional state. Whereas obviously a robot can’t do that on top of, you know, I think like some versions of genuine understanding that it’s not that language can capture, but there’s a creativity involved that I don’t think is. replicable yet in, in any AI models, even though it’s kind of correct. Like, you know, when we’re playing around with Shakespeare sonnet, give me a sonnet in Shakespeare’s voice.
Yes, those imitations are really impressive and fun, but they’re not yet actually Shakespeare.
[00:21:01] Luke: I’m obviously very excited about what we just saw. But it’s easy to forget all of the other deficiencies in large language models. Like they’re not. they’re not conscious, they’re not intelligent, they’re just mathematical models. And I assume ChatGPT 4. 0 is still going to hallucinate to some extent in the same way that all the other LLMs do.
And if you were, yeah, so if you were asking a really deep psychotherapy question, or you were maybe asking a really complex math problem that involved multiple, multiple steps, it’s probably going to hallucinate and make stuff up along the way that could be quite unhelpful. So you’ve got got to have that watch word in mind and maybe That becomes even more dangerous because now it’s based on the demo.
We saw, at least it seems so human and so engaged with the, the human participant that you’re probably even more likely to forget that what you’re talking to is a highly intelligent inverted commas, potentially pathological liar, who’s just kind of making stuff up as it goes along. So as long as you keep that in mind, uh, but still revolutionary technology.
[00:22:06] Krzysztof: yeah, yeah. Another example came to mind. Imagine legitimately experiencing a crisis of identity or existential crisis, which I think all humans are heading toward with this kind of technology, by the way, you know, like, what is the meaning of my life when all of a sudden, all these jobs are no longer available? And, and, uh, this is why we have philosophy and literature and so forth.
So imagine asking, uh, a language model like this, is there a God, is there a God? And you get some sort of really intelligent answer, right? That’s going to have some kind of feeling for you. But you know, it’s coming from, like you said, a mathematical process. Now imagine asking the same kind of question, of your Zen teacher or your priest or your philosopher friend, right?
They might give you the same answer or a different answer, but you know that’s coming from a certain kind of lived experience. And I’m, I’m, all I’m saying is we’re not, we’re cyborgs cause we’re dancing in this, We’re both we have our phones all the time now, but I just don’t think people will allow themselves to have the same takeaways from from the GPT as they would from a human, and that that gap that gap of bridge ability. is still years and years away, if not longer. even if the exact words are the same,
[00:23:45] Luke: Yeah, you’re absolutely right. Like these things, despite now having an even better illusion of feeling human, they’re not. They are just mathematical models. And, uh, like we’re talking now about.
Like coaching and psychotherapy, clearly there is a very significant difference in talking to a human coach as opposed to talking to, um, a piece of technology. It does make me reflect on the movie Her, , Scarlett Johansson and, uh, who’s the guy?
[00:24:16] Krzysztof: uh, uh, Joaquin
[00:24:19] Luke: Joaquin Phoenix. Yeah, great movie. Um, this definitely feels like it’s taken us yet another step closer to the technology in that movie.
[00:24:27] Krzysztof: It’s a movie I think I’ll be rewatching way sooner than later after today’s talk because my memory of it, it’s been many years ago, but I remember being deeply moved by that, by that film, because I think it wasn’t a caricature, even though however many years ago, it felt like, Oh, yeah, this, this is, this is like from another planet can’t possibly be a reality, whatever, now that it really is a reality, the way that, yeah.
The way that humans are, we’re exceptionally good, obviously, at having adapted to a world in which digital things appear real to us. I mean, we’re talking digitally now, right? but our, But the way our minds work, we, it’s not as foreign as, as we thought. And of course, COVID and Zoom.
Made all of that accelerated all of that even more so, you know, to be slightly optimistic about this, there is a lot of loneliness out there for all kinds of reasons and any tool neutral in and of itself that could potentially bridge some previously unbridgeable gaps because of their higher capacity to, uh, feel human, I think is an important technology to take note of and the consequences of it, you know, we’ll find out as they happen, but it’s not all dystopian, right?
[00:25:50] Luke: No, I agree. Like with any step forward, especially this feels like a serious step forward actually. I love the way it’s making technology invisible. , and I think that has very beneficial outcomes potentially, but also comes with a whole bunch of risks like cybersecurity all sorts of other things.
I’m all for embracing these new technologies, but we have to do it with a dose of caution at the same time.
[00:26:14] Krzysztof: And as investors, I think the takeaway here is as with our duolingo example, there will probably be many companies in the crosshairs. And I like, I always like thinking back to Kodak that, you know, Kodak is a giant company for a long, long time until all of a sudden a new thing like digital cameras shows up and you do not want to be the stock price of Kodak.
I mean, it’s almost like overnight the market realizes this is a dead end. So there will be many companies like that. So caution, extreme caution is warranted when investing, I think, in today’s landscape, especially if you’re in the obvious crosshairs of learning language models.
[00:26:58] Luke: And one thing we didn’t say, but we’re all going to get a chance to play with this stuff. So I know you and I used to be paying subscribers to chat GPT for I think we both retired our memberships. I’m now mostly using Gemini. Um, but when GPT 4. 0, the new version gets released, it sounds like it’s going to get released to everybody, to the public, as well as the paying users.
So and available in 50 languages and well, I mean that attempt at Italian sounded pretty emotional and human too. So this technology is going to be very widespreadly available, exciting stuff.
[00:27:30] Krzysztof: Okay, I think that’s, uh, that’s all that we intended to talk about, right?
[00:27:35] Luke: Uh, yeah, let’s check back in once this stuff’s live and we’ve both had a play with it. Maybe we can bring, uh, ChatGPT onto the podcast as a guest, have a chat with it.
[00:27:43] Krzysztof: Oh my god. Fantastic. So this has been episode 27 of Wall Street Wildlife. We would really appreciate a review if you received any value from our conversation. If you have not, let us know in your comments and questions because we aim to improve, but leaving a review on Spotify and on Apple would really be of great benefit to us and encourage us to keep going.
[00:28:15] Luke: Absolutely. Uh, if you want to chat with us, best place to find us is on Twitter slash X. I am at seven Luke Hallard.
[00:28:23] Krzysztof: I am at seven flying platypus. Leave us a comment on the YouTube channels as well, if that’s more of your thing.
[00:28:30] Luke: Very good. Are you ready to become a beast of an investor? Your journey starts here.
