GPT-5 | The Brainstorm EP 97

0:00 Welcome to the brainstorm episode 97 and it's all about GPT5 and we're going to look past the charts that were mislabeled because we don't know that's already been talked about. We're going to dive deeper um and talk about is chat GPT5 the future? What it's done? What are some of the flaws? Um Brett, do you want to lay it out first? And what was the big was there a big step here? I think that's even the the the question. Is it a big step to have a unified model? Um or is this kind of incremental catching up to technical debt that's been embedded as they've kind of spedrun

0:48 the business? I think for me as a user it's a meaningful improvement just like bar none using 03 and and I'm like a power user so I use these a lot I try all the different models uh GPT5 is clearly a step change in user experience now on some benchmarks people say oh well you know it's like 89% accurate on this benchmark versus 88% and so then therefore it's not a meaningful change. And I think that the uh it's clear that we're moving to a spot where um the benchmarks become harder and harder to use to assess the actual utility of the tools. Uh so all of the teams seem to be

1:36 optimizing on this is a great great mathematics wizard which is not really practically useful in any meaningful sense for knowledge workers or this is a great coding agent which is very practically useful for a lot of knowledge workers. Um and it's how is that presented to users in a way that they actually use it. And so the the most important benchmark here I think is actually the active users using the tool in the real world there. Open AAI is clearly um doing amazing. They have 700 million weekly activives and seeming to accelerate or at least maintain kind of their growth rate. And the reason this advanced the the major thing they did is um prior to GPT5 when you use chat GPT

2:24 there was like a little selector in the upper left and you had to select exactly what model you wanted to use and if you were paying for the service it was 03 40 mini 4.5 4.1 research preview you know it was like a menu of items that you had no idea kind of like which was the right one and which was more useful and what context. So I as a power user would swap between 03 and 40 which one does thinking on the back end for you and the other would just provide you an instant response. Practically now it's all just GPT5 and they have a router on top that tries to guide you towards the model that's useful. What that means is for users a lot of users even paid users weren't using the thinking model. they

3:09 were just using the default model. And so users will get a more agentic type experience within chat GPT. Uh and then the free users, the people who weren't paying had no exposure to the thinking models at all. And so those users will suddenly be like, "Oh, this thing can do more complex tasks for me. Uh it can take its time and come back with a fully considered answer. It can search the web to verify what it's saying so it doesn't produce hallucinations." And so I think the the the major advance is actually a clear product improvement in how this is presented to end users that will then allow people to understand what these systems are capable of and begin to integrate it more completely in their daily lives. >> And so then >> Brett >> wait >> one I just I want a self I I want a

3:55 Chiron to pop up that just says Brett Winton self-proclaimed power user. I'm not but I mean I'm but this this is the whole point like power user as in I use it for work and in my personal life at a high degree of engagement and I have no idea if I'm using it well relative to the average person and I know I've gotten better at using these tools over time. >> Well, I think this is actually interesting in describing, you know, I think you're Brett, you're a power user. Um, I think we all use it a lot at work, but it does seem that the community by and large was very split on the usability of chat GBT5. I'm looking at a a tweet from an Arc alum, James Wang. He

4:42 posted this at 5:00 a.m. Uh, today GPT5 is just no fun to talk to. frowny face. And I think as as as shallow as that take may be, I do think that seems to be a general consensus because if you look at what's happened since the launch, you've had some of the backtracking from OpenAI. So much to the point that when Sam went to Reddit, he did an AMA and ask me anything, he got bombarded with you need to bring back 40 or we riot type of responses. And they did. they did they did end up listening to the community and I do think or I guess the question that is out there and we don't

5:28 have the data you know of those 700 million weekly activives how many are core users for enterprise using this daily for work versus how many people are coming back to chat to talk to I have a post here that I was looking at that I sent to the team uh last week when we were talking about this during brainstorm and it was someone talking about how >> uh GBT5 just doesn't have that same humanness that 40 displayed and they were using it to get them through dark times as a therapist. And so there is like a element to this which is it isn't actually about the performance of the model and what it's capable of doing. It's how it makes you feel which is

6:13 crazy that we're already at that chasm. But I do think that this launch Sam Alman described it as bumpy after the fact. Like it does put something on display here that we haven't talked about a lot, which is the element of, you know, what this model makes you feel when you use it and how it responds to you. Like there's so much nuance in the sentences it gives back to you that do differ. And maybe we haven't placed enough importance on this in the way and why people choose GPT versus Gemini versus Grock versus MEAI. >> Well, Nick, I think you're what you're saying is like the qualitative side, but it ties in with what Brett has said where it's like people even quote power users are operating pretty blind

7:02 relative to how they should be using it. And it's because OpenAI doesn't know how you should use it either, right? It's like >> right >> and uh to some extent it's like I'm surprised that every prompt isn't metaprompted, right? It's like when you type in a prompt, it feels like the model should just automatically ask itself, how would you improve this prompt? Because that's what power users do anyway, right? It's like you put in a prompt and you say, how do you make this prompt better? and then you might put it into a different thing and you're like, "Okay, can you add these things to the prompt so that you get a good prompt?" And then you get there. >> But that's kind of that's kind of what test time compute is. It's taking a prompt and it's kind of like injecting synthetic data and and their router at

7:47 the top end is saying, you know, if I ask like give me advice on grilling fish, like I don't need it to spend three minutes trying to figure out the answer. I basically just want the average of whatever it has embedded uh there versus if I ask like here's a chunk of code I need help with this piece. I do want it to spend three minutes to like give a fully considered and reasoned answer. And so what's happening at the top end is it's taking that there's a router that's saying oh we should do more test time compute or not um based on that and then kind of injects it both into the right model and potentially provides qualifications for it. uh and I I think under reportported but they also provided the with GPT5 you can select between different personalities. You can select a more sarcastic one, a nerdier one. I'm sure

8:33 people haven't played with that. Like it is interesting that they have hundreds of millions of people who were only using GPT40, including people who had ongoing relationships with this model that when it was released, they they thought it was even overly sycopantic and overly kind of um solicitous to the end user. Basically, like, oh, I agree with you. You're so great. Uh, etc., etc. That can feel really good to people. And so if they like, you know, overnight your girlfriend that's an AI system gets replaced by another one that's like kind of nerdier and more to the point, you can understand people being like, "This was not this is not the therapeutic relationship I had in mind. Give me the droid I was looking for, right? Like I I wanted my old girlfriend." Uh, and so I

9:20 think that there's a one they can see in the data because they have such a large sample of users. I'm sure not only were people complaining on Reddit, but they were seeing that engagement was dropping for some subset of users. So on an emergency basis, it's like, oh, we need to provide them with 40 or, you know, and then over time provide kind of like can we, you know, kind of move that capability or that relationship into a new model. Or maybe people just want dumb AI girlfriends, you know, maybe that's they'll stick with that. That could be a business model. Um, but I think that there's and like you said, we don't we all don't know exactly what these systems are capable of. Um, you know, how people are using them and how we can use them to best effect. And so

10:06 they're making judgments and packaging that into a product. Uh and it's that particular like capturing why people like 40 is not something you could capture in benchmarks other than the benchmark of this is engagement against the user base and people are engaging with it. Um, >> which also >> well, wouldn't it have been chatbot arena, right? That would have been the one that tried to capture this. But I I agree. I think there is something happening here that I would have thought maybe played out like in the next decade, not this fast, which is the personality of these models starts to really impact usage. And it's you I think this is very you know it's it's showing and it's evident in this launch

10:51 and you know the backlash it received because think about you know we've often compared this moment in time to search. How many times do we think Google has changed their search algorithm? it probably is tweaking on a daily basis almost and you never really hear about backlash. But here, you know, you're building these models, you're putting them out into the wild and then people are having this human-like relationship with it to a certain extent and then you can't just rip and replace or tweak it because people pick up on it because the response is not link-based. It's, you know, you're trying to give back the most natural response possible. So people start to intuitively, you know, attach a personality to it and get emotionally

11:37 attach and maybe physically. I don't know. This is like it's so weird to me. >> This gets into an interesting convexity issue in the future of the business. But when on like Google, people get upset. It's like Google changes their search algorithm and it's like business >> businesses get upset. Certainly. Yeah, businesses 100%. But users, I mean, I'm sure there's been plenty of examples and people can probably find this where people have been, you know, outwardly, you know, uh, against some changes that Google have made and maybe it has >> there's also a difference in that Google is, you know, Google AB tests those changes in a more precise way versus, you know, Open AI is trying to take everybody at once from one model and kind of like port it over. And so you

12:24 could imagine their response could be, okay, now we know that there's this pocket of 40 users and we need to figure out like a system prompt level thing for GPT5 that kind of personality matches to those 40 users and so begin to like shift them over into this. There's probably efficiency reasons why they want everybody on this latest model over time, but they don't have to immediately. But this gets so Nick to what you're saying is I think this is the first time maybe just because it's scaled up to this point but it's like now there's a question it's like to open AI chat GPT5 is the better model but from an engagement standpoint it's not and so this leads to a business

13:11 >> well we don't know like it could be on an immediate term basis you know like the the free user tier for chat GPT is not it's only lead generation into the paid user tier right now. Like they have the beginnings of merchandise, merchant partnerships and stuff. They have an idea of how they're going to indirectly monetize that. But that user base is just a pool that right now on their monetization structure they're trying to convert it to paid. And I think that to me it's pretty evident that kind of like chat GPT5 for people that are going to pay for this on a subscription basis is the better better model. I think the the consumer side of the business almost needs to be thought of differently from the enterprise side of the business. Right now they're very much overlap. >> What's the split between the enterprise

13:57 and API calling side versus like the these are the people chat bonding away? >> You mean in terms of percentage? >> Yeah. >> Or well they're uh Nino I mean they're they're paid is a small percentage of overall users. I can't remember what's publicly disclos disclosed but um you know the majority of engagement is probably on the free user tier and then the API side is its own separate that's like the other interesting thing that's happening here is the in this whole discussion illustrates how much subjective and product judgment there is into creating this chat GPT agent layer versus you can try to roll your own by like doing the API calls even calling chat GPT5 API,

14:45 but you're having to decide, do I enable search tool calling? Do I kind of like do I use the mini model versus the like more expensive but longer to and longer to respond model? Um, and you know, OpenAI is making that decision on behalf of users within ChatGpt. uh and they're providing the underlying kind of modules that you could plug in for yourself if you're an enterprise or you're somebody trying to wrap an application on top of it to try to play with. Uh and that's like even a whole se and and there that's more anthropics business. Uh the other I thought meaningful thing that happened at least in the launch event is uh um that cursor's CEO was heavily featured. Cursor is a coding platform

15:30 that is shifting a lot of token volume to enthropic as the back-end coding agent. Uh and uh this indicates that you know OpenAI may may take that slot and that's a whole it's almost a whole separate business to consider. >> Yeah. Brett, what do you think about the model performance improvements generally? I mean, maybe we have been ruined by the past performance jumps we've been seeing on all of these benchmarks. This one didn't feel like it had that hockey stick type of performance across the board. It had it in certain categories maybe, but it didn't. And that's why I think you saw a lot of >> But but there's Well, there's there's there's like a Yes. And there I noticed

16:17 in the presentation there was like a weird Y scale thing. I was like what? Yeah, this doesn't make sense to me. But but the one um all of the benchmark results are presented on the wrong scale which is they present zero to 100 and what percent did it cross? But actually the hard part you need to say what is the error rate and then what is the log error rate as in going from n a 10% error rate to a 1% error rate is as hard as going from a 1% error rate to a 0.1% error rate. Mhm. >> Uh but if you present it on a 0 to 100 chart, it looks like, oh, we went from 90% to 99 to 99.9 and it's just like a little bit of an improvement, but that's actually as meaningful an improvement. So there's that as an issue, but then

17:03 separately it's like look at the they got to 100% on the AIM math benchmark. It's like who cares? Nobody I'm not I'm not using this to try to do theoretical math. they're trying to illustrate some underlying capability that matters to users and so they had you know some internal benchmarks that they presented which taking all of kind of like the chat GPT volume what is the hallucination rate of this model versus others and there there was real meaningful performance improvement so they didn't talk about in the presentation but if you gave 03 which is the previous best-in-class model at this task uh a it's basically like a scientific figure and asked it a question about it. It's okay at it and uh chat GPT5 is

17:49 better at it, but this is still a capability that they're improving on. If you propose the same query to 03, but you actually didn't give it the figure you just said you were going to, but you passed it a a blank JPEG. I think it was 80 plus% of the time, the 03 model would give you a very confident answer. Okay. So, GPT5 only 8% of the time gives you a very conf like 90 plus percent of the time. It's like there's I can't make a judgment here because there's nothing here. >> It seems worse because it's being more honest. >> Yes. Well, it's it's like just practically like what do you look for in an employee? You don't look for an employee that's going to say, "Oh, yes, I've already done that every time." you want the employee to be like, I can't do

18:35 this because I don't have the right underlying material. So, in terms of like practically using this thing in the world, there's a lot of stuff that's hard to benchmark and you have to have clever ways to test it to understand whether or not it's actually performant. and and that's that because of that, you know, it means the user will kind of try some of these things and they I think they'll switch when they get really frustrated with one and then they'll try another and then if it performs on that specific thing, they'll be like, oh, I should try this more. But it's actually I think it's hard to displace users once they are in kind of like I'm in iteration loop with this thing. It's failing at some things. It's working with others. I'm gonna occasionally try others. But I I do think that there's a gravity to to just I've I'm you know

19:23 these things aren't perfect but they are performing and they're performing better. And I think that um unless you have an evolving set of benchmarks that matches the the actual use cases these things are failing at. So you can see the improvement vector on those then you'll always run into this situation where it's like oh it's 95 now it's 97 who cares and it's yes but that's like demonstrating an underlying capability where it's getting much better. >> Do those charts exist because I've heard you mention this multiple times. It's like has has anyone published those or have we published the error rate based? >> Yes. are we do when we produce performance on these things we always do it on a log error basis and I just can't

20:10 convince the world that it's how they should look at it apparently um but yes you can go to like you know Frank and Joseph's like publications and research and kind of the newsletter and and and blogs and that's how we produce the benchmark data if we're transforming if we're just grabbing a screenshot of somebody else's then it'll just look in this you know I would a dumb way to present performance data, but it is the right way. It is also I do think there's an interesting question which is like the pot of gold at the end of the AI race is agi whatever you want to call it but it's like the path to get there of engagement it's like the average person

20:55 doesn't want to talk to a group of PhDs right and it's like it's like the business model to get there might be far more engagement baiting than people want to think about to Nick's point, right? And it's like so far it's been better model, better engagement, like people like it. And it's like there is going to be a split potentially between best model and oh actually this is a mid-curve model that people love to chat with and making it better isn't actually good for engagement. Well, I think this is really the hard question because you need

21:40 performance to get power users to pay incrementally more money per month, right? So, you know, Brett's willingness to pay is going to go up by the amount of work that he can offload to chatbt. But how much uh individual that's using this from a comforting standpoint from just you know talking to asking general questions I don't know that performance >> could ever get you to get you know 700 million weekly activives to pay you know incrementally more money a month but it will get you engagement which then you can you monetize in different ways it doesn't have to be a direct subscription so you have to go after both but I do

22:27 agree with you, S. I I want to know what is the most queried, you know, question out there because we've never gotten this. I've never heard and I have to imagine it's just spellch check. And that's not to say that these are just fancy spellchecking tools, but I my it's more of to the point of, you know, the majority of these people are using this for very basic inquiries. And that's not a bad thing, right? That's like the on-ramp into making it part of your everyday life. But you but I think Sam you're also hitting on this point which like that engagement baiting is going to be very important and one thing I have been very surprised about I check the app store rankings for productivity almost every

23:14 other day. Gro has held up surprisingly well and my guess is because Ewan has aligned it with spicy mode and all of these very engaging formats um that no one else has. And so if you're looking for that and you're looking for it from a free offering, that's a way to attract a ton of users and get them emotionally attached to these models. And this is like this element of emotion. I just don't think we've maybe discussed enough or never had like a proof point to really be like there is this ele like this something that you can't put your finger on this personality check that these models need to get over to then be able to reach this mass audience. And this is like putting it right in in my

24:00 opinion like right in the center of the conversation now because of how much backlash there was to this clearly more performative model. I but I think that actually the the model personality is not super sticky until you move into voice mode and then I think it'll get really sticky as in like you know you can almost proxy like you can I I'm confident that open AAI will be able to take these 40 users and create a chat GPT5 experience that does more emojis and is more um kind of solicitus and stuff to them. I think what's interesting like they're demoing this set of personalities on GPT5 um that is a initially only available via text but then will be invoice and

24:46 Grock is already doing it with these little video avatar things right where you have it's it's kind of the first time in history where like the spokesperson is actually the product for the company uh and and I think that there's a way to think about this is you are making a hire into an organization, you are interviewing that employee. To what degree do you care about can this person do the job versus is this person a culture fit with me that I'm going to interact with? If you're in a big enterprise, you know, it's kind of like, yeah, it matters to the team, the personality, but really it's like, can this agent perform? And so, uh, OpenAI is clearly designing towards not $200 a month, but $2,000 a month for agents that you plug into the enterprise and they can do meaningful work overtime

25:31 asynchronous to the person and they're good at following instructions and they come back honestly like I failed in this thing because this and this were missing rather than like overwriting the things that are missing with fake data to be like look I did it you know. So there's like plenty of like ways in which they need to push the model in order to successfully operate that those functions that's independent of that personality um issue which is more consumer-based and I agree like on the consumer side it's engagement and it's will I trust this thing to begin doing purchases for me. So there is like a performance characteristic like of kind of like how much am I going to trust this thing to like buy my airline tickets or book my meals or you know manage my subscriptions for my various things you know whatever kind of uh

26:17 purchasing agent um ambitions these companies have and users are willing to give to them and then there's like this bleedover layer where it's like well I need to be in conversation with the thing that's managing my calendar including my corporate and my family calendar and that's kind of like has access to that data and in order to trust it I like have to have a feeling that it's like a some kind of thing I can you know yell at or be soothed by or whatever people want and so there is like a bridge from one function to the other um >> but you know I I agree with you Brett but assume that performance is a given right like in the shopping example you gave assume that the integration >> it's like everyone's buying doesn't matter.

27:03 >> Yeah. Across models. What then makes me as a consumer choose chat GBT over Grock over uh MetAI over Shopify assistant? Is it is it personality? Is it the model that's the nicest to me when I'm like, hey, does this look good? I think it I think it actually might be. I actually I'm starting to have this like awakening of like, no, the personality of the model matters way more than we could ever truly understand. >> This friend who lost their 40 girlfriend in the room with us now. >> I mean, you just spend some time on Reddit and read the It's like Yeah, it's like it is actually people lost

27:48 something so dear to them. And I again I maybe I just haven't been in the weeds enough or I don't know maybe I'm not emotionally attached enough to a model yet. >> Well >> um >> just wait till voice mode. But this is my >> No, you're totally right. Yeah. Yeah. Once we get to voice mode it's going to be like a family member, right? And like the loss of a you know it hits even worse. >> It and it's just I mean >> Yeah. I think it's it's it's fascinating. Like the the the most exciting thing about the chatbt T5 launch in my opinion was the backlash and the reason for the backlash. I think that was the most like insightful thing that came out of it was actually oh no there's like a different vector of competition and user attachment that

28:33 we're not placing enough emphasis on. And it really does come down to the personality of the model. And that's something I I said a few times, but I don't think I really appreciated that enough. >> Yeah. Or tangential and then I'll go to closing point is it's like we're starting to see backlash to AI writing. Like Bology has been tweeting about this, right? It's like somewhat uncanny uncanny valley where it's like incredible writing that you know is not human and it grinds people's gears. Same thing. It's like I've seen on Instagram like very funny automated AI generated videos and people are mad. They're like I don't care if it's funny. It's like this is AI slop, right? And so it's like okay

29:20 there's an interesting lots of interesting dynamics going on. Maybe for another day, but Brett, where do you learn to improve your model usage skills? Are you watching YouTube videos? Are you just trial and erroring? what's what's the strategy? >> Uh I mean for me it's trial and erroring and then asking the model directly how do I do this better. So it's kind of like snake eating its own tail. But it's you know um I I think that there are definitely it you know depends on the workflow. It's like what's the best workflow for like doing code and then coding into like a deployment environment. And I've definitely um I think there the presented capability of

30:06 these systems always is somewhere past the actual practical capability of these systems. So for me personally like do having a coding agent that's doing everything for you results in a lot of time wasted because it kind of like starts making changes that you don't realize. you don't have enough of an internal understanding of what's going on that you can um catch it in lies and errors. And so I'm more kind of like keep my chat separate from my code type thing. But it varies, you know, I I think that even just doing system prompts like I have a system prompt in that's like when you present answers, give me like your internal state of kind of like how confident you are in this. what's the probability just to try to get it to condition some of its answers.

30:53 Um, and just to give you a sense of like how it on coding questions specifically, it is typically in the like 95% this solves your problem and it is totally misdiagnosed the problem and it's just like it is way overconfident in its own abilities at least as currently configured. That was within 03. So, um, you know, I but I don't know. I mean, I I think I think every depending on what you're doing, you need to at least be experimenting with the different models and and kind of like pairing activities against them and also trying to like figure out ways to do it better. Clearly powerful tools. >> We need like a micro dirty jobs uh

31:42 podcast, but it's just you just go to different jobs. It's everyone chatting to their to their AI girlfriend, >> right? >> I mean, yeah, but I think >> Yeah. >> I think that's a big opportunity because it's like I I again it's like we're all in the same organization. I have a sense of how you use it and at the same time I have zero idea how my prompting relates to the prompting that you're doing. >> Right. And I think at an organization level there is there's going to be this constant war across consumer and enterprise between like how well does um the default settings that open AI or or XAI or enthropic present to the enterprise? How well do they map to the set of functions the enterprise wants to

32:28 do versus you know are you doing custom GPDs? Are you doing your own system prompt? Are you trying to like have it take your own data within a specific container? Are you rolling to your own APIs all the way down to like on the consumer side? How much like data are you giving it to like manipulate? And it's still so much open season and and all of this stuff. Um except that OpenAI has this mass of users that gives it this feedback in real time that then they can respond to. Like you said, they made 40 still available until they figure out a way to bridge these users. Um, >> yeah. >> All right. Should we I want to end on reading one Reddit post about ChatGBT that I believe is going viral on X right

33:14 now and it's it's titled ChatGBT induced psychosis. My partner has been working with ChatGBT chats to create what he believes is the world's first truly recursive AI that gives him the answers to the universe. He says with conviction that he is a super superior human now and is growing at an insanely rapid pace. I've read his chats. AI isn't doing anything special or recursive, but it's talking to him as if he is the next messiah. It goes on a little bit further, but I think that summarizes what is maybe happening here. >> Well, one, I mean, who knows? Maybe it could be the next messiah. too. The the >> good good pass out wager. >> Yeah. Yeah. I think that the um

34:02 these are mirrors for people in some profound way. You can see in them what you want to see. Uh and and just like you know if you you know so they will cause people to spin off into Yeah. maniacal infinity if they have an underlying weakness towards that. Um, and they'll, you know, solve a lot of people's like therapeutic problems as well. Um, yeah. Well, willing to put their their their human relationship on the line because this person goes on to say if she doesn't get on board or he doesn't get on board, they're going to leave them. So, that's the the gist of this. You know, either you start using ChateBT like I start using it or I'm I'm going to leave you.

34:49 >> All right. >> You know, two Wait, wait, what? One last point. Do you remember uh Okay, so when Star Wars was first released, this is a non nonsequator, I promise. Um there was after it got so much attention that uh somebody tried to convince George Lucas to start a religion based on the force. Okay, that's just on like a single two-hour movie and kind of like the amount of consumer attention that had. Well, this is a thing that is likely to be used by call these artifacts are going to be used by billions of people. I think it's pretty likely that you'll have some kind of like religious cult that emerges on the basis of interactions with these agents. I think that that is like >> spin up the poly market. Spin up the poly market. Let's get that betting line

35:34 going because I 100% agree with you. >> Yeah. >> All right. So, there's a good good opportunity. I think uh being the founder of a religion has paid well in the past. Not not for everyone. May may result in death, but um all right, we're three episodes away from episode 100. So, we're excited about that. Keep tuning in. Keep telling us. And if you're using chat, GPT or Grock or other in a unique novel way that's useful, please reach out. Let us know how you're doing it. Thank you all.