ChatGPT finds it’s voice! Join our hosts, Paul Roetzer and Mike Kaput, as they discuss OpenAI's cautious rollout of voice features for ChatGPT, assess Chris Miller's insights on the emerging "cloud war," and examine Microsoft's latest AI-driven financial results. Plus, take a look into Perplexity’s Publishers Program, concerns behind California’s AI Bill SB-1047 and Google’s experimental version of Gemini 1.5 Pro in our rapid fire section.
00:04:02 — OpenAI Begins to Roll Out Voice Assistant Feature
00:19:26 — Global Chip/Cloud War
00:30:54 — Microsoft AI Stats and Usage
00:41:54 — Meta Earnings and AI
00:44:54 — Perplexity Introduces Publishers Program
00:48:35 — Concerns over California AI Bill SB-1047
00:53:35 — Writer’s New Models Designed for Healthcare and Finance
00:58:14 — Google has released an experimental version of Gemini 1.5 Pro
01:03:47 — Microsoft and OpenAI
OpenAI Voice Assistant Feature for ChatGPT
OpenAI is doing a limited rollout of its voice assistant feature for ChatGPT after previously delaying its release to address potential safety concerns. The company is making this feature available to a small group of paid ChatGPT Plus customers starting Tuesday.
The voice assistant will offer four preset voices but has been designed to prevent impersonation of specific individuals.
In response to copyright concerns, OpenAI has implemented new filters to prevent the generation of copyrighted audio content, such as music.
This voice feature was initially showcased in May as part of the GPT-4o update, which enhanced the model's ability to handle text, audio, and images in real-time. The company had originally planned to release this feature in late June but postponed it to implement additional safety measures and improvements.
OpenAI intends to gradually expand access to all ChatGPT Plus subscribers in the fall. However, some features demonstrated earlier, such as video and screen-sharing capabilities, are still in development without a set launch date.
Global Chip War Evolving Into a Cloud War
The global chip war may be evolving into a cloud war, according to a recent analysis by Chris Miller, author of the popular book Chip War, which is about the geopolitical rivalry to control AI chips and infrastructure.
As AI systems become increasingly important to the global economy, the data centers that train these systems are starting to be viewed as strategic resources by governments worldwide.
As a result, many countries are now seeking to establish their own AI infrastructure through data centers built on their soil. Saudi Arabia, United Arab Emirates, Kazakhstan, and Malaysia are among those investing heavily in data center capacity and AI development.
Now, U.S. cloud companies see lucrative opportunities in this trend, arguing that if they don't take contracts from foreign governments for "sovereign AI" infrastructure, China will. The U.S. government also views this as a way to promote American technology and limit Chinese influence.
Miller concludes that the tech competition that began with semiconductors is now expanding into new layers of the computing stack, with chips, clouds, and data centers becoming increasingly interlinked in the global race for AI dominance.
Microsoft AI Stats and Usage
Microsoft is having an interesting week in AI.
First, the company released its fiscal fourth quarter earnings, revealing mixed results that led to a drop in the company's stock price. The tech giant beat overall revenue and earnings per share expectations, but fell short on cloud revenue, particularly in its Intelligent Cloud segment which includes Azure services.
Microsoft's Intelligent Cloud revenue came in at $28.5 billion, below the anticipated $28.7 billion. But, despite the cloud revenue shortfall, Microsoft's overall revenue still grew by 21% year over year, with Intelligent Cloud revenue increasing by 19%.
The company noted that AI services contributed 8 percentage points of growth to its Azure and other cloud services revenue, which saw a 29% increase.
Now these developments come just as Microsoft researchers have also released their second report on AI and productivity, titled “Generative AI in Real-World Workplaces.”
This report synthesizes findings from a number of recent Microsoft studies on the impact of generative AI, specifically Microsoft Copilot, on productivity in real-world work environments.
This week’s episode is brought to you by MAICON, our 5th annual Marketing AI Conference, happening in Cleveland, Sept. 10 - 12. Early bird pricing ends Friday. If you’re thinking about registering, now is the best time. The code POD200 saves $200 on all pass types.
For more information on MAICON and to register for this year’s conference, visit www.MAICON.ai.
Disclaimer: This transcription was written by AI, thanks to Descript, and has not been edited for content.
[00:00:00] Paul Roetzer: We're just trying to figure out how to use ChatGPT and like find the three use cases, but you have to understand that governments and tech leaders are looking way beyond that, they're seeing an explosion of intelligence that we are not currently set up infrastructure wise to support.
[00:00:17] Paul Roetzer: But they envision intelligence being absolutely everywhere. And that's the part that I lose sleep over.
[00:00:26] Paul Roetzer: Welcome to the Artificial Intelligence Show, the podcast that helps your business grow smarter by making AI approachable and actionable. My name is Paul Roetzer. I'm the founder and CEO of Marketing AI Institute, and I'm your host. Each week, I'm joined by my co host. and Marketing AI Institute Chief Content Officer, Mike Kaput, as we break down all the AI news that matters and give you insights and perspectives that you can use to advance your company and your career.
[00:00:56] Paul Roetzer: Join us as we accelerate AI literacy for [00:01:00] all.
[00:01:03] Paul Roetzer: Welcome to episode 108 of the Artificial Intelligence Show. I'm your host, Paul Roetzer, along with my co host, Mike Kaput. We are recording this one on Friday, August 2nd. We're doing it a little bit early this week because I'm out of the office on Monday. I'm actually going to be with my friend, Joe Polizzi, who's got his Orange Effect Foundation, an amazing foundation.
[00:01:23] Paul Roetzer: and so we're going to, we have a fundraiser all day long. So we'll be out on the golf course with Joe and that whole crew. so check out Orange Effect Foundation. It's an amazing organization. So we have, um, some fascinating stuff to get through today. Like the, the voice stuff from OpenAI is like, I'm having a hard time processing the potential impact of this.
[00:01:46] Paul Roetzer: we've got some new stuff on the chip war. We've got some productivity studies from Microsoft. We have earnings from Microsoft and Meta and others. So a lot's been going on this week. Oh, and then, figure humanoid robots, Mike. We're going to talk about it. I [00:02:00] don't even know if you saw my note about that one.
[00:02:01] Paul Roetzer: It just happened this morning. So I'm going to talk a little bit about, humanoid robots because They may be coming sooner than we think. So, today's episode is brought to us again by the Marketing AI Conference, MAICON, our fifth annual. MAICON is coming up in Cleveland, September 10th to the 12th. There are 69 sessions, I counted.
[00:02:21] Paul Roetzer: Last week I said on the podcast, more than 60. So I counted this morning, there are actually 69 total. We have 33 breakouts across two tracks, so Applied AI and Strategic AI. 16 AI tech demos. 10 main stage general sessions and keynotes, 5 lunch labs, 3 pre conference workshops, I'm teaching one, Mike's teaching one, our friend Jim Stern is teaching one, and 2 mindfulness
[00:02:45] Paul Roetzer: sessions,
[00:02:46] Paul Roetzer: with Tamra Moroski, our Director of Partnerships.
[00:02:50] Paul Roetzer: who also teaches yoga and mindfulness. So, it's going to be an amazing lineup. Like I was actually sitting there looking through, I was like, if I was an attendee, I wouldn't even know how to pick [00:03:00] which breakouts to go to.
[00:03:02] Paul Roetzer: So, so if you're connected to me on LinkedIn, I actually put up just a sample of a bunch of the sessions, but you can go check out the agenda.
[00:03:08] Paul Roetzer: So go to maicon. ai, that's M A I C O N dot A I. You can use pod 200 to save 200 off of your MAICON pass. again, that is coming up September 10th to the 12th, about 30 some days, depending on when you're listening to this. We are only about 30 some days out from the event, and I know Mike and I are both already starting to work on the workshops and the talks, so I'm doing the road to AGI as my opening keynote.
[00:03:38] Paul Roetzer: Mike's doing one of the closing talks, 30 AI tools in 30 minutes, which is always a fun one. It's. It's, it's action packed. So check that out. And again, we'd love to have you join us in Cleveland, September 10th to the 12th, MAICON. AI to get registered. And don't forget to use that POD 200 promo code to save $200.
[00:03:59] Paul Roetzer: All right, Mike, [00:04:00] OpenAI and voice was the big news of the week.
[00:04:02] Mike Kaput: Yes, it was because OpenAI is finally starting to roll out its voice assistant feature for ChatGPT, which had been previously teased and then delayed, due to addressing some potential safety concerns.
[00:04:18] Mike Kaput: So the company is making this feature available to a small group of paid ChatGPT Plus customers, and it is going to eventually be rolling this out to all ChatGPT Plus users. Now this voice assistant has four preset voices, and in response to copyright concerns, OpenAI has made sure it is going to prevent those voices from, doing any impersonations of specific individuals.
[00:04:44] Mike Kaput: We had a little bit of controversy with that right around the release of the, or the announcement of this feature. And they're also implementing features to prevent that. the generation of copyrighted audio content, such as music. Now, if you recall, we actually covered [00:05:00] this feature when it was initially showcased in May as part of the GPT 4.
[00:05:04] Mike Kaput: 0 update announcement, which enhanced the model's ability to handle text, audio, and images in real
[00:05:11] Mike Kaput: time. So the company at the time had initially kind of planned to release the voice feature in late June, but then they postponed it. They said there were safety measures and improvements that they needed.
[00:05:22] Mike Kaput: to make before unleashing this on the broader world. So we are going to gradually see people getting access to this as we get into the fall here. However, there are some features that were demoed during the initial event, like video and screen sharing capabilities that appear to still be in development without a set launch date.
[00:05:44] Mike Kaput: So Paul, first up, like what have been kind of your initial thoughts? of this new and improved voice mode, you know, the initial demos are pretty interesting, either from things you've seen that users with access have been posting or kind of your own experience so [00:06:00] far.
[00:06:00] Paul Roetzer: Yeah, the demo was certainly impressive back in May. And the question is always with these demos is how, how good is it really going to be when we actually get the product? And I have seen people posting videos that I would say are somewhat shocking in how good this thing is. So, if you haven't seen these videos, the thing is able to change its tone, inflection, accent, integrate sound effects into its voice, convey emotion, Continue when it's interrupted, you stop it and it immediately picks back up.
[00:06:32] Paul Roetzer: It's somewhat uncanny, like how human like it is. And the more I see these examples, the harder time I have believing that the current model, this 4o model that we're all using in ChatGPT is actually what's powering this thing. It just seems like it has way more capability than anything we've interacted with before.
[00:06:57] Paul Roetzer: So I want to rewind back for a minute to [00:07:00] May 13th, 2024, when OpenAI published their Hello GPT-4o article. And so this was introducing the model, Mike, as you were kind of highlighting. So I'm going to read a few excerpts from this and I want to add a little context as to why I think we're seeing something maybe more significant than what people realizes happening.
[00:07:21] Paul Roetzer: So these are the excerpts. I'm just going to kind of read from this. So it says, we're announcing GPT-4o, again, this is May 13th, 2024, our new flagship model that can reason across audio, vision, and text in real time. GPT-4o, the O stands for Omni, is a step towards much more natural human computer interaction.
[00:07:43] Paul Roetzer: It accepts as input any combination of text, audio, image, and video, and generates any combination of text, audio, and image outputs. When we get into the voice component, it said it can respond to audio inputs in as little as 232 [00:08:00] milliseconds, with an average of 320 milliseconds, which is similar to human response time in a conversation.
[00:08:09] Paul Roetzer: said GPT-4o is especially better at vision and audio understanding compared to existing models. Prior to GPT-4o, and this is explaining how the voice model is able to do what it's doing. Prior to GPT-4o, you could use voice mode to talk to ChatGPT. So you and I can go in right now and we can talk to ChatGPT, but we don't have this new version.
[00:08:34] Paul Roetzer: So it's saying the one that all of us have, and have been using, talk to ChatGPT with latencies of 2.5 seconds when you're using the GPT-3 0.5 model and 5.4 seconds for GPT-4 on average. So way slower, obviously than 232 milliseconds
[00:08:54] Paul Roetzer: to achieve this, and this is why it takes five seconds. Voice mode in [00:09:00] a is a pipeline of three separate models.
[00:09:03] Paul Roetzer: So when you and I talk to ChatGPT, and I would imagine this is. Probably kind of how Siri works. I haven't really thought about that, but historically how these voice models have worked. So one simple model transcribes the audio to text. So you and I speak, ChatGPT turns that audio into text. So there's a model to do that.
[00:09:24] Paul Roetzer: GPT 3. 5 or 4 takes it in text and outputs text. And then a third simple model converts the text back to audio. So there was three models that were running behind the scenes to do a single response to your query, basically. This process means that the main source of intelligence, GPT 4, loses a lot of information.
[00:09:47] Paul Roetzer: It can't directly observe tone, multiple speakers. Background noises, it can't output laughter, singing, or express emotion. So they're very clearly explaining the limitations of prior voice [00:10:00] technology, basically. With GPT-4o, it says we trained a single new model end to end. across text, vision, and audio, meaning that all inputs and outputs are processed by the same neural network.
[00:10:13] Paul Roetzer: Because GPT 4 was our first model combining all these modalities, we're just scratching the surface of exploring what the model can do and its limitations. And then it talks about how they've gone through extensive red teaming. Now, again, this was back in May. And what we now know is they went through a lot more red teaming since May to bring this to market.
[00:10:33] Paul Roetzer: had experts in domains such as social psychology. Bias and fairness and misinformation to identify risks that are introduced or amplified by newly added modalities. We use these learnings to build safety. then it goes on to say we recognize that GPT-4o's audio modalities present a variety of novel risks.
[00:10:53] Paul Roetzer: Today, we are publicly releasing it. Again, this is in May. Publicly releasing text and image inputs and text outputs. [00:11:00] So when they announced GPT-4o in May, so the model was done, they only actually made available text and image inputs and outputs. So the audio and the vision capabilities were not released in May, even though they were obviously done.
[00:11:14] Paul Roetzer: Over the upcoming weeks and months, we'll be working on technical infrastructure, Usability via post training and safety necessary to release the other modalities. For example, at launch, audio outputs will be limited to a selection of preset voices, which is exactly what you just outlined, Mike, and will abide by our existing safety policies.
[00:11:38] Paul Roetzer: We will share further details addressing the full range of modalities in a forthcoming system card. the capabilities will be rolled out iteratively with extended red teaming. So here's what I can't shake, Mike. When we think about what is GPT 5 and it's like everybody just keeps waiting for it. The more I think about this, the more I think we're [00:12:00] already seeing it.
[00:12:00] Paul Roetzer: So if you, if you piece together Everything that OpenAI has released or talked about over the last, you know, six months in particular. We have video with Sora. We have voice, which we're now experiencing. The image generation, DALL E is kind of a joke when you compare it to MidJourney right now, but my guess is it won't be when they release the next version of DALL E.
[00:12:24] Paul Roetzer: We have Project Strawberry. We have, like, whatever that is, math, reasoning, whatever it is. The ability to solve more complex problems, go through reasoning, assist in decisioning. We have search, which was, we talked about on last week's episode. We have vision, which they haven't talked more about yet, but the ability for your device to see and understand the world around it.
[00:12:44] Paul Roetzer: They have already previewed basically all the modalities that are talked about in the 4o announcement in May is individual pieces. But when you put those pieces together. And this would fit with the iterative deployment that [00:13:00] Sam Altman has talked about. I think that we already have a Pretty clear picture of what GPT 5 is going to be.
[00:13:08] Paul Roetzer: And I think 4o is literally just a naming convention until all the pieces are in place. And then once we go back, so let's go back one more time and read that opening. GPT-4o is a step towards much more natural human computer interaction. It accepts as any input or as input, any combination of text, audio, image, and video, and generates any combination of text, audio, and image outputs.
[00:13:33] Paul Roetzer: Take that, give it advanced reasoning capabilities, give it the ability to see and understand the world around it, and I think you have GPT 5. And so that's the thing that I can't shake, is like the voice on its own is just mind blowing. I mean, truly, like if you haven't seen these, and I don't say that lightly, like I don't usually use hyperbole when I'm explaining these technologies, like it really is hard to comprehend the [00:14:00] impact a technology like this can could have on business, on marketing, on our personal lives.
[00:14:07] Paul Roetzer: And so when you look at what OpenAI is doing and you actually compare it over to what Google is doing, I think it becomes quite clear that Google and OpenAI are in, in a realm all their own right now. So yes, Anthropic Matters and yeah, 3. 5 Sun, it's awesome, but. Last I checked, like, Claude's still not connected to the internet.
[00:14:26] Paul Roetzer: It doesn't do images. I haven't heard anything about vision stuff with them. so they're doing reasoning, they're doing language models really well, but it doesn't seem like they're playing in the same ballpark. And everything we're seeing from Google and everything we're seeing from OpenAI seems like They're differentiating themselves in this pursuit of whatever this smarter form is going to be.
[00:14:48] Paul Roetzer: And then the last note I had is, also this week, and we'll have a little bit more of a link to this in the newsletter, but Apple Intelligence started rolling out to select users, developers, I think, have [00:15:00] started getting access to that. And so people are starting to now experiment with the new Siri. that is being talked about.
[00:15:07] Paul Roetzer: So we touched on this, let's see, episode 102 on June 12th, we talked about Apple intelligence and the new operating system that's going to come out in October, isn't that last week? We said it's, it got delayed till October. And so I, the thing I like when I see this open AI voice capability, the thing I immediately think about is that Siri is just going to be obsoleted.
[00:15:28] Paul Roetzer: Like it doesn't appear like, if, so if you fast forward a month or two from now and all of us can have access to this OpenAI voice capability, why would I ever open Siri again? Like, if I can just go, like, it just seems like it's so far ahead. And so I was asking myself this question as I was prepping for today's podcast, like, what would I use Siri for?
[00:15:50] Paul Roetzer: Like, if it really does this stuff. And so I went back and just took a look at how
[00:15:55] Paul Roetzer: Apple was describing Siri, because if you remember, Apple's doing a deal with [00:16:00] OpenAI to integrate ChatGPT, but it didn't sound like they were integrating the voice technology. And so I think the play Apple's going to have to make here is accept that someone probably built the better voice assistant, the thing they always envisioned building.
[00:16:12] Paul Roetzer: I think OpenAI just didn't. is going to win at that. And the Apple Intelligence component is Siri is your trusted assistant that works within your device that keeps everything completely private and has access to all of your apps. And it becomes more of like an on device AI agent that enables you to work with all of your applications.
[00:16:33] Paul Roetzer: So, if you go to the Apple Intelligence page, it literally says like um, Surya has all new design that's even more deeply integrated into the system experience. it has expansive product knowledge, so it knows all about your phone and everything within it. It has a richer understanding and enhanced voice communicating ability, but it talks about on screen awareness.
[00:16:53] Paul Roetzer: So Siri is going to know everything you're doing on your screen. OpenAI won't have that. It'll know everything you've [00:17:00] communicated, the personal context. OpenAI won't have that. It'll seamlessly take action across apps. OpenAI won't have that. And so that's what I start to wonder is like, can Apple successfully delineate those two things?
[00:17:14] Paul Roetzer: This on device AI agent that, that knows you and everything within it. And that's how you're going to use that. And then there's a separate voice thing that doesn't have access to all your apps and doesn't know what you're doing on your thing on your devices. So it's just going to be, it's going to be fascinating once we get both of these by October.
[00:17:32] Paul Roetzer: So I assume Apple intelligence story will be readily available to everyone with an iPhone 15 pro and above. And I assume voice from OpenAI will have rolled out and be done with all the red teaming by October. And so this fall, it's just going to be a different world where voice becomes maybe like heads toward a true user interface.
[00:17:54] Paul Roetzer: Like we've never really had it, had that. So we had computers with keyboard and mouse. We had touch with [00:18:00] different devices. And now we're moving into a generation where voice truly becomes an interface and that's going to be wild.
[00:18:07] Mike Kaput: Yeah. And voice, it strikes me more than any of the other interfaces really. is natural and anthropomorphizes almost the models we're interacting with.
[00:18:18] yeah, I mean, I, it's so weird, like AI rights activists. Like, I think you're going to see that become a thing. I've thought about this recently, the situational
[00:18:26] Paul Roetzer: awareness stuff when I was listening to Leopold
[00:18:28] Paul Roetzer: in interviews, I thought about this idea of AI rights activists.
[00:18:32] Paul Roetzer: I'm not going to lie. Like when I was watching, like Ethan Mollick had a demo.
[00:18:36] Paul Roetzer: I think Allie Miller, I watched a demo of her with OpenAI's voice, Where they interrupt it kind of rudely while it's talking. I get the same feeling in my stomach I do when someone does that to another human.
[00:18:49] Paul Roetzer: And like, and I know it's an AI. I have, I have feeling toward that. Like I have an emotion that's triggered in me when someone talks rudely to [00:19:00] someone.
[00:19:00] Paul Roetzer: And it happens when I see people doing it to these, and I get that same feeling about like when someone mistreats an animal, like, I know that may sound weird. And I don't know if I'm like alone on this, but. It's going to be a problem. Like, it's very bizarre how you do, as you're saying, like you start to just like assign human like qualities to these things, even though we know it's not there.
[00:19:26] Mike Kaput: So in another big topic this week, we are seeing the possibility that the kind of global war over chips, over AI infrastructure may be evolving into a cloud war, according to kind of a recent take by Chris Miller, who is the author of an important book called Chip War, which is about the geopolitical rivalry.
[00:19:48] Mike Kaput: to control AI chips and infrastructure. So basically the idea is that as AI systems become increasingly important to the global economy, the data centers, the chips [00:20:00] that train these systems are starting to be viewed as strategic resources by governments all over the world. And so what Miller does in this article is he kind of draws parallels between the systems we have today And the situation that we experienced during the Cold War era when supercomputers were tightly controlled due to their dual use nature for both civilian and military applications.
[00:20:24] Mike Kaput: So like for instance, today's AI systems have similar dual use potential. I mean, they do everything from optimize food delivery apps, but also analyze satellite imagery and. Direct drone strikes. So as a result of this kind of strange dynamic, like many countries are now seeking to establish their own AI infrastructure through data centers and infrastructure that's built on their soil that they control.
[00:20:51] Mike Kaput: So for instance, countries like Saudi Arabia, the United Arab Emirates, Kazakhstan, Malaysia, they're among those that are investing very [00:21:00] heavily in data center capacity and AI. infrastructure. Development. Now, in this article, they kind of talk about US cloud companies see tons of lucrative opportunities in this trend.
[00:21:10] Mike Kaput: And the companies that sell all these cloud type of services argued that if they don't take contracts from these foreign governments for quote sovereign AI infrastructure, another superpower like China will. So the US government also views this as a to promote American technology and limit Chinese influence.
[00:21:32] Mike Kaput: So this, of course, just raises all sorts of kind of concerns and compromises that a lot of security experts in the U. S. are concerned about. Things like Microsoft has a partnership with The United Arab Emirates owned G42, which has historic ties to Chinese companies like Huawei. And basically where Miller gets to is that conclu the conclusion that the tech competition that began way back when with [00:22:00] semiconductors is now expanding into new layers of the computing stack with chips, clouds, data centers becoming increasingly important, increasingly interlinked in this global race for AI dominant. So, Paul, let's first talk about maybe the big, big picture here. you know, this, this this. touches on a ton of points that, you know, Leopold Aschenbrenner touched on in situational awareness and a few of the other topics we've covered recently. Like, why does this matter globally to the direction and speed of AI innovation?
[00:22:33] Paul Roetzer: So, from a simple perspective, I think for listeners, for your own career, for the business, the industry you're in, What this indicates is that the, the big tech companies and leading governments are preparing for an intelligence explosion, like that, whether economic economists are telling us this or not, whether the latest research report is telling us this or [00:23:00] not, whether the stock markets movements are telling us this or not, when you look out over a three to 10 year horizon, everyone is preparing for A world where intelligence is omnipresent.
[00:23:13] Paul Roetzer: And what I mean by that is right now, what we've had, and we've talked about this before, is all of this demand for NVIDIA's chips to build these data centers, to, you know, build what Microsoft and Google and OpenAI and Anthropica are all trying to do for the most part. It has been to, to train these very large models.
[00:23:37] Paul Roetzer: And now we're trying to find, you know, the applications for those models and trying to infuse them into businesses and find these generative AI use cases. But generally speaking, the intelligence hasn't diffused throughout society yet. Like, Apple intelligence is coming this fall. The voice from OpenAI is starting to come out this fall.
[00:23:56] Paul Roetzer: Like, we have been using the, the, the data. [00:24:00] and the computing power to build the models. In the future, the need is going to flip. The majority of the use of these data centers and the chips that power the AI and the energy that makes it all possible is going to be used for inference. Now, we've talked about this idea before, but, you have The training of the models, and then you have the inference that occurs when you and I use the models.
[00:24:26] Paul Roetzer: And so what's happening is as intelligence is starting to be built into everything we use into your phone, your glasses, your computers, your household appliances, your cars, that's going to be the inference compute that's needed. And so what we're seeing is, as he's saying, it's going to evolve to the more like who can serve.
[00:24:49] Paul Roetzer: The, the amount of energy and data and computing power needed to provide all that inference. And so that's where this is all going. And we talked on [00:25:00] episode 83 back in February, 2024 about Sam, you know, the rumor at that time was he was trying to raise 7 trillion and he was talking with UAE leaders and, you know, he was thinking about putting data centers in other countries, that's all kind of related to this stuff.
[00:25:17] Paul Roetzer: There was a ton in the earnings calls that just this week about CapEx spend and infrastructure and how much was being put into data centers. So this is like a huge ongoing topic, but phones and computers where we're going to use AI and have access to intelligent assistants is one thing. The other thing we talked about back on episode 87 is, and this is where it gets weird, is humanoid robots.
[00:25:43] Paul Roetzer: Um, but just this morning, FIGURE, which is one of the key players in that market, so FIGURE, if you'll recall, has a deal with OpenAI, where they're
[00:25:53] Paul Roetzer: embodying these large language models into the robot. And so the robot, all those 4o capabilities we [00:26:00] just talked about, imagine a humanoid robot. That can talk to you.
[00:26:06] Paul Roetzer: So now you're not just talking to your phone and little, you know, cute little bubble that's going, you're looking at a humanoid robot that can talk to you that way. So Brad Haddock or Brett Haddock, the founder of Figure, he tweeted. This morning, or was it last night, only recently has time opened a window of opportunity to scale billions of intelligent humanoid robots.
[00:26:32] Paul Roetzer: This was not feasible five years ago. We are fortunate to be in 2024, where The first year in history when this is possible. Life is about to turn into a sci fi film. And then they announce that they are launching Figure 2, which is the new humanoid robot, on August 6th. So next week we will get a preview of this thing.
[00:26:53] Paul Roetzer: Um, He then went on and tweeted, No doubt the biggest company in our lifetime will be within [00:27:00] humanoid robotics. A humanoid robot company will exceed one trillion of value purely from operating in commercial factories. So doing things like lifting stuff that humans can't. But they build them to be humanoid so they, because the world is designed for humans.
[00:27:15] Paul Roetzer: And so these things need, Five fingers and toes and all these things too, cause they have to function like we do. a robot in the home will bring about a 20 trillion or greater company. And figure robot is working on both as is at Elon Musk with their optimist at Tesla, as are other people. So. The thing we have to keep in mind is we're looking at the near term future as business leaders, as practitioners, as knowledge workers.
[00:27:41] Paul Roetzer: We're just trying to figure out how to use ChatGPT and like find the three use cases, but you have to understand that governments and tech leaders are looking way beyond that, but when I say way beyond, I'm talking like three, five
[00:27:54] Mike Kaput: Right, where they're in, they're seeing an explosion of intelligence that we [00:28:00] are not currently set up infrastructure wise to support.
[00:28:03] Paul Roetzer: But they envision intelligence being absolutely everywhere. And that's the part that I lose sleep over. It's just like,
[00:28:13] Paul Roetzer: what when this all, you know, comes to be? Or even again, like I always say, even if it's just like 30 percent of this vision, like, even if we don't get to billions of humanoid robots, but you know, five, seven, 10 years from now, we got 250 million humanoid robots walking around.
[00:28:29] Paul Roetzer: Like, what is that world like?
[00:28:32] Mike Kaput: Yeah, that's so interesting to think of. And I think that's a really kind of sneakily important point is that you, there's
[00:28:41] Mike Kaput: tons of hype in this space. There's obviously tons of wild claims and like, Fantastical thinking about what can and will happen in the future, but if you look at all the signals you'd use to evaluate if someone is legit or not in any other business decision, like look where the money is going, look [00:29:00] where the effort is going, and look at where, look at what people are actually doing in addition to what they're saying, and that seems to indicate That you're absolutely spot on and again, if it's 20, 30 percent of what we're talking about here, that's, that alone is sci fi.
[00:29:17] Paul Roetzer: Yeah, and again, I just, like, I think you can ground back into like the GPT-4o conversation of
[00:29:25] Paul Roetzer: everything they put in that article in May. They already have, like, they're not publishing that stuff and previewing it if they don't already have it. And it's in red teaming, or if they don't know technologically it's achievable by the end of the year.
[00:29:40] Paul Roetzer: So that wasn't like some three to five year vision thing. That was. This is the model we have built and are in the process of red teaming. And, and then you fast forward a year or two from now, like we talked about, I think on the last episode, this idea of just think 12 to 18 months out, like you're going to have the next version.
[00:29:58] Paul Roetzer: And what is that going to be able to do? [00:30:00] And that's where I just. I think the best businesses, the, the companies that truly differentiate themselves in the next five to 10 years, the leaders who become so valuable, the practitioners become extremely valuable in their careers. They're going to look 12 to 18 months out.
[00:30:15] Paul Roetzer: They're going to look at the signals, as you were saying, like look at the indicators of where this is going. And then And plan, contingency plan for them being right. You're gonna, most businesses are going to plan based on what is true today. If you're the one that says, okay, we'll plan on what is true today, I'm pretty confident that, that the other assumptions may be true.
[00:30:39] Paul Roetzer: So let's build a contingency plan. That they're right. And we do have this intelligence explosion. And let's be ready to be there before our competitors and before our peers. And I think that's the opportunity in front of people right now.
[00:30:54] Mike Kaput: All right.
[00:30:55] Mike Kaput: So our third big topic this week is kind of about an interesting week that [00:31:00] Microsoft is having. So first they released their quarterly earnings. These revealed, kind of mixed results that actually led to a drop in the company's stock price. they actually beat their overall revenue and earnings per share expectations, but they fell a little short on cloud revenue.
[00:31:17] Mike Kaput: which came in about 0. 2 billion lower than expected. However, despite that, their overall revenue grew by 21 percent year over year. interestingly, and this is kind of the important point here, they noted that AI services contributed 8 percentage points of growth to Azure and other cloud services revenue, which overall saw a 29 percent increase.
[00:31:41] Mike Kaput: Now, as part of these earnings announcements, Jameen Ball, who is a partner at Altimeter, released some really interesting stats on kind of where Microsoft is at today, when we think about where the rubber's really meeting the road in terms of AI generating Revenue and usage. So, he had tweeted out [00:32:00] or post on X rather that Azure AI services, which is what was responsible for that eight percentage point bump in the earnings has a 5 billion runway, which is up 900 percent year over year.
[00:32:14] Mike Kaput: It has 60, 000 customers, up 60 percent year over year. GitHub is at a 2 billion run rate. That's up from a billion just a couple of years ago. GitHub Copilot, which is the company's AI coding assistant, has 300 million in ARR. And also, Jameen Ball said that he, the daily users of Office Copilot have doubled quarter over quarter and Office Copilot customers grew 60 year over year.
[00:32:42] Mike Kaput: Now, to kind of tie all this together, these developments come just as Microsoft researchers have also released their second report on AI and productivity, which is called Generative AI in Real World Workplaces.
[00:32:59] Mike Kaput: What this [00:33:00] report does is it synthesizes a bunch of findings from a number of recent Microsoft studies on how these tools are having an impact in the real world, both generative AI as a whole and Microsoft Copilot specifically.
[00:33:15] Mike Kaput: So here's a few highlights and Paul, I know you had quite a few thoughts on some of the other stuff we're seeing in this report as well, but some of the highlights include one of the studies was a large scale randomized controlled trial across 60 plus organizations It found that. Copilot users read 11 percent fewer emails, spent 4 percent less time on emails, and edited 10 percent more documents using the tool.
[00:33:41] Mike Kaput: In another survey, interestingly, Microsoft found that 78 percent of knowledge workers out of 31, 000 surveyed said they used at least some AI tools not provided by their organization. And another survey on copilot usage showed that the perceived benefits from copilot increased. with [00:34:00] longer usage duration.
[00:34:02] Mike Kaput: So, Paul, to kind of like pull all these threads together, it seems like AI is actually contributing real business value to someone like Microsoft. The adoption appears to be increasing according to the studies. Like, what jumped out at you about what they were finding in terms of how people were using either co pilot or generative AI?
[00:34:23] Paul Roetzer: So, I want to, like, laser in on the one where they studied 60 organizations and 6, 000 individuals across industries. So, you talked about what were pretty underwhelming. Results like people with Copilot within this control group read 11 percent fewer individual emails and spent 4 percent less time interacting with them.
[00:34:45] Mike Kaput: Right.
[00:34:48] Paul Roetzer: Is email really the interesting use case here? Like it was focusing obviously on, on products and capabilities that, 365 enables. So they looked at like meeting time and emails [00:35:00] and stuff like that. I think the one they talked about was summarizing emails. So they hypothesized that people were spending less time on emails.
[00:35:08] Paul Roetzer: Because maybe they were using the summarize feature. I would think they'd be able to monitor that. But, I always like kind of laugh at the summarize feature. Again, maybe it's just me, but who uses that? Like, it's the first thing everybody's felt like Zoom has a summarize this thread. It's like five, there's like a total of 50 words in the thread.
[00:35:27] Paul Roetzer: I don't need you to summarize it for me. And like, how long are your email threads that you need the summarize email button? Like just. Read the last email in the thread usually. So I don't know, like, I feel like they were. Assessing this against features that I don't find that interesting. And maybe this goes back to the Chevron conversation where we were saying like they had 20, 000 copilot licenses and they weren't sure of the ROI.
[00:35:49] Paul Roetzer: Well, if, if the features, the AI capabilities you're using are summarizing emails and meetings and like, that's [00:36:00] it. And you're not individualizing use cases to people and like providing training and education of how to use them. then okay, yeah, I guess maybe Copilot is pretty useless. I don't know. Like, so I was a little bit sort of surprised at how they conducted the study.
[00:36:18] Paul Roetzer: Like, if you're Microsoft and you want to show the value of Copilot, giving it to 6, 000 people across 60 organizations and waiting to see if they sent fewer emails or if they spent less time in meetings, if that's your measurement of whether or not they got value from Copilot. then I think you've got a bigger problem on your hands as Microsoft.
[00:36:43] Paul Roetzer: So, I see the problem, and I kind of highlighted this in our last conversation. 6, 000 people got Copilot with, I, I, I read the doc and I scanned it to make sure I didn't miss something. I keyword searches for training, education, and onboarding. They don't appear in the document related to [00:37:00] any of this. So it doesn't seem like they taught anyone how to use Copilot for one.
[00:37:05] Paul Roetzer: They didn't use any imagination in the use cases for it. They just literally sat back and waited to see if they sent fewer emails. So they said, what about. Teaching people like a marketer, like how do you use Copilot to do marketing more efficiently? So there's just like these homogenous applications.
[00:37:20] Paul Roetzer: Like we're just going to assess the use and value of Copilot based on emails, meetings, and documents. Okay. Well, what about I'm a marketer? I'm a salesperson. I'm a customer service. So give me five use cases, teach me how to use Copilot to do those five use cases, benchmark my performance before and after, like, why do we have to do these like blind studies and just see how many people figure out how to use Copilot?
[00:37:41] Paul Roetzer: So, I don't know, like, Iwas kind of excited to read this and then I got into it and I was like, oh, like this, this is what they spent months doing? Like, this seems like a totally backwards way to do this. So my takeaway here is,
[00:37:54] Paul Roetzer: as we've said before, like, research reports can be really valuable. [00:38:00] Don't rely on these to know whether or not Copilot is going to create value for you and your company, or ChatGPT, or Google Gemini, or whatever it is.
[00:38:09] Paul Roetzer: Do your own work. Individualize, personalize the use cases, train people on the platform, create benchmarks, run 90 day studies to look at the improvements. This is not adoption done the right way for generative AI is basically what I'm saying. And I would expect Microsoft would have had the foresight to do this study differently, because it's their product they're trying to show the value for, and yet they just, I don't know, it didn't do it.
[00:38:36] Paul Roetzer: Like, I'm sorry, a 10 percent increase in, or reduction in time spent on emails with some hypotheses as to why that would be occurring, such as maybe they're using the summarize function, like. I'm not buying 10, 000 licenses for Copilot based on that
[00:38:51] Paul Roetzer: data. I don't know. Did you, what did you think? Like, I was, I don't know if I was being harsh or what.
[00:38:57] Mike Kaput: it did make me feel a [00:39:00] little crazy because I was like, okay, this is like mildly interesting. I'm glad we're starting to ask some questions around real world applications versus some of the research they previously kind of just done in a lab or in a really isolated use case.
[00:39:15] Mike Kaput: So I appreciated that, but I was like, we're sitting there day in, day out talking to people, running workshops, doing speaking engagements where it's like, yeah. Even like on a few ounces of creativity is like, look what you can do with document analysis. Look what you can do when you start using this for ideation or as a strategy assistant.
[00:39:34] Mike Kaput: And I'm not saying like, Copilot can do every single possible thing we can envision. Using generative AI tools for, but even I just think of like the slides we show all the time in intro to AI or my applied AI talks or your, your strategic leader talks like these are way more creative
[00:39:52] Paul Roetzer: Yeah.
[00:39:53] Mike Kaput: than any of the things in these studies.
[00:39:55] Mike Kaput: And I'm like, I don't think they're even rocket science either. So I'm [00:40:00] curious as to like, where the disconnect came from. Like, did Microsoft. want to play it safe? Did they just not think of? I can't believe they didn't think of these things.
[00:40:10] Paul Roetzer: I I know, and I think my biggest problem was like, when you read the introduction, it literally says, we believe this research is the largest AI deployments to date. Okay. If you're going to make that claim, this better be good. This better not be, you think people are using the summarize feature in emails to save time.
[00:40:27] Paul Roetzer: So again, yeah, not to kill Microsoft for this, but this
[00:40:32] Paul Roetzer: is, is not what it was supposed to be. Like this could have been way, way better if. They took people, like the BC, I always go back to like that BCG one with the consultants. They took, even that one, they didn't train them how to use GPT 4. They just gave it to them.
[00:40:46] Mike Kaput: Right.
[00:40:47] Paul Roetzer: So I think the opportunity is if we re phase this, we believe that this research is the largest controlled study of productivity impacts in real world generative AI. Someone has the opportunity to do the largest one that [00:41:00] actually. Customized use cases, train people how to use the platform and benchmark before and after.
[00:41:05] Paul Roetzer: Now, I'm reading. Like, so if someone does that study or knows of that study, please let Mike and I know.
[00:41:11] Mike Kaput: Yeah, for sure. And I, you know, I, again, you're right. Like I'm not trying to bag on their really great work here.
[00:41:16] Mike Kaput: Or like, I'm sure it's took a lot of effort, but it feels like they might have, and I understand why, focused on, we want to do the largest controlled study and didn't control for doing an interesting one, you know? So it's like, I'd rather have a study where it's like, Hey, we found 50 people and that have used AI to 10X their output.
[00:41:34] Mike Kaput: And I'd be like, okay, I've read, I'll read that. I mean, just, I realize it's not controlled across all the groups, but that's like way more interesting to me. I guess Maybe it's just
[00:41:43] Mike Kaput: me. I don't know.
[00:41:44] Paul Roetzer: I, I'm with you. I mean, we need to run that study. I don't know.
[00:41:47] Mike Kaput: Yeah. Right. Then maybe that's next state of the industry, report. All right. Let's dive into some rapid fire this week.
[00:41:54] Mike Kaput: So first up also, we saw some, Quarterly earnings come [00:42:00] out from Meta, which saw its shares rally a bit, um, after reporting some strong second quarter earnings. And it also kind of came concurrently while CEO Mark Zuckerberg is kind of reassuring his investors about the company's significant investments in AI.
[00:42:16] Mike Kaput: So at the time Meta's shares jumped more than 7. 5 percent in pre market trading following That earnings report, because their net income for second quarter climbed 73 percent to 13. 47 billion, which exceeded expectations. Now, this marked a turnaround from, the reaction to Meta's previous earnings report in April, when Zuckerberg's announcement of slower growth and increased AI spending sparked a bit of a stir.
[00:42:44] Mike Kaput: of a sell off. Now, what's really interesting here is that Meta has increased the lower end of its CapEx range to account for AI projects. So, we talked about that the previous week, that these companies are making massive CapEx [00:43:00] expenditures on AI. And Zuckerberg actually emphasized these investments are beginning to pay off.
[00:43:05] Mike Kaput: Not to mention, another point we mentioned, we The, you know, emphasized when we talked about Llama last week, he stated that the costs are outweighed by the company's strong performance in its core advertising business. So Paul, this really hits on this point that Meta has a bit of a unique advantage in the AI arms race because they don't have to make money selling AI models, which is kind of why they're going this open source route.
[00:43:31] Mike Kaput: And it sounds like they also have room to subsidize AI because, They have a whole other core business that does really, really well. Like, is that how you view earnings like this?
[00:43:40] Paul Roetzer: Yeah, and, you know, there was excerpts from it, like different quotes I saw from Zuckerberg where they basically said that it's like, listen, we're, I don't know, it was like 55 billion or something like that they're spending. Um. on research and development and stuff. And, and which I think is more than anybody else at this point.
[00:43:56] Paul Roetzer: And he's pretty nonchalant about it. It's like, yeah, like [00:44:00] we're, you know, we'll figure out how to make money down the road. Like we're not too worried about it. And you know, we're going to keep doing what we got to do. We're going to keep investing. Like he's, I think he said something like the risk of under investing is far greater than the risk of over investing.
[00:44:12] Paul Roetzer: And so we're going to pour money into this space. Um, we'll put in the show notes. I haven't had a chance to watch this video yet, but it's on my list to watch this weekend. Um, Jensen and Zuckerberg, so Jensen, the CEO of NVIDIA and Zuckerberg, did an interview together, kind of like a mutual fireside chat, if you will.
[00:44:31] Paul Roetzer: And it's got 1. 9 million views. It was just posted three days ago. So AI and the Next Computing Platform, was their topic. So. we'll put, again, put the link in the show notes. It's on YouTube. It's readily available. NVIDIA published it on their YouTube channel. so that'll, that's probably gonna be a fascinating listening.
[00:44:49] Paul Roetzer: And if we hear anything interesting in it, we'll, we'll pull out some excerpts for next week's show.
[00:44:54] Mike Kaput: So next up, Perplexity, which we talk about more and more these days. The AI powered [00:45:00] search engine is launching a new publisher's program to share ad revenue with publishing partners.
[00:45:06] Mike Kaput: And this kind of comes as a response to all this controversy surrounding the company's use of content from various publications.
[00:45:14] Mike Kaput: They're being kind of tagged for showing a bunch of content that feels like it's a bit plagiarized or used without proper or clear attribution in their results. So, the first batch of Partners in this program includes some prominent names like Time Magazine, Fortune, Entrepreneur, and Automatic, which is the company behind wordpress.
[00:45:34] Mike Kaput: com. And as part of this deal, publishers are going to receive a share of ad revenue And their content is featured in Perplexity's responses to user queries. Now, the revenue share structure is a multi year agreement with quote, double digit percentage that is consistent across all publishers, according to the verge. So, Paul, we've talked time and time again about how AI [00:46:00] companies are kind of starting to pursue these licensing deals with content publishers because they're trying to avoid issues with being sued for training on copyrighted material. This is like a slightly different twist on that. I mean, are ad revenue sharing programs like this a viable way forward for perplexity given its current controversies and woes?
[00:46:23] Paul Roetzer: mean, just, it's just It's probably helpful from a legal perspective. I mean, I think they're still going to get sued by people, but yeah, I mean, this is the deal publishers have to make. It's, it's either we're going to sue you and, and try and get the money, or we're just going to accept that this is the future and we're going to play the game and it's not just the training data.
[00:46:42] Paul Roetzer: It's the inference too. It's when, when the question is asked, do they have access to real time data from these publishers to pull it? I didn't read the full article, so I don't know if they talked about like, what about. Brands, like you're training and using our data in real time too. Like, can we be part of the publisher program?
[00:46:59] Paul Roetzer: Or [00:47:00] do you have to be a big media company who can sue them before you're allowed to be in the publisher program? The other question I would have is what does this do to the quality of Perplexity's results or outputs? Like, will partners get preference in citations? Will I know it's a partner that I'm getting the citation for and maybe not actually the best quality source?
[00:47:19] Paul Roetzer: Like with Google, at least I know what the sponsored links are. Am I going to Is the citation number going to be blue if it's a sponsored citation? Like what? I don't know. I don't know how the model exactly is going to play out. So I think it's interesting. It's probably something they have to try to alleviate some of the legal pressures around their model.
[00:47:39] Mike Kaput: right.
[00:47:40] Paul Roetzer: whether or not it works long term, I don't know. And calling it like ad revenue is interesting since they don't have ads. So that's why I'm wondering, like, are they paid citations then? Or it's just If their model happens to organically surface a citation to a company they have a deal with, then they'll give them money.
[00:47:57] Paul Roetzer: And then we're just going to trust perplexity to actually, like, [00:48:00] report properly on that. I don't know. I've, I obviously have way more questions about this model than I have answers,
[00:48:05] Paul Roetzer: But I do think we'll see more experiments like this run that, that are probably going to keep evolving over time. But I do wonder about the brand side of this.
[00:48:14] Paul Roetzer: I mean, we all, everybody's listened to this, probably works for a brand or owns a company. I mean, We all spend money publishing content like this podcast. And if our transcripts start showing up as answers to questions and, you know, they're citing us like our, do we get money? Like, can we join this program?
[00:48:31] Paul Roetzer: I don't know. I'm not sure how it's going to play out.
[00:48:35] Mike Kaput: All right, so we have a quick update on a piece of U. S. legislation at the state level, which several months ago we had mentioned was kind of worth keeping an eye on. this law is in California, and it is called the Safe and Secure Innovation for Frontier Artificial Intelligence Models Act.
[00:48:56] Mike Kaput: It's a mouthful. it [00:49:00] is, I don't know if that's what SB stands for. That's probably like, I don't know, like Senate bill or
[00:49:04] Paul Roetzer: Okay. I was saying, where's the B in that?
[00:49:05] Mike Kaput: But yeah, that would be my guess. but it is for short SB 1047, which is how a lot of people are referencing it online.
[00:49:14] Mike Kaput: Basically, this is a law designed to regulate large AI models and prevent potential catastrophic harm from those. And so it is just past the California Senate and is expected to face a final vote in the state assembly in August. Now, the reason we're talking about this is that because it's in California, it's drawing a ton of attention from Silicon Valley and the tech world, because there's a lot of controversy around SB 1047 because it requires quite a few, um, mandates around how to regulate the safety of these models. So
[00:49:51] Mike Kaput: companies developing, models of a certain size have to implement testing procedures and also systems to prevent and respond to [00:50:00] safety incidents. Um, these are things like. The bill focusing on future AI systems that might autonomously engage in harmful behavior. And so it mandates that companies training large models have the capability to promptly shut down their systems.
[00:50:16] Mike Kaput: This was termed a few months ago, a quote, kill switch by some people analyzing it. Now, there are some supporters of this bill, including prominent AI researchers, Jeffrey Hinton and Yoshua Bengio. They are pro the bill because they argue it mitigates risks that could be down the line posed by advanced AI systems.
[00:50:38] Mike Kaput: But there is a lot of talk in Silicon Valley and among AI companies of criticism, just like really categorizing the bill as more of what we might call the quote AI doomer argument, you know, assuming that AI systems will be largely autonomous and capable of harm. At a wide scale. Now, Paul, [00:51:00] there's a lot of pushback from AI leaders on this.
[00:51:05] Mike Kaput: one of the top kind of critics of it has been Andrew Ng, who's a huge voice in AI, and he's basically arguing that SB1047 stifles innovation. It goes way overboard trying to hold AI companies liable if someone uses their models for harm. So, It really does seem like some of the top voices in AI are coming out and worrying that this bill could be over reached. Like, what do we need to keep an eye on?
[00:51:33] Paul Roetzer: Yeah, interesting timing. The White House released a report on July 30th. So as we're grappling with this, you know, bill that's moving so quickly in California and the headline of the AP story was White House says no need to restrict open source AI, at least for
[00:51:50] Paul Roetzer: now. And so that was in reference to a Department of Commerce, National Telecommunications and Information Administration.
[00:51:58] Paul Roetzer: Talk about another mouthful. [00:52:00] They issued policy recommendations embracing openness in AI while calling for active monitoring of risks in powerful AI models.
[00:52:09]
[00:52:09] Paul Roetzer: Uh, Assistant Secretary of Commerce for Communications and Information, Alan Davidson, said the openness of the largest and most powerful AI systems will affect competition, innovation, and risks in these revolutionary tools.
[00:52:24] Paul Roetzer: The report recognizes the importance of open AI systems and calls for more active monitoring of risks from the wide availability of model weights for the largest AI models. Government has a key role to play in supporting AI development while building capacity to understand and address new risks. Open weight models allow developers to build upon and adapt previous work.
[00:52:45] Paul Roetzer: Broadening AI tools availability to small companies, researchers, nonprofits, and individuals. This would seem to counter the SB 1047 bill, I would think. It was almost like the government, without being able to step in, sort of making it [00:53:00] clear that they don't want to stifle. innovation at this point that they don't think we're at a point where risk is so severe that we should halt things.
[00:53:08] Paul Roetzer: So yeah, I just, I mean, it's continually interesting to me how this will play out, especially given that this is in the state where most of AI innovation is happening, like Silicon Valley is the epicenter of AI in the world right now. So the fact that that is where this is occurring is just so interesting.
[00:53:27] Paul Roetzer: So yeah, I mean, definitely worth keeping an eye on. The federal government's obviously got a voice in all of this. Um, interesting how it plays out.
[00:53:36] Mike Kaput: So next up, an AI company that we mention often on the podcast, Writer, who is a valued partner of Marketing AI Institute. They have actually launched. Two specialized large language models designed specifically the healthcare and financial services industries.
[00:53:55] Mike Kaput: So these new models are called Palmyra MED 7DB and [00:54:00] Palmyra FIN 7DB. MED for medical, FIN for finance. And they're now available as open source offerings on major AI platforms. The company says that these specialized models outperform larger generalized AI models like GPT 4. in domain specific tasks. So, for instance, the healthcare model Palmyra Med 70B, has achieved an average accuracy of 85.
[00:54:25] Mike Kaput: 9 percent across all medical benchmarks in zero shot attempts, which surpassed some competitors. The company claims that the Palmyra Fin 70B model is the first AI model that is capable of passing the CFA Level three exam, which is considered one of the most challenging tests in investment management. So, Paul, I want to kind of dial in on this idea of AI being applied to specific domains of knowledge work.
[00:54:58] Mike Kaput: Like right now we have [00:55:00] broadly capable general models. Like what happens when people start trying to make models that are really good at specific industry roles or tasks?
[00:55:10] Paul Roetzer: I mean, I think we're going to see a lot more of this. I will say, like, while we know Writer, and have, you know, worked with them for years and been aware of them and their product and their leadership, I have no firsthand knowledge of how these models were trained. And so if I objectively look at this, the thing I always wonder is like, how are you Training competitive models without a billion dollars in funding.
[00:55:37] Paul Roetzer: That's like always my first question is like, are you building on top of someone else's open source and then like fine tuning it for a minute? Like, and I don't know how these are done. I think that they're doing their own. I just don't know how they're doing that. When we think about vertical solutions, and so here's med and finance, but you could do this in any, any industry you want, what is going to make these models [00:56:00] unique is.
[00:56:01] Paul Roetzer: To my understanding, two key components, proprietary data to train on, high quality proprietary data to train on, that's specific to that domain. so for example, doing deals with healthcare organizations where you get access to proprietary data that you can use that isn't publicly available to these frontier models.
[00:56:19] Paul Roetzer: um, and then reinforcement learning through human feedback. Having medical professionals who know the field. work to train these models post training. So you do the run and then you have your core model and then you have, say, 10 doctors and those doctors work with the model and look at the outputs. And then they say, this output is better than this output.
[00:56:40] Paul Roetzer: And they follow guidelines of like what the output should be. And then they use their domain expertise to, to train them after the initial training run. And so data plus reinforcement learning through human feedback with experts in that domain. could give you a model that's better than the frontier model in theory at [00:57:00] doing a specific industry.
[00:57:02] Paul Roetzer: And so if that stays true, if GPT 5, GPT 6 doesn't just like eliminate the need for these vertical models, that it's just better at everything than any fine tuned model, then those are the things you would look for. And again, I think if You're doing a custom model in your own company. Like you're thinking about this for customer service or things like that.
[00:57:23] Paul Roetzer: Then you may take like a Llama 3. 1 model, an open source model, and then use your own proprietary data and hire experts or take your internal exports and do reinforcement learning through human feedback, or LHF. To make that model specific to your company or your industry. And so that's what we're seeing here.
[00:57:43] Paul Roetzer: I assume, again, I don't know how they did this, but I can't imagine any way they would do this without proprietary data and RLHF. and I think we may see more of these kinds. I know I, we talked to a lot of companies. This is what they ask. How do we train our own model? Like we don't, we don't want to use open AIs.
[00:57:59] Paul Roetzer: We want to like, [00:58:00] we want to build our own and own it. and we actually, we have a. Yash Ghat, I think, from Ringer Sciences is doing a talk on custom models, with your own data at MAICON, if I remember correctly seeing that title today.
[00:58:14] Mike Kaput: So next up, with very little formal fanfare, Google has made an experimental version of Gemini 1. 5 Pro available for early testing and feedback in Google AI Studio and through the Gemini API. And it's making some waves for Pretty great performance. So this experimental version of the model is called Gemini 1.
[00:58:39] Mike Kaput: 5 Pro Experimental 0801,
[00:58:43]
[00:58:43] Mike Kaput: it has now actually claimed the top spot on lmsys. org's chatbot arena leaderboard. It surpasses currently GPT-4o and Claude 3. 5. Now that leaderboard changes quite a bit, so That could change, but it also, [00:59:00] for the time being, hit number one on the lmsys. org vision leaderboard.
[00:59:05] Mike Kaput: And this is the first time a Google Gemini model has actually topped this leaderboard. So Paul, I thought like this observation from Ethan Mollick really kind of encapsulated this story. He had posted, companies are just releasing next level models without telling anyone about it. Just quiet, somewhat baffling rollouts.
[00:59:26] Mike Kaput: Gemini 1. 5 Pro Experimental 0801, really rolls off the tongue there, is at the top of the major AI leaderboard. What's different about it? Who knows? He says. Like, what's going on with this behavior from major companies? This feels like something that you would not expect of someone like Google.
[00:59:44] Paul Roetzer: I think what's going on is they're so focused on, and this is just a theory. And honestly, I hadn't even thought about that. You just asked this question. I think what's going on is they are building the next version of the frontier [01:00:00] model and they have some element of it that they can just go ahead and put out safely into the world and they know if they put it on,
[01:00:12] Paul Roetzer: how do you say that?
[01:00:13] Mike Kaput: lmsys. org is, yeah,
[01:00:15] Mike Kaput: no kidding, I honestly, I don't think I've ever heard anyone say it, I'm just going based on
[01:00:19] Paul Roetzer: I see it all the time.
[01:00:21] Mike Kaput: the acronym, so
[01:00:22] Paul Roetzer: So they, they know they have a captive place where there are testers who have fun doing this. It's like a game to them, going and play with the latest models so they can get user feedback. I don't think it's intended as a release for you and I, and for businesses. Like we're just, because of the popularity of this stuff and how fast it's changing, we're kind of just.
[01:00:46] Paul Roetzer: Exposed to the inside of how this all happens, like where people are giving access to early test of models to provide feedback and stuff so they can build that feedback into the real thing, [01:01:00] which is when they will actually make a big announcement from Sundar, not from whoever on the Google team that tweeted it.
[01:01:07] Paul Roetzer: So I think that's what's happening is they're all focused on the big thing.
[01:01:14] Paul Roetzer: and. In, in the process, they release a bunch of little things that, yeah, they could make an announcement about it. They could tell you what's different, but they don't really care that it's just here. Have fun with this. Like, oh, okay, great.
[01:01:26] Paul Roetzer: It's now number one. So to us, it's like, holy cow, like Google just released a model. That's now number one on this leaderboard ahead of GPT 4. What does that mean? Does that mean it's like. Is now the most powerful model in the world. And Google is probably sitting there laughing, like, I don't know, whatever.
[01:01:41] Paul Roetzer: Like, and because they've probably already got Gemini 2 extreme, whatever they're going to call the next one. It's probably already been in testing for three months at Google. And like, they're just laughing at us because we think 1. 5 is fun. And they just think it's some cute little thing. That's like.
[01:01:57] Paul Roetzer: Sam's talked about that before, how they would just release things that they didn't [01:02:00] think were going to be a big deal because they'd been using it already. And, and then people were like freaking out about it. So I think that's what's happening. I think we're just getting an inside look at them releasing some stuff to experiment with it and see how people react to it and do some testing with it.
[01:02:15] Paul Roetzer: And then two months from now, they'll drop the real thing on everybody with, you know, an announcements from Demis and Sundar, a joint blog post from Demis and Sundar. Like, I don't know. I assume that's what's happening.
[01:02:28] Mike Kaput: And that's exactly kind of why I wanted to quickly mention this topic and talk about that Mollick
[01:02:34] Mike Kaput: Post really
[01:02:35] Mike Kaput: quick is because like, I can very well see like plenty of really savvy business leaders that we know and talk to, you see the headline like, Oh, Google like secretly releases.
[01:02:45] Mike Kaput: New model that tops the leaderboard. And you're like, Oh no, like, do we make the wrong move? Like, you know,
[01:02:49] Paul Roetzer: We just built on top of Llama?
[01:02:51] Mike Kaput: not really following this in certain ways, it's really easy to get caught up and be like, like, you cannot sit here and wait for the [01:03:00] right model
[01:03:00] Mike Kaput: to come out and you need to kind of. Dive a little deeper into the headlines, I would
[01:03:04] Paul Roetzer: Well, that's the whole idea last week I set up, like, are we just in the stage now where there's just parallel innovation? Like just everybody's like every other week somebody's been testing a secretive model and all of a sudden they gave it a name and oh my, Anthropic Claude 3. 82 is, is now like the, like, I just think people, it's fun.
[01:03:22] Paul Roetzer: Like it's, it's cool to like, see what's happening. But it doesn't change your life, like, as a business person, as a marketer, like, whatever you do. Like, it, it's just interesting to know, but it's probably more of a prelude to something big is probably coming. And whether or not the big thing that's coming is a leap forward or a parallel innovation is really what we're trying to keep the pulse of.
[01:03:47] Mike Kaput: Alright, our last
[01:03:49] Mike Kaput: topic for this week is that Microsoft has officially added OpenAI to its list of competitors in its latest annual report. [01:04:00] So this change comes shortly after OpenAI announced a prototype of a search engine called
[01:04:05] Mike Kaput: SearchGPT, which
[01:04:06] Mike Kaput: we talked about last week.
[01:04:07] Mike Kaput: And the tech giant is now Calling OpenAI a competitor in AI offerings and in search and news advertising. Now, obviously, this is a little notable, given that Microsoft has a big investment in OpenAI. It's reportedly around 13 billion, and they've got ongoing partnerships together. Now, despite this
[01:04:29] Mike Kaput: new stance, an OpenAI spokesperson said that nothing about the relationship between the two companies has changed.
[01:04:36] Mike Kaput: But,
[01:04:37] Mike Kaput: I guess I gotta ask, Paul, how likely is it that we actually see Microsoft and OpenAI fall out of favor with each other? Like does this actually Threaten at all? Like the dynamic of Microsoft having access to OpenAI's technology and models?
[01:04:53] Paul Roetzer: I mean, unless Microsoft's planning on loading their Whatever percentage they own, I think it was 49 percent of open AI [01:05:00] and the rights to the first hundred million in profits. Like, I don't see a breakup happening and there's no way Microsoft would be allowed, from federal regulators to acquire open AI, which would seem like a natural thing if they wouldn't get caught, up up in, You know, investigations into illegal practices or unfair competitive practices if they were to acquire them.
[01:05:24] Paul Roetzer: So I don't really see a way forward other than they're going to just keep working together. But man, is that awkward. Like I was trying to play that out in my mind, as you were talking of like any industry other than technology, where it was scenario like this would play out where someone who's considering you a major competitor, who you're building products to compete with.
[01:05:46] Paul Roetzer: Owns 49 percent of your company, but can't buy you. Like, what, where, I, how would that be normal? Like, you're just sitting in strategy meetings together, like showing them the next model because they have the rights to [01:06:00] not only see the next model, but to build products on top of it for anyone else. And you know that they're like your biggest competitor.
[01:06:06] Paul Roetzer: It's so weird.
[01:06:07] Mike Kaput: Yeah,
[01:06:08] Paul Roetzer: And maybe it's not, but it sure seems weird.
[01:06:11] Mike Kaput: It does.
[01:06:14] Paul Roetzer: All right, well, quick note. So as Mike always says, we have the This Week in AI newsletter, and I just want to give you a sense of like the stuff Mike and I didn't get to today. So we have Apple Intelligence Foundation Language Models. A new paper came out where Apple shared exactly how they built like their Apple intelligence.
[01:06:29] Paul Roetzer: Fascinating. We talked about the White House, and open source AI. So we've got the AP article linked into there. Meta, I wanted to talk about this one, but we just don't have time. Scraps their Celebrity AI chat bot, which was extremely predictable. And we talked about that when it first came out. apparently Google's Dear Sydney ad on the Olympics isn't hitting, as intended.
[01:06:52] Paul Roetzer: I have not seen it yet, but I have heard lots of very, visceral reactions to it. I think it's like outsourcing creativity was [01:07:00] kind of the sense I got of it. Elon Musk apparently is considered buying character. ai, which actually makes perfect sense for the testing of his Grok bot. AI startup Suno is, is actually like going back at the music industry.
[01:07:15] Paul Roetzer: so they're attacking the people, attacking them. And then Salesforce released a massive, Mint 1T dataset that, is talked about in VentureBeat as possibly disrupting the AI industry. So just to give you a sense, like when Mike and I do this every week, and that's cut down from probably 45 articles,
[01:07:33] Mike Kaput: Yeah.
[01:07:33] Paul Roetzer: podcast is great, but subscribe to the newsletter, you get all kinds of other stuff every week.
[01:07:38] Mike Kaput: Yeah, and you can do that by going to MarketingAIInstitute. com forward slash newsletter. And as I will always also say, if you have not yet left us a review on your podcast platform of choice, we'd greatly, greatly appreciate it.
[01:07:51] Mike Kaput: It is the number one thing that helps us get in front of more people and improve the show. So, Paul, thanks again for breaking [01:08:00] down this week in AI.
[01:08:00] Paul Roetzer: Thanks, Mike. I will. I'll see you next week,
[01:08:05] Mike Kaput: See you next week.
[01:08:06] Paul Roetzer: Bye everybody.
[01:08:07] Paul Roetzer: Thanks for listening to The AI Show. Visit MarketingAIInstitute. com to continue your AI learning journey and join more than 60, 000 professionals and business leaders who have subscribed to the weekly newsletter, downloaded the AI blueprints, attended virtual and in person events, taken our online AI courses, and engaged in the Slack community.
[01:08:31] Paul Roetzer: Until next time, stay curious and explore AI.