They Might Be Self-Aware Podcast (TMBSA) - EPISODE 45 SUBSCRIBE NOW OR BE A PAWN IN THE AI CHESS GAME! This week, @Microsoft and @OpenAI’s “bromance” seems to be heating up to the boiling point! Is their epic investment saga about to reach its AGI-induced climax? And @Baidu’s bold declaration – are AI hallucinations really a non-issue now, or is this just another CEO pipe dream? Meanwhile, @Anthropic unveils an AI that operates your computer like a caffeinated intern. Is this the dawn of an AI oligarchy with the Titans holding the reins, leaving us mere mortals scrambling for technological crumbs? In the cinematic corner, the showdown of the century: SORA versus Mochi AI! Who’s mastering the art of text-to-video creation? Spoiler: @MochiAI may still be catching up, but the open-source revolution could tip the scales in its favor. Join us on @Spotify, @Apple Podcasts, and all major platforms as we charge toward episode 50. Will we defy the stats? Stick around to find out! 🌟 Your future might depend on it! For more info, visit our website at https://www.tmbsa.tech/
00:00:00 - Intro
00:03:23 - Microsoft And OpenAI’s Tension Over AGI
00:07:03 - Baidu’s Bold Claim: End Of AI Hallucinations?
00:11:47 - Are The Titans Circling The Wagons?
00:22:26 - Video Generation Showdown: Sora Vs Mochi-1
00:26:14 - Open-source AI As A Check On Tech Giants
00:27:39 - Wrap Up
Hunter [00:01:40]:
I've got a topic for us today. You do the, I'm calling it the Great AI Power Play. There's the corporate dynamics and the future of innovation and they're, they're going, they're going, they're going after each other. Or another way is to think about is how, how the tech titans out there are playing chess while the rest of us, we're all just pawns in this game.
Daniel Bishop [00:02:27]:
Is this like how after Mistral or any of the other players release a new model, OpenAI has an announcement the next day? Pretty much every time it's a, it's.
Hunter [00:02:37]:
A little bit about that. It's also one of the great things about AI is its potential to democratize access to so many things. But at the same time, as you know, is inevitable. There's a power play of large companies trying to suck all that up so that us little guys don't necessarily get a chance. And so, I don't know, trying to pull back the velvet curtain and reveal, oh, it's definitely velvet. It's a pretty curtain. But take a look at some of the messy and competitive and sometimes we could probably label it as cutthroat world of AI development. You and I spend a little bit of time in and it's a bunch of time.
Hunter [00:03:23]:
It's not all silicon and sunshine. As I hate to burst your bubble.
Daniel Bishop [00:03:28]:
Wait, we're in a bubble?
Hunter [00:03:29]:
What we don't mention the bubble. That's the first rule of AI.
Daniel Bishop [00:03:34]:
Don't talk about.
Hunter [00:03:36]:
Yeah, okay, so I thought start talking a little bit about the Microsoft OpenAI saga. The so called bromance. It's been heating up. Once upon a time these two were the, I don't know, the dynamic duo of the AI world. Because Microsoft just poured billions into OpenAI.
Daniel Bishop [00:04:02]:
Some are claiming like a 50% stake basically.
Hunter [00:04:06]:
Well, yes, but a 50% stake that reverts back to OpenAI the moment that.
Daniel Bishop [00:04:13]:
They achieve AGI and money becomes irrelevant.
Hunter [00:04:16]:
With OpenAI gets to define what AGI is. So.
Daniel Bishop [00:04:19]:
Oh well. Shoot.
Hunter [00:04:20]:
Yeah.
Daniel Bishop [00:04:21]:
We did it, ladies and gentlemen.
Hunter [00:04:23]:
And by the way, there are rumors that they are close to claiming that because it gets them out of this agreement.
Daniel Bishop [00:04:30]:
But is that just like you were really excited about Strawberry a month or so ago? They did it.
Hunter [00:04:35]:
We're there.
Daniel Bishop [00:04:36]:
We finally got it. It's only going to cost you $1,000 a month.
Hunter [00:04:39]:
I was ready to pay the thousand dollars a month just like I would. I don't know that I would keep.
Daniel Bishop [00:04:44]:
Paying but like it was all hype.
Hunter [00:04:46]:
But again, I've said this before, I'll purchase pretty much anything. Like I'm a very easy sell.
Daniel Bishop [00:04:52]:
Right. You're one of the eight people that bought the Vision Pro.
Hunter [00:04:55]:
I am. And apparently they're stopping production on those at the end of this year. That's another thing. They are still working on new versions. In fact, there probably will be a new version if we're going to go on tangents, but a lower cost version that won't be as specked as high. And then potentially in the following year.
Daniel Bishop [00:05:12]:
It'S going to be called the Air. Right. Just like they have the MacBook Air.
Hunter [00:05:16]:
Possibly. I do know also their Vision Pro.
Daniel Bishop [00:05:18]:
Vision and Vision Air. Vision Air. Oh, that's fun.
Hunter [00:05:21]:
All right. All right. Back to Microsoft and OpenAI. So this.
Daniel Bishop [00:05:24]:
Yes. Right.
Hunter [00:05:25]:
This relationship is starting to sour a little bit as there are rumors. No one's, no one's officially on the record here, but that Open AI is feeling the squeeze from Microsoft, Microsoft's grip. There's financial pressures. I mean we just Saw them go out and raise their $6.6 billion.
Daniel Bishop [00:05:44]:
Are they not wildly profitable and the.
Hunter [00:05:47]:
Leaders in the market, any good tech startup? Oh, there's been lots of disagreements over the compute resources as we've seen OpenAI going outside of Microsoft's cloud a lot in order to train their models.
Daniel Bishop [00:06:00]:
Right. They said, oh, we don't have enough GPUs, we're going to go play with Oracle also because we need even more.
Hunter [00:06:07]:
And there's an angle, I think, where people, they're suggesting that maybe Microsoft is intentionally trying to stifle their growth because. Right. If they, if they do achieve AGI, do they also, you know, replace Microsoft as they get out of their agreement with them? So, you know, maybe Microsoft isn't the benevolent, the benevolent giant they once. They once claimed they were.
Daniel Bishop [00:06:34]:
I don't think Microsoft would be trying to stifle its development just offhand because there are enough other major players out there that Microsoft will be in more of a losing sort of situation if OpenAI doesn't get to slash stay at the top of the game. And I do think they've been pretty much at the top of whatever lists you look at for quite some time.
Hunter [00:06:59]:
Except that they lose open AI the moment they actually achieve the top of the game.
Daniel Bishop [00:07:03]:
I mean, we lose everything as soon as we have true AGI.
Hunter [00:07:06]:
Yeah, but it's not going to be true AGI the moment that they claim. And it's not. The world's not going to change in a.
Daniel Bishop [00:07:13]:
World's changing a whole lot.
Hunter [00:07:14]:
It's not going to change even more the moment that Sam says, okay, we achieved AGI, so the Microsoft deal is dead. I mean, I think it's pretty well known that Microsoft is very actively building their own competing models that are separate from OpenAI.
Daniel Bishop [00:07:29]:
Should be five models.
Hunter [00:07:32]:
Yeah. So I don't know. Control is the name of the game. And Microsoft isn't the only player flexing its muscles. I don't know if you saw this in the news last week. Baidu's Baidu. For some reason, that doesn't sound right when I say it, but we'll go with Baidu. Their CEO, Robin Lee declared that AI hallucinations are now a thing of the past.
Hunter [00:07:57]:
Largely resolved, I believe was his direct quote.
Daniel Bishop [00:08:00]:
Right.
Hunter [00:08:00]:
He said a thing of the past.
Daniel Bishop [00:08:03]:
I guess largely resolved is true, but this really feels like a tech CEO trying to drum up interest for their thing. And when I say largely true being correct, there is a hallucination leaderboard that talks about hallucination rates or factual consistency rates. And OpenAI 01 Mini is right there up at the top. And then GPT4O and GPT4O Mini and GPT4 Turbo and GPT4 and then GPT 3.5 Turbo. You get into deep seq. Microsoft, a lot of the major players are well under a 5% error rate. But in our last episode we were talking about like what happens if something gives you incorrect tax advice or medical advice. 1% of the time being wrong about the dosage of a medicine is a lot of people that are going to be really sick or worse.
Hunter [00:08:58]:
Well, speaking of 1%, another thing Robin has suggested is that we are on the edge of a dramatic culling of the AI startup herd. He says, you know, only 1% of the companies that exist today are going to survive this impending bubble burst. I guess he did, he did mention the bubble.
Daniel Bishop [00:09:19]:
I think that that's true. But it's not going to be quite like the.com era. The.com era saw a dramatic reduction in all of these service providing, I guess, websites and I'm not going to say the whole Internet sort of imploded. But there were a lot of people that thought that there was gold in them thar hills. And it turns out, yes, there is and was. But a lot of people were not accurately set up, I think, to properly monetize. And I do think that that's the case for a lot of these AI startups that we have right now because a bunch of them are basically just wrappers around GPT4 or Claude 3.5 or whatever. And so we've talked about not having a moat, a tech moat or any kind of like real value that that actual company brings, other than having done some of the technical work to make this API a little more consistent or easy to use, I think a lot of those are going to go away.
Daniel Bishop [00:10:18]:
And we saw a round of that already happen like about a year ago now when GPT4. No, yeah. I think it's for GPT4 that ChatGPTs became available where you could set up your own set of system prompts and so on to get ChatGPT to respond in a specific kind of way for a specific sort of Persona. And the technology has gone sort of leaps and bounds since then. A lot more capabilities coming up. But like someone comes up with a cool thing, oh, I got an LLM to do this new agentic thing.
Hunter [00:10:52]:
Cool.
Daniel Bishop [00:10:53]:
One month later that's built into any of the big players.
Hunter [00:10:56]:
Sam likes to talk a lot. Well, a lot of startups are to go out of business with this new release. He's pretty, pretty, pretty upfront about it.
Daniel Bishop [00:11:02]:
I completely agree with him on that.
Hunter [00:11:04]:
I don't know that I've really seen it though, yet as I think back, like, have a lot of companies gone out of business because OpenAI exists.
Daniel Bishop [00:11:12]:
I don't know if it's. They've gone out of business. I wonder if it's just we're going to see less of the going into business.
Hunter [00:11:18]:
Yeah.
Daniel Bishop [00:11:19]:
And since everything is a subscription or a contract these days, maybe we'll see a lot of contracts not renewed another year or so from now. And then we start to see a bunch of companies going under.
Hunter [00:11:31]:
So the question is whether these SaaS companies, big players, the Microsoft, the OpenAI, the Baidu, etc. Are they circling the wagons? Are they preparing to consolidate power while all of us small fries get, get, get officially fried?
Daniel Bishop [00:11:47]:
So the viability of a business in general. Yeah, I think, is you're providing a product or service that people want and are willing to give you money for. Like that's, that's capitalism, baby. Right?
Hunter [00:12:01]:
Yeah, that's part of it. Yeah, sure. A lot of when you start a business, there's a very low percentage of success rate.
Daniel Bishop [00:12:08]:
Yeah.
Hunter [00:12:08]:
By the way, for podcasts, apparently it's like only 2% of podcasts get to 50 episodes. And I think this is like, maybe we're on 44 now.
Daniel Bishop [00:12:19]:
And if you want us to get to episode 50, make sure to like and subscribe and tell your friends.
Hunter [00:12:24]:
Yeah, these next six are going to be intense, apparently, because we only have a 2% chance of making it to 50. Yes. Most people fail at most things.
Daniel Bishop [00:12:33]:
Sure. And the point that I really want to make about LLM based companies here is that because most companies don't actually have a data mode, they don't have something that can differentiate them from the market other than just we're the ones who programmed a thing to make that accessible that is wildly different from nobody could program this thing and the outcomes that it provides you, the customer with. Because I think in many cases it's not technical excellence, it's that you have data that underlies something that can provide the outcomes that are most relevant to whatever your business is. You know, let's say that you want a, a transcription. LLM or something like OpenAI has released multiple whisper models which underlie a whole bunch of different transcription services. There are a bunch of companies whose basically whole deal is we'll take in your audio and give you transcriptions and maybe a couple of Extra little features on top. And it costs money to do those things. But let's suppose that you're a company that already has a million hours of high quality phone conversations or, you know, podcast episodes that your network already had accessible and you already had humans manually labeled the transcriptions for each of those things.
Daniel Bishop [00:14:03]:
You could then fine tune a model with your data that isn't accessible to other people and make a better product. And that product can be sold for something that isn't. Just like, I'm the first person to say I took podcast transcripts and made those easily accessible for whatever reason. You can make a business out of that, but it's way less defensible than I made something that other people can't.
Hunter [00:14:29]:
Yeah, AI is definitely becoming more autonomous. I don't know if you saw Anthropic's latest release. They now have a. I guess it's an app that you run on a computer, but you're able to. You're able to give it a broad task that needs to leverage multiple applications on your computer. Basically use your computer as a human would and you can prompt moving the.
Daniel Bishop [00:14:52]:
Mouse around, not just clicking.
Hunter [00:14:56]:
Running different systems. Yeah, and not in the way. I don't know if you ever wrote scripts back in the day. I'm thinking of like auto hotkey, but I know there were like lots of different scripting languages where you would automate things on your computer. But you had to be like, go to these coordinates of exact pixel 252.
Daniel Bishop [00:15:12]:
Click here and wait 172 milliseconds.
Hunter [00:15:15]:
Seconds. Yeah, for. For three seconds. Then. Yeah, move mouse over here, hit this key combination. No, just say in their demo, I think they ask it to create a report by going into one of their backend systems and looking up some information and then copying that information over into a form that's open in another application. But really just using a computer like a human would use a computer.
Daniel Bishop [00:15:40]:
Or getting distracted like humans do. They mentioned on the press release that or at least one of the people talking about it, that in one of the demos it got distracted and started looking at pictures of vacation destination, I think is what it was. Which talk about emulating humans. You know, you're given this task and you just sort of mentally wander off and start doing something else.
Hunter [00:16:02]:
It's. I don't know. It's happened to me before.
Daniel Bishop [00:16:03]:
Maybe they are getting self aware.
Hunter [00:16:07]:
Wink. I don't know. The. I think the question is like, are. Are we witnessing the dawn of an AI oligarchy where only those who play by the rules of the Titans. Even with Anthropic's latest release, there's a whole bunch of things you can't do. You're not allowed to tell it to post on socials. There's one example I don't think you're allowed to tell it to log into any systems.
Hunter [00:16:34]:
They're creating all the rules, they have all of the resources. Training these foundational models is, you know, quickly moving outside the grips of the small players.
Daniel Bishop [00:16:45]:
Well, GPT3, let alone 3.5 and 4 were slash are not accessible to people.
Hunter [00:16:54]:
Yeah.
Daniel Bishop [00:16:55]:
Like outside of the company. Someone's going to make an open source version of this and it'll be available two months from now. People have already made something kind of like this. But you're, you're asking are we on the cusp of something? And I would say absolutely yes. We've talked previously about what would it mean to be like a being for an AI and I do think that embodiment or at least the ability to sort of affect the world in some way, shape or form around you can be a important part of that. To say nothing of people who are like locked in for example. They're obviously still people. But for an AI to have some sort of embodiment or some sort of ability to not just be chat but rather to actually do something to the world around it, I think is going to be closer to irrefutable evidence that this is a being even if we don't classify it as the same as a human.
Daniel Bishop [00:17:50]:
So giving it access to a computer is one step previous to we gave it one of those big industrial arms and just that that's its arm now.
Hunter [00:18:00]:
I guess I, I wonder if. So if these large companies control the cutting edge tech, will they inevitably start holding back the innovation in order to maximize their profits? They don't want to release the thing that eats their own company. Which makes sense happens today. I don't know. At least it gets accused of in pharma more than anything else. Or at least that's where I see it the most. Where people hold, want to hold back things because we got to get the most out of the current solution first. We got to.
Hunter [00:18:38]:
It's got to run its full cycle.
Daniel Bishop [00:18:39]:
And then wait for 20 years before the patent becomes.
Hunter [00:18:42]:
Yeah. So we hold back innovation.
Daniel Bishop [00:18:45]:
Yeah. I don't see why tech companies wouldn't do that. Is that the right thing for humanity? I would argue no. Same for the pharma companies, etc. And there's a famous example, true or otherwise But I know I've heard it a bunch of times of like a pantyhose manufacturer having bought the patent to an obviously superior set that just like didn't tear and just buried the patent because they didn't want other people or even themselves making something that people wouldn't have to keep buying more and more of when it comes to AI yeah. If there's something that would put themselves out of business because it's too capable, I think from a purely like capitalistic perspective, they have plenty of incentive to not make that readily and easily available.
Hunter [00:19:31]:
I had not heard the pantyhose story. Not well versed on pantyhose history, but maybe all I need to buff up on that.
Daniel Bishop [00:19:40]:
So.
Hunter [00:19:40]:
Yeah. So, all right, so what do we do about all this? I think awareness is the first step in trying to be have your eyes open of what's happening. All the open source stuff we see is certainly a good step, but we have to recognize the dynamics at play and question them. And hopefully we question them to some degree on the show. I guess we only have four episodes left according to the statistics. But yeah, we've got to support policies that promote open access to data, to computing resources, and encourage more investment in independent research and development.
Daniel Bishop [00:20:20]:
Now we need to talk to the folks at Nvidia, I think is what it is, because if they keep releasing 12 gigabyte video cards, we can't, we can't keep up. And speaking of a company not wanting to eat into its own profits, you've got these big bulky A1 hundreds and H1 hundreds and so on with, what is it, 48, I think, plus gigs of video RAM. If that became the consumer norm, then they couldn't sell the really big ones for $10,000. So I think that's a very good example of the exact kind of thing that you're talking about, where one way of making these sorts of things more accessible to the masses is software to distribute the training and availability of large, large, like the really big, large language models, but then also having hardware accessible that even could do that. And the main player right now, Nvidia, depending on which rumors that you're Looking at, the 5000 series models are not exactly going to be a quantum leap forward, or maybe it is quantum, but the actual meaning of very, very, very small in terms of what amount of video RAM is like the norm.
Hunter [00:21:30]:
Those are coming out next year, early next year.
Daniel Bishop [00:21:32]:
Next year.
Hunter [00:21:33]:
Yeah. I've been waiting to build a new system until they come out.
Daniel Bishop [00:21:36]:
Maybe you should just get one of the Mac M4s, once those are available.
Hunter [00:21:39]:
I'm going to do that too. Yeah. But they don't support Cuda and a lot of the open source projects require that.
Daniel Bishop [00:21:46]:
Right. Some of them are starting to support. What is it? Mlx.
Hunter [00:21:48]:
I think that's their metal version. So assuming you don't want to go into the cloud, you've got to give Nvidia a few. A few bucks.
Daniel Bishop [00:21:59]:
A few hundreds to thousands of dollars. Yeah.
Hunter [00:22:01]:
Yeah.
Daniel Bishop [00:22:02]:
To tens of thousands. To hundreds of thousands, depending on what sort of scale that you're working at.
Hunter [00:22:07]:
Yeah. And it's certainly not, not all doom and gloom out there. There are organizations and individuals pushing back, advocating for open source AI. There's new open source. Awesome, Incredible. Open source AI releases every single week, every single day, probably every single hour if you really spend enough time.
Daniel Bishop [00:22:25]:
Yeah, there's a lot.
Hunter [00:22:26]:
One that caught my eye this last week was Moki 1.
Daniel Bishop [00:22:30]:
Mochi. Mochi.
Hunter [00:22:32]:
Mochi. Mochi. It's probably Mochi. That sounds better than Mochi. We'll go with Mochi. It's open source Sora, for lack of a better phrase. It's. It's not of the quality of Sora quite yet.
Hunter [00:22:44]:
It's not quite there. I've seen a number of demos of it. I don't know, I guess maybe not everyone knows what SORA is.
Daniel Bishop [00:22:50]:
The number one. Sorry, go ahead and say what SORA is and then I want to make a very specific comparison.
Hunter [00:22:56]:
SORA is a generative model for text to video where you describe a scene and it renders that scene out as a video that you can watch. Yep.
Daniel Bishop [00:23:08]:
And it is not available to the public. There have been some backs, backs and forths with industry folks, industry being like in the film industry specifically. And the first demo video that I think pretty much everybody saw was of a stylish woman walking down the street in Tokyo and there's reflections off of. Off of the street and so on. On the genmo almost a year ago.
Hunter [00:23:34]:
Now, by the way, that demo. Almost a year ago.
Daniel Bishop [00:23:37]:
Yeah. And usually what they show off becomes available a lot sooner than that. But so for Mochi 1, this open source model, I think they say it's going to be open source and it's under an Apache 2.0 license, which is like really broadly accessible. I don't think it's actually downloadable at the moment, but you can use it on their site. Go to genmo. I'm going to say GNMO AI, the very first demo video that they have in the upper left corner is of a stylish woman walks down a Tokyo street filled with warm glowing neonanimate site. I'm pretty sure it's the same prompt as what was on Sora to show look, we're making something that's basically as good and I'll say right now doesn't look as good as Sora did almost a year ago and it's not a minute long video, but it looks really good. I'd say better than anything that I was seeing coming out of Runway for example.
Hunter [00:24:34]:
And it's shipping and anyone with something actually available, some technical ability and one of those Nvidia GPUs can run on their own.
Daniel Bishop [00:24:44]:
Yeah, granted, I do believe that one of the limitations for this right now is that it's only in 480p and it's only X number of seconds. Most of the models that make video only let you do up to like five, six seconds, something like that. But also if you look at how most movies and TV shows are shot these days, a lot of shots are just a couple of seconds long anyways. So if you want to, as long as you can handle consistency. I do think that there's something to being able to cut together multiple portions of something and this looks really good. And if it's going to actually be open Source and at 480p, look at where we were with Image Generation 2, three years ago. Tiny little, I don't remember when exactly it was, but like some of the earliest image generation stuff of the precursor to the current age was like a 16 by 16 tiny little stamp that kind of looked like something if you really squinted at it. Right.
Daniel Bishop [00:25:47]:
And now look where we are. So I feel like this is a very important release and depending on how fine tunable it is or how much people can understand what actually went into making it, and maybe if there's even training data made available, which I don't think they said there will be. There's every possibility that this is going to be a huge step forward for video, text, image, video.
Hunter [00:26:14]:
Great example where we have, we now have open source models that compete at the same level. I feel comfortable saying the same level as the commercial solutions. And that will only continue to accelerate and maybe double clicking out for a second. Back to this concept of the great AI power play and the AI oligarchy, et cetera. It's not just about who wins and loses at the the corporate level, but it's about the future that we all want to build. I hope we all want to build a future where innovation is driven by curiosity and this collective good. And not just important, but not just profit margins and market dominance, et cetera. I have this saying, I think it's my saying, I'm going with it, my saying, which is refuse to be artificially limited by technology.
Hunter [00:27:12]:
But at the end of the day, I definitely think that technology should serve humanity, not the other way around. And I definitely think that there's a path, there's a path for that. I think we veer off that path sometimes, but as long as we keep that as our North Star, it's all going to be great. Until everything becomes self aware and then it's anyone's game. There's no telling, I don't know, dig a hole, maybe go down below for 50, 60, 100 years and come back.
Daniel Bishop [00:27:39]:
The computers run everything. We go back to the agrarian societies.
Hunter [00:27:42]:
Yeah. And. And during that time we'll continue broadcasting more episodes of they Might Be Self Aware. So make sure you like and subscribe this one so that we can make it past the episode 50. I don't even know what happens when you hit 50 episodes, but something 51's.
Daniel Bishop [00:27:57]:
Got to be a big one, right? We're gonna have a big celebration.
Hunter [00:28:00]:
We did it.
Daniel Bishop [00:28:00]:
We beat the odds.
Hunter [00:28:01]:
Yeah.
Daniel Bishop [00:28:02]:
You stuck with us this long, now look at us. That'll be the day.
Hunter [00:28:06]:
And you can see that day on pretty much all of the major platforms. Yeah, Spotify. Good one.
Daniel Bishop [00:28:14]:
TikTok. We have TikTok shorts video.
Hunter [00:28:16]:
We're not currently uploading our video there, but apparently you can. I don't know, I have to look into that. Sure. TikTok, there's some shorts. There's more shorts coming. I have a backlog of shorts to get posted. They're all pretty good.
Daniel Bishop [00:28:27]:
We're getting into the winter. I mean, people aren't wearing shorts right now. That's what you meant, right?
Hunter [00:28:32]:
That's exactly.
Daniel Bishop [00:28:33]:
Branded. They might be self aware shorts.
Hunter [00:28:36]:
Yeah. I couldn't have said it better myself. And I guess with that until next time.