Last week's episode primarily focused on challenging developments in generative AI companies and this week we take it to the groundbreaking tools reshaping the AI landscape. In Episode 90 of The Artificial Intelligence Show Mike and Paul explore topics like empathy in AI following Hume AI's new demo, the influence of AI on journalism, Claude 3's skill surpassing GPT-4 on the Chatbot Leaderboard, and much more.
Listen or watch below—and see below for show notes and the transcript.
Today’s episode is brought to you by rasa.io.
Rasa.io is the ultimate platform for AI-powered newsletters. If you’re looking to transform your email newsletter into a powerful, engaging tool that truly resonates with your audience, rasa.io is the game-changer you need. Join the 500+ organizations already making their newsletters smart.
Visit rasa.io/maii today.
00:02:32 — Hume AI Demo Is Turning Heads
00:11:48 — AI Journalists / AI + Journalism
00:24:33 — Claude 3 Opus Now Beats GPT-4 on Chatbot Leaderboard
00:30:55 — Stability AI CEO Is Out
00:33:58 — HeyGen Raises $60M at $440M Valuation + Avatar in Motion
00:39:22 — Databricks Releases Major Open Source Model
00:43:16 — Apple Set to Unveil AI Strategy at Developer Conference
00:46:18 — OpenAI Courts Hollywood Studios
00:50:01 — Anthropic + Cohere Funding
00:58:22 — Elon Musk Releases Grok 1.5, Teases Grok 2
Hume AI Demo Is Turning Heads
An AI startup called Hume AI just turned heads with an incredible demo of stunning new—and potentially creepy—AI technology.
Hume AI’s core product is its Empathetic Voice Interface, or EVI. EVI can carry on a conversation with you simply by speaking and responding via voice chat. But that’s not what has everyone paying attention…
When you converse with EVI, the system understands your emotions, so it can react to those emotions with language that addresses your needs. That means EVI can understand your pitch, tone, and facial expression in order to conduct a more emotionally intelligent conversation with you.
Hume AI states that EVI's exceptional empathetic features, include: responding with human-like tones based on facial expressions, addressing emotional needs with appropriate language, utilizing state-of-the-art end-of-turn detection to determine speaking moments, pausing when interrupted, and continuously improving by applying learned reactions to enhance user satisfaction over time.
Along with releasing both a video demo and a self-serve one you can try out yourself, Hume AI also announced it has raised a $50M Series B round.
AI Journalist
We just got a sneak peek at one potential future of journalism in the age of AI. We’ve been playing around with a new AI tool called AI Journalist from our friends at HyperWrite.
AI Journalist is an AI-driven research and writing tool that generates articles on any topic that you give it.
It does that by: searching the internet for articles and real-time details, picking the best sources and reading them, and then writing an in-depth article based on what it finds. It also hyperlinks all its sources, right into the article itself.
The result is an article with factual information and real-time sourcing that sounds like it could have been written by Reuters or The Associated Press.
This comes amidst a couple other high-profile efforts to reinvent journalism with AI.
One that’s on our radar is Google’s generative AI news product, which it’s reportedly testing with a very small group of news outlets.
According to reporting from Alex Kantrowitz in Big Technology:
A journalist first selects a ‘seed’ source like a city council, parks department, local school, etc. they plan to cover. With a link from that single source — whether that’s a press release, 500 page report, or even a tweet — Google’s generative AI software produces a first draft of the story, complete with a lede, nut graf, quotes, and the rest. (The vision is to eventually allow for multiple input sources.) The idea is that a journalist can add to that draft with reporting and fact checking and eventually publish a full story.
Claude 3 Opus Now Beats GPT-4 on Chatbot Leaderboard
Claude 3 Opus has just dethroned GPT-4 as the most powerful language model out there, according to one major leaderboard.
The website LMSYS.org, aka the Large Model Systems Organization, is a group that, as part of their work, evaluates large language models through their Chatbot Arena.
Chatbot Arena uses both crowdsourcing and Elo ratings, a popular way to evaluate how skilled players are at zero-sum games, to rate the capabilities of LLMs.
Their leaderboard is considered a trusted industry source of which models are most powerful and capable at the moment.
And, as of recording, Claude 3 Opus now sits at the top, just slightly beating out one version of GPT-4 that previously held the #1 rank.
Now, this is just one set of rankings. And they change constantly. But that’s sort of the point here: Claude 3 Opus came out this month. And very quickly displayed incredible capabilities that made it a contender for your use cases and workflows.
Disclaimer: This transcription was written by AI, thanks to Descript, and has not been edited for content.
[00:00:00] Paul Roetzer: What I realized was that a lot of humans aren't very empathetic.
[00:00:05] Paul Roetzer: Like, they actually struggle to be empathetic. And so what I realized was, that
[00:00:11] Paul Roetzer: AI may actually enhance human empathy by synthesizing it for people.
[00:00:17] 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.
[00:00:47] Paul Roetzer: Join us as we accelerate AI literacy for all.
[00:00:54] Paul Roetzer: Welcome to episode 90 of the Artificial Intelligence Show. I'm your host, Paul Roetzer, along with my co host, Mike [00:01:00] Kaput. coming to you on, we're recording this on Friday, March 29th. It is a morning edition on Friday, March 29th. So again, doing
[00:01:08] Paul Roetzer: this a day early this week. We usually record on Mondays, but it is spring break for some of us.
[00:01:13] Paul Roetzer: So. I am heading out. And so we're squeezing this one in before we go. And we had a lot of tools this week
[00:01:21] Paul Roetzer: I say like the last week's episode was all about like large language model platforms. I feel like this week
[00:01:27] Paul Roetzer: there's a whole bunch of interesting tool, AI tool So we're going get into a lot of that.
[00:01:32] Paul Roetzer: Speaking of tools, our episode today is brought to us by Rasa. io. It is the ultimate platform for AI powered newsletters. If you're looking transform
[00:01:42] Paul Roetzer: your email newsletter into a powerful, engaging tool that truly resonates with your audiences, Rasa. io is the game changer you need. Their smart newsletter platform personalizes content for each and every subscriber, ensuring every message
[00:01:57] Paul Roetzer: send is highly relevant and incredibly [00:02:00] engaging.
[00:02:01] Paul Roetzer: The tool also allows you to automate away tedious tasks that go into newsletter production, stay ahead of the curve, and make your emails a must read. Join the 500 plus organizations already making their newsletters smart. Visit RASA.IO/MAII Today, M A I I is Marketing AI
[00:02:23] Paul Roetzer: So rasa.io/MAII today.
[00:02:26] Paul Roetzer: Okay, Mike, get started with our, AI tools the week.
[00:02:32] Mike Kaput: Alright, Paul, so first up, an AI startup called Hume AI has just started turning some heads with a pretty impressive demo of some stunning new and potentially creepy AI technology. Now, this startup is. co-founded by a former Google DeepMind researcher and their core product is something called its Empathetic Voice Interface, [00:03:00] or EVI.
[00:03:02] Mike Kaput: EVI can carry on a conversation with you simply by speaking and responding via voice chat. But that's not really what has everyone paying attention to this. When you converse with EVI, the system actually understands your emotions. So it can react to those emotions with language that addresses your needs in the tone that dress, addresses your needs.
[00:03:25] Mike Kaput: So that means EVI can do things like understand your pitch, your tone, your facial expression in order to conduct a more emotionally intelligent conversation with you. So. Hume AI said in an announcement, EVI has a number of unique empathetic capabilities. It responds with human like tones of voice based on your expressions.
[00:03:47] Mike Kaput: It reacts to your expressions with language that addresses your needs and maximizes satisfaction.EVIknows EV knows to speak because it uses your tone of voice for state of the art, end of turn detection. It [00:04:00] stops when interrupted, but can always pick up where it left off. And it learns to make you happy by applying your reactions to self improve over time.
[00:04:09] Mike Kaput: So the company released both a video demo and a self serve one. You can go try the tool out for yourself. And at the same time, Hume AI also announced it had raised 50, 000, 000 in Series B funding. So Paul, I played around with this a little bit and it seemed pretty impressive, but I can already hear the internet comments explode with some strong opinions on this type of technology.
[00:04:33] Mike Kaput: Like, did you find this cool? Creepy? A little bit of both?
[00:04:38] Paul Roetzer: my first reaction was that I was kind of surprised. This was first real demo we're seeing of this.
[00:04:45] Paul Roetzer: Like, it seemed like such an obvious. thing, that would come that when I first saw it, I was like, is this really the first time we're seeing this?
[00:04:53] Paul Roetzer: Because this is, we've talked about empathy on the show before. And my, my thought [00:05:00] is that this is really interesting. Uh, it is definitely potentially creepy, but I think we have
[00:05:06] Paul Roetzer: take a step back and about the role of empathy in
[00:05:10] Paul Roetzer: human communications. And years ago, when I was first starting to really think about the that
[00:05:16] Paul Roetzer: I remained uniquely human as AI became more intelligent, more advanced.
[00:05:21] Paul Roetzer: I always had empathy on the list. And then there was a point a years back where I realized that probably wasn't going to be the case. And we've talked about this. I was
[00:05:31] Paul Roetzer: 20, 30 episodes ago. I don't remember what the tool was we were talking about at the time, but what I realized was that a lot of humans aren't very empathetic.
[00:05:41] Paul Roetzer: Like, they actually struggle to be empathetic. And, so so while it is sort of a unique human trait, there are many humans who don't have
[00:05:49] Paul Roetzer: that trait, or it's not a very strong trait of theirs. But it's so important in communications and relationships to have empathy. And so what I realized [00:06:00] was, um,
[00:06:01] Paul Roetzer: that
[00:06:01] Mike Kaput: can
[00:06:02] Paul Roetzer: AI may actually enhance human empathy by synthesizing it for people.
[00:06:08] Paul Roetzer: So like if, what I mean by that is if someone who isn't normally very empathetic
[00:06:12] Paul Roetzer: and struggles to relate to and convey with other people actually writes
[00:06:19] Paul Roetzer: something and then tells the AI, make this more empathetic, the AI can do that every time. Well, that's what it was. We were talking, um, episode
[00:06:26] Paul Roetzer: last year about, a that was run for doctors where.
[00:06:30] Paul Roetzer: The, patients found the communications with AI more empathetic.
[00:06:35] Paul Roetzer: And what you realize is like, wow, okay, a lot of doctors probably struggle to make those connections with people and to be empathetic toward them. So while I think it's, a little bit unnerving that technology like this exists. I do choose to kind
[00:06:51] Paul Roetzer: see the positive side in that I think more empathy in society is always good.
[00:06:58] Paul Roetzer: And if some [00:07:00] people need to rely on their AI assistant to help them be that way, So be it. I think can be
[00:07:08] Paul Roetzer: a good thing. I also see it having a lot of positive impact within business. So, you know, I always want to try and like connect this back to how do we use this in business?
[00:07:17] Paul Roetzer: So if you think about like HR, you know, when you have to communicate bad news from an HR perspective, there's a lot of people probably aren't very good at that.
[00:07:25] Paul Roetzer: So to be able to put that through an empathy layer, you know, always making things more empathetic. Think about customer service, you know, and understanding,
[00:07:32] Paul Roetzer: know, what people are going through. Think about patient communications. We just talked about marketing, sales, like all of these things, empathy is able to enrich the communications and the relationships that occur.
[00:07:43] Paul Roetzer: And so, I think what happens is, this is, again, since not prevalent already, I can't imagine a near future, one two
[00:07:50] Paul Roetzer: years out, where every AI chatbot doesn't have an empathy component to it. both be your emotions as you're communicating with it, but then [00:08:00] able to communicate back to you in an empathetic way.
[00:08:02] Paul Roetzer: So, think this is probably the very early, you could certainly imagine siri. You know by fall of this year, actually understanding the tone of your voice and, and not just communicating to all of us in the same way, but communicating based on how you're feeling, how you're acting, you how, how your voice sounds.
[00:08:19] Paul Roetzer: and, and certainly we've talked a lot in the last few episodes about robotics. You could imagine, you humanoid
[00:08:25] Paul Roetzer: robots. It's going to a whole different level if, you know, they're embodied with not only intelligence, but empathy. So, I do still feel that empathy is a uniquely human trait.
[00:08:39] Paul Roetzer: the machine doesn't actually feel anything, but it can synthesize empathy at a superhuman level.
[00:08:47] Paul Roetzer: And that's a weird thing to wrap our minds around, but, I think that's just a fact like it's it's going to be able to to be empathetic or synthesize empathy
[00:08:58] um, in all [00:09:00] modalities and that's that's going to be something we're just gonna have
[00:09:02] Paul Roetzer: adjust to as a society but again I think I generally choose to view this as a positive it's applied in the right ways.
[00:09:10] Mike Kaput: Yeah, I would agree and it's worth mentioning to your point about use cases and seeing this stuff everywhere that while you can play with the personal AI assistant on their website, the really the big focus here is the EVI API itself, which they're encouraging others to build on top of and bake into basically any application you could think of.
[00:09:33] Mike Kaput: So that can be very positive from that sense as well.
[00:09:36] Paul Roetzer: Yeah,
[00:09:36] Paul Roetzer: I think it's a good point, Mike. And also, a reminder, like, you don't have to wait for access this
[00:09:43] Paul Roetzer: to use AI tools to help you be more empathetic. You can use it chatGPT,
[00:09:46] Paul Roetzer: Gemini, Claude, whatever you're using, when you're writing something, ask it to make it more empathetic and it'll do it.
[00:09:53] Paul Roetzer: Like you, you can do this right now. I've seen people use this for like, you know, customer communications and patient [00:10:00] communication. So yeah, use the AI to help you with empathy. Like I, I
[00:10:04] Paul Roetzer: that's a valuable use case for it. Like I said, start, we
[00:10:07] Paul Roetzer: need more empathy in society. And if we learn to be more empathetic through the machines, great.
[00:10:12] Mike Kaput: Yeah, that's a really interesting point because I find myself more and more using it in that respect of saying, well, now I've got this 24 7 assistant that's not going to judge me, that I'm not going to feel embarrassed talking to somebody about, and that has all these capabilities. I find myself saying, well, What do you think of the tone of this email or this communication?
[00:10:33] Mike Kaput: I can sometimes be a little too blunt when I don't mean to be like, how might this be perceived? Or how could I better understand what this person might be going through?
[00:10:41] Paul Roetzer: Yeah,
[00:10:42] Paul Roetzer: I think the, it's reasonable to assume these models in the not distant
[00:10:46] Paul Roetzer: future will automatically detect that sort of thing. So, you know, let's say you go in and you, you need to, you
[00:10:53] Paul Roetzer: lay some people off and you draft a letter chatGPT where you're doing layoffs. it would [00:11:00] detect that a negative connotation to content and
[00:11:04] Paul Roetzer: recommend a more empathetic approach to it. You know, that's the kind of thing, again, I think as these things become true assistants, and like you're saying, Mike, always there 24 7, it's not just, reactive
[00:11:18] Paul Roetzer: us asking for more empathy, but it's actually by assessing what's being created and how tools are being
[00:11:25] Paul Roetzer: used and actually bringing that kind of value to it saying, Hey, based on the tone of this, you may want to consider it to be more empathetic.
[00:11:33] Paul Roetzer: Here's a version of it that you could, you know, take a look at. And I think that's, again, like the, Sort of the positive outlook
[00:11:40] Paul Roetzer: what co pilots or assistants can be is that they're proactive in helping us be better professionals also.
[00:11:48] Mike Kaput: Alright, so this next one is a little crazy as well. So we just got a sneak peek at kind of one potential future pathway of journalism in the age of AI. So [00:12:00] we've actually been playing around with a new AI tool called AI Journalist from our friends at HyperWrite. And this isn't AI driven. Research and writing tool that generates articles basically on any topic you give it.
[00:12:15] Mike Kaput: But what's interesting is that when you give it a topic, it goes and searches the internet for articles and real time details about the topic. It actually looks at and starts to try to evaluate and pick the best sources and understand them. And then it actually writes you an in depth article based on what it finds.
[00:12:35] Mike Kaput: And using the tool, it will actually automatically hyperlink All the sources that it's been using right in the article itself. So what you get in just a few seconds is an article with at least its attempt at factual information and real time sourcing that looks and sounds honestly like something written by Reuters or the Associated Press.
[00:12:59] Mike Kaput: And this kind of [00:13:00] comes amidst some other discussions we've been having on this podcast about some high profile efforts to kind of reinvent journalism or rethink it with AI. And one that's also on the radar that we heard about again this week is Google's generative AI news products. So this is something we talked about on a previous episode that it's reportedly testing with a very small group of news outlets.
[00:13:25] Mike Kaput: And we saw a report from Alex Kantrowitz in big technology, where he kind of dived deeper into this because he was very interested in how this tool or purported Gen AI initiative would impact journalism? The way he described this was a journalist first selects a seed source using Google's tool that they're piloting, like a city council, a parks department, a local school, something they plan to cover.
[00:13:53] Mike Kaput: And then with a link from that single source, whether that's a press release, A 500 page report, a tweet. Google's generative [00:14:00] AI software produces a first draft of the story, complete with a lead, a nut graph, quotes, and the rest. And the vision is to eventually allow for multiple input sources. So this idea is that a journalist can then add to that draft with their reporting, their fact checking, and eventually publish a full story.
[00:14:19] Mike Kaput: So, you. These are two separate tools and initiatives, but both appearing to kind of do the same thing, which is augment the work of journalism using artificial intelligence. Now, Paul, we both have backgrounds in journalism. Like, Hyperwrite's AI journalist tool, I'll be honest, blew me away. It's pretty, pretty interesting.
[00:14:40] Mike Kaput: Are human journalists in trouble due to these tools?
[00:14:45] Paul Roetzer: I think you can take that perspective Uh, for sure, I mean, you can look at this and experiment with it and realize it's, it's pretty good. And it's, obviously
[00:14:57] Paul Roetzer: pretty raw. And I'm, if I'm not mistaken, Matt built [00:15:00] in like a day or over a weekend, through Claude aPI. API. So, yeah.
[00:15:07] Paul Roetzer: This is just like a V1, not even probably. It's just an alpha basically of the concept and it's available. We'll put the link in the notes, but like you can go it out for free, I believe, Mike, right?
[00:15:18] Mike Kaput: Yeah, yeah, with HyperWrite, so they did two pathways here. Matt built an open source version, so if you are familiar with how to run that yourself, you can basically just connect your API key for Claude to the tool and run everything on your own. But there's an easier to use version that I think most people would probably find easier to adopt, which is on HyperWrite's website.
[00:15:43] Mike Kaput: So you can use that, try that out for free. They give you like credits based on your usage.
[00:15:48] Paul Roetzer: Okay, so, couple of just thinking out loud here one we still have to deal with the fact that, uh, if you want to own the content, you don't if the AI [00:16:00] writes it.
[00:16:00] Paul Roetzer: So if you're a media outlet, if we're talking about impact on journalism, this doesn't change the fact that U. S. copyright law mandates that a human has to create something to, get a copyright. And I don't know of a publishing outlet that wouldn't want copyright
[00:16:16] Paul Roetzer: on their material, so I don't see this as all of a sudden journalists are just going to start using this stuff.
[00:16:21] Paul Roetzer: tool or media companies will start using this and put journalists out of work. That's not viable. the second thing that comes to mind is great journalism is sourced. Like
[00:16:33] Paul Roetzer: where where you go do the work, you talk people, you, you do interviews. So all this tool is doing is going and pulling from existing third party content and sources.
[00:16:42] Paul Roetzer: There's no first party stuff. no finding that source and getting great story. Like you and I love the information as an example. The information pumps out great stories.
[00:16:52] Paul Roetzer: because they're sourced within these tech companies and they get, they, they break stories that you're not going to find on the
[00:16:59] Paul Roetzer: So it's [00:17:00] not replacing that kind of journalism. i, I, like you and I talked about this first day we saw it, Mike, and the couple of things jumped
[00:17:08] Paul Roetzer: out to me was the massive potential value for internal documents and processes. So like we were on a call last week and something came up, or it was on Monday, And it was a new program that I wanted to kind of research and so I mentioned on the call and so Mike like in real time goes and
[00:17:26] Paul Roetzer: tests this thing out and so they write me an article about this and
[00:17:29] Paul Roetzer: it does and now we have to verify the facts of it but it just literally saved us Like an hour probably of doing research and writing a brief, which is something we do all the time just
[00:17:41] Paul Roetzer: write briefs on topics and people and ideas. And it just wrote it and it sourced it and like gave the links to it. So my immediate thought was, wow, this could really be valuable, um, within internal processes and documents that we don't need a copyright on that are never going to see the time of day. I could [00:18:00] see it. From a journalist perspective, if we want to stay on that line of journalists
[00:18:05] Paul Roetzer: and writers, that you could use it to get stories started and then layer in the interviews and like the original research and reporting that goes into these stories.
[00:18:17] Paul Roetzer: But I also thought about, the The, the decimation of local media. So again, as, as people with backgrounds in journalism and the media industry, for years, we've been watching this industry just get crushed and it had nothing to do with AI. And so you and I have talked about this many times, Mike, all the loss of, of journalists, like, you know, I think I shared the story when my dad.
[00:18:43] Paul Roetzer: so my dad, worked for the Cleveland Plain Dealer and so he, he delivered papers for like, know, 15, 17
[00:18:50] Paul Roetzer: years and then he was a manager of a depot for a number of years. And when my dad used to take me into press room at downtown Cleveland, like when I was kid,
[00:18:59] Paul Roetzer: we would go [00:19:00] in and still smell like the ink of presses. There was hundreds of journalists. in, in the Plain
[00:19:05] Paul Roetzer: Dealer newsroom. Now there's like, I don't know, 20, 30. So like that whole industry has just gotten wiped out because of their reliance on advertising.
[00:19:15] Paul Roetzer: like classified ads used to like what made these local media outlets go. And then classifieds, obviously everything just kind of went online and they had to find like a next thing.
[00:19:24] Paul Roetzer: So they've gotten crushed. because of their ad supported model.
[00:19:28] Paul Roetzer: And so you and I, Mike, for years known this has nothing to do with AI yet. What happens when AI comes for this industry? And I think we're here. Like, you know, for years I've been talking to journalism schools and, and heads of these, communication schools and, and saying like, Hey, it's going to come.
[00:19:45] Paul Roetzer: There's going to be a day where this thing does what we're these students to do. And I, I think, you know, We're, we've arrived at that. But the positive potential is society needs local
[00:19:58] Paul Roetzer: Like we want to know the [00:20:00] local sports stories. We want to know what's going on with local
[00:20:02] Paul Roetzer: We, we want to be informed citizens.
[00:20:05] Paul Roetzer: And because the business model doesn't support that happening anymore, we have communities that have no verifiable
[00:20:15] Paul Roetzer: coming, other than like Facebook boards and stuff like that.
[00:20:18] Paul Roetzer: So my, again, positive on this is what if we can use these tools to fill the gaps in local news?
[00:20:27] Paul Roetzer: What if a future newsroom is an editor or two that works with the AI to write the stories of what's going on in the community? And, you know, curates like the reliable sources to like source those stories. And then the editor does editors do,
[00:20:45] Paul Roetzer: which is verify facts, like make sure the information is reliable.
[00:20:48] Paul Roetzer: So And so I can actually imagine a scenario where local media re emerges in an entirely different, much more streamlined model, and doesn't replace [00:21:00] jobs or eliminate writers, it may actually create new writing and editing jobs where the AI and the human are working in tandem, and that to me is the The possibility of all of we can
[00:21:10] Paul Roetzer: get stuck on the negative and be like, Oh, I was just going to replace everybody.
[00:21:13] Paul Roetzer: But like, what if, like, what if we just think about reinventing that industry, reinventing local media, um, expanding, you know, the stories that are being told through liable sources in your industry, you know, trade magazines are the same thing. Like we came up in B2B marketing, Mike, at the agency. And like those trade magazines are completely dependent upon advertising dollars.
[00:21:35] Paul Roetzer: And so you have all of these trade pubs that are risk
[00:21:38] Paul Roetzer: of going away just because of the ad model, not because of AI. Well, maybe AI is the answer there too. So again, I don't want to be like, pretend like there aren't negatives here, but I think that if you step back, there's a, there's a chance kind
[00:21:51] Paul Roetzer: of re imagine that whole industry and using these as true tools to accelerate that.
[00:21:56] Mike Kaput: Yeah, I, I tend to agree on that. I think that there's a [00:22:00] massive unexplored opportunity here, which will be interesting to see play out and kind of one final note, even if you zoom out a little further and just say, What does this do for your ability to research anything? You mentioned internal briefs and documents, and to me, this is all upside.
[00:22:17] Mike Kaput: I try to tell people, I get geeked out about it. I'm like, do you realize you have research superpowers now? It's really cool. And to your point about the example of the topic we were looking up during a meeting, like, That just saved us a bunch of time that we then used in other ways and told us, gave us the ability to answer a question that was really important to our business.
[00:22:40] Mike Kaput: Now we can do that over and over and over again.
[00:22:42] Paul Roetzer: Yeah and the other thing You know, I thought of this week was, you know, you know, we talked about in last week's about how
[00:22:49] Paul Roetzer: you have to, you can't make your plans based on what these models are capable of today. And the one big thing that always, you know,
[00:22:56] Paul Roetzer: talk about with large language models is the hallucinations or the fact they make [00:23:00] stuff up, there's a, you know, a higher error rate.
[00:23:03] Paul Roetzer: The indications are that that's going to go away. And so there's a paper we'll talk more in depth about next week, but DeepMind, Google DeepMind just published a paper this week called, Let me see if I have it. Long form factuality and large language models.
[00:23:18] Paul Roetzer: was a collaboration between Stanford University and Google DeepMind.
[00:23:21] Paul Roetzer: And Ethan Mollick tweeted this
[00:23:24] Paul Roetzer: and summarized like three quick points. And then we'll just kind of leave it here. We'll come back to this next week. Large language model agents can achieve superhuman rating performance on fact checking when given access to Google. So if you, if you mix search in. AI can
[00:23:39] Paul Roetzer: actually be more reliable than humans at reporting facts. The second point was bigger models are more factual, which
[00:23:45] Paul Roetzer: exactly what what we said last week. As you get gPT 5, Gemini 2, the bigger these models get, it just sort of seems to solve for the reliability and factuality of content.
[00:23:55] Paul Roetzer: And third, large language models are 20 times cheaper than humans.
[00:23:59] Paul Roetzer: So Google [00:24:00] DeepMind's own research is basically saying, We're going to solve for this hallucination
[00:24:04] Paul Roetzer: These things are going to become more factual and reliable.
[00:24:07] Paul Roetzer: And so you're going be able to rely on them more in newsrooms and publishing and media outlets, and even on brands, like as you're telling stories doing research.
[00:24:15] Paul Roetzer: So yeah, I mean, I think the AI journalist thing is really interesting. But I think in it, in a more
[00:24:21] Paul Roetzer: macro level, it just is an indicator of where this is all going and how we're going to become more and more reliant on these tools as an integral part of our workflows, whether you're in HR, finance, research, product, sales, doesn't matter.
[00:24:33] Mike Kaput: Alright, so in our third big topic today, Claude3 Opus, the most powerful version of the new Claude3 family of models, has just dethroned GPT 4 as the most powerful language model out there, according to one of the major Leaderboards. Now, the website leaderboard we're talking about is called lmsys. org, aka the Large Model Systems Organization, [00:25:00] and this is a group that we've mentioned a couple times.
[00:25:03] Mike Kaput: that evaluates large language models through a section on their website called their chatbot arena. Now, the chatbot arena uses crowdsourcing and ELO ratings, which are a popular way to evaluate how skilled people are at zero sum games. They use it in chess a lot. And they use this methodology to rate people.
[00:25:23] Mike Kaput: The capabilities of large language models. So this leaderboard is one of the trusted industry sources of which models are most capable and powerful in any given moment. And as of us recording this on Friday around 9, 10 a. m. Eastern, Claude3 Opus now sits at the very top, just slightly beating out one version of GPT 4 that had previously been number one.
[00:25:51] Mike Kaput: Now, this is just one set of rankings, and the rankings change constantly, but that's actually the point here, and why I think this is [00:26:00] important to talk about, is that Claude3 Opus, like, came out this month, and very quickly started to displace other very, very powerful models, and it quickly displayed incredible capabilities that you That could enable and empower you with certain use cases and workflows you might not have been able to totally dial in before.
[00:26:22] Mike Kaput: So, Paul, can you walk us through, like, how do you think about all these different models and, like, how fast things change and how to bet on singular models?
[00:26:34] Paul Roetzer: I mean, I don't know how you, I don't know. it is so fast moving. So I was, I've been testing Claude3 Opus perplexity. So I I have like perplexity paid plan and you choose which model you use.
[00:26:47] Paul Roetzer: So I'll kind of bounce around test the different ones. But I did finally just start paying directly for the Claude license. So
[00:26:54] Paul Roetzer: it uh, just, again, I think it just shows how dynamic this [00:27:00] space is. As you mentioned, like new models every week. There's a couple other models I think we're going to talk about in the rapid fire items.
[00:27:05] Paul Roetzer: this is not the only news for the week of of the leaderboard. But the one thing that keeps coming back to me is everybody
[00:27:14] Paul Roetzer: keeps talking about like beating GPT 4 and like this monumental thing that they're now ahead of GPT 4. GPT 4 came out in March 2023, a year ago.
[00:27:27] Paul Roetzer: It was training stopped six months before it came out.
[00:27:32] Paul Roetzer: So, everyone is now beating a model that is like 18 months old.
[00:27:40] Paul Roetzer: OpenAI has not stopped building and training. A more powerful model. So if all of these companies are just catching up gPT 4 and
[00:27:50] Paul Roetzer: in some evaluations, just beating them out slightly on different evaluations, what in the world is GPT 5 going to look like?
[00:27:58] Paul Roetzer: Like it's [00:28:00] just the thing that I keep coming back and we keep talking about. How you can't get comfortable with the current technology
[00:28:08] Paul Roetzer: is like, claude3 is amazing. Like our early testing is really impressive. You talked about some of your experiences with it last week. but but it's being compared to an 18 month old model.
[00:28:20] Paul Roetzer: And so, like I even, I got an email, Yesterday, I got to go check and see where the details are, but, it said that DALL E3 is now going have editing ability.
[00:28:30] Paul Roetzer: Like, you're going to be able to edit images chatGPT.
[00:28:33] Paul Roetzer: So, I think, as Sam talked about in the Lex Friedman podcast, they're just going to, like, gradually introduce a bunch of capabilities that collectively may be a part of the GPT 5 story or whatever they end up calling the next version.
[00:28:48] Paul Roetzer: and so that, like, Definitely play with Claude, pay the 20 bucks a month if you need to, and just experiment with it. But, I just,
[00:28:56] Paul Roetzer: keeps building up my anticipation for what [00:29:00] 5 and Gemini 2 are be. And then I don't know, Mike, you, but I thought it would be interesting to note people.
[00:29:04] Paul Roetzer: So I actually, in Chrome, I have an AI tools group saved. And so my current Browser that has that. I have
[00:29:14] Paul Roetzer: Runway. These are are all the license I'm paying for I have up all the time that I have just bouncing and test depending on what it is. So I have Runway going which I have for video and images. I have now my Claude license active.
[00:29:26] Paul Roetzer: I have ChatGPT team and we have both a workspace and I still have the personal one. We
[00:29:31] Paul Roetzer: Google Gemini.
[00:29:32] Paul Roetzer: I perplexity, playground. ai
[00:29:35] Paul Roetzer: ai I love for images, I have the leaderboard, the chat arena you just mentioned,
[00:29:40] Paul Roetzer: And then I now have Google AI Studio up where I can text, test Gemini
[00:29:45] Paul Roetzer: Pro. So if
[00:29:45] Paul Roetzer: if anyone's ever like, people always ask me like, what tools do you use every day?
[00:29:48] Paul Roetzer: Those are the two, four, six, eight tabs that I have active all day long that I may
[00:29:55] Paul Roetzer: in at any time. What about you? Do you have any other ones? that you keep active all day long
[00:29:59] Mike Kaput: Yeah, those are, [00:30:00] my, my stack that I go to daily is very, very similar, I would say. yeah, all the, I think that's worth noting, like, all the paid versions of the major assistants, Gemini Advanced. Claude3, ChatGPT, Team, Plus, whatever, just so you can get the most powerful version of this stuff. I think Gemini 1.
[00:30:21] Mike Kaput: 5, you mentioned, is always open as well, just based on how interesting some of the capabilities are with the context window. I would say honestly just on top of that Hemingway, Grammarly, just because I'm writing a lot. But yeah, very similar stack. But you'd be so shocked how game changing even that stack can be if you're keeping this open and using it for all these things in your life
[00:30:46] Paul Roetzer: yeah. very
[00:30:55] Mike Kaput: big first rapid fire item is that, [00:31:00] hot on the heels of last week's story that we reported, where three out of five of Stability AI's top researchers resigned from the company, we have another resignation. The company's CEO, Emad Mostaque, has also now resigned.
[00:31:14] Mike Kaput: The company has appointed their chief operating officer and their Chief Technology Officer as co CEOs in his absence. So the company said Mostaque stepped down to pursue quote decentralized AI and Mostaque appeared to confirm that in a separate post on X. He said quote not going to beat centralized AI with more centralized AI all in on decentralized AI.
[00:31:40] Mike Kaput: Lots more soon. Now, by decentralized AI, it appears he's broadly referring to an approach to AI development where computation, decision making, it's all distributed across A network of devices not essentially centralized in a single set of servers owned by a [00:32:00] single company. So, Paul, given your knowledge of stability, AI, and Mostaque's kind of background and history, like, what was your read on this situation?
[00:32:12] Paul Roetzer: Um, it just seemed inevitable. I mean, obviously if you listen to last week's episode, I shared some of my perspective on this, this company as a whole. there's there's probably a dad in
[00:32:23] Paul Roetzer: here somewhere about renaming the company Instability AI, that would that would make my daughter cringe, but, Yeah, I mean, the fact that he, I don't know how they phrased it, that he left, or resigned.
[00:32:39] Paul Roetzer: I think resigned is probably a generous term here, given everything that had happened. So I have, I don't know, like, we'll see where they here.
[00:32:49] Paul Roetzer: the decentralized AI thing, it just I don't know, like, I, I get it. Like, I, I think [00:33:00] it's an important thing. I,
[00:33:01] Paul Roetzer: it just seems like a quick sleight of hand to like, Hey, look at this.
[00:33:05] Paul Roetzer: It's let's forget about the fact that we just blew through a hundred million in funding and don't have a model
[00:33:11] Paul Roetzer: that's viable. And let's go talk about something else. So we'll see, see where he goes, see what he does. He posted a tweet where was, having a chat. Team's meeting with Satya Nadella,
[00:33:23] Paul Roetzer: You know, bragging about that. So we'll see where that goes. Satya's sort of like collecting all the pieces, it seems.
[00:33:29] Mike Kaput: No kidding.
[00:33:30] Paul Roetzer: So, again, like I follow him on Twitter. He shares a lot of interesting stuff. He's patched into a lot of people. He generally seems to, he likes
[00:33:40] Paul Roetzer: tweet ahead of time to let you know he knows things, so he's a good one to know when new models are coming.
[00:33:45] Paul Roetzer: Cause he'll tell you, he knows new models are coming. so I follow him. hope,
[00:33:51] Paul Roetzer: I I hope he does good things, but
[00:33:54] Paul Roetzer: not shocking at all that this is how it ended at Stability.,
[00:33:58] Mike Kaput: All right. So [00:34:00] next up, HeyGen, which is a leading AI startup that is for creating synthetic voices and videos, just raised 60 million at a 440 million valuation. They have drawn a ton of attention for hyper realistic AI generated videos.
[00:34:17] Mike Kaput: So these have. Extremely lifelike virtual avatars, or you can make these pretty stunning deepfakes of real people and have them say whatever you want.
[00:34:28] Mike Kaput: Now at the same time they announced this funding, they also released a new feature called Avatar in Motion 1. 0, which we have a video linked to in the notes here, it's worth taking a look at. Now according to HeyGen, this feature allows you to quote, move your hands, make gestures, use unique tones of voice, and HeyGen will flawlessly track.
[00:34:48] Mike Kaput: Translate and lip sync your video with any input text. So, Paul, there seem to be quite a few ways this technology can benefit companies, plenty of ways it can go wrong. Some [00:35:00] of this stuff is just increasingly pretty stunning. What are your thoughts on HayGen and kind of the overall trend here of what they're trying to do?
[00:35:07]
[00:35:07] Mike Kaput: trying to
[00:35:09] Paul Roetzer: their site to try and like see this and there was nothing there. The only thing I I could find was tweet, but they did, uh, tweet again last wow. this blew up, we release a demo to the public?
[00:35:19] Paul Roetzer: Comment if you want access. And then they replied, said, okay, the people have spoken. We'll try and whip up a demo tonight, which would have been Thursday night, March 28th, that can use tomorrow, which
[00:35:28] Paul Roetzer: been friday, March 29th. So by the time you're listening to this, There is likely a further demo of this technology on the HeyGen site or on the HeyGen x slash Twitter account.
[00:35:38] Paul Roetzer: yeah, I mean, it's definitely moving us further and further down the path of you don't know what's real. And that is a very, um, scary premise for society. But, I mean, just the quick demo they showed, I mean, it's
[00:35:55] Paul Roetzer: it, you would have to really, really study that to not know it's [00:36:00] fake.
[00:36:01] Paul Roetzer: Um, so the fact that the guy's walking through the park and, you know, gesturing hands, like everything just looks legit.
[00:36:07] Paul Roetzer: So yeah, I think that's it. Fascinating tech. Absolutely. can certainly imagine a bunch of practical business cases for this. You're certainly going to see this used in.
[00:36:19] Paul Roetzer: marketing and sales stuff, whether disclose
[00:36:22] Paul Roetzer: it or not, I have no idea. this probably gets back the importance of having generative AI policies for your company that, um, states when you're allowed to use this kind of technology, whether or not you disclose it to people that you are using it.
[00:36:37] Paul Roetzer: So I think it brings us back to the importance of having those policies within your organization. And then also
[00:36:43] Paul Roetzer: making sure that if you choose use this technology, you
[00:36:47] Paul Roetzer: clearly understand their terms of
[00:36:50] Paul Roetzer: Like, what are you giving them the rights to once you create a deepfake of yourself on their technology?
[00:36:57] Paul Roetzer: And we saw this again with like Inflection last
[00:36:59] Paul Roetzer: [00:37:00] This lesson of you have to trust these companies that you're giving this information to or letting them create these deepfakes because what if HeyGen is the next one that crumbles and somebody, you know, scoops up the technology? Or who are the investors in HeyGen?
[00:37:15] Paul Roetzer: Like all these things that. You don't really think about because
[00:37:19] Paul Roetzer: just think,
[00:37:19] Paul Roetzer: wow, this is really cool tech. I'm going to go deepfake myself. And then all of a sudden you're wishing you hadn't given that intellectual property to somebody else. So again, just kind of like the cautionary tale of make sure you check the terms of use.
[00:37:32] Paul Roetzer: If it's from a business perspective, make sure legal is involved in assessing this before you go deepfake, deepfake. Deepfaking people in your company without their knowledge or deepfaking yourself.
[00:37:42] Paul Roetzer: Um, just, it's just where we're going to be with all this stuff from audio to video to images. You got to be aware of, how
[00:37:50] Paul Roetzer: this stuff works and who you're giving this to, because this
[00:37:53] Paul Roetzer: is the one I see most often. All the time. Every conference I go to people are, you know, Hey Jen, you deepfakes themselves.
[00:37:58] Paul Roetzer: I have not done it yet. I [00:38:00] don't know if you've tried it yet, Mike,
[00:38:01] Mike Kaput: I haven't done it yet.
[00:38:02] Paul Roetzer: but I haven't done it for that exact reason. Like I don't understand their terms.
[00:38:06] Paul Roetzer: I don't know who I'm giving my information to, and where it could go from there. If somebody, if it changed, if Hey Jen changes hands and now all of a sudden I've passed on my permissions to whomever acquires them or buys them or what.
[00:38:18] Mike Kaput: Yeah. And we've belabored this point quite a bit on this podcast, but also just understanding That, from my understanding of reading HeyJen's website, any of these sites, they need a couple minutes of footage of you, so I'm not saying people should be paranoid all the time, but like, there's enough footage of you probably on the internet to be able to do this, so it's probably good to start acting accordingly in your own business, things like contracts, the usage of your likeness, things like that are probably going to become quite a bit more important than they used to be.
[00:38:51] Paul Roetzer: Yeah. And I think, you know, for your communications team, legal team, you probably need to.
[00:38:57] Paul Roetzer: You know, be monitoring for deep fakes of executives things[00:39:00]
[00:39:00] Paul Roetzer: like that. Board members, again, this is, this isn't meant to be fear stuff. This is the reality of where we are. Like it's not hard to deep fake anybody, without permission. So,
[00:39:10] Paul Roetzer: Yeah, mean your communications
[00:39:12] Paul Roetzer: whoever
[00:39:12] Paul Roetzer: does your media monitoring and social media monitoring, they need to be monitoring for deep fakes of key people within your organization
[00:39:22] Mike Kaput: Alright, so next up, Databricks, which is a popular data intelligence platform, they just released a new open source foundation model called DBRX, and they say it quote, outperforms all established open models in quality benchmarks and allows you to quickly build your own custom LLM. on your data. What appears to be noteworthy about this model is that DBRX beat, says it beats at least, Metas, Llama2, Grok, and Mistral on language understanding, programming, math, and logic tasks, and it says it beats GPT 3.
[00:39:58] Mike Kaput: 5. on [00:40:00] most benchmarks. Now on that last bit, this is the important part. Databricks says, quote, this is important as we have seen a major behavioral shift in the last quarter among our 12, 000 plus customer base. Enterprises and organizations are increasingly replacing proprietary models with open source models for better efficiency and control.
[00:40:23] Mike Kaput: In our experience, many customers can outperform the quality and speed of proprietary models. By customizing open source models on their specific tasks. So Paul, this relates to a topic we talked about last week about how enterprises are embracing multiple models to achieve their goals, including open source ones.
[00:40:42] Mike Kaput: So when you read about something like this, like how are you thinking about the advent of these new powerful open source models for enterprises?
[00:40:53] Paul Roetzer: Well, I mean, two things immediately come to my mind. One, there's [00:41:00] probably going to have to be a role within enterprises that is exclusively focused on this stuff.
[00:41:06] Paul Roetzer: Like, I mean, the more I think about it, so I think chief AI officer is going to be, you know, an emerging role in enterprises. You know, we're already starting to see that,
[00:41:15] Paul Roetzer: But you could start see the formation of an org chart underneath that. And I could see managing like the symphony of large language models as. A full time role or multiple people working on this. Like you might have like someone who's focused on the open source stack. Someone
[00:41:33] Paul Roetzer: focused on proprietary tools. So who figures this out is basically what I'm saying. We have talked to a lot of very big enterprises and I have yet to find one. that has someone who solves for this. It is like falling
[00:41:49] Paul Roetzer: the CTO, cTO, CIO, Chief Digital Officer in some cases, but they have no training and background in how to do this.
[00:41:57] Paul Roetzer: So I feel like this is an emerging role. The [00:42:00] complexity of these choices is becoming so great that it is not something you're going to get out a weekly podcast.
[00:42:08] Paul Roetzer: Like me and you, Mike, are not going to like be able to give the recommendation to people of what you do. with as dynamic and complex as this space
[00:42:16] Paul Roetzer: becoming. So one, I think internal structure is going to start having people dedicated to this. and
[00:42:21] Paul Roetzer: two, uh, outside consulting companies are, are like, have do this.
[00:42:27] Paul Roetzer: Like, and I know people like McKinsey Deloitte and, people like that are
[00:42:33] Paul Roetzer: are doing this, but it's, it's moving so Um, So that's the thing is like, if, if this is overwhelming you, you're like, what? There's
[00:42:42] Paul Roetzer: like another open source model and it's now beating this. I thought last week you guys said this.
[00:42:46] Paul Roetzer: Yeah, exactly. Like it's moving so fast, that you don't even have time to get in and test these and
[00:42:53] Paul Roetzer: like around with them. So I think that's my main takeaway is. Again, shows how fast this is [00:43:00] accelerating and the need start
[00:43:02] Paul Roetzer: getting people in place who focus on this stuff. It's, it's really hard to keep up
[00:43:09] Paul Roetzer: So if you're feeling overwhelmed, join the club. It's, it's tricky right now. And it's only going to get crazier.
[00:43:16] Mike Kaput: All right, so all signs point to the fact that Apple is going to unveil its long awaited AI strategy at its upcoming Worldwide Developers Conference in June. Now this conference is set to take place from June 10th to the 14th, and Apple has not revealed what it plans to announce related to AI, but it's clearly Teeing this up and playing into the media frenzy around it.
[00:43:43] Mike Kaput: So, Bloomberg actually reported that several people familiar with the matter say that Apple's kind of big presentation at the conference will focus heavily on AI, and Apple SVP of Marketing, Greg Joswiak, posted on X this week, quote, [00:44:00] Mark your calendars for WWDC 24, June 10th to the 14th. It's going to be absolutely incredible.
[00:44:07] Mike Kaput: And he very, very intentionally capitalized the A and the I in absolutely incredible. Paul, we obviously don't have a ton to go on yet, but you're a long time Apple watcher. Like, what would you expect to see or love to see at this event?
[00:44:24] Paul Roetzer: We've talked about the obvious plays making siri actually intelligent functional and
[00:44:29] Paul Roetzer: probably an agent of sorts where you can just say what you want and it's connected to all your apps and all your data already.
[00:44:36] Paul Roetzer: And that could certainly lead to like explosion in value creation through intelligent so siri is the most obvious thing. Other than that, it's a
[00:44:45] Paul Roetzer: a complete guessing game right because tim Cook and Sundar Pichai, who, you know, CEO of Alphabet, were photographed having dinner together recently.
[00:44:53] Paul Roetzer: Um, they're, they're having conversations with Google. They're probably having conversations with a lot people. [00:45:00] and so the question becomes like, Is, is Apple going to build their own foundation model or are they just going
[00:45:06] Paul Roetzer: partner with Google like they did on Search or Maps? Apple has a history of not, of being a follower, of, of waiting around to see how things play out and partnering with other people in the meantime.
[00:45:20] Paul Roetzer: So your iPhone had Google Maps long before it had a really powerful version of Apple
[00:45:24] Paul Roetzer: but now Apple Maps is really good. so I think there's a lot of examples like that apple
[00:45:29] Paul Roetzer: is very willing to be patient and, partner with other tech companies, to figure out what their long term play is.
[00:45:38] Paul Roetzer: And so it's possible we something like that, like maybe Apple focuses on smaller models can run on your phone
[00:45:47] Paul Roetzer: and they partner with someone like Google to do bigger models, the larger foundation models. We have no idea, but I am, I am going to block my calendar off on June 10th to hear [00:46:00] what they have to say And as an investor in Apple for a long time, I am very, very anxious to see what they
[00:46:09] Paul Roetzer: and try and sort of project around the corner a little bit if I think it's going to be received well by Wall Street, I would say.
[00:46:18] Mike Kaput: So, rumor has it that OpenAI is now taking meetings with Hollywood studios to convince them to use their new AI video generation model, Sora, in their work. Now, this is according to reporting from Bloomberg, which cited anonymous sources.
[00:46:36] Mike Kaput: sources who are familiar with some of these discussions.
[00:46:40] Mike Kaput: Those sources also said these upcoming meetings are just the latest round of outreach from OpenAI. This actually comes right on the heels of a blog post that OpenAI published called Sora First Impressions, which shows some stunning early short films being created by artists [00:47:00] who got access to the tool through a program OpenAI was running.
[00:47:04] Mike Kaput: Now Paul, we have covered a lot of AI tools and developments over the years. I've seen very few of them seem to capture the imagination and cause concern like Sora is. I mean, I've heard about it from a lot of people in the marketing world, the agency world, the creative space. Everyone has a lot of opinions.
[00:47:26] Mike Kaput: Not all of them are great. I mean, Is this like an effort by OpenAI do we think to manage and message like the disruption this tool is potentially gonna cause?
[00:47:37] Paul Roetzer: Yeah, there's definitely a PR aspect to this, a very heavy PR to
[00:47:42] Paul Roetzer: this. I mean, that blog post you referenced about them with the like three or four, I mean, it's so blatantly
[00:47:49] Paul Roetzer: trying to put positive spin on this. Yeah, yeah. Uh, there wasn't much, depth to that. I, so, you know, I do think they're trying to play [00:48:00] both sides here.
[00:48:00] Paul Roetzer: They're trying to talk about it as like creative tool gonna enhance, you know, what creative people are capable of doing,
[00:48:07] Paul Roetzer: But it's hard not to see the alternative side. And so I got, what part of me is like, isn't this what the Hollywood strike was all about was not letting this stuff.
[00:48:16] Paul Roetzer: Come into there, but I don't know. We don't know enough to really get deep and comment
[00:48:21] Paul Roetzer: on it. But, I do. I mean, Sam specifically has said many times, including on the Lex Fridman podcast recently, they often preview technology
[00:48:32] Paul Roetzer: give society the time to prepare for it. And this Certainly seems like case of that where they putting this out there.
[00:48:41] Paul Roetzer: They're showing the capabilities and they are trying to prepare people for what this thing is going to be able to do. And I don't think
[00:48:49] Paul Roetzer: we're seeing half of it
[00:48:50] Paul Roetzer: yet. Like I, my impression is that Sora is way more powerful and capable than we currently understand. And so [00:49:00] I think people are.
[00:49:02] Paul Roetzer: rightfully
[00:49:04] Paul Roetzer: concerned or uncertain about what exactly this thing is going to be able to do because all appearances are that
[00:49:13] Paul Roetzer: it's going to have a significant impact on creativity and publishing businesses and media companies.
[00:49:20] Mike Kaput: yeah, I don't mean to be too skeptical or cynical, but putting my marketing and PR hat on any time I see Something like this Sora blog post. I'm like, okay, we're trying to overcompensate. Like it's so blatant that it's cool Like there's cool examples, cool art being made. No doubt about that. It's awesome.
[00:49:38] Mike Kaput: People should check it out But when you're calling out these huge quotes about how it's going to Augment, not replace this type of art. I'm like, okay, we are
[00:49:48] Paul Roetzer: working a little hard here.
[00:49:49] Mike Kaput: Yeah,
[00:49:50] Paul Roetzer: Yeah,
[00:49:51] Paul Roetzer: it was not, it was not a, two sided blog post that did not address the concerns in the room. It was, it was just, everything's going to be [00:50:00] sunshine and rainbows.
[00:50:01] Mike Kaput: Yeah. Alright, so Anthropic just got a nice boost to its bank account to the tune of $2.75 billion. This is because Amazon just completed the final round of $4 billion in funding that it committed to last year to fund. Anthropic. Now this is part of a kind of unusual funding deal that we reported on at the time.
[00:50:25] Mike Kaput: Amazon basically agreed to invest up to four billion in the company and it appears they're exercising their option to do just that in this latest move. interestingly at the same time, a huge stake in Anthropic worth more than a billion dollars or about 8 percent of the company is up for sale because that stake is owned by the failed cryptocurrency exchange FTX and it's being sold as part of that company's bankruptcy proceedings.
[00:50:54] Mike Kaput: One potential buyer of that stake has been turned down. CNBC. is [00:51:00] reporting that Anthropic reportedly has ruled out taking money from the Sovereign Wealth Fund of Saudi Arabia that showed an interest in the stake over national security concerns. Anthropic isn't the only other big model company raising money cohere is as well.
[00:51:16] Mike Kaput: The information reports that the company is raising at least 500 million in a round that values Cohere at around 5 billion total. That round is being co led by a Canadian pension fund called PSP Investments. Now, Paul, we talked a fair amount last week about Anthropic, Cohere, kind of the overall landscape of Foundation models.
[00:51:40] Mike Kaput: Now, Anthropic appears to be a far healthier company from a financial perspective, but both are raising a fair amount of money. Like, what are the differences between these two companies? Like, what does this mean for the overall competitive landscape here?
[00:51:57] Paul Roetzer: You mentioned the Anthropic
[00:51:59] Paul Roetzer: So there's [00:52:00] a Forbes article and this is a little technicality, but just so people familiar. again, it's not like amazon just stepped
[00:52:06] Paul Roetzer: up and did two and a half billion dollars or whatever. So the paragraph from forbes says, Amazon's investment disclosed in a quarterly filing included a 1. 25 billion. Uh, deployed in September 2023 in
[00:52:17] Paul Roetzer: form of a convertible note that converts in this round. Convertible note is basically debt that converts into equity. Uh, Amazon also has an agreement the company wrote to invest up
[00:52:26] Paul Roetzer: 2. 75 billion in a second note that would expire Q1 2024. A source with knowledge elaborate that either Anthropic and Amazon
[00:52:36] Paul Roetzer: decide on before march 29th, the last business day of the first quarter, to trigger the larger second note, which would then convert into a funding round. So that's what's happening here.
[00:52:46] Paul Roetzer: And so again, the technicality, if you don't understand how convertible notes work, Amazon, gives a loan for 2. 75 million or whatever that number is. And if they don't choose to convert it into equity, then it's payable back as debt [00:53:00] with principal and interest. So in essence, what happened is they had an agreement last year for up
[00:53:04] Paul Roetzer: four billion dollars through these convertible notes. And within that, there's probably some sort of cloud compute like, know, versus not all cash, but they're basically just activating the note that they already had.
[00:53:16] Paul Roetzer: That, that's kind of the technical way.
[00:53:18] Paul Roetzer: the difference, I mean, Anthropic is, certainly appears to be thriving and will remain a major player in this space.
[00:53:29] Paul Roetzer: You know, trying to compete with OpenAI and Google um, and others. So I feel like they're a major player. Cohere seems to be really going much more toward the working directly with enterprises to build and tune custom models. They're not playing in the
[00:53:46] Paul Roetzer: It doesn't seem like in the ChatGPT competitor game where you're going to like have up
[00:53:51] Paul Roetzer: all day long, just testing it with, you know, Anthropic and, OpenAI's, ChatGPT, but Cohere seems like they're going to do [00:54:00] more like distribution channels.
[00:54:01] Paul Roetzer: And I think we talked about, or we'll talk about like their deal accenture where they're trying to basically work with the big consulting firms. And so the big consulting firms who have
[00:54:10] Paul Roetzer: relationships, the big enterprises. would then basically bring Cohere in and build on Cohere. that seems to be their models, more of a distribution through existing consulting firms.
[00:54:21] Mike Kaput: Gotcha.
[00:54:23] Mike Kaput: So we have another AI startup that's drawing some attention, this time due to the high profile of some of its investors. This company is called Asari AI, and it just launched with the backing of people like Eric Schmidt, the former CEO and executive chairman of Google, and Jeff Dean, who is the current chief scientist at Google DeepMind.
[00:54:44] Mike Kaput: So the company described its mission in a statement on X and they wrote, We're building a new type of AI agent and tools that help us imagine and create 10X better solutions, products, and services that we can trust, including for [00:55:00] applications that we might not have even thought of yet. To do so, our AI agents learn, reason, plan, and verify at scale.
[00:55:08] Mike Kaput: And we're reimagining how people can seamlessly collaborate and create with AI. Now on their website, the company elaborates a little further, saying they're building intelligence to, quote, co invent a future, quote, where people and powerful AI agents interact and collaborate in unseen ways to create and discover.
[00:55:28] Mike Kaput: And they emphasize again that we're building AI that plans, abstracts, verifies, and discovers and discovers. new skills and knowledge. Now, Paul, we do not know a ton yet about this company. The people involved, though, are notable. The trend of companies building AI agents is notable. Like, can you talk a bit about why this company caught your
[00:55:51] Mike Kaput: eye?
[00:55:51] Paul Roetzer: So I've shared before, I have a AI tech watch on twitter. That's list.
[00:55:58] Paul Roetzer: We'll drop the link [00:56:00] again in the show notes. There's 84 accounts that are tracking there. Again, they're all tech companies, model, or application companies within the AI tech space. So the question I
[00:56:12] Paul Roetzer: I get a lot is like, well, how do you. How do you choose which vendors you're going to feature?
[00:56:16] Paul Roetzer: know, we've had PR people reach out and can we you on vendors?
[00:56:19] Paul Roetzer: And my general response is, yeah, go ahead, but we probably won't cover them. the way that we figure out
[00:56:27] Paul Roetzer: what matters or which ones to watch, it comes down to a few things. But in essence, we're trying to find the signal through the noise. And so
[00:56:34] Paul Roetzer: signal here is I get alerts from Jeff Dean and Eric Schmidt. Like on Twitter, I, I, subscribe for notifications anytime they tweet.
[00:56:42] Paul Roetzer: And so I do that with probably about a hundred to 120 different people, researchers, entrepreneurs, investors, journalists. That's how I determine what
[00:56:53] Paul Roetzer: so a lot of the tech companies Mike and talk about on this show, that's where it comes from. aren't sponsors. These aren't we're getting [00:57:00] pitched on. This is pure, like, Journalist work, I guess. Research analyst
[00:57:05] Paul Roetzer: Like, we are just monitoring the space in real time. And so the way we decide what matters is who are
[00:57:11] Paul Roetzer: are the founders, who are the investors. We track research paper writers.
[00:57:16] Paul Roetzer: So if there's important research papers that have been written, like attention is all you need paper, I track those eight people in real time, all the time. So like,
[00:57:23] Paul Roetzer: when Jeff Dean, the chief scientist of Google and Google DeepMind, and Eric Schmidt, former cMO, who's a major player in defense and intelligence in the U. S. government right now, when they tweet the same company in the
[00:57:38] Paul Roetzer: five minute period, that gets through the noise.
[00:57:41] Paul Roetzer: And so that's what happened here. I know nothing about the company other than what Mike just shared. but they're
[00:57:46] Paul Roetzer: on the AI tech watch list now because the two major investors are major former players.
[00:57:52] Paul Roetzer: One former player of Google and one current player of Google. And that's what I would advise you all to do. Like you, you have, if you want to [00:58:00] really stay on top of what's happening in the AI space, the only way to do it is to monitor
[00:58:05] Paul Roetzer: key players in real time. And that's Hopefully what do for you each week, like that is basically what I spend my life doing, is monitoring all stuff.
[00:58:14] Paul Roetzer: and so that's, that's how we determine which companies are worth talking about and which aren't, along with some other variables that we'll maybe get into at another time.
[00:58:22] Paul Roetzer: So
[00:58:22] Mike Kaput: Alright, our final news item for this week came in just before we started recording. Elon Musk and his company XAI have announced Grok 1. Grok 1. 5, is their latest AI model. They claim that Grok 1. 5 can process long contexts up to 128, 000 tokens, and that this model has displayed big improvements in coding and math tasks over the initial year.
[00:58:51] Mike Kaput: Model. Musk posted on X that Grok 1. Grok 1. 5 be available on X next week, and he also said, Grok 2, quote, [00:59:00] should exceed current AI on all metrics in training now. Paul, what were your initial thoughts of Grok
[00:59:08] Paul Roetzer: So I woke up this morning, saw this tweet, and as I'm making coffee, I go grok and start around with it. Now, anyone who's listening to the show, I don't have a favorable impression of grok to
[00:59:17] Paul Roetzer: this point, but I've always said just, I assume they'll figure this out, but Grok has been useless up until this point.
[00:59:23] Paul Roetzer: so so I go in and I, I, I what version
[00:59:27] Paul Roetzer: of Grok are you? And it said, I'm Grok version 2.0. And I was like. Well, that's interesting. Next Generation AI Assistant built by XAI with inspiration
[00:59:36] Paul Roetzer: the hitchhiker's Guide to the Galaxy jarvis from Iron Man.
[00:59:39] Paul Roetzer: My primary goal is to provide you with accurate, informative, and engaging responses to your questions and tasks.
[00:59:44] Paul Roetzer: So then I said, what can you do differently or better than the version?
[00:59:47] Paul Roetzer: So I'm assuming I now have at least 1. 5, but maybe I do have
[00:59:51] Paul Roetzer: I have no idea. So it says Grok 2. 0 offers several improvements and enhancements over the previous version, enhanced reasoning capabilities, It has improved
[00:59:59] Paul Roetzer: reasoning [01:00:00] abilities, allowing it to provide more accurate and insightful answers to complex questions and problems.
[01:00:05] Paul Roetzer: Increased context length, although it said it was only like 8, 000 tokens, which I asked it about and it was, I don't know, I thought it was supposed to be more.
[01:00:12] Paul Roetzer: Um, Custom training framework, real time knowledge, uh, better performance on benchmarks, and an increased focus on providing helpful and informative answers.
[01:00:21] Paul Roetzer: so then I, I asked it, I was like, all right, let's get some real time stuff. So I said, what happened in the Cleveland Guardians game last
[01:00:29] Paul Roetzer: So the opening night for the Guardians was March 28th. They were at 10 p. m. Pacific time against Oakland Athletics, and it wrote, uh, what appeared be an extremely accurate summary of that game.
[01:00:42] Paul Roetzer: And it was almost like a like a news summary,
[01:00:44] Paul Roetzer: and I didn't even ask for that. So I was like, oh, okay, well, this is kind different.
[01:00:48] Paul Roetzer: then I asked, can you give me a, I asked about Elon Musk's personality, like based on his tweets and did a pretty good job. And
[01:00:55] Paul Roetzer: then I said, can you give me a summary of the last 10 tweets that that Paul Roetzer has shared?
[01:00:59] Paul Roetzer: [01:01:00] And it seemed to just like, glitch, like it just working.
[01:01:03] Paul Roetzer: And I said, did you glitch? And nothing happened. And then I asked it another question and nothing happened. And I was like, Oh, that's kind of weird.
[01:01:10] Paul Roetzer: it seemed like we were onto something here. It was faster. It had real time knowledge. It was not showing me Twitter cards, thank God, because it would surface like the
[01:01:16] Paul Roetzer: random twitter users as representations of where its knowledge was coming from, so it seemed like we'd.
[01:01:21] Paul Roetzer: done away with that. and then I, then it just stopped. And then I went back in and everything was gone again. And so this has been my experience with Grok to date is there's no conversation
[01:01:34] Paul Roetzer: There's no way to like go previous chats or threads. It all seems to be a single thread, but
[01:01:41] Paul Roetzer: thread always just disappears. And so there's no way to track of any conversation you've previously had. So I
[01:01:47] Paul Roetzer: I don't know if I actually had. a more advanced version for a moment. Like they rolled something out and then they pulled it back because now when I go in I can't do these things.
[01:01:56] Paul Roetzer: And so I feel like I may have had [01:02:00] 1. 5 or 2 temporarily, but now I have usual Grok back and it seems, it seems to have all the The normal
[01:02:10] Paul Roetzer: But, but my biggest one is it wipes away your history and
[01:02:15] Paul Roetzer: can't go back and look at like different so, and there's no search function. So if I had a conversation today and a week from I I wanted to back
[01:02:25] Paul Roetzer: and find the thing that it helped me do. I would have to just scroll through this thread in theory if the thread didn't erase itself.
[01:02:32] Paul Roetzer: So, I don't know.
[01:02:34] Paul Roetzer: We'll wait and see. We'll if it's getting better. It seems like the model itself definitely improving. The user experience is still useless in my opinion. Like, would
[01:02:43] Paul Roetzer: not be using this instead of ChatGPT or Claude because it's
[01:02:47] Paul Roetzer: it doesn't seem to built for that. It's like,
[01:02:50] Paul Roetzer: I don't so that's my take is that it seems like we're heading in a direction of it being more useful. but I, I, I don't think they have to solve the user interface thing. [01:03:00] And it's, it's just,
[01:03:00] Paul Roetzer: it's not valuable
[01:03:02] Paul Roetzer: until they do that. I'll keep testing it. I'll keep paying my 22 a month for all of you who don't want to pay the 22 a month. I'll let you know what I learned.
[01:03:10] Mike Kaput: That's what we're here for. All right, Paul, before we go, two really quick announcements. Um, first up, we have. Tons of people, regularly, who reach out and say how much they like the podcast, thank us for some of the work we're doing here. If you're one of those people, we have a favor to ask you. If you have found value in the podcast, could you please go rate the podcast and leave a review, if possible?
[01:03:42] Mike Kaput: We'd love to start sharing further with new audiences, through reviews, through ratings, how much everyone is enjoying the show. So please, on your podcast platform of choice, go ahead and leave us a really quick rating and a review. If you can't second, [01:04:00] we say this every episode, but it is still there. The Marketing AI Institute newsletter this week in AI is an extremely valuable weekly written brief to all the topics we discussed today with further analysis of those topics.
[01:04:14] Mike Kaput: And all the things we don't have time to get to in each and every episode, which is often 6 to 12 to even more other topics of what's going on in AI. So if you have not signed up for the newsletter yet, please go ahead and do so. Go to marketingainstitute. com forward slash newsletter and check that out.
[01:04:35] Mike Kaput: Paul, until next week, thank you so much for watching. Breaking down the crazy, wonderful, sometimes scary, but always interesting world of AI for us.
[01:04:45] Paul Roetzer: Yeah, thanks again to you, Mike, as always, doing a great job curating everything and leading the conversation. And anyone who is out celebrating spring break week, have a wonderful
[01:04:54] Paul Roetzer: time away. Hopefully, we'll You know, if you're listening this, it's all you're doing this week from a work [01:05:00] perspective. So enjoy your time away your friends and family.
[01:05:03] Paul Roetzer: Be safe and we will be back next week to do it all again. Thanks, Mike. Thanks everyone for listening.
[01:05:08] Mike Kaput: Thanks, Paul.
[01:05:11] 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.
[01:05:34] Until next time, stay curious and explore AI.