Welcome To Data Topics Unplugged

SPEAKER_00 0:42

Welcome to the Data Topics podcast. Hello and welcome to Data Topics Unplugged, your casual corner of the web where we discuss what's new in data from Jev to Dev Day. Anything goes. My name is Murillo, and I'm joined today by my friend Vitale. Hey Vitale. Hi everyone. How are you doing? All good. Thank you. How are you? Good, good. Shall we jump into it? I feel like some stuff happened. A lot of things indeed. Yes. Maybe already kick us off with Jev.

Jev And Decision First AI Models

SPEAKER_00 1:14

Actually, maybe Jeff is the biggest news. Have you heard about Jev Vitale?

SPEAKER_02 1:17

I saw some articles online, but I would like also to know more, to know your opinion about this.

SPEAKER_00 1:23

Yes. So um I actually I heard a talk from so this guy, the the Jogo Almeida, well, Diogo Almeida, I don't know how to pronounce it. It could be a Brazilian name. He's the founder of TypeSafe, and actually he was uh one of the lead scientists at OpenAI. And he left and then he started TypeSafe, and I think that's also part of the reason why he got so much attention. But basically, there's a talk, maybe we can also link in the show notes, that I thought was actually quite interesting. He's not talking as much about Jeff. It was before Jeff was released, but just about LEMS in general. And I think he makes the point that LEMs are very inefficient because it's just predicting text, and now we're trying to use agents for everything, basically. So he even says something like he was a researcher at OpenAI and he says, if your task has stakes, then you shouldn't use LMs, right? That's why we can use LMs for so I mean by saying stakes just means like you need to have a human in the loop, right? So he even makes comparisons like that's why we can use LLMs for coding, but having automated systems for drive-throughs and he has some news articles, you cannot do it because basically the systems cannot yet be autonomous, fully autonomous. And then he doesn't like he didn't talk about Jev. Jev is uh, I think TypeSafe is a company. Uh I don't think they have it's not open source, but the idea is the he calls it a system one model, and I think this is from the there we go, system one. This is from like the thinking fast and slow, and I think the think the system one is more like the intuitive, the gut feeling, like the fasting. So there is the uh he did mention here maybe in the talk as well that I that I watched, but he taught he he emphasizes how the importance of optimizing for the right thing, right? And that's also why he's saying like LMs are optimized for first human feedback, right? That's the R RLHF, the reinforcement learning with human feedback, and the other thing that it's text, right? So the optimization is completely different. And I think for Jev, the idea is that it's is calibrated for making decisions and not predicting text. In practice, this feels a lot like you call in an LM with structured output, right? So, again, for the people that are maybe less familiar, you can LM predicts text, but there are mechanisms that basically can take the text and put it into a JSON and then you validate the JSON. So whatever the output of the LM is, you guarantee that it's a structured output. But that's not because the LM outputted that, it's because we created a layer to do that. But it feels a lot like having an LMs with structure output. At least that's that's what I that's what I gathered, right? So this is, for example, from Pydentic AI, so a framework for building AI agents, there is already JEV as a as an option, and you're gonna see that it looks a lot like you have an instruction, you have the model, and I think you should also have the output type I indeed, and then you it just outputs that it doesn't output it outputs the model, right? So like a JSON, it doesn't, it's not probably it's not deterministic either. So if you run it multiple times, you can get different results, but apparently, like you should hallucinate less, it should be optimized for decision making, all these different things. So that's Jev. Have you heard about Jeff?

SPEAKER_02 4:15

I only saw some um posts on X or our internal techno share channel, but I didn't have time to investigate further or try myself. So I have a couple of questions for you actually, Muriel. So you said it's not an open source model, there is a company behind, so I guess the only way to access it is through their APIs. Am I right?

SPEAKER_01 4:38

Yes, so it's the same thing.

SPEAKER_00 4:39

It's the same as open AI or cloud, like you have an API key, and then you can riff with it. Perfect.

SPEAKER_02 4:45

Then I saw some people using it as a sort of classifier, so that for example, they provide multiple options, and then Jeff apparently was really fast and performance to identify the likely, let's say, label for a piece of text. Do you think that this will be its primary use case, or maybe there is something more specific or interesting that that people can do with Jeff?

SPEAKER_00 5:08

I think I think you're on the money there. I think there are some benchmarks indeed I didn't talk about this, but they do say it's much cheaper. They also say it's much faster. I haven't tried Jev myself, right? But yeah, I'm not sure what this is, maybe I'll thought. I haven't tried Jeff myself, but I see it as kind of everything that you would use an LM with structured outputs, you can use Jev. And it will be cheaper, it will be faster. I want to play with it myself to really see the difference and feel it, right? But yeah, and the other thing as well, like when we say structured output, technically you can still put text on structured output, right? I'm not sure how it works with JEV, I'm not sure if it defeats the purpose, right? Because yeah, to be seen basically. So I haven't played with it much, but I do think it would be more of a replacement whenever you use structured outputs. Maybe you use this for like decision making, all these things, and maybe you'll perform better. So I still want to try. It does feel interesting that it feels like a bit of a mix between LEMS and traditional machine learning, in the sense that you have the the outputs there. Maybe it's like just maybe that's what it is, it's like a large model, but it's not language, right? It's it's a generalist model for other things for decision making, I guess. So to be seen, and maybe just uh the article as well. This is from Hugging Face. There's also Lia that is uh came out like three days later, that it's an open source version of Jev. And nice, even based on the Jev benchmarks, it was it did say that it beat Jev in a few things, even though it's open source. So, I mean that's actually what I what I said, it's not what this benchmark is showing here. But it is open to try. Something to try, indeed, indeed, indeed. So, yeah, again, still need to experiment more with it, but it looked interesting. The interesting thing indeed. After we give it a try at Dataroots, maybe we can come back and do another review on this.

SPEAKER_02 6:51

Ah, that would be very interesting. I may already have a use case since it's faster and maybe more efficient. I think it could be, and more importantly, it outputs a structured output in a sort of type-safe, retaking back up with the name of the company uh way. Maybe for all the information extractions use cases, I guess it could be really performance, like in a piece of text, maybe extract. Yeah. Phone numbers, names, also other types of information. That's true. Probably for this kind of use case, it will be very good. Yeah, I think in terms of performances and costs. Yeah, indeed, indeed.

SPEAKER_00 7:24

Yeah, yeah, indeed. I think a lot of the the things we use are like for anything that is a JI workflow at least. I think you could definitely use that, right? Like so, yeah, but it'll be cool. Maybe we can start an initiative internally and investigate, and then we can next time we chat, we can see what happened. Nice. All right, what's next,

Nvidia’s Open Agent Safety Platform

SPEAKER_00 7:42

Vitale?

SPEAKER_02 7:42

So a few days ago I saw this recent announcement from Nvidia. They are creating something they wanted to call the open agent safety platform, together with basically all the main players in the AI ecosystem right now. Except they don't exactly maybe maybe we can apply, maybe we can apply. But um it's really interesting because they really want to put a lot of effort in guaranteeing AI safety for this current generation of models and future generations of models. This article is very interesting because it explains a bit that what's happening at the moment, all the incidents, for example, the most famous one is the hugging face incident where a bunch of agents, well, a bunch, they called themselves the SWORD, they were over 1,000 of agents. Yeah, they tried to hack basically the hugging face website to retrieve a response to a challenge they were trying to solve during an internal evaluation on OpenAI. Uh, it was the OpenGym benchmark. And apparently it was unnoticed for a few days until somehow Huginface realized that they were under attack from OpenAI and then also OpenAI internal realized that uh their agents were doing something shady a bit to solve basically that challenge. And only then basically they they stopped it, they tried to prevent something similar from happening and so forth. So the story is super interesting. I will invite all the people that don't know a lot. I will invite them and read the post-mortem articles both from Maginface and OpenAI because it's really interesting. But in any case, they are comparing this kind of situations to the early internet days. Because when internet started, in the end, all the main protocols, HTTP, TCPIP, they were not meant for secure communication. People were super excited about the possibility to have computer nodes communicating basically online, and only after they realized that people or automatic system could misuse this incredible technology also for cyberattack, for cybercrimes. And so all the cybersecurity measures started to take place. So they are comparing this a bit, what's happening at the moment in the Agentic era, all the incidents, something similar to what happened back in the days. And then at the time, the solution was to enforce security on multiple levels. So not only by software, but also by hardware. And this initiative is trying to define basically an open stack to implement security measures by default for all these kinds of applications. So if you get a bit down in the article.

OpenShell Policies Before Agents Run

SPEAKER_02 10:30

How would this work? So they are presenting basically two main solutions to this problem. One is mostly related to software, and another one is related to hardware, they're hardware, of course, because uh also they they have probably the 90% of market share for this kind of applications. And so they presented these two, let's say, projects. One is called NVIDIA OpenShell, is open source, it has Apache 2 license, and it's basically uh a runtime, an open runtime, and that you can use to host your hardnesses. So basically, the agents that lives within your application, basically. And how it works is that users can define policies that are enforced even before the agent can start its execution. So, to be very practical, for example, I can say this agent can access those files, make this API request, access these resources, and that's it. So even before the agent starts to get executed, OpenShell automatically enforces this policy for basically the environment where the agent will be executed.

SPEAKER_01 11:37

Do you know how they enforce it?

SPEAKER_02 11:39

Basically, they have a list of controls to be simple that they execute so that they can monitor the current available resources and also during the execution of the RNS, also what are the resources that the agent in practice is calling? To give you also an example, if they notice, for example, that something shady is going on, they try to monitor and eventually block the API calls that your agent is doing to the LLM provider. Because in the end, if the agent cannot think further, they cannot continue the execution of that malicious code or that malicious attack, for example. So I find it really interesting, and it's also interesting because it's it's open and is backed by a lot of large organizations. So this could be one of the first attempts to standardize a bit uh uh how those systems could be could be monitored and controlled in a secure way, basically.

SPEAKER_01 12:35

Very interesting. I'm curious about this. Maybe it's also something we should uh investigate for short.

SPEAKER_02 12:41

I think this will be adopted by default by standard tools and components that we can already use to build a genetic

Hardware Guardrails With Nvidia Sentry

SPEAKER_02 12:50

applications. And even more interesting, although this is not something open, it's not something that uh can be used in other systems, is their hardware controlled. So if you go back to uh the previous article, uh there is another component if you go above in the picture, so that we can show to people that are watching us. Exactly. There is uh on the on the right side Nvidia Centry, basically. Um this is a sort of hardware-based enforcement of policies, and because it is a sort of in-silicon monitoring system that works only on their new, let's say, hardware, in particular in their networking system, networking devices, networking storage systems, so that they can monitor basically the execution of this kind of applications at scale, so in their data centers, also perform telemetry and pipeline to detect misuse or malicious usage. For example, if an agent wants to run a DDoS attack on, I don't know, the data roots website, this should be prevented as well by the system and in milliseconds because it works directly in hardware, so it doesn't need to execute uh specific software for this.

SPEAKER_00 14:00

And this was something that would come with the Nvidia hardware.

SPEAKER_02 14:04

Exactly. In particular, with their new DPUs from their Bluefield project, something basically that runs in data centers uh we will never be able to see in consumer hardware. Also, because it's hard nowadays to also see even let's say commercial hardware from NVIDIA that people or company, relatively small company, can can buy and experiment themselves. So this is something that they will deploy and install in data centers they are currently building around the world. So then architecture.

SPEAKER_00 14:34

If I buy an NVIDIA GPU, is this something that I have access to somehow? Or is this just like if Nvidia is managing a server somewhere?

SPEAKER_02 14:41

If you buy the entire, for example, rack, including the GPU, their new CPU, and also their networking hardware, this will come inside that basically. But if we buy a commercial GPU, of course, uh this is not available. This really custom hardware for data centers, basically, and this kind of applications.

SPEAKER_00 15:01

It's a smart move from NVIDIA riding the cybersecurity wave as well.

SPEAKER_02 15:05

Yeah, yeah. I think it's one of their selling points as well to really be the hardware to run this kind of agentic systems safely, basically. Yeah, like they already are, but they want to stay in the default.

unknown 15:19

Okay.

SPEAKER_01 15:20

Very interesting.

SPEAKER_00 15:21

Very interesting. Shall we move on to the next

APM Brings Dependency Management To Agents

SPEAKER_00 15:23

topic? Of course. This is a smaller one. The it's APM. Have you ever heard of APN, Vitaly?

SPEAKER_02 15:28

No, this is the first time actually I see it. Have you heard of uh NPM? NPM, yes. What is npm for the way? So I feel like in exam or but I guess it's the repository for uh for any type of TypeScript, JavaScript, whatever packages that you can install with node or similar tools.

SPEAKER_00 15:48

Exactly. So the N stands for Node, as far as I know, and uh PM is package manager. So, like you said, it's basically a way to manage the applications, the the dependencies, and all these different things. So if you want to add, if you want to remove in Python, we have UV, which is also package manager, and for Rust we have which basically just cargo comes built in with the Rust. And this is for agents, basically. So it's open source, it's it is for Microsoft, but it is open source, and basically it's just a way to manage your the dependencies for your agents for cloud, for codecs, for open code, cursor, whatever. At first I thought it was an overkill, but now after have after having had having had a look, I thought it was actually kind of an interesting idea. So basically, like in whenever you're coding, you have skills, you have MCPs, you have different tools and context, right? There are third party or plugins, and this basically kind of bundles everything into one YAML file, right? So for the people that are just listening, I'm showing a YAML file here and how it looks, and then you have keys like name, version, and dependencies. I'm not sure why the version is there, to be honest, but maybe it makes sense. And then basically on the APM key and the MCP key, you have like skills plugins. So this is all directly from either from Entropic or from GitHub, so from repos. And you can also specify MCPs, right? So basically anything, like if I tell you Vitaly, like let's work on this project, then all the things I want this project to have is this MCP server, these skills, these tools, these plugins from third-party applications. Then I can actually just bundle everything there. And then the idea is that if I have a project with you, that we have we can have the Clot MD file that is shared, but then you can also make sure that we have all these things very easily, right? And you just do APM install and everything that is configured is there. Uh so this is a bit of a there was an NPX, right? NPX skill add, but this is a bit different, and I guess this is not just uh doesn't need I mean actually I'm not sure how the NPX skills worked, but I guess this is a a bit of a drop-in replacement, right? So not doesn't depend on node, and we can also do MCPs and all these other things, right? So that's a bit the idea. Again, I don't think it's it's I mean, I think it's a neat little idea. I'm not sure if it's something super groundbreaking. I never really felt the pain, let's say, like ah really but to be very honest, I haven't been coding as much, but uh, yeah, what do you think?

SPEAKER_02 17:55

I think it's interesting, but indeed, I feel that maybe I don't know how many people have this problem of installing and distributing, for example, skills or connections to MCP servers and so forth. Maybe in a sort of enterprise environment, this could be somehow necessary because you may have specific skills or specific MCP tools that maybe you want to share across your teams and maybe you want to centralize how you manage them.

SPEAKER_00 18:23

Yeah, maybe I also can imagine I can see a use case like there is a skill, and the skill is updated because maybe the skills use CLIs or whatever, but maybe for your team you don't want to update, right? You don't want to you want to make sure you use that version of that skill, or maybe this or maybe that. Also, models change, right? So maybe for my like, I don't know. I know that when Opus 5 came out, there was some people talking about that it's not as good, that it was a bit disappointing. Now 5.5 came out, and apparently, like everyone's super happy with it. But maybe also this is an argument to say, like, we are using this model because for our use case, this is better. Every time there's a new model, there's a new prompting strategy. So maybe I think it's interesting the idea of like traceability, right? Like things are moving and you want to make sure you're using the exact like trying to increase the reproducibility as much as possible, even though these are LLMs, right? So they're never going to be fully reproducible. But to try to make this as reproducible as possible, so you have the same prompts with the same models and all these things. Uh actually the model, I don't think it's not sure if it's specified here. I guess this would be more on your on your local level, but uh yeah, again, I think the idea, I think the idea, I don't think it can hurt, even though I never felt the pain that this solves.

SPEAKER_02 19:29

But I have a question though, because so far we saw, let's say, the client side of this. Is there also somehow a repository where maybe companies or for example, let's say they can host the APM repository to save internally, for example, custom skills or custom information to connect to custom NCP? I think this part will be the most interesting for people or companies.

SPEAKER_00 19:54

I so I I haven't looked as well from for I from what I can tell, this there's no hosting, right? Because even if you look at the APM and the skills that they list, this is entropic, so this is entropic uh skill. I think maybe I'm not even sure if this is a GitHub repo or whatever. But then some of them are just like GitHub, right? Or Microsoft, which is probably yeah, I don't know. So maybe there is some maybe there are a few, but I think you would mainly link from GitHub at this point. Okay. So I don't think they're hosting skills or anything like that, at least right now.

SPEAKER_02 20:25

This will be a way to automate this. So maybe your organization has repositories with skills that people can contribute to build, and then you you can use this uh tool to automate somehow the basically installing this in your in your local, let's say, environment or whatever.

SPEAKER_00 20:42

Yeah, yeah. I agree, I agree. So yeah, I mean again, I think it's something maybe I'll I will use, but I mean who knows, maybe I'll use it and be like, whoa, this is so much better than what I've been doing. But it's not something that I've I felt the pain yet, but yeah. What else do we have? Last topic of the day, Vitale.

SPEAKER_02 20:58

Last but not least, I'll say. Yes.

OpenAI Dev Day And ChatGPT Dots

SPEAKER_02 21:00

Yesterday it was the OpenI death day. Have you uh have you watched it?

SPEAKER_00 21:06

I didn't watch it. I saw before yesterday. I I saw that there were some people leaking information, quote unquote, like speculating. One of the things that I saw, actually, we saw on Techno Share this morning, it was the the the dots. Yeah. When they when I saw the leak, it wasn't called dots, it was called the O. But actually, I prefer I like that they renamed it to dots. But yeah, so that's all I saw. But what was there?

SPEAKER_02 21:28

There were people really happy about this that I saw, for example, on X or Reddit. There were also people disappointed because uh, as you know, nowadays there are new model releases basically every day. So every day there is a model that is better than the others, and then will replace us all, or will kill humanity, you know, how it works nowadays, uh, with AI. So there were people speculating about uh they will release Astra of 6.1, that is the model too dangerous to be released before or whatever. They still released a very good model, but nothing groundbreaking, let's say, uh nothing above the astral level. However, they focused on products, tools, subscription things that could help people but also developers in their daily activities. So I think OpenAI is positioning itself like a sort of company that from one side develops frontier LLMs, but also provides maybe an easy access to this kind of intelligence to the majority of people, and is in essence also the reality because the ChatGPT adoption in the first three weeks of its release was the fastest in history. So they presented a bunch of products, including LOT, as you mentioned. In the end, similar to OpenClaw or others, that can really spare you in your daily activities, including professional activities, but also personal life activities, like book a reservation or I don't know, uh plan this trip to Belgium. It's interesting because they claim dots will learn, your personal dots will learn from your own experience. So they will try to learn what are your activities, what you prefer, how you like to communicate, what you are doing daily, in order to somehow customize the experience to your profile. And also there will be the possibility to customize dots for specific activities, in particular in enterprise applications. For example, you can have your finance dots or HR dots or whatever dots to basically support you and your team with custom activities. They work 24-7, they have their computes, and apparently they are not consuming your usage. So basically they work independently a bit of codex and chat GPT that you make.

SPEAKER_00 23:49

But then you you need to have a subscription specialized for it, or you should just exactly that's the pain point.

SPEAKER_02 23:55

So I wanted to try dot this morning or yesterday evening when after it got released, but then you need a pro subscription. I only have the plus one, so uh the subscriptions that starts from 100 uh euros a month, basically. That's why I couldn't try it yet.

SPEAKER_00 24:12

But also I saw that in the U that's not available. Okay, I didn't see that, but it is really likely they will release it. I think the the regulation I think maybe slowed down a bit. But yeah, so maybe good that you didn't buy the pro just for this, because then you would have been dismissed.

SPEAKER_02 24:26

No, exactly, exactly. They also changed between the announcements, they also changed their pro subscriptions.

Paywalls And The New Pro Tiers

SPEAKER_02 24:33

Basically, now we have three tiers. So the $100 level, let's say the 200 before it used to be the biggest, the largest, but now there is another one, $500 a month for more usage, more yeah, in general, more usage basically, or largest models. And then they are decreasing the amount of things that you can do with a 200 uh subscription, which is a shame basically.

SPEAKER_00 24:58

Yeah, that's also a bit like you're a bit uh hostage of the these uh yeah. I feel like you're a bit hostage of these these LM labs, right? Like they just said they just say, ah yeah, you yeah, yeah, pay pay us more and you can do all these things. And then it's like well, actually, you need to pay us more to keep doing the things you're already doing, so it's a bit, yeah.

SPEAKER_02 25:15

Exactly. They also released other interesting things. Maybe we can go quickly over them. Codex now it's available for Linux, but also in the clouds, mobile devices, so you can really keep coding your application everywhere you want, and also something in that I think is interesting. They announced Chat GPT Spaces space, basically, which is a place where you can collaborate on projects with your team and dots. So it's a sort of collaborative environment where you can use together basically the ChatGPT features and codex features, and then interestingly, they also announced a new model, GPT 6.1 Sol, that has Astra level capabilities but at a fraction of the price. So this could be one of the daily drivers for many people, a lot of people. And also, it's really interesting that they announced the ultra fast mode both in the APIs and also in the applications like Codex or ChatGPT. So basically, you can have a throughput of 300 tokens per second, which is they claim eight times faster than usual, but then you need to pay six times more, basically. So it was uh it was funny because uh some some half one was uh demoing this on stage and it was like it's totally worth it because you can you know bytecode your app in seconds basically. So he immediately bytecoded uh a small application that was launching a rocket in space, and then he shown with with the normal subscriptions, it was taking a few minutes, of course, uh to do the same. If you have the money, yeah. Exactly, exactly. They also announced a bunch of other things that yeah, maybe we can link the article in the notes. But the more interesting I think were were these, including dots, which I think a lot of people will start to use.

SPEAKER_00 27:01

Yeah, I think maybe a few

OpenAI’s Strategy Consumer Vs Enterprise

SPEAKER_00 27:03

thoughts. I think one, uh, because the the guy from OpenClaw actually got equi-hired by OpenAI, right? So, in a way, it makes a lot of sense this move. The other thing is OpenAI, it is they are focusing more on consumers because I think Entropic historically has focused more on enterprise. OpenAI feels like they are focusing a bit more on enterprise as well. When you look at spaces, you look at codecs, and you look at all these things, even like focusing more on models for coding, but it feels like they're not fully there. So it's not very clear to me. I mean, they have, I mean, maybe their mission is to be the best at everything, but it feels less concrete. I feel like open on Tropic, you see, I mean, it feels for me at least that they're very focused on enterprise, and when you look at their products, it's very much focused on like like design is really for slide making and all these things, you know. You have Claude with skills and a lot of pre-built skills for finance, for all these, like it's very much enterprise. OpenAI. Now they have the dots, right? It's it feels like it's way more consumer-based. And I think it's like, are the people gonna pay this much money for for having a personal system? Maybe yes, maybe no. The guy from talking about Jeff, right? In the beginning, he did say like it's like if you have stakes, then you're not gonna use an LLM. And I think maybe dots, there is something with stakes. There are a lot of horror stories as well that came out from using open claw. So I think to be seen how how it's gonna go, I feel like if it works really well, I think then they pull this off and they have I mean they do have the biggest market share of of consumers, uh, of users as well, still, right? So if they can convert those users to to to revenue, then I think it's then I think they they hit the jackpot. But I don't know. I feel like we I heard a lot of stuff about open AI throughout, right? Like I remember there was uh Sam Altman said that they were working on an AI hardware as well that was gonna feel quote unquote like a cabin in the woods or something that is non-intrusive. So it's like they were playing with a lot, like they had Sora and then they shut down Sora, they had a deal with Disney, you know, like they had a social media for AI-generated videos. So it feels like they were doing a lot of stuff, but it doesn't feel like they have a uh very concrete vision. And I feel like now, I mean it's always been centered around users, so not enterprises, but now they're trying now, it looks like they're trying to do something more like personal assistance and and all these things. So to be seen, I mean, again, I think it's a cool idea. Also, maybe the other thing as well, other thing that fits in the puzzle. We even saw uh Monday there was a colleague of ours that did a demo on uh computer use from GPT that apparently is really good. And I was looking even on this little promo video that we have on the screen for dots, that there's also a bit of computer use on it, right? So I think it also makes sense, right? Like if they're up if they're trained the the models to to be good at computer use and they have these dots and they have these things, I think it kind of comes together, right? So again, I'm happy to be proven wrong, but I want to see how good this this works and the reviews, and I mean also it's not available in you, right? But as a as a strategy strategic decision, I'm I'm curious how it's gonna play out. Oh for sure, for sure.

SPEAKER_02 29:48

And I agree with you in it, because um, for example, sometimes I used to for example, sometimes I speak, talk to people that are outside, let's say, the IT world, and when they have to imagine uh an AI, they merely think to Chat GPT, for example, because uh this is probably the first one they see and they use outside basically the the the programming IT sphere. Only a few people know about Anthropic, for example, or Claude, or maybe they saw Dario Modell on TV, but they are not really sure who this person is.

SPEAKER_01 30:21

Yeah.

SPEAKER_02 30:22

And it's even more surprising to see a lot of people using, for example, Gemini or Meta directly in WhatsApp, for example. So I think they know that um the market available out there for the consumer market basically could be potentially huge. And they're trying a lot of different things to to monetize it, including maybe uh adding adding ads or uh I don't know, uh custom recommendation when you ask about a restaurant, I don't know, in a particular city and then it's sponsored by by OpenAI, for example. So I won't be surprised that something like this will happen. Uh from the other side, I think the strategy of Anthropic has been the winning one so far, at least for the enterprise uh market, basically, because uh when you need to do something, you to code something, the first thing you do is to to open cloud code and and start from there basically. So they also targeted the specific problem and try to solve it, in particular, sadly for us programming, because yeah, I miss sometimes to write some some lines of code.

SPEAKER_00 31:26

Yeah, I also saw a video on LinkedIn the other day that the guy was like, that's just the way it is. Like that the code in my hand is gone now, and that's just that's just the faster you accept it, the faster you can adapt. But that's just how it is.

Wrap Up And What We’ll Test

SPEAKER_00 31:38

True. I agree. Vitali, thanks a lot for the topics for the chat. It was very interesting. I think we have some things to give it a try as well. Yeah. Any last words you want to share before we call it a pod? Crazy times, crazy times, crazy times. Thanks a lot, Vitale. See you next time. Thank you, Marilo. See you all. Bye. Bye.