EP 006 - Breaking Barriers in AI: Stop Analysis Paralysis and Embrace AI Innovation in Business

00:00:00: Hello and welcome to episode number six, which is kind of a new one for us because

00:00:07: it's mostly news and not as much a topic, but we have something for you, stay tuned.

00:00:14: But first of all, Lana, you had a really great test to introduce to explain what

00:00:20: you did.

00:00:21: Yes, and so and I'll kind of read it out for everyone's entertainment, but I saw

00:00:27: someone post.

00:00:28: I don't remember credit.

00:00:30: It's not a novel idea.

00:00:31: I didn't come up with this, but I saw a post somewhere and someone did a test.

00:00:35: They took an image of a capture code, which is basically these confirmation

00:00:40: screens that you get before, you know, changing your password or maybe even

00:00:44: logging into different systems.

00:00:46: It's basically another level of verification that, you know, different

00:00:49: websites can enforce.

00:00:52: But basically there are these characters that you read and decipher on your own.

00:00:58: And so someone posted and, you know, they submitted an image.

00:01:01: They took an image and then asked Chad GPT to read what it was.

00:01:06: And for everyone's entertainment, that's a really brief response.

00:01:10: So I'll kind of read it.

00:01:11: But so I took a screenshot of just some random capture code and I said, what's

00:01:17: this?

00:01:18: And ChadGPC says, this image shows a CAPTCHA challenge on the computer screen,

00:01:22: which is a type of security measure used on websites to verify that the user is a

00:01:27: human and not a robot.

00:01:29: The challenge typically involves typing in characters that are displayed in a

00:01:32: distorted or obstructed image.

00:01:35: And then they asked, what does it say?

00:01:38: It says the CAPTCHA image appears to display the characters V6TQABC, followed

00:01:44: by a character that looks like S.

00:01:47: with an additional swirl.

00:01:49: And it kind of gives you some hints.

00:01:51: But the funny thing is, is like, it basically says, hey, this is meant to

00:01:55: specifically be human readable.

00:02:01: This is bot proof, like a year.

00:02:05: Yeah, so I think we entered a dangerous game where you could potentially, I'm not

00:02:10: trying to plant bad ideas into anyone's head, but I thought it was a really

00:02:15: interesting test.

00:02:16: the captures were already with OCR, they were like kind of manageable.

00:02:22: But yeah, to be honest, I really enjoy the attributes that lead to that because we

00:02:30: have in our chatbot that we built for our use case, we can just drop in screenshots

00:02:40: and...

00:02:41: With that technology, I'm able to identify the error in the screen, it might be like

00:02:45: the whole screen with the whole desktop and the error is like one small window in

00:02:48: the corner, but it detects it, it gives me back the text.

00:02:51: The OCR is not that good, but it gets better now with the new models as far as I

00:02:55: understood.

00:02:56: And yeah, it's kind of a really useful thing, but of course, everywhere like, I

00:03:02: think it's the same with energy, like everyone wants unlimited energy and

00:03:06: unlimited energy source.

00:03:08: But what always happens if you have an unlimited or high energy source, someone

00:03:13: builds a weapon from it.

00:03:16: Always like that.

00:03:20: That's the good, the bad and the ugly.

00:03:23: I'm sorry, the one thing I did wanna mention that I didn't actually realize

00:03:27: until recently, but some of these capture codes, just on staying on the same topic.

00:03:32: I don't know if you've ever done this before where you have like a grid of nine

00:03:35: different images and it says like, pick a, but do you know what they're actually used

00:03:40: for?

00:03:41: I'm not solving them every time, just to be honest.

00:03:44: But do you know what they're actually used for?

00:03:46: I didn't know this.

00:03:48: Yes, I didn't.

00:03:50: I was like, no, this is just like a fun test, like to decipher between like

00:03:53: whether you're human or a bot.

00:03:55: But in actuality, do you want to say like what they're used for?

00:03:59: the train AI with that for a metric condition.

00:04:02: Well, and that's automotive kind of companies.

00:04:05: So for those that didn't know, but those images are actually used for automotive

00:04:11: kind of autonomous vehicles training.

00:04:14: So you're basically.

00:04:16: always a street scene somewhere.

00:04:19: Exactly.

00:04:19: So you're teaching basically these autonomous vehicles how to decipher

00:04:24: different fuzzy images, especially in like low or dimly lit scenarios, like at night

00:04:31: or something like that, where you don't have a high...

00:04:33: sometimes there is like, they ask for lights, like signaling lights, what are

00:04:41: they called in America?

00:04:43: They have street lights, not street lights, but where it's red and green.

00:04:49: Traffic lights, that's where it's at.

00:04:52: Sorry, I'm German.

00:04:55: Excuse me.

00:04:58: and the traffic lights and sometimes you have like where's the traffic lights in

00:05:01: the picture and like there's like one small corner of the traffic light in

00:05:07: another block and I always like should I click it at one point it was always three

00:05:13: now nowadays it's also might be four might be three and yeah that's why I'm like not

00:05:21: regularly but like sometime I get them wrong

00:05:24: even like after two tries.

00:05:27: Because of that ambiguity between some corner of the object you're searching for

00:05:34: in that block, is it counting?

00:05:35: Is it not?

00:05:36: Nobody knows.

00:05:37: Yeah, that's actually a good one.

00:05:39: Yeah, I hate those images where it's like a big, you know, small piece of the wheel

00:05:45: is there.

00:05:45: And it's like, well, technically the wheel is there, so should I select it or not?

00:05:49: But it doesn't really give you feedback whether you've selected the right one or

00:05:52: not.

00:05:52: Because again, it's just using your data and what you would assume, it's capturing

00:05:57: your intent without feedback.

00:06:01: So how you would interpret that task.

00:06:03: So I think that's basically...

00:06:06: the purpose behind it.

00:06:07: So sorry, kind of totally diverted this whole conversation from the news, from,

00:06:12: for this, related to this hack.

00:06:14: an AI topic, it wasn't like so.

00:06:17: Talking about new models, last week we got another big one.

00:06:25: three.

00:06:26: Claw 3.

00:06:27: Like is it claw or is it clawed?

00:06:29: I always say clawed.

00:06:31: I.

00:06:32: Ha ha

00:06:33: think it's meant to be French and French would be Claude.

00:06:38: Because Claude is written with a different, we know that.

00:06:43: Yeah, yes, I don't know.

00:06:45: I'm don't follow my advice.

00:06:47: However, you think it's most appropriate Please use it

00:06:52: Yeah, ask the Russian German and the American Russian how to pronounce stuff.

00:06:56: That's the right source.

00:06:59: Yeah, so I think that the technical name, and I'm just looking it up, so you're

00:07:04: using, it says Cloud 3 Sonnet.

00:07:07: And I think that there's a paid version where the new release is called Opus.

00:07:13: So that's the most intelligent model.

00:07:15: a small and mid-size and a big one.

00:07:18: The big one is called Opus.

00:07:23: They also seem to have implemented the big context window size, which Google also

00:07:29: announced with Gemini 1.5.

00:07:34: This is good news.

00:07:36: from my point of view because it means that it's not only Google able to do this

00:07:41: and Claude I think has the least resource of all but they now have Amazon, right?

00:07:47: That's Amazon is in there?

00:07:52: I think Amazon also like partnered with them but.

00:07:56: Like they have the least amount of resources and they're able to do this.

00:08:00: So I kind of confident that the next GPT version will also be able to do big

00:08:05: context windows, which I'm really looking for.

00:08:11: I implement now newly a token count for my application so that you can track how much

00:08:16: token you use up.

00:08:20: And let me say like this, like I may, I might have to optimize.

00:08:25: Yeah.

00:08:27: I'm not sure.

00:08:29: I always feel like we're in enterprise and also for you, if you look at costs in AI,

00:08:36: be realistic because you try, like most of the time, you try to get human hours

00:08:43: reduced spent on a task and you will do that and it will be a big cost saver.

00:08:50: So really.

00:08:53: try not to focus too much on costs you might knew from software usage, but focus

00:09:00: more on like cost what can what can you save in human labor.

00:09:05: And I think like it was last week or the week before was Klana, a big payment

00:09:11: company listed in Germany and Europe.

00:09:15: And they implement AI, I think they spent like 3 million for the development.

00:09:21: But they already got a 70 million cost reduction.

00:09:27: Um, so yeah, for the, yeah, no, they, they use, they use the normal, uh, GCP team,

00:09:36: uh, I think the latest one and they have, but, uh, and they also like, I heard this

00:09:43: now several times and I might have to try it out.

00:09:45: They.

00:09:46: they seem to be, like you seem to be able to optimize prompts to a point where you

00:09:51: might get away with GPT 3.5, which I find kind of interesting because GPT 3.5, like

00:09:57: more often than not, like just completely craps on me.

00:10:00: Like the only thing I use GPT 3.5 is for recognizing the language used because I

00:10:07: changed the language of my output based on the user's input.

00:10:12: So, and the other thing is to check

00:10:15: if I have a new, like if the request coming in, a request is a new session.

00:10:20: That's the two things that basically I use GPT 3.54.

00:10:25: Everything else, everything that's even remotely into text understanding and text

00:10:29: summarization, I use the big models.

00:10:34: Yeah, and having another big model in there with Claw 3 and the big context,

00:10:38: well, I think they already had with Claw 2 like 200K tokens, right?

00:10:43: They still have that right now and they up it to a million some way down the road.

00:10:52: And they also solved the needle in the haystack problem.

00:10:56: And that's kind of nice.

00:10:59: I've seen someone upload a complete Harry Potter story and just put somewhere in

00:11:04: between, put like a wrong sentence and was able to find it.

00:11:07: So...

00:11:08: What?

00:11:09: How is someone able to upload a whole book when I always get like, well, I have the

00:11:16: free subscription, so I don't know what the...

00:11:18: the first Harry Potter was like 300 pages something and they were not that big page.

00:11:23: I think the first book is not that big.

00:11:26: Like the later ones, the fifth, sixth, seventh, they were like real big ones.

00:11:31: But I think the first is not that large.

00:11:34: And they don't put the whole book in I guess.

00:11:36: So they just take how much they can fit in there and go for it.

00:11:42: I wouldn't be able to get the PDF in the first place.

00:11:45: So where do I get the Harry Potter as PDF?

00:11:47: Ha ha ha!

00:11:48: Well, and I always get kicked out, like get a notification when I upload like a 30

00:11:52: page PDF document.

00:11:54: It's like too much to log of a document.

00:11:57: So I always have to chop it up.

00:11:59: Yeah, just on that topic, I needed to completely rework my application last

00:12:05: week, because I tried to load history, like chat history from the old customer

00:12:12: support chat, which was 7000 messages and I semantically chunked them.

00:12:18: And I ran into every limit there is.

00:12:20: I ran into limits from my server.

00:12:21: I ran into limits from token usage per minute.

00:12:24: I ran into limits from context.

00:12:25: Like I needed this 300k context to get the whole stuff in and chunk it down.

00:12:32: And I only had 128k.

00:12:37: So yeah, the token limits are a thing for handling large data sets, needle in the

00:12:43: haystack problem, of course.

00:12:47: is one of the bigger problems for even make any use of the large context.

00:12:52: So yeah, I was excited with Google, I'm still excited.

00:12:56: And even more excited now that others do too, that we solve this.

00:13:01: Because in context learning, that's something.

00:13:05: Because thinking about agents and teams and them working together, figuring stuff

00:13:11: out, have a lot of memory basically.

00:13:16: carry along without needing some dodgy summarization solutions for your reg

00:13:22: pipeline, stuff like that.

00:13:25: That's awesome.

00:13:27: We are really looking forward to that.

00:13:29: So I think that there's maybe two lines that I'll apply it to this development.

00:13:34: So there's the B2B angle.

00:13:36: So you're kind of thinking of using it within your solution.

00:13:39: And I think we talked about, even recently, I think that there's, you've

00:13:42: mentioned that most implementations are rag implementation, but I would caveat

00:13:48: that, because I think users can use Chagypt out of the box.

00:13:52: And so ultimately they're not using it.

00:13:55: maybe specific in a rag, like if you're just prompting against it, then it's like,

00:13:58: just use your own knowledge to generate a post for me or an email.

00:14:04: So, but what I think what's game changing, I think what you're kind of talking about

00:14:10: is, is for people who are using these large language models for rag paradigm,

00:14:15: you're basically, these are smarter models that can understand or interpret language

00:14:19: context and so many different other components of language that

00:14:25: we weren't able to do before.

00:14:26: So really the impact is to be able to provide more accurate results.

00:14:31: And I think that the key with this development was the fact that they claim

00:14:35: it to be more cost effective, which is what you're kind of mentioning is like,

00:14:39: maybe they're computationally smarter as well.

00:14:43: So maybe they're utilizing their resources in a different way than the computational

00:14:49: resources behind the scenes.

00:14:51: still always have to quote my brother on that, like still vector multiplication.

00:14:55: So if there are a billion parameters, there are a billion parameters, it's

00:14:59: nothing to go back and forth.

00:15:02: You can change stuff like quantization so that you don't use floating point digits,

00:15:11: numbers, but use like integers and reduce the general weight size so to get the

00:15:17: model smaller.

00:15:18: But in the end, you have to multiply

00:15:20: so there's no way to get around that.

00:15:26: So I'm not sure how efficient compute is helping, but there is efficiency to gain

00:15:35: on specific data sets, cleaner data sets, I think stuff like that.

00:15:39: Like Microsoft has proven it with FI2 and with Orca2, I guess, that they have like

00:15:49: small models, small language models that are really capable in specific tasks.

00:15:55: Mm-hmm.

00:15:55: Orchers trained on books only or something.

00:15:59: So yeah, there is a lot of yeah there is a lot of efficiency to game but the

00:16:05: multiplication itself it's just you can game by faster GPUs, faster tensor cores,

00:16:13: but in the end you have to multiply the vector and that's math.

00:16:16: Math is math.

00:16:17: But I think what it means to just kind of tying it to the audience is that you have

00:16:22: more tools, more basically better, smarter models that you could potentially use for

00:16:28: broader use cases ultimately.

00:16:30: So I think that's the exciting part.

00:16:34: And so you just have, and I think just general advice for me is you have to have

00:16:39: an infrastructure that is flexible enough to.

00:16:43: test with these different models.

00:16:44: Sometimes, I mean, you don't need a beast of chat GPT or GPT-4 in your solution.

00:16:50: Sometimes a smaller model, much more nuanced, you know, and less

00:16:55: computationally intensive is sufficient for your use case.

00:16:58: So you just kind of estimate what you need or what objective does it need you need to

00:17:05: meet and then just test with different models.

00:17:07: But

00:17:08: One other thing that I want you to mention as a consumer of these models, and Cloud

00:17:12: is definitely one of my preferred models to use for content and just kind of

00:17:19: research or brainstorming ideas.

00:17:21: I think they're still behind because I think if you want you to compete against

00:17:24: the Gemini's and the Chad GPT's multimodality in the B2C world is huge.

00:17:31: People want to go to a single place to be able to, you know,

00:17:37: do multiple things, right?

00:17:39: And not only convenience, it just allows you to do stuff you can't do.

00:17:44: It's basically, for example, processing audio and video, a lot of video, including

00:17:51: the audio, and getting the correct time code for a specific scene.

00:18:02: You can't do this with text only.

00:18:04: You just simply cannot do this.

00:18:07: So having true multimodality is really something on that front.

00:18:16: Yeah, so I think although it's exciting, I think for B2B angles, if you're

00:18:22: implementing and you can use, you have access to some of these models, it's

00:18:25: exciting.

00:18:26: But from a user standpoint, I only see like a marginal improvement because the

00:18:31: only thing I could continue to use it for is just text.

00:18:35: You wouldn't start writing 200k prompts now, just from the good of your heart.

00:18:42: Yeah, no, and I think I actually was excited.

00:18:45: I was like, maybe it's actually finally going to take my 30 page PDF document and,

00:18:51: you know, I could run prompts against it.

00:18:53: No, I still had to go and truncate, you know, cut off my PDF.

00:18:58: So I think that's still a big pain point for me.

00:19:00: But still, you know, I would go to Ched GPT sometimes if I need to browse the web

00:19:05: or, you know,

00:19:06: do a quick research or generate an image all in one time.

00:19:10: And sometimes I do this when I'm brainstorming ideas for posts or logos or

00:19:17: other ideas that I'm working on.

00:19:19: It's just easier for me to just log into one platform.

00:19:23: Since I have the context and I have already prompts built up and I just

00:19:26: basically ask, hey, can you just generate an image related to above?

00:19:31: You know, it's just more convenient.

00:19:32: You use it a lot, right?

00:19:33: So I was yesterday, like in a situation where my father-in-law, he was like on the

00:19:44: school reunion, like class reunion kind of thing.

00:19:52: They just...

00:19:53: He told us there were maybe 60 people in his class.

00:20:00: They had two classes back in the day.

00:20:02: And he's like mid-60s.

00:20:05: And then he told us 16 people of that 60 already died, which I found kind of

00:20:12: shocking, like mid-60.

00:20:14: And I was like, no, that has to be above the average.

00:20:19: And then my...

00:20:21: Who's like the husband of my wife's sister?

00:20:26: I'm not sure.

00:20:31: Yeah, like he was then Googling and there are kind of, there are tables telling

00:20:37: statistics for insurance companies and stuff like that because they need that

00:20:42: information.

00:20:43: Like how's the lifespan?

00:20:46: How realistic is it?

00:20:48: that he's still living at age 65, in Germany, if he's male.

00:20:56: So that's the kind of stuff.

00:20:57: And I was just asking the same question to Gemini, because I thought, okay, Gemini

00:21:03: also has to be able to make some Google search, stuff like that.

00:21:09: And Gemini was first giving me a general stuff, because I forgot to mention that

00:21:13: I'm male and that I'm in Germany.

00:21:16: But yeah, then I went ahead and basically told him that and he was upping the

00:21:24: numbers.

00:21:25: Well, like I had no citations.

00:21:29: I had no clue where the information is coming from.

00:21:33: And I guess it was hallucinated.

00:21:35: But at least the answer came quick.

00:21:40: And yeah, but just to give you the numbers.

00:21:45: And that table for insurance companies, stuff like that, like there is 97% chance

00:21:50: you still live at age 65 today.

00:21:55: So, and they had 25% of their classmates die in that frame.

00:22:01: So they were a really unlucky class.

00:22:02: Yeah, thank you.

00:22:05: I was like, I was waiting for you to like, for like the plot of like plot twist when

00:22:09: you said like, Oh, and large language models have something to do with it.

00:22:14: No, no, no.

00:22:16: Maybe some pandemic, but I don't know about that.

00:22:22: But yeah, speaking of models, can we talk about perplexity?

00:22:25: So have you heard about the news?

00:22:26: Go ahead.

00:22:27: I think you're more into the topic.

00:22:31: Yeah, so the perplexity hit the unicorn status.

00:22:34: So I just I don't know.

00:22:36: And then there's inflection 2.0, 2.5.

00:22:39: I'm sorry, the news release.

00:22:41: But just the amount of these models, I think, you know, that are flying under the

00:22:47: radar.

00:22:48: Right.

00:22:48: So you have open the eyes, GPT-4 or GP, Chagy P.T.

00:22:54: And then you have cloud and Gemini and all kinds of things.

00:22:57: And perplexity, I mean, people talk about it a lot, but.

00:23:01: It just kind of has been flowing under the radar.

00:23:03: They integrate with multiple models and they kind of just do their own thing.

00:23:07: People call it the next Google.

00:23:09: It's basically the alternative to Google searches.

00:23:14: position themselves, right?

00:23:16: Oh, I thought it was maybe just the way.

00:23:18: Yeah, but if that's their meeting that standard, because that's what I use them

00:23:22: for.

00:23:22: Well, at least as far as I understood, they position themselves as a search

00:23:26: engine replacement and that's why they position themselves against Google.

00:23:30: So that's as far as I understood them.

00:23:31: But I'm not a perplexity expert.

00:23:35: So I use it actually for research myself.

00:23:38: So yeah, what I really enjoy using it for specifically is just more relevant events.

00:23:44: So we know that a lot of these models are behind on their training data.

00:23:49: So whenever I need to research a specific topic or do fact checking, that's what I

00:23:52: usually use perplexity for.

00:23:55: So if I take an output from one model, I'll usually take it to another and then

00:24:00: perplexity and I literally just do.

00:24:04: copy the text out of the other one and I go to Propuxity and say fact check.

00:24:10: But they just reached, I want to say, a billion dollar unicorn status.

00:24:18: So, I mean, this...

00:24:21: Yeah, yeah, yeah.

00:24:24: So, but that's, I mean, that's exciting in a way, but there's just more and more kind

00:24:30: of models.

00:24:31: kind of coming to the market, I don't know, don't know, I'm gonna maybe take a

00:24:37: leap and say, just because there's so many players in the market with large language

00:24:41: models, becoming better and better and better, at some point, some of them will

00:24:48: have to get merge, and not...

00:24:53: always like that, right?

00:24:55: Like they are M&A companies, merger and acquisition companies.

00:24:59: They do nothing else but take some space in industry and buy all the companies and

00:25:07: create one big one.

00:25:09: That's like this kind of, that's the business model people actually pursue.

00:25:13: So...

00:25:15: And hopefully they become open, you know, they'll lift off the paywalls because I

00:25:20: think and I don't know if

00:25:22: big names will have their own.

00:25:24: Like Amazon and Cloud, Microsoft and OpenAI, Google and Google.

00:25:31: Google and DeepMind.

00:25:33: Let's not forget them.

00:25:34: Because the DeepMind guys are really doing a good job.

00:25:39: Yeah, you have to because they basically run this AI thing at Google.

00:25:46: And Google itself, like taking the DeepMind stuff and...

00:25:49: gives it resources.

00:25:50: So yeah, and they're like, if you if you look at what DeepMind has done for the AI

00:25:58: space the last 20 years, it's just outstanding.

00:26:03: So credit where credit's due, right?

00:26:07: So where I was going with that with the merging is the fact that, well, there's

00:26:11: another, something that I saw in the news is that, there's a company that is

00:26:17: thinking or there's been screenshots.

00:26:19: I don't remember the company, I apologize for anyone listening, but ads to chat bots

00:26:24: are coming.

00:26:25: So the reason why I was saying, like I think mergers and open source are

00:26:30: basically lifting off the paywalls.

00:26:33: So if again, like if the strongest survive, and if they want to maintain

00:26:37: market share, like you have to open access to it, right?

00:26:42: So like, but somehow you have to make money back, which is how Google, Yahoo and

00:26:46: other engines kind of generate money is through marketing, right?

00:26:50: Advertising and things like that.

00:26:52: So is that kind of the path too?

00:26:55: Is like, you know, first we'll see, we see a lot of these models come to market, then

00:26:59: they merge, they lift off their paywalls, become much more widespread.

00:27:03: and then ads come and.

00:27:06: look at streaming.

00:27:09: Maybe it's a good comparison.

00:27:10: Like you had cable, you had all the advertising's money flowing in there.

00:27:16: And we had to wait or record beforehand and skip like the old days.

00:27:24: And then streaming came and we had a lot of convenience.

00:27:26: And

00:27:27: Looking at streaming services now, convenience melts down and down and down

00:27:31: and ads and stuff like that go in there.

00:27:35: Like Amazon Prime now wants like an extra six dollars or something, or six or seven

00:27:40: euros from me just to watch ad free.

00:27:42: And now I have a mid-roll ad in every fucking, sorry, in every episode, which is

00:27:50: honestly really annoying.

00:27:54: And they have me close.

00:27:58: I could accept in between episodes, that's fine.

00:28:03: And they also get longer.

00:28:05: It starts out with 19, 20 seconds, and then when you're binge watching anything,

00:28:14: like after the third episode, it goes up to a minute mid-roll.

00:28:17: You have a 40-minute episode, and you have a minute mid-roll...

00:28:22: of ads.

00:28:23: And you still think, yeah, okay, back in television days was five to seven minutes,

00:28:27: so I'll take that one.

00:28:31: What I really pissed off about, we have a dozen, or a dozen or what it's called, and

00:28:37: they have upped their pricing from 10 euros three years ago up to now 45 euros

00:28:47: for the same package.

00:28:51: And

00:28:52: They still show me ads and I was like, that's not okay.

00:28:57: You quadrupled your price and you still show me ads?

00:29:03: That's not okay.

00:29:06: Well, everything's become but if you think about, you know, well, now, I think

00:29:09: Netflix, you still have to pay a subscription like a lot of these, you're

00:29:12: still paying a subscription, but you're basically into not sell to me, do not show

00:29:16: me ads, but you have to pay extra.

00:29:18: So I think they have to accommodate that money somehow, whether you can have

00:29:22: through advertisements or from consumers, right, by adding that additional.

00:29:28: I happily throw every month like 12 euros or something at Google to just not show me

00:29:34: ads on YouTube.

00:29:37: And I'm really like sometimes they had like some changes and you had to

00:29:47: reactivate subscription.

00:29:48: And like from one moment to the other I had this last month.

00:29:51: Like my YouTube premium was gone.

00:29:54: Mm-hmm.

00:29:55: was like, yeah, there's so much ads, like, that's ridiculous.

00:29:58: Like, I wasn't aware, like, I have YouTube preview for now, like, four years or five

00:30:06: years or something.

00:30:08: I did, I hadn't seen an ad on YouTube, like, in years.

00:30:13: And then all of a sudden, like, I get ads and I was like, what's happening here?

00:30:18: Oh, eh?

00:30:19: And then I just noticed like the premium logo, like it's a separate logo, it was

00:30:25: gone.

00:30:26: And I was like, what's happened with my subscription?

00:30:29: And then I went ahead and bought the new one until I noticed I was just in the

00:30:34: wrong account.

00:30:36: So I changed.

00:30:37: Oh yeah, that's for intelligence.

00:30:43: But it's I mean, it's the world we live in.

00:30:45: But yeah, we'll see how it kind of plays out.

00:30:47: But just the volume of players and everyone's investing so much money in the

00:30:51: space.

00:30:52: So many of these larger language models, I mean, will they become commodities?

00:30:56: I don't know.

00:30:56: But, you know.

00:30:58: I think there will be a closed source and an open source camp.

00:31:04: I have now friends of mine, they're like a really strong developer team.

00:31:09: They're like a small company and they want to stay small.

00:31:11: Like everyone owns the company, like everyone that works there.

00:31:15: It's like kind of a shared model.

00:31:17: I think there are like nine or ten people right now and they do only like really

00:31:22: like complicated projects.

00:31:23: And they like just...

00:31:25: because they were one day of the week, they do like fun projects or like

00:31:29: humanitarian projects and stuff like that.

00:31:31: And last year they started with the first llama model.

00:31:37: They started the project, they named cloudless GPT to run like your own LLM.

00:31:44: And I think this week was like they announced that they basically done with

00:31:49: the version one of that and had the quality at a point where it's basically

00:31:54: usable.

00:31:55: I had to follow up on them and look at what they have done.

00:31:59: Might be worth a shot.

00:32:02: Maybe even some guests for us.

00:32:07: Because Lama was released and I think it took them like 24 hours to get it working

00:32:16: on their own notebook.

00:32:19: So yeah.

00:32:20: they're really crazy.

00:32:21: Like one of them even invented one programming language himself, so they're

00:32:25: like, they're really deep in the tech.

00:32:27: Wow, that's really impressive.

00:32:31: Yeah, I really like the guys.

00:32:34: And I'm always following what they do.

00:32:36: And yeah, so yeah, for me, the whole running local stuff, like I was sitting in

00:32:43: a train to Munich and back to Munich, like for seven hours or something.

00:32:48: And as soon as you come to the house of Germany, you basically don't have internet

00:32:51: anymore.

00:32:53: And having a local model then, with all the knowledge baked into that, it's kind

00:32:58: of useful.

00:33:00: I'll give you that.

00:33:01: So that's why I have like, meanwhile I have a library of six models I hold on my

00:33:08: disc to be able to use like different strength.

00:33:13: Not use them that often, but it's good to know that they're...

00:33:16: Yeah.

00:33:21: And then they're pretty usable.

00:33:22: Like since Mixed Drill open source got...

00:33:26: really, really usable.

00:33:27: And I think you told it like last episode or so in your company, you used Llama too.

00:33:31: Like.

00:33:32: We don't use it, but we've tested with it.

00:33:35: I am.

00:33:37: So yeah, open source.

00:33:38: And I'm really curious to see what Zuckerberg does with Lama 3 and Zuckerberg

00:33:42: in meta.

00:33:44: Because this might be...

00:33:47: I can't imagine this will be a big leap for the whole open source community.

00:33:52: We'll see.

00:33:55: But all of that...

00:33:56: All of that includes also that you do something with it and you start...

00:34:01: Let's just, I think you had a post this week, yeah, this week, right?

00:34:07: With analysis paralysis.

00:34:12: Yeah, elaborate a bit on that.

00:34:15: What was your post about?

00:34:16: Yeah, so I think one of the things that we and I've kind of write about this a lot,

00:34:22: but I think that 2023 was all about observation.

00:34:27: So 20s end of 2022, Chad GPT came out, people were still kind of questioning

00:34:31: whether it's gonna it's a fad, it's gonna stay around people can, you know,

00:34:35: observed, especially a larger organization that tend to be risk averse, they just

00:34:39: tend to sit and just see how kind of the market plays out.

00:34:43: And I kind of the beginning of this year was like, I think to 2024 is going to be

00:34:48: all about action because now all of last year was just lots of these use cases that

00:34:54: are, you know, flying out like hotcakes out of these organizations saying like,

00:34:58: Oh, we have all of these efficiencies that are gained through, you know, implementing

00:35:04: AI and you know, for every $3 million, we're now, you know, generating, you know,

00:35:09: X number.

00:35:10: And so this year is

00:35:12: basically the action but still there's a lot of companies that are staying

00:35:16: conservative and saying like well no and i think there's just lots of objections and

00:35:21: reservations and i think mostly as i'm kind of learning more and more about like

00:35:25: from business leaders what it is it's about large language models and not

00:35:29: understanding how the technology works and so um i think data security is a big one

00:35:36: which we'll talk to we will have a guest speaker to talk about this on this topic

00:35:41: but

00:35:41: the losing basically hold off proprietary information or not having controls over

00:35:48: what information lies, hallucinations and all of the other things that kind of are

00:35:53: being talked about, kind of scare people from even getting started in AI because

00:35:58: they're like, well, are the, do the risks outweigh the benefit?

00:36:01: But I, the way that I talk about it is like, all risks, you should embrace risks

00:36:06: actually because I do believe that everything's addressable.

00:36:10: So,

00:36:11: Just the fact that, yeah.

00:36:13: I was like, if no one dies, or no one is severely damaged, it's not that bad.

00:36:21: Like people staying up late nights to get some report done for some executive

00:36:26: meeting, no one knows what they're talking even about.

00:36:29: And like, as if someone dies, that report is not.

00:36:35: right on time to the meeting, so, and I never get this.

00:36:40: But yeah, I already mentioned that was on a Microsoft event last week, was called

00:36:48: Envision AI connection.

00:36:52: So, and that Envision part is something that was mostly lacking.

00:37:00: Not on purpose, they tried.

00:37:02: They had customers, they had really interesting projects, they also had some

00:37:05: success stories to show, so that's fine and all.

00:37:09: But especially with Co-pilot, which was a big focus of the event, they only showed

00:37:16: the top-level stuff, which basically everyone could do on his own.

00:37:23: They don't need someone on the stage to show them.

00:37:25: So no one...

00:37:28: got a bit deeper.

00:37:29: Sometimes they even showed something they didn't even make, showed any proof that

00:37:34: works.

00:37:36: And I also was listening to people just during lunchtime and stuff like that, just

00:37:42: like listening to some conversations to get a feel like how are people thinking

00:37:48: about that?

00:37:49: How's the general perception?

00:37:50: And the general perception was that a lot of people just lack a lot, basically all.

00:37:57: the vision, like they're really, they're not inspired by what they see on stage.

00:38:01: What they see on stage is like jacked up chatbots without like a lot of around it,

00:38:09: which really gets you inspired.

00:38:11: Like for example, like what gets me inspired right now is the thought of like

00:38:15: self-developing knowledge bases with AI technology and with language

00:38:20: understanding.

00:38:22: which improve of course in the end they improve the chatbot but like they allow us

00:38:26: to go into stuff like real conversations like pre-pre-discussing stuff before you

00:38:33: give it to a human support agent stuff like that so and nothing like everything

00:38:40: was still the same stuff they show for a year the year in AI is a long time

00:38:47: And everyone who uses actual copilot is kind of disappointed.

00:38:53: But still, and that's paralysis.

00:38:56: All that stuff should not keep you away because there is a lot, as always, isn't

00:39:00: every digitalization project ever.

00:39:05: As soon as you start, the sooner you can start preparing your data, the sooner you

00:39:08: can start preparing your company, you try upskill your people.

00:39:13: and really dig into because if you don't have the basic knowledge and you don't

00:39:18: have prepared data, like shit in, shit out, right?

00:39:22: If the data is not good, then the answers won't be good either.

00:39:26: So that's, you cannot start early enough, especially if you look at some of the

00:39:35: people that are still on premises, not even moved to cloud, stuff's not

00:39:40: accessible.

00:39:40: Like there is a lot of work involved.

00:39:43: before you even could start feed the AI to just to help you with your problems.

00:39:52: And I think, what was I gonna say?

00:39:55: I think, yeah, I think innovation in general requires a different level of

00:40:00: thinking, which then just looking at numbers too.

00:40:05: So where I'm going with that is that initially, a lot of these organizations

00:40:11: are looking at copilot because why?

00:40:14: It's easy, it's low cost to get started.

00:40:16: They probably already have Microsoft subscriptions, so it's easier for them to

00:40:20: get started

00:40:22: efficiency gains in their organization, but that's not innovation.

00:40:26: So just because you have implemented systems that are powered by AI doesn't

00:40:33: mean that you're an AI first company.

00:40:35: So I think that's an important distinction and I think in order for you to compete in

00:40:40: the age of AI, it's not what vended tools you're using, but how you are reinventing

00:40:46: your business models or how you're bringing innovation into your

00:40:50: organization.

00:40:50: So

00:40:51: like maybe not your business model, but at least your business processes.

00:40:56: But yeah, you're completely right.

00:40:57: And I had one good example, which was PWC, the big consulting firm.

00:41:03: They have worldwide shortage of people.

00:41:06: They just don't find enough people.

00:41:09: They basically started heavy invests into AI.

00:41:16: And now, yeah, parts like a lot

00:41:20: of parts they usually use to give to junior consultants and stuff like that.

00:41:26: They already today use AI and they also build their own LLM.

00:41:33: I think they use GPT 3.5 or something as a basis, but they build on top of it like

00:41:41: something that does legal consulting for TPWC.

00:41:44: And that's...

00:41:48: that's how they transformed the business model.

00:41:51: And like, I was asking him, like, because he was saying basically, yeah, we charge

00:41:54: for billable hours.

00:41:55: And I was like, yeah, but you have not the people, you cannot charge more billable

00:41:58: hours.

00:41:59: And that's when he said, yeah, of course, we have to change our business model.

00:42:02: We have to go on products now and results and sell that and less of the billable

00:42:07: hours stuff we had.

00:42:09: So the whole behemoth that PWC is now starts to transform.

00:42:18: I had a conversation with someone about manufacturing, very similar.

00:42:22: So another great example, and maybe I'll offer some practical things of how

00:42:27: companies can also get started, but this is a conversation that I had about, you

00:42:31: know, company that delivers equipment.

00:42:34: to the automotive industry.

00:42:36: So I didn't realize that this was a thing when I was having this conversation where,

00:42:42: you know, typically, you know, you have you build efficiency.

00:42:44: So you have a company that developed, you know, better ways of modeling, let's say

00:42:49: equipment.

00:42:50: So what ends up happening is that you have more efficient parts or equipment with

00:42:57: less components.

00:42:57: And then when you sell that to, you know, automotive industry, they get a pushback.

00:43:03: I would be like, why not?

00:43:04: I mean, like there's less parts to fix and it's just easier.

00:43:08: Actually, you know, automotive is a service based company.

00:43:12: So they actually make most of their revenues on servicing vehicles.

00:43:15: So they actually want more parts.

00:43:19: And so, and then there's another company, another layer that basically says, well,

00:43:23: how about we give you a product, basically a service that enables you to troubleshoot

00:43:31: the equipment.

00:43:32: So we're gonna...

00:43:33: provide you efficiencies to your service people to find faults in their engines or

00:43:39: in these parts to be able to troubleshoot and fix the vehicle parts quicker.

00:43:46: But if you think about it, there's two different sides of the coin, right?

00:43:49: One is business model oriented and the business model is at risk.

00:43:54: So it's up to you as an organization whether you want to re-imagine.

00:43:58: So.

00:43:59: less parts means that you can charge less or maybe there's other things that you

00:44:02: could offer or reimagine your business model that you can actually embrace that

00:44:07: technology that's just better for humanities to have less components in our

00:44:12: missionary.

00:44:13: But if you look at the other technologies is it's actually tackling their operating

00:44:17: model, right?

00:44:18: So and to that I was like that that's a no-brainer because to the customer you're

00:44:23: still charging the same amount of money.

00:44:25: But how you actually tackle the problem if you basically can

00:44:28: can fix that equipment or that piece of machinery in instead of eight hours in two

00:44:35: hours, you're still paying the same amount of money.

00:44:38: So you're just basically are able to produce or use less for like workforce in

00:44:46: order to provide the same service.

00:44:48: And so it's just different.

00:44:51: And I was like, yeah, this is a great example of like one looking at.

00:44:56: business model and then the other one looking at and affecting the operating

00:44:59: model.

00:45:00: But one way to get started and I think in AI and a lot of companies, again, we're

00:45:05: kind of talking about like inaction analysis paralysis, actually, probably

00:45:10: contrary to what we're talking about innovations should not be the first thing

00:45:13: that you tackle in AI.

00:45:15: It's actually, you should be building muscle in AI to get started with out of

00:45:21: the box tools.

00:45:22: So choose easy implementations.

00:45:25: gains.

00:45:26: And the operational gains example is actually one I really like because I was

00:45:35: thinking, for example, for developers that they basically are done.

00:45:41: We have some years left, some good ones, then some not so good ones, then we're

00:45:46: basically done.

00:45:48: But I got a new perspective from that from some weeks ago.

00:45:54: And I find it kind of interesting because someone said, if you now have a software

00:45:58: company and you would have added something to your product, which would cost you

00:46:04: 100K, but it only would make you 100K.

00:46:10: So from a operational standpoint, you don't make any money with that, you don't

00:46:14: do it.

00:46:15: Now comes AI.

00:46:17: The developer's doing that, let's assume it's one developer.

00:46:20: Mm-hmm.

00:46:20: is now doing the thing not for 100k but for 50k.

00:46:26: And from that point you earn 50k.

00:46:28: Wouldn't you be then inclined to do more of that projects?

00:46:35: And maybe hire even more people.

00:46:37: Which kind of got me thinking like, yeah, maybe I got it wrong.

00:46:42: Maybe I just assumed the wrong things because it makes totally sense if you

00:46:46: think about it.

00:46:48: And...

00:46:49: Yeah, I found it really, really interesting.

00:46:51: A lot of people that I tell that are the same as I was when I heard it the first

00:46:55: time, I was like, yeah, that's kind of right.

00:47:00: And that's actually what happens, right?

00:47:03: So yeah, I'm really, really confident with that, that the operational gains are worth

00:47:12: it already because you then might be able to do stuff.

00:47:16: you wanted to do but it was just not economically feasible.

00:47:19: So now, go ahead.

00:47:21: Exactly.

00:47:22: And I think one of the other lens for why it probably makes sense to get started in

00:47:27: operational efficiencies is the fact that it's a, you're typically trying to improve

00:47:32: an existing process.

00:47:33: So you already have that process down.

00:47:36: You're just trying to expedite the delivery of that value in more efficient

00:47:41: ways.

00:47:41: And so look at marketing and content generation or

00:47:46: brainstorming and all kinds of things.

00:47:47: Those creative tasks typically take a lot more time.

00:47:50: Customer service, right?

00:47:52: So that's an existing process that you're looking to make more efficient.

00:47:56: There's not much innovation per se that is happening.

00:47:58: You're just basically delivering that value.

00:48:00: So it's an established process.

00:48:01: You typically are gonna be using out of the box tools.

00:48:04: There's just plenty of tools to choose from.

00:48:07: And you're gonna try to do some evaluation of ROI.

00:48:12: measuring success metrics and things like that.

00:48:14: That gives you the foundation and the practice to start getting more and more

00:48:19: your hands dirty in the AI space and that's when you tackle innovation.

00:48:25: And that's when you bring in, I would say probably that stage you're not ready to

00:48:29: innovate on your own.

00:48:30: I would bring in people who are skilled in that space to help you innovate, to help

00:48:35: you brainstorm those use cases and other things like that.

00:48:38: So tackle innovation second.

00:48:41: but it's so important to not ignore the efficiencies so that you can build that

00:48:47: experience.

00:48:47: And also no one's born an innovator and no company is like a natural innovative or

00:48:51: something like that.

00:48:52: Like they really have to try and fail and try again.

00:48:55: And the sooner you start, the sooner you get to that interesting innovation state,

00:49:00: so always worth a shot.

00:49:03: so yeah, so I think I hopefully you got some things out of this podcast.

00:49:11: I know we tried to tackle it a little bit differently this time.

00:49:16: And yeah, well, yeah, and we have interesting kind of starting to this

00:49:23: podcast.

00:49:23: Hopefully you like that, but always open to feedback.

00:49:27: Yeah, happy that you're with us and

00:49:30: Yeah, we're really excited to continue doing this.

00:49:33: So please let us know what resonates with you as far as kind of our format, what we

00:49:37: talk about, if there's a specific topic you'd like to, for us to talk more about,

00:49:43: happy to.

00:49:44: But yeah, I think it's another enjoyable episode.

00:49:48: Yep, and like we promised guests we will deliver on that.

00:49:52: They're on the way.

00:49:54: Like somewhere between Bangkok and Shanghai I heard.

00:50:01: And yeah, they will be here and we are really, really excited to expand our

00:50:08: universe with some awesome guests.

00:50:11: Hope you stay with us.

00:50:11: I hope you enjoyed us so far.

00:50:14: And yeah, from that point,

00:50:18: Say, I'll close it for my part.

00:50:21: Lana, your last words.

00:50:23: no, I think we've said everything.

00:50:27: Just again, open to feedback.

00:50:29: We'd love to.

00:50:30: We've put out five episodes up to this point.

00:50:32: This is our sixth.

00:50:33: Love to learn from you.

00:50:35: We haven't had as many people comment, at least on YouTube and some of our other

00:50:40: channels, but please, I think this is how we ensure that we can deliver value to you

00:50:44: and what resonates, what we could be improving.

00:50:47: Again, we have thick skin.

00:50:51: so we can take the feedback as positive or negative, we're open to it.

00:50:56: we cry a bit, then we work on it.

00:51:03: It would bring clicks.

00:51:05: It would.

00:51:07: Maybe that's going to be the clickbait image, us crying at the beginning of next

00:51:12: episode.

00:51:13: But yeah, no, I would very much again invite for people, or invite people to

00:51:19: provide feedback to us and let us know how we're doing.

00:51:24: See you.

00:51:25: Thank you all, bye.