Apertus: Swiss Open Source LLM Demo

Watson Tech World

💻 Launch your websites with Cloudways — managed cloud hosting made easy: https://go.watsontechworld.com/cloudways --- A walkthrough of Apertus, a fully open-source Swiss AI model with open source code, open source data, and multilingual support. Demoed on chat.publicai.co. --- #Apertus #AI #OpenSourceAI #OpenSource #SwissAI #PublicAI #TechDemo #Switzerland #OpenData #LLM #OpenSourceSoftware

Transcript

00:00Hi everyone. In this video I'm going to do a demo of a really awesome new AI large language model

00:11or class of large language models from Switzerland. And I've tried it out and I think it's fantastic.

00:19It's basically open source and the data is open as well. And honestly I think the quality is

00:27extremely good. And I think it's comparable in quality maybe to GPT-4 but I can't say for certain.

00:38But I love that it's from Switzerland and I love that the quality is very good. So let's get started.

00:46Just do this quick intro about it. So the model itself is called Apertus, Apertus, Apertus.

00:57And that's a Latin word that means open. And so the idea is that the code is open source and also the data

01:06is open source. And it's trained from scratch on 15 trillion tokens. So it's quite a lot of tokens.

01:14And this is unique in that they insisted on the data being open. So I think like if it's public domain or

01:27data that's allowed to be downloaded. A lot of the other big LLM, M's, maybe some companies scraped

01:39many, many, many websites but they didn't get permission from sites to use their data.

01:45So this one is unique in that they insist on the data being open and also the code is open.

01:52And they have a hugging face page where if you're so interested you can check these out.

01:58So for example, this one is the most popular, 70B. So I think that's 70 billion parameters.

02:06And yeah, so it gives you some information here. It says fully open model, open weights,

02:14open data, and full training data, full training details, including all data and training recipes.

02:20It's massively multilingual. So it's got over 1,800 supported languages, which I think is fantastic.

02:30And yeah, so that's a bit of an intro. Now what I wanted to show you is this site.

02:39It's publicai.co. This is one place where we can actually test this out.

02:46So Hugging Face is where you could download the code and model on your own computer.

02:54But if you don't want to run it on your own computer or you don't have enough,

02:58you don't have a powerful enough computer, you can use publicai.co.

03:02And what you can do is you can make an account and you'll confirm it.

03:13And then once you confirm it, you should see something like this.

03:17It'll be a blank.

03:21And this looks pretty similar to the design of like ChatGPT or Google Gemini or Cloud or many other,

03:29or even DeepSeek, if you go to DeepSeek.com.

03:34So it's got a very similar design.

03:37And I'll show you this is an example.

03:42How can one learn to become an entrepreneur?

03:46So I forgot how long this took.

03:50It might have taken maybe five to ten seconds,

03:52but it gave a very comprehensive answer.

03:56And let's just read the top.

03:58To become an entrepreneur,

04:00one should consider the following personal and skill development strategies.

04:04Learn the foundational skills.

04:09Develop a solid understanding of finance, accounting, and basic business operations.

04:15Gain experiences, and so on.

04:19Launch and iterate.

04:21Okay, this is fantastic.

04:22Stig, it talks about MVP, minimum viable product.

04:27And, yeah, so I think this is an excellent answer.

04:32And here we have the options to edit.

04:36You can read it aloud.

04:38Say it's good or bad.

04:39And I'll say it's a good response.

04:41And we can continue.

04:44And you can also regenerate.

04:47And this is an option read aloud.

04:49I didn't try this out much.

04:51But, uh...

04:52Okay, so I want to test this out.

05:00Honestly, I didn't try it out before.

05:03Okay, so maybe that's having a problem, but that's okay.

05:20And, uh...

05:20We'll exit here.

05:23And, uh...

05:24I wanted to go back here.

05:27So, you can go to new chat like this.

05:29So, again, it's a similar design to ChatGPT or Google Gemini or whatever.

05:34And this is an interesting thing.

05:37If you know German,

05:39you can search in German, and I think that's fantastic.

05:43And this actually says Swiss German.

05:46So, I won't go over this, but, uh...

05:50If you want to practice German or if you know German,

05:52uh...

05:53This is pretty fantastic.

05:56And they make some funny jokes.

05:59So, up here it says

06:01sometimes it makes mistakes.

06:04And if you know German,

06:05this is, like, the double S.

06:09And that's kind of funny.

06:10And it doesn't...

06:12And for some reason,

06:13they like to make fun of Bavarians.

06:15I think that's kind of funny.

06:17Anyway, uh...

06:19So, you can search in German,

06:21and that's kind of interesting.

06:23And very briefly,

06:24I wanted to mention about the models.

06:26So, you can choose different models.

06:28I'm just going to stick with the Switzerland one.

06:32The Apertus model.

06:34The Singapore model,

06:36uh...

06:37I think I'll talk about in another video.

06:39You can see this is Swiss...

06:40Swiss...

06:41dash AI Apertus 70b Instruct.

06:46And this one is the Gemma Sea Lion.

06:51And, uh...

06:53Yeah, I...

06:54I'll talk about this one later, hopefully.

06:56But I'm going to stick with the Swiss one.

07:00And, uh...

07:01So, let's continue...

07:04with, uh...

07:05Let's see here.

07:09And here we can...

07:10For some reason,

07:10I don't know why I did that,

07:12but you can name the chat...

07:14Learn...

07:15Let's say...

07:16Learning...

07:17Business.

07:19So, you can name the...

07:20Name the tabs,

07:21kind of like in ChatGPT.

07:23Or I'll just call it Business.

07:26Okay.

07:27Uh...

07:27Maybe that's a bug.

07:29But that's okay, anyway.

07:32So, uh...

07:34But again,

07:34this is really, really, really amazing,

07:36in my opinion.

07:37And even just two years ago,

07:39like, this might even be better

07:42than ChatGPT 3 or GPT 3.5.

07:45So, AI is really, uh...

07:48Massively increased

07:49in a short amount of time.

07:53Um...

07:54And so, uh...

07:55This is something

07:56I didn't try out much,

07:57but I want to try out

07:58the Dictate

07:59to see if it can, uh...

08:02I want to ask it a question.

08:09Teach me how to start

08:11an online business

08:12run 100% remotely.

08:14Okay, so that's fantastic.

08:24So, it transcribed it perfectly.

08:26Teach me how to start

08:27an online business

08:28run 100% remotely.

08:36And you can see,

08:37I wanted to show you

08:38this is the actual speed.

08:39I think it gets really fast,

08:41so I just want to show you

08:42what it looks like.

08:44And you can see here,

08:45starting an online business

08:46and running it 100% remotely

08:48as far as a statistic blend

08:50of planning,

08:52the right technology stack,

08:54and effective management skills.

08:57Identify market needs.

09:01Define your business model.

09:02And, uh...

09:05Yeah, so it gives lots

09:07of very useful information.

09:11And it gives specific information,

09:14uh...

09:15like if you're, uh...

09:17in Switzerland.

09:19And, yeah, this is super comprehensive,

09:21and I really love this, actually.

09:26Okay, and there's another thing

09:27I wanted to show you,

09:28which is copying the output.

09:29Just like in ChatGPT or Google Gemini,

09:33you can copy the output.

09:35So here is copy.

09:37And I believe it copies

09:38to Markdown code.

09:41So this is Markdown code.

09:43And generally,

09:44I don't find Markdown code

09:45useful by itself.

09:47I usually will convert it

09:48to HTML.

09:51So you can copy this.

09:53And, for example,

09:54you can use a website

09:55such as Markdown2HTML.com.

09:59You can paste that in

10:00and copy this.

10:03So that's the converted HTML.

10:05And let's say

10:05you're using Google Docs.

10:07You can paste this in.

10:10And see, look at this.

10:11This is wonderful.

10:11It gave six whole pages

10:13of output,

10:14maybe within 10 to 15 seconds,

10:17which I think is really amazing.

10:20And this is really a great answer.

10:22I think it's...

10:23Yeah, it's very high quality.

10:25So let's continue here.

10:36And we have this Notes option.

10:38I haven't actually tried that out.

10:41I wonder what this actually is.

10:45Okay.

10:46I'm not sure what that is.

10:48And let's go here.

10:57Workspace.

11:00Okay.

11:01So they have different tools.

11:02I don't have time to go over these,

11:06but these seem to be kind of like the...

11:09some tools in ChatTPT,

11:11which I think is really cool.

11:13And I think you can search through your...

11:18You can search through your chats.

11:24And this is something,

11:26at least as far as I know,

11:27it doesn't exist in ChatTPT

11:29or Google Gemini.

11:31The search feature,

11:32I think that's very useful.

11:33I think that's very useful.

11:36And...

11:36All right.

11:38And let's test out...

11:40I want to test out the quality.

11:43I've asked it some questions about business.

11:46Now I wanted to ask it some questions

11:48I ask of almost...

11:50of many different...

11:52large language models.

11:54So I wanted to ask it...

11:57basically two complicated questions

11:59and see how it...

12:00how it answers.

12:02So I'm asking...

12:04Describe the process of photosynthesis in plants.

12:08Okay.

12:08So this is a simple question

12:10or a simple command,

12:11but the answer is rather complicated.

12:14And let's see how it goes.

12:18Okay.

12:18So you're asking key components of photosynthesis.

12:22It's a process by which plants,

12:23algae, and certain bacteria

12:24convert light energy into chemical energy

12:27stored in glucose

12:30through the conversion of carbon dioxide and water,

12:33releasing oxygen as a byproduct.

12:35Okay.

12:35That's excellent.

12:37And you see here,

12:38it actually links to Khan Academy,

12:41which I think is just great.

12:46I'll just call this plants.

12:53And you can see,

12:56I think this is really wonderful.

12:59I gave a really great answer.

13:01Again, here's a link to Khan Academy.

13:05And I want to ask it one more question.

13:10Let's ask it this question.

13:12Describe the impact of artificial intelligence

13:14on job automation,

13:17or let's just say on jobs.

13:20Describe the,

13:21on jobs.

13:24Do you think

13:25millions

13:26of people

13:28will lose

13:30their jobs

13:32because of AI?

13:35So many people think this is the case.

13:37I think this is the case.

13:39But I'd like to know

13:40what this AI thinks.

13:41Okay.

13:44Analyzing the impact

13:48of artificial intelligence,

13:49AI on the job market

13:50involves considering multiple factors,

13:52including

13:53technology trends,

13:55socioeconomic conditions,

13:56and human adaptability.

14:02So job redistribution,

14:04job transformation.

14:07So it's going to create

14:09a whole new category of jobs.

14:12And some people

14:13are going to need to upgrade

14:14in upskill.

14:18So yeah,

14:18I think this gave

14:19a really great answer.

14:22And

14:22yeah,

14:24so I wanted to do

14:25to do basically

14:26a conclusion,

14:28do a little bit

14:29of a review here.

14:30So this video

14:31was about

14:32this model

14:37from Swiss LLM.

14:43And the word means

14:44open in Latin,

14:47apertus.

14:49And this is from Switzerland.

14:50You can get it on Hugging Chat

14:51or you can use

14:52the publicai.co site

14:55to use it.

14:56Okay,

14:56I hope you enjoyed.

14:57Please subscribe to

14:58and follow this channel.

15:00So yeah,

15:01I hope you enjoyed it.

Category

Transcript

Be the first to comment

Recommended