#2 - Giulia Solinas - Exploring Agents in AI

WEBVTT
00:00:00.000 --> 00:00:12.080
Music.
00:00:12.219 --> 00:00:17.959
2D Time for AI podcast, the podcast that brings you the best AI conference speakers
00:00:17.959 --> 00:00:21.099
and allows them five minutes to summarize their presentations.
00:00:21.379 --> 00:00:24.959
My name is Peter, and I'll be your host today as I've been in the previous episodes.
00:00:25.199 --> 00:00:29.179
If you want to know more about AI, you can follow me on TikTok, on LinkedIn.
00:00:29.819 --> 00:00:33.699
Before I go to interviewing Julia, can you do me a favor?
00:00:33.959 --> 00:00:39.279
Can you go and send this episode to a coworker or a friend that you know that
00:00:39.279 --> 00:00:42.559
is interested into AI? That will help a lot. Thank you.
00:00:43.239 --> 00:00:48.499
And now going to my today's guest, Julia Salinas. Hello and welcome to the podcast.
00:00:48.979 --> 00:00:52.359
Thank you very much, Peter. I'm very excited to be here.
00:00:52.759 --> 00:00:55.559
Thank you a lot. Thank you. Thank you for being on the podcast.
00:00:55.759 --> 00:01:00.919
Julia, you're located in Munich, Germany. Do you have snow there?
00:01:01.239 --> 00:01:03.719
Not yet. We had it already, though.
00:01:04.279 --> 00:01:09.159
So a tiny snowman was already growing in our garden a couple of weeks ago.
00:01:09.879 --> 00:01:14.159
Did you go out and try any glue wine on the Christmas markets already,
00:01:14.179 --> 00:01:18.319
or is that still planned for you? Already checked and enjoy that.
00:01:18.739 --> 00:01:20.699
So I can visit us next time.
00:01:21.579 --> 00:01:27.899
Excellent. Munich is one of the cities that, in my mind, is big on Christmas markets.
00:01:28.279 --> 00:01:31.559
So I still have to go and visit the Christmas markets in Munich.
00:01:31.699 --> 00:01:39.679
I haven't done that. Julia, you are the AI and data science consultant at Kroos, a creation company.
00:01:39.879 --> 00:01:42.719
What does the company do and what do you do there?
00:01:43.059 --> 00:01:50.179
Kroos is B-STEC consulting, meaning that we provide a number of services related
00:01:50.179 --> 00:01:52.659
to both business and technology.
00:01:52.799 --> 00:01:56.179
Because these two things actually need to go hand in hand together.
00:01:56.179 --> 00:02:03.319
When it comes to tech, Crows started as an engineering solution provider,
00:02:03.319 --> 00:02:08.719
and then it slightly moved into its services into operation and then combined
00:02:08.719 --> 00:02:11.779
with the strategy type of services.
00:02:11.779 --> 00:02:17.399
Because actually they observe that it's an all-around package what companies usually need.
00:02:17.579 --> 00:02:22.059
It's not just about providing a single solution, but it's a solution that fits
00:02:22.059 --> 00:02:27.859
with the people who use the technology and the organization that then is going
00:02:27.859 --> 00:02:29.939
to use that solution itself.
00:02:30.239 --> 00:02:32.799
So that's, I would say, is the...
00:02:33.374 --> 00:02:38.494
Description of Cross, what I do at Cross, I'm a data scientist,
00:02:39.034 --> 00:02:46.594
and my daily tasks relate to programming solutions that somehow enable LLMs
00:02:46.594 --> 00:02:49.034
to work with business requests.
00:02:49.334 --> 00:02:56.274
That's what actually there is, the increase in demand, and try to make it effective
00:02:56.274 --> 00:03:02.774
and productive. In the past, you moved from the educational part to the business part.
00:03:03.054 --> 00:03:04.954
Can you compare the work in those two?
00:03:06.214 --> 00:03:09.834
I'm not sure if I can ask which one is better, but what are the differences
00:03:09.834 --> 00:03:11.854
and what did you enjoy in both of them?
00:03:12.374 --> 00:03:15.914
Well, academia is a bit different than business. In academia,
00:03:16.094 --> 00:03:20.794
the business is publishing papers and trying to push research.
00:03:21.474 --> 00:03:29.474
That is, I would say, my old hat. and then providing a good teaching contents to students.
00:03:29.854 --> 00:03:35.234
There are many things, however, that from academia, I was able to bring into consulting.
00:03:35.454 --> 00:03:40.974
First, well, the capability to, I would say, try to explain concepts like scientific
00:03:40.974 --> 00:03:43.754
concepts to business context.
00:03:44.474 --> 00:03:49.954
And that is quite important when it comes to LLM, Gen AI, because there is a
00:03:49.954 --> 00:03:51.474
lot of research underneath.
00:03:51.474 --> 00:03:57.314
And this is content that should be understood by the head of HR,
00:03:57.614 --> 00:04:00.794
by someone that is down the line in production.
00:04:01.134 --> 00:04:05.334
And I would say this is like the bridge that is important to maintain between
00:04:05.334 --> 00:04:09.794
academia and business, try to explain those concepts to entrepreneurs.
00:04:10.248 --> 00:04:16.688
People from every background. You have a LinkedIn newsletter that is called AI a la carte.
00:04:16.948 --> 00:04:20.108
As an Italian, you like good food a la carte.
00:04:20.408 --> 00:04:24.708
As an AI expert, you like good data a la carte.
00:04:25.068 --> 00:04:29.508
What are the things that you post in that newsletter and why should people subscribe?
00:04:29.728 --> 00:04:34.508
AI a la carte is like my little creature within cross projects.
00:04:34.888 --> 00:04:41.508
And as you well spotted, there is a bit of Italian food culture in it.
00:04:41.688 --> 00:04:47.548
And the idea would be promoting different types of pragmatic AI.
00:04:47.768 --> 00:04:52.788
And with pragmatic AI, I mean, all the workflow that goes from the ideation
00:04:52.788 --> 00:04:57.908
of an AI project to the launch in production into small bites,
00:04:58.228 --> 00:05:00.708
like small pieces of food that you.
00:05:06.768 --> 00:05:10.768
AI LaCarte is really starting from scratch of an AI project,
00:05:10.988 --> 00:05:16.908
try to have a bullet list of the points that everyone should have in mind when
00:05:16.908 --> 00:05:23.968
starting a project related to predictions models or large language models to
00:05:23.968 --> 00:05:25.528
more sophisticated one,
00:05:25.748 --> 00:05:31.348
like how do we need to do when we want to fine tune a model or what do we need
00:05:31.348 --> 00:05:35.688
to check if we want to comply with AI governance?
00:05:35.948 --> 00:05:39.908
So I want to bring it up all these different topics and make it.
00:05:41.035 --> 00:05:44.575
Up for good and enjoyable readings.
00:05:45.135 --> 00:05:49.915
And they are. I'll put the link in the description of the podcast so people can go and subscribe.
00:05:50.535 --> 00:05:55.495
Julia, I've invited you to the podcast because you were a speaker at the Weekend
00:05:55.495 --> 00:06:01.855
2024 Media Festival or the part of the Rovin conference that was called AI Weekend
00:06:01.855 --> 00:06:03.535
with your presentation,
00:06:03.755 --> 00:06:05.975
Unlocking the Power of AI Assistant.
00:06:05.975 --> 00:06:11.055
This was the first version of the AI Weekend Conference, I think.
00:06:11.255 --> 00:06:14.875
How did you like the conference and how did you like Rovin? Have you been to Rovin before?
00:06:15.475 --> 00:06:20.195
Surprisingly, being in Italian, it was my first time to Rovin.
00:06:20.455 --> 00:06:27.795
And I found it lovely. It was like being back at home for me and really enjoying the beautiful scene.
00:06:28.015 --> 00:06:33.455
I was very much impressed by the lineup of the conference and by the very thrilling
00:06:33.455 --> 00:06:36.955
conversations happening backstage. stage after the presentations.
00:06:37.315 --> 00:06:38.795
So I had a great time there.
00:06:39.295 --> 00:06:43.195
Yeah, I've just checked the website of the conference.
00:06:43.375 --> 00:06:46.535
The 2025 has already been announced, at least the date.
00:06:46.715 --> 00:06:53.855
So I was there too. I think it was really great. I was also surprised with the
00:06:53.855 --> 00:06:58.195
quality and I think it was really good. So people should go and check it out for next year.
00:06:58.495 --> 00:07:03.035
Here is the time for you, Julia, to present us a summation of the presentation
00:07:03.035 --> 00:07:05.495
that you have. Julia, here are your five minutes.
00:07:06.675 --> 00:07:10.635
So, in a nutshell, what I presented was related to,
00:07:11.301 --> 00:07:19.201
putting AI assistants into production and make it usable for users in organizations.
00:07:19.201 --> 00:07:24.901
This is, I would say, a very topical issue. If you consider that 97 of business
00:07:24.901 --> 00:07:29.321
owners believe that ChatGPT will help their business in some way,
00:07:29.481 --> 00:07:34.621
or that actually co-pilot users can increase their productivity by,
00:07:34.621 --> 00:07:37.621
I would say, something around 30 to 40%.
00:07:37.621 --> 00:07:42.701
The thing is that there is gen AI for everyone in the organization,
00:07:42.701 --> 00:07:48.241
and it's like the type of AI that can go with a pilot subscription.
00:07:48.541 --> 00:07:55.901
And then there is the AI supported by the assistants that instead comes with customized solution.
00:07:56.181 --> 00:08:00.601
And this was about this thing that I was talking about in Rovin.
00:08:00.721 --> 00:08:08.721
Why is it that's so important? Because actually, when we think about a doc solution in business,
00:08:09.021 --> 00:08:14.381
we have a couple of things to double check before signing a subscription with
00:08:14.381 --> 00:08:19.101
whatever provider of large language model would support this solution.
00:08:19.281 --> 00:08:23.621
First thing, who are the users who are actually benefiting from the assistance?
00:08:23.961 --> 00:08:29.081
Are they internal? Which kind of limitations would we need to check in order
00:08:29.081 --> 00:08:33.441
to make the assistance usable for those users? Second point,
00:08:33.621 --> 00:08:34.961
actually very, very relevant.
00:08:35.321 --> 00:08:41.801
What is the infrastructure in which the AI assistant will nest it into? It's this legacy.
00:08:42.041 --> 00:08:47.581
It's not legacy. Should we use any compatibility software in order to integrate
00:08:47.581 --> 00:08:52.861
the assistant with the current infrastructure? A crucial point then refers to scalability.
00:08:53.081 --> 00:08:59.041
Usually when we develop a solution for an organization, we start with one single
00:08:59.041 --> 00:09:02.701
department, one single use case, but it's just the start.
00:09:02.921 --> 00:09:06.521
We want to make sure that instead this solution will be applicable.
00:09:06.935 --> 00:09:12.115
To many other use cases, not just in that department, like the HR department,
00:09:12.335 --> 00:09:17.015
but they will scale up also in other departments. What are the issues that we
00:09:17.015 --> 00:09:21.415
face when we have this scalability goal to achieve?
00:09:21.735 --> 00:09:25.615
And the last but not the least, governance and compliance.
00:09:25.655 --> 00:09:31.475
We all know we are, at least from our side, we are in the European Union.
00:09:31.475 --> 00:09:38.835
We need to fit with the GDPR compliance very soon, also with the EU AI Act.
00:09:38.975 --> 00:09:44.195
What do we need to take care of in order to have our data, the data that actually
00:09:44.195 --> 00:09:48.555
this assistant will use, compliant according to GDPR?
00:09:48.715 --> 00:09:53.955
Do we need to anonymize them? Do we need to mask them? Where do these data leave?
00:09:54.175 --> 00:09:59.055
Will they stay in Europe? Can they run on the cloud? All these issues then are,
00:09:59.355 --> 00:10:03.155
I would say, the ingredients, since we talk about AI as a food,
00:10:03.475 --> 00:10:08.975
are the ingredients that we need to consider when we develop an AI assistant.
00:10:09.435 --> 00:10:15.635
And when I was actually talking in Rovigny, I was bringing up one use case that
00:10:15.635 --> 00:10:24.135
we developed for a very large bank. The assistant actually was supposed to support the HR department.
00:10:24.435 --> 00:10:30.695
And the typical case is that the HR person receives a request from an employee.
00:10:31.295 --> 00:10:38.295
Please help me in finding out about how many days of holiday I can take or if
00:10:38.295 --> 00:10:41.095
I'm entitled of that mobile phone.
00:10:41.235 --> 00:10:46.995
And then the HR person spends lots of time looking into the documentation,
00:10:46.995 --> 00:10:52.275
the profile of the person, and then tries to formulate a suitable answer,
00:10:52.475 --> 00:10:58.975
why not having a system that helps in saving time in scanning all through this
00:10:58.975 --> 00:11:04.615
documentation and then supports the HR person with the suitable answer?
00:11:04.835 --> 00:11:10.435
So for that specific use case, we had one challenge that refers to one critical
00:11:10.435 --> 00:11:11.775
issue that I mentioned before.
00:11:12.095 --> 00:11:17.675
The HR person needs, so has a lot of documents to screen on and needs to make
00:11:17.675 --> 00:11:22.755
sure that the specific answers relate to the specific employee.
00:11:22.915 --> 00:11:25.295
If we would ask the LLM to.
00:11:25.696 --> 00:11:31.216
Just go through all the documentation, maybe this LLM would actually pick up
00:11:31.216 --> 00:11:35.896
the information that relates not to the relevant employee, but to the other one.
00:11:36.116 --> 00:11:40.896
And that's why it's very important to customize the solution,
00:11:41.156 --> 00:11:47.716
for example, for putting filters or rule base that actually direct the LLM in
00:11:47.716 --> 00:11:52.396
discriminating the right information and then provide the suitable information
00:11:52.396 --> 00:11:59.236
for the HR without compromising the privacy of the employees and still providing all,
00:11:59.436 --> 00:12:04.176
for example, the relevant links to the relevant documentation so that actually
00:12:04.176 --> 00:12:12.236
the HR person can cross-check that the answer provided by the LLM somehow fits
00:12:12.236 --> 00:12:14.256
with the background information.
00:12:14.876 --> 00:12:19.916
What is the benefit of this? Of course, save time for the HR person so that
00:12:19.916 --> 00:12:24.936
actually this HR person can spend more time in frontline situations with the employee.
00:12:26.236 --> 00:12:31.796
Plus, security and transparency about the data of these employees so that actually
00:12:31.796 --> 00:12:38.176
we can comply, as we said, with the regulation and the privacy. All right.
00:12:38.656 --> 00:12:42.976
I think that, you know, AI agents were probably one of the most covered topics
00:12:42.976 --> 00:12:45.256
that we heard about in Ruin.
00:12:45.536 --> 00:12:52.756
And I really enjoyed all the presentations. But what I think I can see with
00:12:52.756 --> 00:12:58.676
companies is that they don't really understand the idea of agents as such.
00:12:58.676 --> 00:13:03.056
Can you give a great example of an HR agent?
00:13:03.276 --> 00:13:08.416
Can you give a couple of more examples of what would be practically usable agents
00:13:08.416 --> 00:13:14.156
for companies so that people can get an idea of what exactly an agent for them could be?
00:13:14.696 --> 00:13:24.056
That's a great question. And actually, it opens the doors of what agentic AI really is.
00:13:24.256 --> 00:13:32.056
So with agents, usually we have, I would say, two components that relate to that concept.
00:13:32.056 --> 00:13:38.376
The first one is an interface that actually, given the answer from a human,
00:13:38.556 --> 00:13:42.776
is able to answer in a way that the human can understand.
00:13:42.776 --> 00:13:48.996
An interface that speaks the natural language of a human. Is it enough to have an agent?
00:13:49.616 --> 00:13:57.776
Actually, no. We need, as background, also the intelligent part of this agent, like the brain.
00:13:58.476 --> 00:14:03.696
And this brain can be, I would say, more or less intelligent depending on the
00:14:03.696 --> 00:14:08.016
task that this agent needs to do. Let me give you an example.
00:14:08.356 --> 00:14:15.576
So, for example, you might have an agent that needs to give answer just about
00:14:15.576 --> 00:14:20.496
orders or scheduling for logistics.
00:14:20.816 --> 00:14:23.156
Do we need a great brain behind?
00:14:23.596 --> 00:14:26.856
Probably not. probably just a list of.
00:14:27.452 --> 00:14:31.012
Let's say, rules behind the scheduling can be sufficient.
00:14:31.572 --> 00:14:37.992
We might have also another type of agent, an agent that needs to give product
00:14:37.992 --> 00:14:45.452
recommendation or needs to direct customers into different escalation solution.
00:14:45.692 --> 00:14:51.852
In this case, on top of, let's say, this rule-based, we might have the need
00:14:51.852 --> 00:14:59.192
of an LLM, a more sophisticated model that supports the agents in giving the
00:14:59.192 --> 00:15:00.532
right answer to the users.
00:15:00.752 --> 00:15:05.212
So I would say that the type of agents really depends on the type of question
00:15:05.212 --> 00:15:08.412
and task that the final users need to get.
00:15:08.692 --> 00:15:14.312
And probably if we look ahead, since we are at the closing days of 2024,
00:15:14.772 --> 00:15:23.032
2025 is promising E to be even more intense on these agents' traits of AI or
00:15:23.032 --> 00:15:25.032
agentic AI, as they said,
00:15:25.432 --> 00:15:31.192
in which actually the model that is behind, or we might even have multiple models behind,
00:15:31.612 --> 00:15:37.092
will chat with one another, integrate with one another in order to, for example.
00:15:37.912 --> 00:15:44.792
Give an answer, provide a first answer, but then also cross-checking that this answer is correct.
00:15:45.132 --> 00:15:50.612
Let's say having a judge LLM checking that the previous LLM is giving the right
00:15:50.612 --> 00:15:57.172
answer so that actually we are increasing the complexity of this brain underneath the agents.
00:15:57.632 --> 00:16:03.112
Is it easy to put all these agentic AI in production? My hint is no.
00:16:03.452 --> 00:16:08.692
And there I would say we should still have the human in the loop.
00:16:08.692 --> 00:16:15.272
And this human interloop will be even more and more relevant as long as the
00:16:15.272 --> 00:16:18.152
complexity of these systems will actually increase.
00:16:18.592 --> 00:16:22.132
I agree. 2035 will probably be agents in AI.
00:16:22.312 --> 00:16:25.672
If nothing else, all of the experts say that, so why wouldn't I agree with them?
00:16:25.672 --> 00:16:32.232
But on the other hand, we have seen recently Anthropik releasing their AI being
00:16:32.232 --> 00:16:40.172
able to record your computer and do stuff via the human interface, right?
00:16:40.372 --> 00:16:46.752
And Google has just released its new model, Gemini 2.0, to which you can also
00:16:46.752 --> 00:16:51.752
screen share your computer and it talks to you about what is happening on that. Yeah.
00:16:52.172 --> 00:16:58.572
How big is the possibility that we actually won't need to create agents,
00:16:58.572 --> 00:17:07.772
but we'll just have AI using the human user interface to do tasks because that's cheaper and faster?
00:17:08.812 --> 00:17:15.152
Not sure if it will be cheaper and faster for extended solutions.
00:17:15.152 --> 00:17:19.852
And here I'm thinking more about business contexts. Of course,
00:17:19.992 --> 00:17:26.432
if it's like a single user that tries to increase her productivity in some way,
00:17:27.047 --> 00:17:34.227
Maybe yes. But once we want really to scale these solutions in production, we need to consider,
00:17:34.727 --> 00:17:39.207
for example, how many calls to the LLMs should be done for this type of use,
00:17:39.347 --> 00:17:44.487
how many tokens will be used, which kind of infrastructure is necessary.
00:17:44.487 --> 00:17:47.507
And these are just things to start with.
00:17:47.667 --> 00:17:51.707
But these are things that then will go on the bill.
00:17:51.887 --> 00:17:58.027
And that's where, for example, organizations have to face increasing cost potentially.
00:17:58.547 --> 00:18:07.787
What I think instead would be a more rewarding approach is having multiple agents
00:18:07.787 --> 00:18:13.987
supported by small language models that will interact with one another.
00:18:13.987 --> 00:18:20.827
And this is actually the way in which we will embrace open source LLM,
00:18:21.007 --> 00:18:24.707
which are smaller, which can be fine-tuned,
00:18:25.167 --> 00:18:31.147
but can still, despite the cost of fine-tuning, might still be more cost-effective
00:18:31.147 --> 00:18:34.867
and be efficient in the answers that they provide.
00:18:34.867 --> 00:18:39.187
At the beginning, when we talked, you mentioned the transfer of knowledge from
00:18:39.187 --> 00:18:44.227
theories or from the academics to managers or to companies.
00:18:44.227 --> 00:18:50.567
I think that a lot of managers or people working in companies currently have
00:18:50.567 --> 00:18:53.107
a question about how much about
00:18:53.107 --> 00:18:58.007
AI or about large language models or what topics should I understand,
00:18:58.007 --> 00:19:05.507
how deep into the theory do I need to go to create or to assist projects in generating agents,
00:19:05.607 --> 00:19:07.527
for example. What do you feel?
00:19:07.647 --> 00:19:09.827
What is the best way to do it?
00:19:09.967 --> 00:19:14.547
Do I need a theoretical understanding of how large language models work?
00:19:14.667 --> 00:19:17.927
Or is it enough to just understand the big concepts?
00:19:18.207 --> 00:19:22.607
Or I don't need any of that and Kroos can come and do everything for me?
00:19:23.818 --> 00:19:31.578
I would say we would be very happy if CROSS can help in increasing the AI literacy, of course.
00:19:32.178 --> 00:19:38.958
Big smile here. Nonetheless, very honest. AI literacy should be a must that
00:19:38.958 --> 00:19:40.618
everyone should master.
00:19:40.918 --> 00:19:46.518
And with everyone, I would say top to down, down to top. Knowledge about AI
00:19:46.518 --> 00:19:52.638
in general, and with AI, I mean not just generative AI, should really spread
00:19:52.638 --> 00:19:55.798
out across the organization. That would be my recommendation.
00:19:56.158 --> 00:20:00.638
And here there are different ways in which organizations can enable themselves
00:20:00.638 --> 00:20:05.158
in getting, I would say, more knowledgeable about AI.
00:20:05.158 --> 00:20:10.278
There are trainings by governments, there are plenty of online resources,
00:20:10.278 --> 00:20:15.278
because these would actually have, in my view, a win-win solution.
00:20:15.698 --> 00:20:21.938
First, employees from all levels will get a better feeling of what does it mean
00:20:21.938 --> 00:20:29.178
working in the world of AI, so that actually we are increasing trust in AI from
00:20:29.178 --> 00:20:30.378
an internal point of view.
00:20:31.118 --> 00:20:37.118
And this will facilitate employees to be acquainted with AI resources.
00:20:37.118 --> 00:20:42.498
They will see the benefits and would not be scared to say, OK, AI will steal my job.
00:20:42.718 --> 00:20:46.878
No, someone that actually is trained on AI will steal your job if you are not
00:20:46.878 --> 00:20:48.658
keeping up to pace with that technology.
00:20:48.998 --> 00:20:55.578
And from a C-level perspective, having a better understanding about AI in general will make.
00:20:56.296 --> 00:21:00.196
C-level decision maker, savior, more
00:21:00.196 --> 00:21:05.776
experts and more knowledgeable about potential use cases so that actually we
00:21:05.776 --> 00:21:12.016
are not falling into the problem of listing tons of use cases that then will
00:21:12.016 --> 00:21:14.656
stick just to one Jupyter notebook
00:21:14.656 --> 00:21:19.856
that just stays in the drawer of a data scientist that will never be used.
00:21:20.076 --> 00:21:27.796
So we need that type of literacy to really get the best of this technology.
00:21:27.936 --> 00:21:30.756
This technology that, in my opinion, will stay with us.
00:21:30.916 --> 00:21:35.716
The last but not least, the more knowledgeable we will be on AI,
00:21:36.096 --> 00:21:40.956
the better we can also understand the ethical issues related to that.
00:21:41.156 --> 00:21:46.016
Something that from a data scientist's point of view maybe remains a bit hidden,
00:21:46.236 --> 00:21:52.856
but instead It's definitely a point to consider because it's about the ethical
00:21:52.856 --> 00:21:58.496
considerations of the users that then make use of the solution.
00:21:58.496 --> 00:22:05.136
So I would say that from an academic point of view, and having read some papers
00:22:05.136 --> 00:22:08.336
about the technical features, ethical considerations,
00:22:08.676 --> 00:22:13.796
we need that transfer of knowledge to go through all the levels of the organizations
00:22:13.796 --> 00:22:19.176
in order to really get a good understanding and a good feeling about AI.
00:22:19.176 --> 00:22:21.996
A last question on this podcast.
00:22:22.396 --> 00:22:27.576
What is the thing that you are currently playing with to have a pet project
00:22:27.576 --> 00:22:30.516
or a pet peewee on AI that you are working with?
00:22:30.696 --> 00:22:33.876
What is the thing that excites you currently in AI?
00:22:33.876 --> 00:22:40.976
Yeah, so I would say that my new patch project is my new GitHub repository,
00:22:40.976 --> 00:22:45.736
which is called personal training on Instruct Lab.
00:22:46.036 --> 00:22:53.596
Instruct Lab is a methodology developed together by Reddit and IBM to fine tune things.
00:22:54.028 --> 00:23:01.988
Small, large language model. And that brings me to the idea of having these open-sourced models.
00:23:02.628 --> 00:23:10.188
The smaller one, to make it clever, fine-tune them in order to have them efficient into production.
00:23:10.488 --> 00:23:14.308
So that's where I want to go in 2025.
00:23:14.468 --> 00:23:18.168
I opened my GitHub repository. I
00:23:18.168 --> 00:23:21.008
already looked into some trainings and that's what
00:23:21.008 --> 00:23:24.128
excited me yeah all right excellent can we
00:23:24.128 --> 00:23:26.928
link to that github is that public i would say
00:23:26.928 --> 00:23:30.488
that actually there is a great open source github
00:23:30.488 --> 00:23:33.288
repository by instruct lab so that's where
00:23:33.288 --> 00:23:39.088
everyone can learn so yes we'll send people there julia thank you very much
00:23:39.088 --> 00:23:42.748
for being on the podcast i think we got a lot of great ideas on what's happening
00:23:42.748 --> 00:23:46.648
currently and what's going to happen in the future we'll see what's going to
00:23:46.648 --> 00:23:49.708
happen in the future thank you very much for sharing with us your knowledge
00:23:49.708 --> 00:23:51.148
and thank you for being on the podcast,
00:23:52.008 --> 00:23:57.848
Petra, thank you so much for this conversation I enjoyed it a lot and it's a
00:23:57.848 --> 00:24:04.048
great podcast that you are editing thank you so much thank you and have a great 2025.