About this talk
In this session, the speaker discusses the transformation of software development practices involving AI, specifically focusing on the reduction of deployment time from conceptualization to production. He emphasizes the importance of identifying the right tools, such as agents and assistants, to facilitate faster and more efficient software processes. The speaker shares insights on automating tasks while stressing that human creativity and governance remain essential aspects of development. He also highlights the challenges organizations face in aligning AI solutions with business objectives, including risk management and defining effective prompts for AI agents. The talk features a live demo showcasing how to set up an AI agent, integrating tools to improve workflow and maximize productivity. The speaker concludes with reflections on the ongoing evolution of AI in software development and its potential impact on future job roles.
Full transcript
Hello. Uh, [applause] I want to preface I'm not a professional by any means. Uh, but I am member of the Quich Yacht Club for those in here from Latvia. [sighs] Uh, long story about that. But today I'm here to talk keep a workshop without tapos without you having laptops and all usual jazz. So we're not going to be probably doing a workshop. We're going to do a
slow demo. So, if you want to, you can follow up with me. I will try to keep you up and like make you be able to do this at home if you want. I'm also recording uh because I usually look at my own presentations afterwards and I upload them to YouTube so you can find it later. If you want the recording, connect it to me in LinkedIn.
So I work as a field CTO in E code and for the last six months I've been helping companies uh transform their organizations software development life cycle around AI. So what does that mean? It's kind of looking like this. So this is the dream come true of any AI transformation in software development. Five minutes from idea to production. So we're kind of hoping that we could actually
deliver something in minutes to the customer. So when we get an idea, hey, we should test this with customer, we could run through the usual secure and controlled loop of software development and get something to production and learn about it. Uh if you're doing this today, you're probably wipe coding and you're probably following bad practices when you're doing it. Uh but you probably have done this. Uh
I have a phone application that works that way. Uh the ideology behind it is that theoretically it's possible to get here. Uh our goal is to convert this into reality as F code and try to get customers getting closer to five minutes. At some customers I would be happy if they would get into monthly releases for example. So if you worked on certain customers getting from quarterly
releases to monthly would already make life a lot easier. So this doesn't mean that we're actually removing work or that we're actually seeing decline on employee employment. So when we automate work, we usually invent more work. That's the beauty of working in IT or any other field in humankind's history. We're really good at inventing new things to do. How many of you have had this situation at
work that you became really good at your job and then someone came and gave you more work? Yeah, that's going to happen this time again because like based on our experience, uh the companies that we have helped end up hiring more people because they can now deliver more and create more revenue. Uh the big beauty of this is that you're not going to be automating your idea
generation. What's your product vision? What are we going to actually be doing? And you're not going to be defining your continuous improvement. How do you actually improve as an organization? You're not going to let AI decide on that. Or you can, you probably seen the YouTube shorts. Amazing shorts. Uh if you are, if you haven't, again, I can send them to you our ideology is that we're
seeing that closing in five different phases. And based on my experience, most of the organizations are about here. So going towards the phase one of having an assistant at work, I would hope to say that they are in the agent or multi- aent pipelines like it's pretty easy. You're going to see today that like getting an agent running is not the hard part. It's being able to
scale it, making it good, making it do the thing that you hope repeatedly, making topics scale, and then uh not getting hacked, which is surprisingly hard in AI. It's al all about like how do we actually start creating proper enterprise data structures. So how do we actually follow who owns the data, how do we integrate to the data and where are we going and basically the ideology
is you also need to start figuring out how do you actually change your people and culture. So when we start creating agents we end up thinking about okay we can make an agent proof of concept very quickly. I'm going to be showing that. But how do we actually then scale it across the organization? How do we measure the impact? That's a very big challenge in many organizations.
Executives love return on investment numbers. They love figuring out what did we save? What's the money impact on this? And you as an employee are like, why would that matter? It makes me it it made me faster. Again, they want to know how much more work they can invent. they want to make a predictable thing especially in uh private equity or uh stock market. Then there is
the whole thing of how do we govern and accept risks. Uh sec security is actually a really funny thing. Security starts from what risks do we accept. In most organizations, when you talk to security, the security policies are actually decided risks that the company takes and which ones it doesn't. So, you're actually taking always a gamble when you turn on a computer because that might be hacked.
When you connect it to the internet, it might be hacked. Uh, it's a risk. But then you start figuring out what are the risks that you're willing to take. Well, you need a computer to be faster. Surprise. It's actually safer to use computer compared to the other risks that doing paperwork does. It's same thing like again thinking about it from the point of view of AI. Okay,
kept working. So the first thing that we need to do when we start creating an agent is to find the problem. You can't imagine how many organizations have a problem on finding the problem. They have a lot of ideas about agents. They have a lot of ideas about we need to solve like speed. We really need to be faster. And then you're like, okay, what do you
want to solve? I don't know. And you're like, okay. Uh, so have you heard of value stream mapping? And they're like, what's that? And you're like, we take a look at where does the requirement come from? And we define how do we get to production with it? And we put the boxes down on what happens in between us are us starting to do getting the requirement to
production. What happens? And that is already black magic to many of our customers today. Then you start defining where do we actually create this bottleneck. So where do we have a stop gap? Where do we start like okay time to go to service now and create a ticket so the change approval board can approve it and then we can be like going to the security and they
will order a pen test and six weeks later we can get the pen tester in place and he will click two buttons and he will say all good and we can release not this one but one other customer had this problem. Another customer had given uh the customer a promise that they the customer can test the software six weeks before they release. So that means every time
they want to release they have to wait six weeks. [clears throat] So try to do continuous deployment that way. uh when we start putting it in this way it becomes more clear where can we actually get the speed what are we talking about where are the challenges and then we can take these red boxes in this case you can do it this many different ways and you
can start figuring out for example okay we have a code review as a quality gate that's where we do the quality checking then you need to start figuring out what are we going to be talking about as a code review does it include our static analysis, does it include our required test coverage? Do we have automations on these? And then you start getting to philosophical discussions. Is
code review even uh current day actually useful if people are not actually reading the code. [sighs] All kinds of conversations in online. You probably hear a lot about those during two these three days. So then we have an idea. So what you can do is you can go into an assistant and you can go and ask it. So here's my Gemini. How many of you have used
Gemini Copilot Chat CPT? I assume everybody. So this is an assistant at the moment. Like you can go here and I can ask uh how do I uh get my myself from um RA to Helsinki West. I can ask it this kind of things like you can ask all kinds of things but it's going to start thinking and it has this thinking process behind it and it's
going to do all kinds of whoops and things and you can see how it's thinking this is all about the way and we will get you will probably hear about that also during these two days but this is the assistant so you will need to do something and it doesn't really have anything so if I would want to take these steps uh I would then have to
start figuring out like how do I actually do these things? So if I want to for example get the bus from Ria to uh the coastal town of Ventabuse, uh I would have to order the bus ticket myself. I would have to do everything manually because it doesn't know how to do anything in real life and that's why it's as an assistant. the next stage is kind
of understanding why your prompt doesn't work because that's all you do at that point. prompting. We're going to get into why this is important because this is kind of the only thing that you need to worry about when you build an agent on top of integrations, data, security, governance, and all the other things that you have on this. But you need to kind of start figuring out,
okay, how do I break down the task? So, what if I have a specific thing that I want to do? I should probably break down that I want to go to Estonia or I don't want to go to Estonia. In my previous example, I want to specify. So, I want to be very specific with the requirements that I have. I might not want to take a bus.
So, I would type I don't want to take the bus. Could I walk? Uh, you want to iterate. So, you want to try multitude of different styles. And you can also ask AI to make you a better prompt. So we we will get to that you can add examples. So most of the time when we start creating agents we start adding examples. So we start telling it
this is my code style. This is the usual thing that I'm doing. This is our usual file format that I'm getting things out. So when I'm for example generating reports to customer with AI, I start by giving it example reports of previous things so that it can follow that structure because I know it works. And sometimes you just want to nuke everything away. Just if you have
ever used uh assistant and you put the document there and then you add it give it an updated document. Most of you have probably had this experience where it just starts hallucinating between the two documents. So it's easier to just start a new chat and then it's you basically start it all over again. Uh if you're a software developer, you have a version control. So you can
just tell it get reset that's as hard all of the all of the things that it did gone. There is also in uh open code undo in um cloud code undo and view auth code version control but version control is still really easy on this. So this is how you prompt better and make the assistant work better. Okay fantastic. How does that relate to anything? It's the
input parser in the agent. So the agent takes an alm instruction and that alm instruction is your prompt. Basically that's what you control. That's what we control when we control an agent. We give it a baseline prompt that defines how does this thing work. The input depends. Uh, if you're smart, don't let the end users write the prompt if you can avoid it. That way, they can't
skip around and do stupid things. But most of the time when we go to agents, we want the end user to be able to kind of ask all kinds of things because we don't really understand what we're asking them to do. So, usually we end up in a chat icon, for example, which allows you to ask questions. And that means that now you have to start figuring
out what kind of instructions do you need to make so it answers correctly because like we can make all kinds of stupid instructions. I'm going to show you a few hopefully if I have enough and then this gets converted into an output at the end of the day. That's when it's an assistant. But when it's a to when it's an agent we give it tools and tools
are external factors. So if we go back to my example of having Ginai open. So in Gina you can actually see these tools nowadays. I could make it make a video, create music, or I can tell it that like uh critical in my Google calendar that prevents me from doing this. So now I'm giving it a tool. I gave it the permission to go to my Google
calendar and hopefully it doesn't leak anything important. Uh but ideology is that now it's reading my Google calendar. It's checking what am I actually doing. It's also possible not with Gemini but other tools that we're going to start seeing tools that we can be doing. So I have important meetings and I have a weekend cottage trip and there's also Mother's Day so I shouldn't probably be doing
this. Uh so yeah I think I'm going to not be able to do that trip travel. So tool is an external it might be internal way of communicating from alm to external word. So when you when the does web search for example that's when it's allowed to do tool calls. So that's how we start creating agents and that tells you that agents are really a lot of
different things. But the idea behind it is that when it can act on things that it finds it becomes an agent because then it can actually do something for you. So examples of agents and this is when I start explaining a little bit deeper on that. So I'm actually going to make it from scratch so you can see how it works. So this is this is just
one of the tools. I can do the same thing on basically any tool that I have. So here is uh I can also for example have another tool here. I could have for example open code which is kind of a crowd code if you know what that tool. So the ideology is that you can basically start doing all kinds of things with these tools. Uh but I'm
going to use an because it gives us a really nice UI where you can see the steps happening. We're going to get to that. You can see what I'm what's triggering things and what where do we end up. This is not the only way to do agents. This is just the logic building so that it's easy for you to follow hopefully. So the very first step is
you define how does it activate. So we could have an external way. Let's move those away. We could have a web web hook. Web hooks are those things that the services send when you start doing things. We could have it scheduled. It could be every day at certain time it prompts read my emails and it would do things like sending you on WhatsApp message with different things.
So I'm going to use on chat message because we're we I want to show you different things across this. Uh so this is chat chat interface now towards the agent. Then I want to give it an AI. We could do all kinds of other things here but for today's sake we're going to talk about AI. Uh I'm going to show you open AI because I know have
that panel open but I can also show you Microsoft Foundry where you have more options but the idea is that you can take an open AI and you have all these options you basically just need message a model because everything else is kind of there for you to be able to upload files doing different things. It's when you want to get deeper. Bunch of YouTube tutorials. Uh
then you need an OpenAI account. So what happens is that you go to this booking page where most of you have probably never been. This is the Open AI's own web page. You probably been in this page. This is CHBT. Uh because you you look for CHBT, not OpenAI. You go to the top right corner and you find API platform to get the API token. And then
you get the token by creating API token. And we should be able to see here on this dashboard my one call that I did earlier today to verify that my token works. And this is when we get into these weird things that I'm going to be talking about. Tokens. Tokens are things when you send words to every separate word and sometimes we're going to say in separate
words because that's the easiest way to understand it is going to cost one token and that's an input token output is going to cost you based on how many tokens come out of the system and when we do those whoops that you saw Gemini doing Ginai is sending your tokens back in and back in. So when the AI is thinking, so doing that thinking process, you're sending
more input tokens in every round. So that's when you start consuming more tokens and it becomes more critical when you have more data sources or integrations because then you're sending more input to the AI model because it needs to know what you're doing. And when you tell the AI what you're doing, you have to send it what are the integrations and data sources that you're giving it.
we're going to get there. You're going to see the number go up. Uh then we define uh that here. So we would create a new credentials. We would put the API key here and we would press save if someone's following along. Then we pick a model. U for today we're going to call chat latest because I don't really care. And then we have a prompt. What do
we want this chat to do? Uh we're going to open a new chat and we're going to have um I am building um what would we build? Re uh RA summer trip planner for 2026. Uh, and I need the system prompt for my And surprisingly, the agent will build you your first one. You can see the role that we talked about. Okay, what are we looking for?
It gives it a wipe. What are we looking for as a wipe? It gives us gives the seasonal topic. What are we looking for? Geographic location. Operational guidelines. So, things that I'm expecting to do. So like it should be walkable, it should be real time aware of things. So it needs to kind of think about topics and I can give it a constraint checklist and then it
gives the example that I was talking about. So if you don't remember my list, just go and ask AI. It will make it better than you do at the first time. Then you can start improving this across these things. So I'm going to take this for today because it doesn't matter if this is wrong. I'm not coming to RAT this summer for summer holiday. Uh I have
other plans. Okay. So now we have a model. So what happens if we tell it that I am ga and I am coming to recut his summer may uh for a weekend. what should I do? So, it starts sending it to the message model. So, you can see uh that I'm sending this chat input. You can see the JSON that it sends back from okay, this is
giving me a really bad responses, but that's fine. So, the big thing here is at the moment Oh, this might actually solve a problem that I usually have. I might have screwed myself over by putting the open AI model. Let's see if it answers call. I am uh I have screwed myself over. Ah, it doesn't remember. Great. So, AI mods, as you can see, it doesn't know
who I am. it just starts talking on to me where should I travel in Burba immediately after I sent the message this is because AI doesn't have a memory so when you're chatting with your gemini chat CPDS etc it doesn't actually remember anything about you before you give it a place or a memory in the back end Google Microsoft etc have given us a memory already so
in here we would have to give it a memory if we want it to remember So we would give it a simple vector database for example or we would give something else so it keeps remembering us because if we don't do that it forgets who we are. So if we go yeah let's save that. So if we get go here to my other example this one for
example you can see that we have added a chat model. So this is ashure chat model. We have an AI agent. It's a customer support manager. You can see it's very much simpler. And we have given it a simple memory. So the ideology is that we're allowing the chat to be remembered by giving it a So if I tell it now I am caller, we can see
that it goes to simple memory. It creates an ashure prompt. So again sends it to cloud. In this case, it could be also a local alarm if you want to run such one. And it starts writing topics. So now it's trying to figure out what should it even answer if I'm sending I'm call uh this is total useless uh token spend but we're going to do that.
Nice to meet you caller. It started telling me about things. Okay. Uh I don't care about that prompt right now. If we remove this now I say say who am I? It's going to do the same prompting thing here. So we can see as your open chat model. It should answer call because it knows nice to meet you call. It's not going to remember who I am.
Hopefully if it does then I'm uh my demo is broken and I am really screwed. But yeah, I don't have access to your personal details so I can't identify who you are. So let's give it the memory pack. We don't want it to be stupid. So that's the kind of ideology behind memory becomes important later. Then we can give it tools. So as we remember AI becomes
a thing when you give it tools. Without tools it doesn't it can't do anything. So we could give it the HTTP request tool. Uh I have an MCP client here because uh I usually do MCP demos but you can have different kinds of tools and you can write these tools yourself when you need to. Uh we're going to add this and it's going to be we're going
to just give it the Google URL and then it can fetch from Google. Without this it can't fetch from Google because again it needs the tool to be able to do something. So when you're defining agents, you need to define the tools that it can access. So if we look at it from source code point of view, you end up having to figure out what are the
tools that you use. I'm not showing you source code today because I only have 45 minutes. So I'm trying to keep it simple and also there is almost 40 of you. So that's So that's a singular agent. So that was one agent. not that complex like you probably made similar things done those kind of things like how many of you have tried to use chatbt or gemini
or comparable for web searching yeah not that complex what you want to then do when you start developing these agents is you need to start figuring out okay what tool do I use should I use nhn gi should I use co-pilot should I use something else Should I use cloud code, open code? Which one is giving me the tokens? That's not today's topic. Uh I might talk
about it a bit tomorrow. You start personalizing it. So you use those custom instructions to make it more familiar to your staff. You start giving it context about the topic. You start writing effective prompts and giving it context. So same as what what we got out of Gemini, you start indexing your workspace. So you start giving it access to your data. How many of you have a
company data sources available on some AI tool? About half the room. So that's when you start being able to for example go to Confluence or Yera and use the robot search for example or you might have a co-pilot for business from Microsoft and you can use those things to search Then we get to the hard part or easy part about AI. You need to start choosing your
AI models. Uh why does this matter is because this defines how smart they are but also how much the AI is going to cost. So we're going to start focus like the more further we go the more these token counts and those kind of topics start arriving and we need to start figuring out how do you actually save money. This is one of the best ways of
doing it. The other side is choosing the right tool. The third one is not giving it too much context or data. That gets really difficult and requires you to understand the business case and you really want to reuse prompts because you don't want to invent these things every time. You want to use that agent or that model call multitude of times. So I mentioned context. So why
do you want to control the context? It decreases the ambuness. So if you use for example in uh AI tool for example in open code we can see how much context we're spending it's 0 to 100%. And after you go over about 65% so I usually say 70% so it's easier to remember the AI becomes hallucinating and the hallucinations come from the fact that it ran out
of space. So it started to add stuff by rewriting things or pushing them down which means that you're starting to create houseations of the data. So if you give your AI a small model and you give it 600 pages of documentation, it's not going to be able to do it. So for example, if I give uh run my local model and give it an MCP, it would
just crash because it doesn't have enough memory to process that. When you use context correctly, you get better results. You have less interaction because it understands what you're trying to do and you have a better experience overall. And it starts by new chats per task. So don't try to use that one window for everything. You probably noticed that you want to give only the relevant information. You
don't need to tell it your mom's birthday and the day like her all daughter related informations, especially her social security number. Uh for it to write you a message for Mother's Day on Sunday. Uh you could also send it yourself. Uh you want to exclude irrelevant information. Again, don't give it too much. Don't give it wrong info. Give it the rows. We were looking at the row
earlier. So tell it who is it? What kind of person are we expecting? Clarify the and clear. So this is when for example there is research that Finnish language is one of the best languages to prompt. Not because it's uh like anything weird. It's just very unambiguous. We're very direct. So it's very direct language at that point. Same in Japanese for example if I remember correctly when
the language is not ambitious it's really easy to get better Finnish is however really bad because of other stuff but ideologically you can be very clear on your prompt. Don't try to be overly nice to the AI because if you're not clear with the work task it's going to do something stupid. and use examples as said before I mentioned MCPS and I wanted to cover them before
the next demo. So it's a way of integrating and I want to point out that there is two camps or actually five camps but uh for simplicity there is two camps there is skills uh client-based integration side and MCP server side and you just need to understand that there is a way of doing standard integrations for AI and you can use few different ways MCPS is one
of them and the other one is skills.md and skills.md requires you to have stuff on your laptop. MCP servers can be central hosted. That's the gist of this. But the ideology is that you can connect your data source to an external thing. Let's say GitHub for example, and it can do topics in GitHub for you based on the tools that you allow. So how does this integration
work? You set you have this protocol. Okay, there you have this protocol usually in your IDE and you send the request to this external server where you have different kinds of possible end bases. It sends those back and puts them into your prompt allowing you to send data over. So client server architectures u we go very deep if we go too deep here but like I'm trying
to be concise. So why do we need that kind of thinking before we go here the next thing you think about well that agent was super simple what do I do next you start doing multi- aent so you will have one agent and other agents below it or you have one agent that goes the next agent that goes the next agent. So the thing is that we
start getting into the point where we want not just one agent to be calling and answering us. We want to have different tasks and different things that it can do. So going back to our demo, we can see that like I have another demo. We don't save it. Um I have this main agent and in here I have add added three different tools for my agent. You
can do it in many different ways in this tool also but idea behind it is that now if I open my autoen agent it tells me that it's going fetcher and it's giving it a prompt like what I'm sending to my main agent. It tells it that it's an atas an expert Wait. So I'm calling an agent from an agent. So if we go back here and
open that fetcher from the agents here. So I have this agent that is allowed by e execution by another workflow. I have set up here an agent that can call the atosen mcp and has a profile of doing being an atosen expert. So this one knows how to get stuff from my Jira. So now if I want to ask about our what's going on in root help
in Jira it's going to start calling that as an and we can see that it sent there a query and if we go inside it we can see that it sent a request forward And we can see what kind of variables we are sending. So this is why we're in the UI and not in the code because this would be hard for you to show to you
guys without that. So we are doing queries against the agent and hopefully my token hasn't gotten old. But at the same time this doesn't show anything because it's happening in here. So we can see that the agent gets an execution. It starts sending requests. It's calling that AI model with different kind of prompts and it's sending those towards the MCP server. It's doing it multitude of times
and it's failing. Demo effect. Fantastic. ah and then it actually answered to me. So it actually gave me the it tells me that it's an error and then it gave me answers. success. Yay. Uh I wasn't expecting that. So ideology being that like I can really easily add this kind of thinking here. So we can see that hey access was to a gitlab repostory root help 105.
But wait a second I have a GitLab agent here also. So if I was a services worker, I could ask it, hey, I have a JRA ticket root help 105. I would like check that access request in GitHub. Can you help me? Let's see if it works. I didn't try this earlier, but at the moment it's going to go and try to figure out, hey, what happened?
And it's going to call the GitLab MCP if all goes well as it did. And it started calling GitLab. And it noticed that my token is old. So I didn't actually put my token up to date token in place. So it started telling like hey there is an existing and there is an in process and a certain person is in charge of that. So you can create
these kind of workflows and we could trigger this whole workflow from the UI itself because we were looking at those workflows earlier. So we wouldn't have to run it in this architecture that I was showing. The workflow is just us adding them instead of below this one agent next to it. Uh I sadly don't have time to demo it today. But just wanted to leave you with
the thought that like when you start building these agents making it isn't really that hard as you could see getting them to work properly and orchestrating them and running them in an enterprise scale becomes the hard part. So you need to start thinking about all kinds of different things on your IT side. So you need to start thinking how do we do permission management like how do
I do that kind of things and it's your IT thinking about all these things and you probably need to figure it out or you probably don't need to figure it out but they probably need to and I am now out of time at this point. So if you are interested about how you actually take these into use, we have a panel starting uh in another room uh
room A and also I have a talk tomorrow about how do you get started with AI. So feel free to join that to hear about bit more about all the other stuff related to this and hopefully you build a lot of agents at home now. Yes, [applause] thanks a lot. Um, so uh while we are waiting for the questions uh on slideo, I have one personal question
to you. I noticed that you had a uh that you had um an open code running in the terminal. So can you please slightly explain us what this open code is, why would you use it and uh how does it work? >> Oh that's a hard question. Not because like you should use it. Uh so cloud code is the current hypest thing that you have seen in
LinkedIn and you're like oh my god I need cloud code to do anything into AI. What happens is that it can call different tools automatically on your terminal. So when you run it on a laptop, it can do things for you automatically. So when you do things, it just works and you don't need to tell it everything. That's the black magic thing. What is open code is
uh you might have read about cloud code reading their source code, but this was the second time they leaked their source code. And when they did it the first time, the open source fanatics were like, wait, we can write this ourselves. Like what's so magical about this? It was about two years ago. So they started writing like getting that functionality into open code and then they kept
working on it and anybody who has isn't working in it. The open source experts are people who are very dedicated to making the thing that they use the best ever. It might be the worst ever for most other people. But if you work like them, it's really good because they will always do changes so that it works better for in their workflows. And open code happens to
be one of those. So when you look at the cloud code week, uh you can for example see that cloud code refers to open code features nowadays that hey open code is doing it this way. We should do it this way in the comments. So [clears throat] it's kind of like that open code is really fast evolving tool based on what the community and people are doing
and it's better than cloud code in my opinion but that's my opinion >> uh when like before we go deeper into the questions that people already asked and I I will abuse my privilege of asking questions and uh I will ask you about one thing and I know that within open code you can use different uh models. So one of them is for example DeepSseek V4. So
um the price of it is very very low. How is it in comparison in results? >> Um in our company deepseek is not allowed because of the Chinese uh training that so I don't know how it compares by personal experience but based on uh author feedback. So if if this get when I release this to YouTube I am not using deepse on my company machine. uh deepseek
is uh basically really good if you need basic things. So if you don't need a lot of thinking or complex AI, you could use deepseek. The challenge for example why we have forbidden it in our company is that you don't really know what data it's trained on or what it's been. So when you do an AI model, you can change the model weights. So for example, you
can make it say certain things or deny certain historical facts and that creates this thing that you no longer know what's the truth. You can you can say the same thing about the American models or the European models that they're all politically biased based on certain things and which is why the foundational models have this. It's a very long talk. >> Okay. Yeah. Yeah. Okay. So let's
let's go to the questions that like you know real people asked. Um agents versus assistants is geminy gem. Oh yes, gems. Who's many people use it? Actually, I use it myself. So, does it have a knowledge base of a document? Um, an agent or an assistant. >> So, based on the fact that it uh depends, but like based on the fact that it's just documents that you
gave it already, it's not going to be an agent. But if you give it a live source from Google Drive, for example, it becomes an agent. So, if it can't actually get the data outside and someone needs to go and update it, it's just an assistant because then you could just put that into a prompt and it's a pre-made prompt Um, okay. So, now let's go to
the last question because we have we're running out of time. So, let's uh probably focus on uh how many No, let's Isn't it your main tool in building agents? If not, please explain which one is. Uh my main tool is uh open code, but I tried to make this so that everybody can follow because the room got full and I tried to stay in very like simple
stacks so that everybody could like follow. Uh, open code is very very programmer friendly but it's not really friendly when I have almost a full room >> and it can work in your terminal and it can work in your on your remote VPS and everything all those crazy and you can write it on telegram right >> yeah it can do all kinds of things I don't recommend
you adding it right on your telegram uh I think I can answer why didn't I talk about context window uh I have 45 minutes I was already going over time if you didn't notice. Uh how many tokens did I spend? Uh that was about 100,000 tokens on the Azure models, but uh I only spent 10 on the OpenAI model because I have more in the Astro side
than I have on the OpenAI side.
