Introduction and Audience Check

So I just want to get a sense of the crowd to ensure that I deliver the messages in the best way. Who here consider himself or herself to have a technical background, so a software engineering background? Half of the room.

What I will do, is, strictly speaking, non-technical.

I have not written a single line of code in my life. I'm familiar with the tools, but I have not coded myself.

So what could be useful for you is to see how a non-technical person with tech savvy can use these tools, at least for my demo.

Gauging Tool Familiarity

Then, here, at this point, is a provocative question. Who has used chargpt? OK.

Who has used Cloud? Okay.

Who has linked Cloud to Gmail and Fireflies and Notion and other tools through an MCP? Okay. Half of the room.

And who here has used Super Whisper? Interesting. Okay. Okay.

I'm not sure you account for this. And who has used other forms of other dictation apps and integrate them in AI workflows? Okay, a few people.

And who has heard or used Gamma? OK, so this just gives me an idea of what different people might like and allows me to tailor the demo a little bit.

Turning Claude into a Personal Assistant with MCP

So what I thought I would show you is how you can transform Cloud into your personal assistant. by bringing together information from different platforms through this protocol that is called MCP, which basically allows LLMs to source information from different databases and platforms. You should think about it as an API.

So I will show you that we will create a project on Cloud that replicates this assistant. And then if we have enough time, we will try to write a proposal using Gamma.

Voice-Driven Workflows with Super Whisper

And you will see that throughout this demo, I will not be typing. I will be using an app that is called Super Whisper that will allow me to dictate what I want to say in any field where I put the space bar.

And this is really useful, not just because it saves you time, because you typically speak faster than you type, but also it changes the way you interact with these tools. It makes them really more and more conversational. You tend to reveal different information when you speak than when you type.

Is this clear? Any questions at this stage?

Why are you not using 4.5? Because I still haven't started. It's a good question.

Switching Models and Kicking Off the Demo

Now, we switch to 4.5. We are ready to start.

Can you look, actually...

Querying Gmail and Calendar for Tomorrow’s Schedule

Imagine you are an experienced personal assistant. Can you please look through my Gmail and give me an overview of my meetings for tomorrow?

So you see the connectors here. And what the model will do is to go through my Gmail and give me an overview of my day for tomorrow.

It's going through, it picked up the right date. It's going through my calendar. Now it's giving me a list of my meetings.

Great.

Why Centralizing Context Matters

Now you might ask, why do I care about this? Well, because once I have the information in Cloud without having to copy-paste it, I can use it for a variety of things, including, for example, ask the model to prepare me for all of these meetings, bringing in information from my note taker.

So you see here that I've connected my Notion, my Fireflies, so this really starts becoming my operating system. The things that I need to work and to be productive, I'll all start being in here.

This is great.

Generating Briefing Notes for External Meetings

Can you write a short briefing note for my external meetings?

So with the SN FCC and with Bluebeam. well I'm not a native speaker you can see so I just need to go through the typing

And so now what the model should be doing is to go through my notion, to go through my fireflies, and to pick and to look on the web and to really come up with a series of bullet points and valuable information that allows me to prepare for these meetings in the best possible way. So it's going through my email.

Great. It's really giving me the context that I need to prepare these meetings in the best way.

Have you any questions at this point? Yes.

That's a very good question. The reason is that here I can also do external search. It's a bit more flexible compared to Notion.

I can do external search. I can take this information and use it on cloud projects. I can link it to the notetaker.

I think cloud is a bit more flexible to do this right now. Might be that as Notion continues to evolve, two months from now, I will switch to Notion.

It's a very good question. Thank you for raising it.

More questions on this?

Automating with Claude Projects

Now what I want to show you is, have you experimented with projects? On cloud, who has used projects? Okay, some people.

Now what I can do is I can ask the model to summarize this in a project so that I can just basically run it in the evening or in the morning to prepare for my day. This was great.

Drafting Reusable Project Instructions

Can you draft the instructions for a project for me that are able to replicate these two things pretty much every morning or let's say every evening? And then I actually want to change it a little bit.

I should tell the assistant, please prepare me for the day. And it should go through my meetings and through my calendar, and then prepare the external meetings as we just did.

OK, so here what I'm doing is, in a sense, I'm maximizing LLM to LLM communication. So I'm telling Claude to draft me project instructions that I can then use on Claude itself. And the reason is it's quicker, but also because the LLM would know best how to be talking to an LLM.

This is pretty good. What I can do then is I can take all of this, pretty much as it is, and normally I would obviously spend some time checking this.

Then I go on projects. Personal assistant, darling. I put my instructions here.

And here I can be a lot more creative and thorough than I'm being right now. On the files I could put some instructions on how I like things to be done, some background on the company. We are just keeping it simple for now.

Prepare me for my day tomorrow.

Relying on Instructions Over Perfect Prompts

And you see that now I'm in a situation where I can deliberately prompt very badly. This is generally speaking would be a bad prompt because the model has no context. I'm not being really specific. I'm being clear, but I'm not giving enough context to the model, but the instructions really allow the LLM to give me what I'm looking for.

And for the sort of things that you do every day or close to every day, this sort of thinking can be really useful.

All right. This seems to be working. Great.

It identified the meetings. It's looking through messages. And it's bringing together the information.

As we wait.

Any questions? Yes?

Q&A: MCP Servers, Context, and Trade-offs

How many MCP servers did you have at the maximum? Because the more you have, the higher the context is used. Yes, there is a trade-off there.

Because the more things I'm bringing in, the more power I'm using to look through all of those things. Or the more clear I need to be about where the model needs to look. And so right now I'm keeping it very light.

I've got Fireflies, I've got Notion, and the Gmail Google Calendar system.

This is now the sort of things that starts becoming interesting and useful because I don't really need to chat. Claude starts really being my personal assistant where I can basically, I could basically say, can you give me the daily briefing, please? And Claude is doing that for me.

We've got an overview of the day. some material for each meeting?

Yes? Did you have any observations? Let's check.

Q&A: Hallucinations and Context Management

So there are two things that drive that, and you probably know. The first would be the length of the chart. the length of the chat, the longer the chat is, the higher the likelihood of hallucinations.

And that's because the model does not have a lot of context, or it loses the context, and so it's more likely to hallucinate. And then the second thing, the second reason why that happens is that the information you have given to the model is not comparable to the context that you have given. So these tools are ultimate people pleasers.

They will try to give you an answer no matter what. If they don't have enough information or, if they don't have enough context, enough information to give you an answer, they will make it up.

So I cannot comment on these specific cases, on every specific use case, but I think these are the two things that would drive the probability of hallucinations.

Was this useful? The length of? I think let's catch up on that separately. I don't want to make this too technical.

I'm really happy to go through it later during the pizzas. Any more questions on this? or any questions in general

Adoption, Skills, and Organizational Considerations

requires some technical expertise to set these things up. For a non-technical vehicle, it's really hard.

I would like to use Notion as well, and I come back to it when I give up, because it's so complicated, so many things to do there. And then I'm like, OK, maybe I will try it again. So even Notion is a good tool.

How do you see this approach?

Non-Technical Builders and Team Models

I think, so as I said at the beginning, I'm as non-technical as it gets. And I think these systems have become sufficiently user-friendly so that a person like me, doesn't have any coding background, can build something useful with them.

I think what happens within an organization would depend on company size. Because in a smaller organization like Mindstone, we would expect people to develop these tools. for themselves and to develop the tools for themselves and then incorporate them in the day-to-day.

As soon as organization size increases, then I think there are probably two models. There is one where you start having non-technical champions. They really own this and create, and I think Borat is gonna show it for us later, some non-technical or non-coding tools that can be used by other colleagues. And then I think there is also world, and this is mainly in large organizations, where you might have the IT team developing ad hoc tools.

Does that answer the question?

All right, this is I will see what's the person who goes and makes the decision. But still, like, even, like, I know this thing, so, but some other person who's talked to me, they look at both outcomes, so we're going to have to connect notion, have to connect achievement, stuff like that, so it's so difficult.

How do you reach those people? How, do you think they are, like, you will put the video out, or someone will do a course, or setting up, or people in the room, people will start using it.

Yeah. I mean this is,

That's a very good question. This is the reason we exist as a company, ultimately.

Tool Landscape and Training Approaches

Because I think this whole AI landscape is a lot patchier than what you had in the 90s, where Microsoft emerged as a dominant player. I think right now, and maybe we discuss it later in the fireside chat, but right now my sense is that this dominant player is not there. And there will be different tools that can do different things.

And so there will be a stronger element of individual agency in people actually having to create their own operating system, their own digital garden. You can choose the metaphor that you prefer.

And I think that leaves a space for different forms of training. like what we do, which is based on live demos and a personalized learning platform, in-person workshops. I'm a bit more skeptical about the traditional e-learning models because I think there is a strong element of mindset shifting, of using technology in a different way that would be difficult to address through just a typical e-learning course.

DIY vs. Services for Setup

Do you think what would be a good thing is some kind of service business or is it easy enough for anybody to just, okay, I want to look

I think it's not a yes or no, it's a scale. That depends on how tech friendly you are as an organization or as a person, and also perhaps how much time you want to dedicate to this. There is definitely a scale, and there will be a market for consulting on how to set this whole system up.

But I think many people will just do it on their own because it has come to the point where even a non-technical person like me can do it. Setting up these integrations literally takes three minutes. It's as simple as it gets.

Yes?

Safety, Permissions, and Change Control

What happens if they start to do changes that you won't like?

For example, you got these appointments and now you will say, You know, create something in Notion or change something in Notion.

And it will look like what you want, but once you approve it, it goes to Notion and, let's say, deletes a folder and creates a new one.

And now you don't have a version control system over this. Or, for example, it goes into Google Calendar and drops a folder or something.

Minimizing Risk with Clear Context and Read-Only Flows

I mean, again, we go back to what I was saying earlier about the importance of context. If you do not give the models the right amount of context, then you are increasing the probability of those hallucinations. which is why it's important to give that context.

Personally, I'm still not at the stage where, I use Cloud as my operating system, but I would not push changes through Notion. It works more the other way around for me right now. And it's still very useful and it still allows me to make use of the data and the information in the best possible way.

No, how am I using it differently? I'm just saying, so I'm just saying that I'm not allowing for that channel

I'm only picking information from Notion. I'm not asking Cloud to make changes there. So I'm treating Notion as my data source, as my database. I'm not treating Notion, and it's a database that I want to remain clean, in a sense.

Well, it's my own definition of my operating system.

Personal Operating Systems and Risk Profiles

If you would like to set something different for yourself, you can. That's the beauty of it, that I think we can be creative in the way in which we use these tools, depending on our risk profile, our trust in technology, our ability to operate with these systems.

Individual Agency in the AI Era

And that's what excites me about this whole revolution, we might say, or evolution, is that I think there is a stronger role for individual agency than there has been in the 90s for the internet, where the technological solutions that everyone used became a lot more standardized.