THE RISE OF THE AI COWORKER

E120 The Rise of the AI Coworker

Transcript

Dalton Anderson (00:00.77) Welcome to Venture Step Podcast. We discuss entrepreneurship, industry trends, and the occasional book review. I'm your host, Dalton Anderson. In this episode, we're going to be talking about Gemini Spark. Gemini Spark is the evolution of these AI chats or agentic workflows. This is the agent as an infrastructure next step. So the next step with the evolution of

agents and this use of LLM models, in my opinion, is every employee at your company is not only hired for their skills, but how they utilize LLMs, which I think people can agree on. And I think that is starting to become more mainstream, maybe in a couple years. And then the next thing is really every

employee has agents. Like they might have five or six agents. So there's less people management and then there's more agent orchestration slash agent management. But for that to work, agents need the right framework, context, and rules to consistently operate and operate at scale. And Gemini Spark is Google's take, first take at least, of

how to do that for a business user without having to utilize their ADK or agent developer kit, which is basically their agent orchestration piece, which is a multi-agent workflow, which I will be trying out and I will be talking about on the episode. But that's this episode. And a couple episodes from now, maybe I talk about that topic. One thing I would like to call out is the elephant in the room.

I completely botched the last episode when it came to sharing my screen. I messed it up. That's my bad. That's on me. Sorry about that. And so in this episode, I will be sharing my screen with some live demos. So I'll double check to make sure that it's correct and working right. That's important. So before we dive in, as always, let's provide a quick background of Gemini Spark.

Dalton Anderson (02:20.44) When was it released, where's it at? If you have access to it, and if not, how do you get access? First off, it was spoken about during the Google I/O similar to last episode, where it was a product that was recently released and it's in beta right now. So it has access to all of your drive, your email.

It can do things on its own. It can look at your emails, constantly monitor them and take agentic actions. It can buy stuff on the internet. It can do reservations, do flights. It can do a whole bunch of stuff, which you have seen before. That's nothing new, but I think for

for an agent to be able to do that within a platform like Google with that amount of context and the ability to search the web through Google Chrome or search and create files on Google Drive and take its own actions of what it thinks best, what it thinks is best for you and what you prescribed, that level of context with compute, it hasn't been done before. And so

Spark has a lot of potential and will be reflective of how other companies attack this same problem. Okay, so that's Spark, what it does, and it's always on. And then the next thing that's interesting is that you can create skills for Spark. So I created some skills as an example. And if you don't know how to create skills, you can just tell it, "Hey, I need to create a skill for this. We need to be consistent here. And to do that, I want a skill and I want the skill to do these things."

And then it will create the skill for you. And it was like, "Okay, well, when I send this email, I'll reference this skill, which is Dalton's language protocol." And so it will use that consistently and it it may you may not know what you want to say. So it's like, "Okay, can you look at the last twenty emails I've sent to you to utilize those emails as a corpus to establish a skill?

Dalton Anderson (04:36.897) That reflects my language usage." Something like that. And it will do it. No problem. It's exactly what I did, 'cause I was lazy. So I was like, "I don't wanna do all that. How about you try it?" And it worked. And I think it was decent. I do think that

Dalton Anderson (04:56.393) It could it could ask you some more questions before it gets started. I think its my main, my main gripe seems like it will get started right away. Like you'll tell it something and it won't intentionally interview you on whatever you're asking for. It will just go. So unless you're very prescriptive on exactly what you want, it kind of just goes off. And that's kind of a problem because you have to pay for the commute.

compute, of course, and you have a certain amount of compute you can spend in a five month five-hour window. Regardless, who has access? People who have the ultra subscription or inner enterprise subscribers?

And it's in beta for both of them. Okay. So this tool under the hood, it utilizes your, it's called a knowledge graph, but your knowledge graph for your account. Basically, it utilizes all your emails, your files, your contacts, everything that it knows about you to try to personalize the output to you. And that's the holy grail of

LLMs. Like for LLMs, they need to have the context to be useful. And so now it has the context to do cool stuff for you. And you don't have to worry about it as much. Like I'm fine with giving away my data because it's such a pain in the butt to prevent people from having your data. But if I'm going to give my data away, I might as well get some big benefits. And this is one of them, in my opinion. So

it's grounded within your account. And so it knows you better than other LMs. And so it can take tasks on for you. It can do things. It can You can ask it questions. You can, you can not tell it all the context and then it will figure it out from your emails or your interactions you've had, which is borderline creepy if you're not into that. But I'm like, "Yeah.

Dalton Anderson (07:12.639) You did it. I'm so proud of you." Whereas you could you could get turned off from that. But I think it's great.

So it consistently uses your context on your Google profile to ground the AI. And another thing it does is it's always on. If you want it to be. And you can schedule tasks, you can ask it to do complex things, multi-step frameworks. You can do a whole bunch of stuff that is, in my mind, pretty cool. And if it doesn't know, it will go and research on the internet and figure it out. It's kinda like a goal task.

When I spoke about Antigravity 2.0, you could do forward slash goal and it will just go until it figures out how to do it or it will make a successful outcome. And I specified that it's important to state what is a way to validate that your goal was achieved, which is important because if you say "Win the game," and the AI just edits the code to infinitely win.

That's not winning, or "Hey, don't die in Tetris," and then AI just pauses the game. That's not what you wanted. Or it's like "Optimize society," and then it's like humans get deleted. That's not what you want. So you've got to specify what the goal is, yes. But then you have to specify, "Okay, how do you measure this outcome?" Similar to how you manage employees, is it's very alike. I mean,

there's a lot more human connection, of course, with humans, but the goal, having a goal, measuring the outcome, and creating a process is is quite similar. Okay, so you have to be careful when you first kick off Spark, as it doesn't ask you that many questions or doesn't ask at all. It kind of just goes with my trial and error. And then it's always on.

Dalton Anderson (09:18.334) It utilizes your Google account as context and you can schedule tasks, you can make multi-step frameworks, you can research stuff, you can find stuff on the internet. It can do a whole bunch of stuff for you. which is great. I'm I'm personally pretty, pretty impressed with it. There's some some of the stuff I asked it. So let's get into it. I'll share my screen for real this time. I swear. So let me try that out.

Dalton Anderson (09:51.638) All right. I verified that I'm actually sharing my screen. So Gemini Spark, let me adjust my mic. Gemini Spark is a platform that is within gemini.google.com. And you can toggle between your regular chats and your Spark chat because they're separate. And one thing that you will notice is you can schedule things, which you can do in the Google chat as well. So I have some scheduled items on my Google chat.

And have stuff scheduled in Google Flow, which we'll talk about later on the episode. But you can have scheduled, and so you click on schedule and you can see what you have scheduled. So it says "manage podcast guest pipeline" and then "scout Venture Step guest targets." And then it says "monitor podcast pitches." So a lot of my stuff is related to helping me manage the podcast and

I think that's applicable for a lot of people that are managing things through their email on their Gmail account. Like email's tough. A lot of requests come through email. What is important, what isn't important? It's hard. And to, to get help, you need to have and provide the right context. And to get the context, you've got to be providing all the information. So your inbox need to be structured correctly.

You need to have AI have access to these things. It's there's a lot of stuff that you've got to think about. And Spark can remove some of that for you, but you've also got to help it out. So anyway, so these are the scheduled tasks. And then there's also skills. So it is "get more perspective" is a skill. I think that's just a popular skill. And then "prep for my meetings" is a popular skill. And generate. These are, my bad. These are recommended. I'm trying.

So if you look at the "Active," you can see mine. So I got the Venture Step vibe, which is aligned with how I correspond to people for Venture Step. I'm not having AI write my emails for me, but it does help to have them draft something where I can edit that where it's like eighty percent and then I edit it myself.

Dalton Anderson (12:15.649) For some things. And then this one I just tried out, but focused my energy.

I guess like focus my energy. It basically tries to specify what you have actually in your, on your plate and what you can't eat. It's if you got too much on your plate and you're trying to eat it all, then I'll be like, "Hold on." So align your align your workload with your energy instead of your calendar. So a lot of people don't have that choice. I personally don't have the ability to completely re realign my calendar whenever I want and how I feel.

Not yet at least, but maybe one day. All right. So "match my writing style" is another thing that I did. So I have this one and it utilized, utilized my voice, quote unquote voice, across the Google workspace. Okay, so then if you go to "Tasks," you can see what I've got going on. So I've got "travel itinerary and active planning."

Which I think was a cool one, and I was asking it questions. But basically it and I'll mark this out. I don't want to share my girlfriend's email. So in this instance, I I'm going on a trip and I might have to blur out all these dates, but we can we figure that out later. But anyways, I'm going on a trip and I

wanted to ask Google, "Hey I'm going on a trip with my girlfriend." I didn't say when. I didn't say who she was. I didn't say anything like that. And so it figured out from my emails where tickets were coming in that I was going on a trip with someone and the dates were X days. And then

Dalton Anderson (14:17.613) I think it looked through my Google Drive and found a file called "Itinerary." And then it read that file and noticed that there was a woman's name on there. And then it read all the emails probably associated with that person and activity on that file. And then concluded that that is my girlfriend and that's who I'm going on the trip with. And then read that file and specified some things that

might be interesting to include on the trip, which is something I asked to do, but I gave it zero context and wanted to figure out if it could

investigate and and discover who my girlfriend was and then figure out where I was going and suggest some activities. So it suggested restaurants that were quite expensive. I mean you can see this. $195 without wine pairing and then $270 per pairing per person. I told it, "Hey, like we're more of a hole-in-the-wall type of group." And so

we got some hole in the wall spots. But what I was trying to test was whether or not AI or this Spark system could figure out a solution with limited context from the user. Could it ground itself within my Google account and navigate a problem with limited, limited input, which is similar how people would use it in the real world.

Not everybody is some prompt expert that knows exactly what to say or know what knows what's important, what's fluff. People just talk about different things. They might leave their mic on and they'll have a separate conversation that's unrelated to the task. Like being able to tell whether or not that's related or not related is important. And then also being able to do more with less is important. And I think that's also something

Dalton Anderson (16:26.018) that people probably don't test very often because when you're testing stuff, you are probably an expert and you're not thinking about these things. So I try to make it ambiguous as possible and eventually figured it out. When I say eventually, it came back like 40 minutes later. I was like, "Okay, this is this is this is who it is, and this is what you guys are doing. And this is all the stuff I found, and this is why I, I think this is so and so." So

that one's cool. So this other one is "Podcast Guest Management." So Venture Step gets quite a few emails for guests on the show, which is great. I'm not complaining or anything, but it's hard to manage because I'm sending messages for schedule. They're sending messages for their schedule. We're trying to schedule each other basically on like a Calendly or something like that. And we've got conflicts or things fall off and or

I need to respond or they need to respond. And so things are constantly changing status. And it's like, where "Where are we at here? Is it a me thing? Is it a you thing? Is it an us thing? W What's going on?" So this is a

Dalton Anderson (17:41.678) pipeline, guest pipeline to manage the guess. And basically what I asked it to do was to architect a, architect a pipeline state controller that will be operating every day. And the thing I'm trying to, trying to do is understand where things sit. I don't want opportunities to be lost. So I want to log the opportunities with the pipeline and then service these opportunities

to me. So it's it's like a rehydration guest pipeline, which is great. So it identified emails that I hadn't responded to and guests that wanted to be on the show. And

They might have an opportunity on the show, they might not. I it kind of depends where where we're at. But what I'm saying is I had a hard time managing this. I'll cut that out.

Dalton Anderson (18:42.2) This hydration, rehydration of the guest pipeline is important because things fall off both on my end and on their end. And it's difficult to keep track because these people are really busy and I'm somewhat busy, probably nowhere near as busy as they are, but they've got stuff going on, I've got stuff going on, things fall off. And to prevent things from falling off without having to pay for a CRM, which I don't need because I'm not at that level yet, and I think that's overkill, is

to build out some kind of pipeline, but I also don't want to manage it. And so that leaves me at AI automation. So I tried to do something like this prior and it didn't work. And the reason why it didn't work was there's a whole bunch of rules. It took a long time. I eventually got it to work. But this took me

ten minutes.

Maybe five minutes. Like I, I just voice typed it, told it what I wanted, and it started. And then I gave it some feedback on some of the columns it wanted. Because as I said, it doesn't ask you any questions. So it just kind of goes. But I did something similar on my and I have to make sure to share on my Google Flow, which worked for some time. It's called Google Studio. I think they renamed it.

Google Studio Flows is what it is. And so if I share my screen, share this tab instead. If I share my screen, I go to Flows here. This is a, right here, automate podcast guest CRM and research pipeline.

Dalton Anderson (20:26.647) Maybe it's, I have two of them. So this one's active. Venture Step's here. So there's a whole bunch of steps here. Like it looks for keywords, it adds labels, it does summaries, it does all this stuff. But the problem with it is very rigid. And you're not coding, but you've got to organize the steps in the right manner. And then you've got to specify the context. Like this is me extracting what the, the email has.

Everything, putting it in this extract function and then giving it to the AI and specifying where it's coming from, from that email with all the tags. So it's pulling out all the tags within the email, the HTML tags. And then from there, it is bringing it back and doing a whole type of prioritization research of the guest so then I can look at their research and then review myself.

And then from there, it extracts what AI stated based off of the historical context of Venture Step like whether or not this person's a good fit. And ultimately, I make the last call on what I want to do with the show because it's my show. But I utilize AI to research the guests if they've got anything weird on their on the internet or if they have an active company or

if they've got something cool that they're talking about that they may have not shared because not all these people are salesmen and like not everybody's selling. And you know, when you're coming trying to pitch to go on the show, you gotta sell your your idea. And sometimes people are just like, "My name's Jim and I, I do car stuff." And I'm like, "Okay, great. That's awesome, Jim."

And then you read about Jim and it's like, "Wow, this is crazy. This is actually insane." And then you're like, "Okay, this person definitely needs to be on the show." So it's it's it I guess the way I would describe it is it normalizes the signal for yourself. Because if you just took those emails at face value, you'll get misinterpreted results because it's just an email. It depends on who sent it.

Dalton Anderson (22:55.105) The timeline of of where they're at, what they're doing at that time, if they're a rush, if they've got a good marketing firm. All of that is important for them to have value extracted from that email that they're sending to whoever they're sending it to. But I don't necessarily want to take their email at face value because if it's being sold a certain way, then what is the truth? Sometimes the

What's being sold is true, and what's and sometimes what's being sold is further from the truth. You never really know until you pull back the cover.

And in that instance, it normalizes the path to

factual information or how I should feel about this offer. And it's much faster because personally it takes a long time to research the companies. And maybe I don't need to do that. Sure. I could just have people on the show or take interviews, whoever. But then I think the show's less interesting. Like I really believe that Venture Step has really great guests and

I put a lot of effort in getting the right guests and so I'm trying to make it easier to one, identify which guests are a good fit and then two, make sure that those guests are in the show and forcing them to get there. That's where I need help. I need help with the transition from yes to get you in the show. So I've removed a lot of friction when it comes to scheduling.

Dalton Anderson (24:41.486) But in this instance, I wrote out this whole thing. It took forever. Like, no joke. Like a lot of testing to do this, this flow to take an email, extract it, do an extraction summary, extract out all the tags, give it to AI, AI to give it back to this extraction function, then draft a reply, and then add a row to this CRM that I have. Whereas Spark on this,

on this other side of the spectrum, Spark is something that's more automated, less rigid, but Spark did it in five or so. So my question really is if you remove all the friction,

does more control really matter that much for a lot of scenarios? And I think the answer is no. So keep that in mind. Because this was such a better, like a much better experience than the Google Studio Flow. And maybe Google Studio Flow rolls into Spark and Spark utilizes Flow to do certain things. I don't know. But currently,

Spark is just leaps and b leaps and bounds beyond this Google Studio Flow. They have different purposes, but there's a lot of overlap. Especially when you think about Gmail and managing emails and then creating folders or files or managing stuff like that. Spark knocks it out of the water. So, anyways, so in this instance, it reads through my emails and then strips.

sees which emails have been sitting around and then it created a spreadsheet of here of pitches and then it determined like who was highly bookable and who wasn't highly bookable and then it has follow ups.

Dalton Anderson (26:54.466) And then it specifies h for hydration, like who needs to be rehydrated, like who needs an, who needs a reply or a follow up, which is great. And then the second thing I was pretty impressed with that I did was this Venture Step guest outreach. So this was a little complicated because I didn't ask for one thing. So I asked for it to r research guests, but what I specified was "Hey, every Tuesday, go on a hunt and and

find some find some talent that could go on the show. Scout LinkedIn, tech news, executive bios, industry news, and find some in innovative take within any vertical. I don't really care what vertical it is or what industry. I just want it to be innovative." And then I said, "Take all that information, find the person, research them, specify whether or not they're a good a good partner for the show by looking at

the RSS feed from Venture Step, and then log that information in the CRM for outbound." So these are outbound requests, and then, "Find the email of that person." So there's a lot of steps of this. And this is something that like you legitimately pay for consistently. And I'm not saying that replaces it, but for somebody who is managing many different tasks to

take this off my plate and allow me to easily find some outbound leads once a week and just look through. And maybe every week for three weeks, they're all duds. And then, you know, four weeks in, I get, I get a couple people that I really want to book on the show. And then it's part of the process of getting them on the show. But I thought those was pretty cool where it specifies

multiple steps and tasks to retrieve research, then draft, then insert data. And so it's multiple things. It's like one, you've got to find where you want to look. The next thing is you got to find good information where you're looking. And then after that, you've got to extract that information of what you determined was good. And then you've got to specify

Dalton Anderson (29:21.29) on different sites these people and do additional research. And then you're asked to take that and then determine whether or not these people in their opinion or their best grounded truth on your context of your account, in this instance is Venture Step RSS feed, whether or not these would be a good fit. And so then it needs to understand the guests that were, were on the show and then circle that back to the guests they're looking at.

And then if they determine yes, then they've got to then implement it input that information into a slide and find their email. Not the slide, a Google Sheet. Find their email and then insert that information into a Google Sheet and then draft an email. I feel like that's a lot. That's a lot of work. Like for you to do that yourself, that that's multiple hours to do multiple

guests. Like if you did 10 guests, that's, that's some time. That takes time to find all that information. Like that stuff's not easy to find. And whether or not the information is true or factual, I think it's a toss-up between human and non-human. If you find somebody's email on the internet, doesn't mean it's theirs. I mean it could be theirs. You don't know until you email them. So I can't really verify whether or not this information is true, but I can say that it did

put the information in and it put it in this this f this file which is nice. And it has who they are, where they're at. hold on. One second. People are trying to get buzzed in. I think t it's Amazon. Our good buddy Jeff's trying to get in the apartment.

Dalton Anderson (31:23.278) Alright, I'm back.

Dalton Anderson (31:27.022) But my personal opinion is this is pretty cool. This use case I thought it'd fail, but it did a great job at creating this

Dalton Anderson (31:44.491) unknown scope, curating it, and then throwing into a spreadsheet. And then drafting emails for me. And then also one thing I forgot to say is if you confirm that you're okay with it sending an email, you've got to confirm for every individual action. But drafts, it can do no problem. It can create files, it can do anything you want within your scope. But once you start going external, you need to confirm that action. So

I sent an email to Sarah about the restaurants. It just says a goose egg to this episode. Or what is it? Not a goose egg.

Dalton Anderson (32:28.514) missing. I'm I'm missing it. I'm blanking out here. Goose egg.

Dalton Anderson (32:36.342) I don't know. I don't know. I you know what I mean. The like a secretive a breadcrumb or whatever. So I wanted a breadcrumb or

Dalton Anderson (32:50.04) Share this tab and said, I accidentally didn't share my tab. So this is what it looks like. But I showed it earlier. But what I mean is If If you're doing anything within your world your world within Google, it can do whatever. It can make files, it can draft stuff, it can research, come back to you. But when as soon as you say, "I want you to book something, I want you to buy something, I want you to send this email."

Then it it has to ask for permission and then it sends the email. Which is pretty cool. It's pretty cool. And I could see how that could get abused in the future, but people already send automated emails anyway, so it's not a huge deal. But that was Google Spark. I once again think the technology is really cool and I think it's closer to the future than it is further away of what

a workplace would look like or what a builder would be util utilizing in their free time. And I'm grateful for the tool for it to be always on. Not sure how many ch compute it utilizes to have three of these tasks running. But very, very cool stuff. And of course, wherever you are in this world, good evening, good afternoon, and good morning. Thanks for listening and listen in next week. Goodbye.