On September 26th, Dave Nelson, CEO of Dialog Consulting shared insight into Unleashing Productivity with Generative AI.

Unleashing Productivity with Generative AI

Introduction: Unleashing Productivity with Generative AI

Good afternoon, everyone. I’m Kelly Paletta. Today, our topic is unleashing productivity with generative AI. Welcome to our webinar. For those of you that don’t know me, I’m the Director of Sales and Marketing at EXP Technical. At EXP Technical, we provide IT support to hundreds of small and medium-sized businesses here in the Pacific Northwest. We’re your local IT support provider, and our mission is to supercharge your business and make it secure the EXP way. This webinar is one of our commitments to our values, which include contributing to the community, sharing knowledge, and sharing success. In this webinar series, for those of you that haven’t attended before, we bring thought leaders and subject matter experts to the attention of the community of local business leaders. That’s what we’re going to do today as well.

Our vision is that even if your path never joins with EXP, you don’t ever have to become a client of ours. We want to provide information that helps empower your business because if the region succeeds, we will all succeed together. I have just a couple of things to mention before I introduce our guest speaker. One thing I’ll say is we have a prepared presentation for you. It’s going to be very fast-paced. The good news is that our guest presenter, Dave Nelson, is allowing us to record this event, which is fantastic because this is going to be the most visually stimulating presentation you’ve seen in a long time. There will be a lot of replay moments that you want to go back and revisit or share with others in your organization, whether that’s with your executive team, operations professionals, HR, or IT. I’m sure there will be moments where you want to play along at home and replay. We’re fortunate that our guest presenter is gracious enough to allow us to record and share that recording. Be patient with me; it takes about a week before I have that available, so you’ll get an email from me about a week down the road.

Now, let me introduce our guest speaker. Today, joining us is Dave Nelson, CEO of Dialog Consulting. The reason why we invited Dave to our presentation today is because I’ve attended presentations he’s given in the past, and it’s obvious that he’s much farther on the path towards the adoption of AI than certainly I am and many of the people that I know. We have a common value of sharing knowledge and sharing success. Dave is willing to share what he’s learned so that we can all succeed. Dave, I think I want to hand the baton to you. Welcome to our event. Thank you so much for being here. Feel free to speak up when you’re ready, and I will yield the floor to you to move forward from here.

Sure, I can help with that. Here is the revised text with punctuation added:

Unleashing Productivity with AI

Alright, good. Let’s get started here, Kelly. I’m happy to interact as we go, so if you’ve got a question, please chat it to Kelly, not to me. Chat to all participants if you like. I’m not monitoring the chat as I speak. Kelly will see them and will periodically interrupt me. Anything we don’t get to today, we will give you a process at the end, so we will ultimately answer all of your questions.

I’m going to use a virtual studio here today, so here is my virtual worldwide headquarters of Dialog Consulting Group and our presentation, which I’ve titled “Unleashing Productivity with Generative AI.” That’s what we’re going to focus on, and I call this a leadership imperative.

Here is our imperative as leaders: our people have well-polished and practiced processes, and they’re happy to show up tomorrow to do what they did today and do it well. Yet, this new thing has come along which can significantly improve productivity, so we need to motivate people to explore this new technology and start to look at how they can evolve their processes for greater productivity. I’m going to try to quantify that for you, but that’s the imperative: motivate people to be curious explorers and maybe within 6 to 12 months make some significant changes.

Now, let’s dive in. You see a ton of AI-generated images today, but I want to back up to December of 2022. This is one of the first articles that I saw in the morning blog about this thing we call ChatGPT or generative AI. “Good morning, ChatGPT gives an early glimpse at what AI could become.” Now, here’s the cool thing: this technology, the interface is a chat interface. You already know how to chat, so you already know how to use this. What to do with it, that we got to talk about, but how to use it, if you can chat with a human, you can chat with this bot.

I want to take you briefly into the rest of the acronym, so let’s talk about the P. This is a neural network that learns very much like we do. It encounters information, it receives positive and negative feedback, mostly from humans. We call this reinforcement learning, and the neural network establishes connections. That is the process of learning.

The thing is, it can cover more ground than all of us put together, so it is pre-trained.

At the time it was originally released, it had already consumed 175 billion internet words and parameters. It went on in the following March, when it became image-capable, to cross over 300 billion. We have a brand new version of this just in the last three weeks as we speak today that reportedly has trained on 1.3 trillion internet words and parameters. So, it knows a lot of stuff, and it can cover a lot of ground, which is great because we’re in a world of too much information, and that information is doubling annually. How do we use our linear brains in this exponential world? We use this bridge.

Now, it doesn’t copy or catalog or memorize people’s content. It learns from it, and so when you ask it to do something with what we typically call a prompt, it generates new stuff that’s never existed before, and it doesn’t copyright what it creates. In fact, what it creates is different every time, even with the same prompt. So, what it creates is yours.

Check it with a plagiarism checker. I like Grammarly, but use what you like. No plagiarism found.

So, it generates new stuff based on knowing a lot of what’s going on out there in the world.

Unleashing Productivity with Generative AI: Ideation

Now, a quick experiment just to open your eyes to the possibilities here.

This is the center row of three rows of students in an innovation and entrepreneurship class. This is part of the MBA program at Wharton, one of our top MBA programs in the country. These students, 200 of them, receive an assignment each year, word for word what you’re reading on the screen: “Generate an idea for a new product or service appealing to college students that could be made available for $50 or less. Write it up, due a week from today.”

Then the professor takes these ideas and puts them into what might be called live market testing. They’re published on a website. They invite students from all over eastern Pennsylvania. I’m coming to you from western Pennsylvania, Pittsburgh area today, and they ask them to rate how likely they would be to purchase this product or service if available at the advertised price.

Any idea that gets at least a 50% take rate, at least 50% of the people want to buy, they call that exceptional. Now, out of the 200 ideas submitted this year in this live testing, sort of a typical year, five rated as exceptional.

But what was new this year is, hey, you can ask the same question to ChatGPT with one minor difference: not an idea, but 200 ideas. So, its 200 ideas go head-to-head, apples to apples, with the 200 ideas from the students. They put the ideas from ChatGPT into the same live market testing, and out of its 200 ideas, 35 rated as exceptional.

What’s going on here? AI can’t be creative. Technology can’t be creative, right?

Well, you got to start to think about that. That’s more a question to reconsider.

Now, humans are creative in so many ways, dozens of ways, and this is just one form of creativity. But let me try to characterize this one for you. There was a quote back in the ’80s: “Innovation or creativity is taking two things that already exist and putting them together in a new way.” And this form of innovation or creativity, because ChatGPT’s library knowledge base is so great, it can see combinations, and this is one of the great promises of AI. It can see combinations that no individual human ever would.

Now, in the ’80s, of course, we already had music, we already had television. It was Tom Freston who said this and did this when he put together those two in MTV.

In that form of creativity, ChatGPT absolutely excels. And I also say this: when you and your team engage with this product, with this technology, it drives more creativity, as if there is another person in that conversation, more creativity in the world after this.

Now, it is generative, and it is pre-trained.

What’s the T? Transformer.

I don’t want to geek out on the technology today because we’re really talking about business applications, but here’s what you need to know about transformers. Invented by Google recently, 2017, so this is a pretty new idea, and Google published it. You can read all about it on Cornell University’s website. “Attention is all you Need.” That’s the name of the paper.

But here’s the other thing about transformers: it’s really about words. If you think about words, they don’t necessarily have fixed or deterministic meanings independent of other words. “The bank is closed.” What came to mind? Some kind of financial institution. It wasn’t the word “bank,” it was the word “closed” that helped you understand what we’re talking about here. I could say something else entirely about a bank: “It’s slippery.” We’re talking about something else altogether, right?

And by the way, thank you ChatGPT and Stable Diffusion for these images. Generating images is an important part of this tool because if we can communicate in images, we are, well, reportedly seven times more effective.

So, all the transformer did, what the core of the technology is, is it’s able to assign probabilities to words to determine relationships and meaning. People sometimes call this a probabilistic word selector. If you give it some words, it can select the next one.

And aren’t we sort of the same thing? Probabilistic word selectors. You ask us a question two days in a row, we’re not going to answer it with exactly the same words every time, and this is the same kind of thing. So, in that way, it varies from other forms of technology you’ve used. It is not deterministic, it is probabilistic. Cool.

Now, I would like to use the word “transformer” in a different way, or at least the root “transform.” Your business processes. So, which processes? How much productivity gain to expect here? And, hey, Goldman Sachs has taken a shot at answering that question. “Generative AI could raise global GDP by 7%.”

We’ve never seen a single invention in the history of humans with this kind of impact on the world’s economy. Two-thirds of occupations could be impacted in time, but here’s where we’re at right now. So, we got the 2017 invention, thank you Google, and we’re at zero down here on the left. In other words, it can do 0% of the job for 0% of the occupations. Now, Google wasn’t really focusing, at least not solely on this. They had a business, but a group co-founded by Elon Musk and Sam Altman and a few others, OpenAI, decided to focus exclusively on transformers. They made version one, version two, version three, and so by the time we get to the middle part of 2022 with GPT-3, we have progressed along the horizontal axis.

This is starting to impact potentially more and more occupations, but the percentage of the job it can do is still fixed at zero. It’s only a theoretical thing until the last day of November, first day of December, we got ChatGPT, a new kind of interface we could actually apply to tasks we’ve got right now, and boom, like a rocket ship, we are sitting right there.

In other words, four in ten U.S. jobs can be about 30% automated.

Now, personally, I’m using the actual language from the report, but I think “automated” is the wrong way to think about this. This doesn’t do something independently of us humans. It’s a tool to make us more productive, more efficient, and more effective. So, I’d like to change their lingo to be “productivity improved.” We can improve productivity for 4 out of 10 of our people.

Oh, and I should say, out of any 10 people, five of them are driving the trucks or running the machines or the so-called essential workers. This didn’t apply to those folks; it applies to knowledge workers.

So, what this is saying is that out of 10, five are knowledge workers. Four out of those five are our salespeople, our marketing people, our HR people, and those of us that are leaders or managers in our organization. We should be using this technology. So, that gives you a sense. Do the math on your corporate headcount or business headcount. Multiply by 40%, or if you got a count on your knowledge workers, multiply by 80%. You got a sense of what we’re talking about here.

That is your leadership imperative.

A New Era

Now, you would put this in the same category as PCs. Remember when those came into business in 1984? Browsers in 1994, smartphones in 2007, blockchain in 2009, and now we have ChatGPT in 2022. And I think you can maybe most of us ignore blockchain for a bit longer, but hey, when you come back and look at these slides 10 years from now, if I didn’t have that on the list, you would wonder why. But that’s a pretty short list over the last four decades of really big developments in business.

And you certainly wouldn’t do business with or wouldn’t hire an employee who came in and said, “I don’t use PCs.” Yes, it’s equally crazy that you wouldn’t use ChatGPT, in my opinion.

Now, I’m going to use the term “ChatGPT” as we dive into this as a generic Kleenex. I’m using it as a generic term. What I really meant was facial tissues, right? So, just bear with me as I use ChatGPT generically because I want to introduce a few others for you here.

Now, ChatGPT is in its fourth version. I think that’s a terrible product name because it’s not the first version of four; it’s the second version, and this is rather confusing. It’s not an 0, not it’s not for that O, it’s O as in the word Omni. What OpenAI is trying to convey here is don’t think about this as just text anymore. It’s images. They are beta testing videos. It can speak and listen. It can start to read our emotions and moods. It is an omni-dimensional chatbot.

I did ask ChatGPT what it thought of the name, and it said, “Well, that could lead to some confusion there, Dave, because typically 4.0 is the initial version of the fourth gen. How about instead 4.14 Pro 4X?” But apparently, no one there in San Francisco at OpenAI actually asked their own product about a good name.

Now, the bad naming continues because they just came out with a new release, right? Version five, right? Or 4.2? No, no, no, it’s 0.1. So, I don’t know what’s going on with their naming person, but the rest of that company’s pretty sharp.

Now, Microsoft has something called Copilot, and they just had 13 billion lying around and decided that it was better off giving it to OpenAI in exchange for the code.

So, what you’re looking at here between ChatGPT and Copilot is essentially two identical twins. So, why are they different, right? If you use both these products, and you’ll probably use both these products, why do they vary? Same reason two human identical twins vary. They may be born with the same exact DNA. They may be equally intelligent, but they have had different experiences in the world, different friends. They studied different things. Their personalities have evolved as they’ve been alive. So, these two are equally smart but essentially identical twins.

Now, it’s funny. Usually, Microsoft copies whatever Apple is doing, but in this case, Apple copied what Microsoft is doing. So, you might have heard recently that Apple has announced something called Apple Intelligence. It’s going to be available with iOS 18.1, not the one that you can get right now, but 0.1 this fall, and it is powered by ChatGPT. So, they’ve licensed the same technology. Siri is essentially getting a brain transplant, and your phone is about to take a quantum leap forward like you have not seen in 10 years.

So, sometime during the fall or next spring, carve out some serious amount of time to relearn your iPhone. Now, bear with me if you’re an Android person. And sort of bad news if you’re an Apple customer: Apple Intelligence needs a heck of a lot of processing power, and the only older iPhone it works on is the iPhone 15 Pro or Pro Max.

Now, if you’re an Apple shareholder, this is great news because everybody has to buy a new iPhone 16. I got mine, but I still don’t have Apple Intelligence until they do the 18.1 release. You’ll see some amazing steps forward.

Now, a whole different code base, but another transformer comes from Google. They saw they got leapfrogged. It took them a year to catch up, but they came out with Gemini. You probably heard some others along the way: Bard, Lambda, etc. But Gemini in December, and they’re back in the game.

And we also have another cool one from a startup called Anthropic. It’s now Claude version 3. Now, some people from here at OpenAI decided they didn’t like the direction of that company. They went and started Anthropic. They prevailed on a guy named Sam Bankman-Fried, who used to run FTX, the crypto exchange, that he wasn’t really utilizing that 500 million dollars of his depositor money. How about giving it to us? And he did.

So, that’s where the single biggest part of FTX funds went.

They did amazing things with it. Google saw that, hedging their bets, they gave them 300 million. Amazon has come in here recently with 2.4 billion. And yes, Sam Bankman-Fried is in prison because you can’t do that with your depositor money. But recently, the failed FTX has prevailed upon the bankruptcy court to let them sell two-thirds of their Anthropic stock in a private transaction, and investors are now getting 118% of their deposits back. I don’t think that’s ever happened in the history of bankruptcy before.

And of course, the failed FTX still owns a good deal of Anthropic and some other assets. Of course, Sam remains in prison.

And then we’ve got, as a fifth choice, something called Perplexity.

Ranking Generative AI Tools

Okay, Dave, too many choices! Which one?

Well, the Wall Street Journal has attempted to answer that question for us with this article, “The Great AI Challenge: We Test Five Top Bots on Useful Everyday Skills.” And according to the Wall Street Journal, at least at the end of May, the winner is Perplexity, followed by ChatGPT, followed by Gemini, Claude, and Copilot.

But if you dive into this, what you see is it entirely depends on what you’re trying to do.

They tested a whole bunch of different kinds of things, and you see that Perplexity, while they called it best overall, isn’t the best for most things, nor is any of these things. They all score first or worst just depending on the task that applied to.

And hey, there’s Copilot in fifth place. But that was last May. With the advent of 0.1, which also goes to Microsoft, my guess is if the Wall Street Journal redid this ranking right now, it would probably look something like this: ChatGPT, then Copilot, then everybody else.

But here’s the bottom line: What we’re looking at is very much like a five-horse race. And while at that time you can see Copilot back there was in fifth place, it was only half a length behind ChatGPT, which is barely a nose behind Perplexity.

Anybody could win, and the race gets rerun every day, and anybody else could win.

Bottom line is you can’t make a bad choice. So, I think what will probably govern your choice of chatbot is your current vendor relationships.If you’re a Microsoft shop, the answer is Copilot. If you’re a Google shop, the answer is Gemini. You’ll probably find occasional uses for one or more of the other ones.

You’re looking at my image here, which, thank you, ChatGPT, for generating a neck-and-neck five-horse race with a dark horse back there. What the heck is the Llama doing? So, this is one to keep an eye on. By the way, some of you are saying, “Isn’t Elon Musk doing something at xAI?” Yeah, Grok. So, there are other ones out there, and there are some coming from China like Baidu’s Ernie.

I’m not trying to be all-encompassing here, but Llama is super interesting because Mark Zuckerberg, Meta, Facebook, Instagram, etc., is making Llama available open source for free. To the extent you might want to incorporate large language model technology into a product you build at some point in the future, I think many of us are likely to choose Llama. Hey, capitalism is a great thing. We’ve got lots of choices.

Practical Applications

Now, let’s get down to the core point of the webinar today: What do we do with this?

Let’s start to dive in. So, chatgpt.com. But, of course, if I say ChatGPT and you’re a Microsoft shop, please just translate to Copilot or translate to Gemini or whatever your favorite is. Generic term.

But here is my chat bot. I just have to type any words or phrase or question, or sometimes we call these prompts, right there, and away we go.

Using AI to Resolve HR Issues

So, let’s pretend for a second I am the director of HR for Acme Corporation.

I receive an email one day from one of my employees accusing another employee of sexual harassment. I begin my HR process. I meet with the players. I counsel everybody on confidentiality. But sooner or later, I’ve got to generate some content.

I’m going to show you a very specific content generation use case. But if you think about it, we all create content every day. Salespeople responding to an important email or creating a PowerPoint for tomorrow’s prospect, or marketing is blogging, or something’s broken on the website, they need some help creating better HTML code, or we’re creating a job description, or whatever. We all create content all the time. This is the number one use case for knowledge workers, and probably all of us do it many times a day. That’s when you should think chatbot, large language model.

Now, I’m going to do it in the sexual harassment scenario, but you can do it for whatever your content generation task is.

I’ve moved the box to the top. I type my question, and you’ll notice how different this is from what a Google query might have looked like. I say to ChatGPT, “I have an employee that is sexually harassing another employee.” Yeah, I should have stuck in the word “allegedly.” Okay, so let’s get this right.

Based on our employee handbook—wait a second, it knows a lot of stuff…1.3 trillion internet words and parameters, but it doesn’t have our employee handbook.

But here’s the first really big idea: So often, the most valuable things you’ll ever do with these chatbots involve giving it something that you’ve got to help answer whatever task you’ve tasked it with. So, I open up my Microsoft Word document. This is our employment agreement. I go to section five, about a thousand words. I copy and paste this whole section that deals with all prohibited behaviors, harassment behaviors in the workplace, not just sexual harassment, but this is the relevant part of our employment agreement. It’s always better to provide more detail.

I’m in Pennsylvania, you’re in Seattle, Washington, or wherever you might happen to be—Texas, California. These states have different policies and procedures. Even if you’re the director of HR, you might not have read something that happened yesterday in terms of an update. GPT has, and the federal policies, and the Society for HR Management (https://www.shrm.org) website with all the best practices. So, this knows a lot of stuff. Be more specific.

This thousand words looks nothing like your typical Google query.

Two seconds later, bam, got my draft.

“Dear harassing employee,

I am writing to inform you….”

Now, we’re not going to send it to the “dear harassing employee.” I’m doing something intentional here. I think it is a good idea to assume a healthy paranoia when dealing especially with ChatGPT. We’ll talk about levels of safety here in a second about potential information bleed. You really don’t need to put in the names of the players, right? So, I am going to substitute once I’ve got this back on my computer, Microsoft Word, Pat’s name. But notice it’s not doing a traditional, you know, fill-in-the-template kind of thing. When it read section five of our employment agreement, it could see we are Acme Corporation and appropriately use that.

While most of this is uniquely generated, it’s never existed quite in this exact form before in the history of words or humans. It read my section five and found one paragraph in section five that was appropriate to extract because that was about sexual harassment. Then the rest of this is uniquely created.

We come to the end. “Thank you for your attention to this matter. Sincerely, your name here.”

Once it’s back in Word, I plug in my name, Dave Nelson, my title, Director of HR, Acme Corporation. I just finished in 2 minutes what might have taken two hours.

But I’m not finished with this job yet because this is only an initial draft.

Here’s the thing: these chatbots can make mistakes, and there’s so much stuff you know, your domain knowledge, that they don’t know. They don’t know the actual players. They don’t know any history. I certainly didn’t provide it. They don’t know the corporate culture. So, I still have a lot of value to add. In fact, I need to go through every single word here because here’s the rule about AI-generated content: it’s not the manufacturer or the service provider responsible for it. It did what it did based on your request. The providers have no liability for what their tool did. We do. So, we as the users, and this is different from every other kind of product or service liability, we as the users are responsible for the safe, ethical, and correct application of this technology. So, you better read every single word and make it your own.

Now, let’s get into some cautions as we go.

Don’t worry, we’re going to keep going deeper here.

Be paranoid about data compromise, privacy, intellectual property, customer data, and financial data, at least until we know more. The very oldest of these products, at least how long they’ve been available to the public, is not even two years old yet. So, we don’t know everything that can go wrong or where the possible weak points are. So, healthy paranoia.

In fact, my guess is even if you’re a Microsoft or Google shop and you use their core tools, sometimes your employees are going to use ChatGPT. This time, you should read the privacy policies per ChatGPT or OpenAI. “When you use our services for individuals, we may use your content to train our models.” Can I translate that for you? If you don’t tweak the settings, we will use your content to train our models.

That seems like a bad idea.

So, how about four clicks? It only takes 15 seconds.

Number one, you can see we’re on chat.com. In the upper right, your account, you click it, and it brings down, you see where I put the number two, settings, click that. Number three, data controls, only four clicks. Once you’ve activated data controls, improve the model for everyone. Notice it’s defaulted to on. You want that to be off. It’s not going to hurt the way the product runs. It just makes you and your people safer. They don’t train on the data they’re receiving from you.

Now, this brings up an interesting point, and I want to shift to Copilot in Word for a second. We need a generative AI use policy before too many more days go by, right? So that everyone knows how to use these tools and which of the tools are right. So, let’s use Word to create our gen use policy.

If you think about it, Copilot has been out there. It’s read everybody’s policy that’s published. It’s read all the websites on advice on this. It’s already done the research. So, let’s go.

Here I am in Word. It looks exactly like the Word you have, but I have paid my incremental $30 per month to get Copilot in the Office tools.

So, our Copilot, or sorry, I should say our Office licenses have generally cost us about $100 a year. Get ready for $460 a year ($100/year for Office + $360 for Copilot). Now, you can see why Microsoft’s market cap is over 3 trillion, and this 13 billion is probably the best investment in the history of capitalism. But now that I paid my $30 incremental second license on top of Office, got a little icon right there, click that little icon, and it will activate Copilot in overlay mode.

So, click. I’m going to try to zoom in and out on the screen so you can read these things clearly in Zoom. So, describe what you’d like to write, says Copilot, including notes, outline, etc. Notice there’s also a button to reference various files that might be useful in this task. Or there’s another way to activate Copilot.

Once you think about it, software, if you’re paying attention to it and looking at it in any detail, most software has five ways to do the same thing. So, you get to choose which of these modes you prefer.

But if you click on the Copilot in the upper right, it generates a side panel.

One thing I want to point out, mostly it’s the same, but there is one important thing about the side panel view. Notice the first icon in the chat box. People ask me all the time, “Dave, what’s the best resource to learn how to use Copilot?” I know it’s a super meta answer, but my answer is use Copilot to learn Copilot. Just click that icon. When you click that icon, you see it says view prompts. Click the prompts, and they built a whole ton of prompts into Word to give you and your people ideas of what you can possibly do. So, you actually just play with the tool. The learning is built right into it. This is really cool. Any question that maybe they haven’t defined a prompt for, just ask Copilot, “How do I?” It’s sort of amazing. So, use Copilot to learn how to use Copilot.

Generating an AI Acceptable Use Policy

Now, we’re talking about generating an AI use policy. So, let’s go.

Here’s the everyday use case, and these are all going to play at actual speed. I have not accelerated or in any way deleted or changed the response time. So, I’m going to click on the icon to go into the popup mode. You see me moving my mouse. It just popped up. I’m going to zoom in a little bit so you can read it. Draft with Copilot. I say, “Good morning. I’d like to develop a generative AI use policy for my company. Kindly take a first cut at a draft, and I’ll provide comments and direction.”

Then I’m going to click the generate button. So, let’s zoom back out again, and let’s watch how this thing goes. So, the generate button right there. Let me get my mouse over it and click it. So, click, and you’ll see the document starting to come together in the background. It’s happening. It’s now scrolled below the screen. Working on it, working on it, working on it. About 10 seconds, and boom, done.

Now, I look at this document. I decide on first glance, before I’ve even chosen to keep it, let’s say it’s maybe a little bit too legalistic. The bottom line is if my team can’t read and understand this thing, it’s pretty simple and straightforward. Well, we’re not going to get what we want in terms of use. So, I’m not going to click on the button to keep this draft. Instead, I’m going to provide some feedback and directions. So, next, I say, by clicking over there in that box, I say, and let’s zoom in here so you can read it clearly, I say, and notice how conversational this is, “Let’s keep the language relatively simple and straightforward so that our team members can read, understand, and apply it.” I mean, I’m talking to it essentially the same way I would talk to somebody who works for me, or a consultant I’m working with, or my labor law attorney, or whatever. Talk to it like a human. You already have the skills to do this. You do not in any way need to change how you normally interact.

Now, I click the little forward button over there at the right. It’s time to generate a new draft. Now, you will see it’s refining. You’ll see it start to paint in the background there, the refined draft. It’s below the page now. Working on it, working on it, working on it. Done. This time, it looks good. It’s a lot closer. Simple bullets, pretty clear language. I mean, there are so many ways you can customize this. You can say, “Make it at a sixth-grade reading level,” or “Make it a little bit more humorous,” or “Make it a little more serious,” or as you like. It really takes two seconds to 10 seconds to do these tone changes. I’m good with this one, so click the button to keep. Now, I’ve got my first draft. Think about that. The old way, you would have had to go out and research this or hire an expert or what have you. Now, let’s pretend somebody wanders by my office and says, “Hey, Dave, I see you working on a gen AI use policy

Using AI to Compare Documents

I got one too that I got from one of my friends in another company. What do you think of this one versus that one? Now, the old way we would have done this in Microsoft Word. I just want to illustrate the changes in behavior and actually the benefits of doing so. The old way is I would have popped up my two documents in Word. So, I got the policy I’ve just generated. I’ve got this one I got from another company. I click into the tools menu in Microsoft Word. I select where it says track changes. I pick to compare documents, and voila, I get this diff marked thing. Here’s my problem with the dimark thing. If you’re a lawyer, this is great, but for all of the rest of us, I don’t need to know every comma change, every format change. If I put in an extra space or corrected or, and every single detail is called out here. You know, I look at this and I say that must have been designed by a team of monkeys

Monkeys!

Now, my college roommate ultimately became president of Microsoft Office products division, so I have access to things that other people don’t have. So, I reached out to him to get a photo of the actual product team that designed Microsoft’s diff marking, and it was a team of monkeys. Thank you, Copilot, for my little joke.

Now, here’s the new way with Copilot. So, I go here, and I type my question. So, we’ve zoomed in. I say, “Please compare our generative AI use policy, the document I just created.” And now, notice the reference a file, right? You can always give it stuff, and this is one of the really valuable use cases. So, I either click reference a file or, nice shortcut, you’re going to get really familiar with this type of slash. Slash triggers the file picker. So, up here, I’m going to say slash, and then notice it brings up recent files. And if I didn’t find what I want, now this third one is exactly what I’m looking for. But if I didn’t find what I want, I type the first letter, second letter, it’s just matching right off of my PC, Mac, off of our OneDrive, off of our SharePoint, right? So, it’s accessing files. But in this case, all I have to do is click from the recent list. I select the company SJ use policy. You see that populate in there. I finish my sentence and let me know what, if anything, is missing from the first document. And then I’m going to again tell it to generate, and voila!

Hey, the missing elements in your policy compared to the other one, Dave, you did not address specifically approved AI tools. You did not address the use of company data. This is way more valuable than the team of monkeys for most of us. I love this kind of approach. It’s really getting down to the core of what we would care about as humans. And again, maybe the lawyers need the other version, but the rest of us, what we really need to do is understand the content level changes and the principles.

Continuing, I say, “Please draft sections addressing approved AI tools and review of AI-generated content,” because I decide I want to add those in what I missed initially. Based on slash, I select the second document for use and slash the first document, and I get my nice draft of two different paragraphs there. And after putting those into the document, I say, “Draft a message to my co-workers asking for their comments on this document by Friday.” And of course, it knows the context of what we’re working on. This is just a really productive way to do so many different things.

So, when I said earlier, 4 in 10 jobs, specifically the knowledge worker jobs, can be potentially 30% productivity improved. Not that you need to try to capture all that on day one, but over the course of the next 12 months, we can learn this stuff. We can evolve our processes. That’s your leadership imperative. That’s what is to be had, and the number will continue to rise.

Back to the HR Issue

Now, I wasn’t quite done with, was a detour back to our sexual harassment scenario. I say next, “Please draft a message to the person submitting the complaint so they know I’m taking action.” Boom! I didn’t have to remind it where we left off. And then, and this is what I would call sort of systems thinking. It’s really good at a different kind of use case. Anything you need that’s processes, procedures, checklists, agendas, job descriptions, it’s great at all that.

And so, in summary, on this first example, if I’m director of HR at Acme Corporation, what I just showed you is super productive. But let me also point out what could go wrong. Let’s pretend I’m director of anything else, and I don’t really understand HR policy, so I don’t have the ability. I like to think of ChatGPT as my summer intern. And if you did have a summer intern, even if they’re super smart, super productive, super fast, you would never share their content with the world without some kind of review. So, if I were director of anything else, let’s say director of engineering, and don’t really understand HR, so I can’t provide the proper review and oversight on something like ChatGPT, I would consider it extremely risky to do what I just showed you. So, that use case makes a lot of sense if it’s your domain, HR. It doesn’t make very much sense. In fact, it’s a pretty big risk if it’s not your domain.

Questions

Hey, Kelly, I’m just looking up at you. If we got any chats, it’d be a great time to jump in. And anyone who has any questions, feel free to chat them to Kelly. I’ll address them, and again, we’ll provide you a mechanism when we’re done. So, if there’s something on your mind, let’s make sure we do chat it to Kelly before we shut down the Zoom bridge. How you doing there, Kelly?

EXP: Dave, there was one that I wasn’t going to interrupt because it was kind of comedic, but it came in through the Q&A. They asked if you use “please” and “thank you.”

DAVE NELSON: I do. You probably noticed the please and thank you here. The reason I’m using please and thank you is because, well, first of all, my mom told me to, right? So, I’m just bringing my same behaviors as if I were working with another person. But I will tell you the second reason: I actually think it creates a far better experience.

What you’ll see is sort of a mirroring back towards you. How you behave to ChatGPT is also how it behaves back to you. So, it’s just much more friendly, supportive, and humorous because I interact with it in that way. It then, you know, sort of mirrors your personality. So, it’s just more fun. Some people like to name it. “Hey, ChatGPT, I’m going to call you Sydney from now on.” Okay, and then you’re Sydney, as you like.

The last thing I’ll point out is this is a continuous learning neural network. On one level, it is evolving throughout interactions, and I just think it’s a better idea to try to do my part to lead it to a better place by being a good human. So, three reasons there for you, Kelly. But please and thank you, I’m very encouraging, and it’s just going to be a lot more fun.

EXP: So, I’m going to condense a few questions that came in kind of rapid-fire here too. They speak a little bit to security. Are all of these tools secure and safe to use? Another one was the same question but asked about, you know, someone noted that Copilot is really good at finding out information in your Microsoft tenant. So, what are your thoughts about it gaining access to, maybe inappropriate access to, HR information or other data that might exist in your network?

The third, I’m just going to sandwich this on top of it because it speaks to Copilot. I don’t have Copilot embedded in my Word. You addressed this, but we should go back and mention it one more time. Is that an additional expense? The answer there is yes, it is an additional $30 per month, paid annually.

DAVE NELSON: So, go on to your Microsoft admin account or have your IT person help you with that. There you go, Kelly. Your IT person…

EXP:  This isn’t a sales pitch, but we can help you with Microsoft licensing. For sure, we can help you with that.

DAVE NELSON: To the other security questions, because there are a couple of interesting questions, internal and external, I will mention in case we don’t get to this today, Microsoft does allow you to use Copilot for free, unlimited use on one app, and that app is on their Microsoft Edge browser. So, if you have Internet Explorer, please delete that straight away. That’s the old one. Get on Edge, and there’s a lot to love about Edge, including free Copilot on that. But for all the other apps, you are going to have to pay an incremental license for Copilot across the Office tools. That’s $30 per month as a license.

EXP: Security, and that’s a great way, if you don’t mind me interjecting, that’s a great way to learn how to get familiar with Copilot by using it in Edge where it’s free.

DAVE NELSON: Exactly, and then you’re going to want it everywhere else. As you saw in that document compare example, that was just one example of how things get better with Copilot compared to the way we used to do things in Word. Microsoft is doing a beautiful job of taking the core technology and integrating it into the things we actually do with our Office documents.

To keep it relatively short, I think you are really secure using Microsoft Copilot compared to ChatGPT, even with the controls I showed you to set. You will notice this when you start to use the Copilot, even in Microsoft Edge. There will be a message: your chats and data are protected in this conversation. Microsoft partitions what you’re doing in a way that’s very similar to how they partition your Office 365 documents from mine. They’re really good at keeping all our stuff separate, safe, and secure. So, I actually think you are materially safer using Copilot as opposed to ChatGPT, even though at the core, they’re the same technology. Microsoft’s doing some beautiful partitioning internally. As you have more and more people access Copilot in Word and the other Office tools, you do have to, from an IT perspective, be very careful about document permissions because all of a sudden people can ask Copilot, and if it’s got open access to you know last year’s bonus data, boom there’s an answer!

So, there are some significant, you know, gotchas or issues when you start to activate Copilot. Now, in a lot of ways, this is super valuable. “Hey Copilot in Teams, tell me everywhere I was mentioned today while I was off-site.” Then you get something very nice from the channels you subscribe to.

But we do have to make sure, especially from document permissions internally, that the right people have access to the right things. Those are great questions and real issues for sure. It’s just further incentive to follow best practices and implement the appropriate controls. I mean, these are issues regardless; it’s just that Copilot is a very good detective that can find and ferret out data when you need it, whereas people might not have gone to the work or might not have the skills to do what Copilot is able to do at the request.

Another question came in, if you don’t mind. This might be a quick question, but aside from actual version updates, have you, Dave, witnessed AI chatbots improve their ability to interpret or accurately answer your prompts over time with your use of it? I mean, I think what the question is, do you notice that it’s learning aside from being updated? Do you notice that it’s getting better as you work with it longer?

Well, it’s getting way better, but probably the majority reason is because the providers of this technology are training it on more data and parameters. If you use the original 3.5, it couldn’t generate or interact with you regarding images. It was really bad at math, and you know, it’s really closed so much of the gap in terms of things we need to do in our businesses since then. So, I’ve been on 3.5, 4.0, and now I’m on 4.1. Again, pardon the stupid naming, but also Copilot since February, I think, and they are definitely improving. Part of it is because of our interactions, but I think more of it is because they’re being trained on more and more data on the back end. They learn from, you know, the internet, our websites, Wikipedia, Reddit, etc. Wall Street Journal, you know, Dow Jones made provisions for these things to read The Wall Street Journal, so they’re learning from all of these sources.

Let me show you one last slide on where we were going. You certainly don’t want to end up like these folks, not reviewing the content that’s generated for you. Humiliated lawyers fined $5,000 for submitting ChatGPT hallucinations in court. The attorney simply had it draft a brief, glanced at it, and mentioned five real judges. Now, sometimes it gets creative when we don’t want it to, and the researchers don’t really know why. We call these hallucinations. The judge passed it to the defense team that didn’t recognize the cited case law and asked some associates to go to the law library to look it up. Those cases didn’t exist. They complained to the judge. The judge fined the lawyer $5,000 and made him write personal apology letters to the five judges he cited with the five fake cases. “In your handwriting, please do not use ChatGPT to generate these apologies.” How cool is that?

So, the lawyer, in his defense, said, “I heard about this new site, which I falsely assumed was like a super search engine. Yeah, I’d like to complete the sentence for that gentleman. So, I used its output without checking it because I’m a lazy, reckless idiot.”

EXP: Yeah, you know, actually that brings up two questions on my side too. One of the advantages when Copilot first became available is it helps you with the hallucination problem because it has source attribution. Maybe I haven’t been interacting with other tools quite as much, but that’s one thing that you can use to link to citations as reference material.

It seemed like there was another question, but you’re nodding, so I’ll just stop there and see if you have comments to address that.

DAVE NELSON: Let me make sure we cover two things in our remaining 15 minutes. Anyone watching has probably figured out that we weren’t going to end this at 1:00 PM; it said 1:00 AM in case you weren’t reading the meeting notice carefully. We can talk about this all day.

EXP: Well, I gotta order pizza if that’s the case.

DAVE NELSON: Exactly. I want to back up and show you the user interface on ChatGPT, and the same exact point applies to Microsoft Edge Copilot. I was zoomed in when I was showing you the content that we were generating for the sexual harassment response, but I want to zoom back out. This part of the product is called the conversation history. Now, if you’re using Edge, it’s going to be on the other side of the screen, but this is the history. It’s on the left with ChatGPT; it’s on the right with Copilot.

Super important because if you simply click to begin a chat, you open up your browser, you open up the app, it pretends as if it knows more or less nothing about you, only some very general stuff. So, if you want to pick up on this sexual harassment issue, say, six months later, you’re going to have to upload section five of your policy again. You don’t want to just start by clicking on ChatGPT or starting a conversation; you want to go find the last time you had that conversation in your history. Click on the sexual harassment policy, and that reloads everything we were talking about last time.

It doesn’t keep everything in its brain for context, but it does keep everything in memory for access. You have to trigger it, but that allows you to resume instantly without reestablishing context. You do not experience that effect with people, right? We always have to remind ourselves where we left off. This remembers better than you do. So, to me, an effective part of using this technology is being organized about these histories. They become very valuable, and you might even want to create some.

I love using this for helping me understand how to use a product or other kinds of tech support issues. “Hey ChatGPT, I’d like to tell you about all my technology and apps and describe what I use,” and then I just go into that one conversation to get answers. Or, “I’d like you to learn to write like me. Here’s 50 writing samples,” and now I can say to it in that conversation, “Please draft a blog post on the top four trends in artificial intelligence and make sure you’re in my style.” It looks at the other things you’ve written and mimics your style. It’s a fabulous mimic, so use it in those ways. You’ll find as you work this history, it gets better and better. You accumulate more and more detail there.

I would say there aren’t very many features yet here with respect to managing your history, so you’re going to have to be sort of organized and strategic. My guess is that OpenAI and Microsoft, etc., are going to bring us a lot more features on those histories. Now, I want to jump down since the topic came up to Edge, and let’s talk about the Edge browser. Essentially, if I say Microsoft, it’s like I’m saying OpenAI or ChatGPT. These companies are really joined at the hip. There’s a technology flow going back and forth. Of course, Microsoft’s putting a lot of feature functionality on top of it with their integration.

Interestingly, Microsoft is even entitled to essentially half of OpenAI’s profits. They don’t want to own the company, I guess, for antitrust reasons. But the code that runs ChatGPT is the code that runs Copilot. So, every time you see OpenAI take a giant step forward, like they did three weeks ago, that means Microsoft is taking a giant step forward.

Now, let me show you two things with respect to Edge, and these are both, I think, pretty exciting deltas relative to ChatGPT. When you first open up the Edge browser, it looks pretty much like a browser. But as we saw with Word, there’s a way to trigger Copilot there in the upper right. Number one, the screen splits, and number two, a message: “Your personal company data are protected in this chat.” That is a huge win.

Now, as I type a question here, and I don’t mean for you to read this, I just want to show you a little bit of the interface. It has an answer. When it answers, it cites sources, numbering those sources, and you see those numbers appearing throughout the chatbot response. This is like, “Yay, Microsoft!” It goes way beyond what ChatGPT did. To the extent that we suffer from the potential of these things to make stuff up, this really helps with the oversight because we very quickly can check sources. In fact, that’s where the split screen comes in. If I want to click on the second source there, I just click that source. Boom, there it is in the side-by-side window. So, I’ve got the web here, I’ve got the chatbot there. I think that’s fabulous.

You’ll also notice that there are suggested next questions just below the sources. It’s even using its AI to help you figure out what you might want to access or ask next, which I think is also a big win. So, a quick summary of Microsoft Edge: you get all of the features for free, whereas with OpenAI, if you’re using ChatGPT very much, you quickly bump into the $20 a month paywall. So, you get the features for free. It’s certainly a good way to start to become familiar with the Microsoft technology before maybe you get that incremental Office license.

Number two, your chat and data are better protected here, even after you did the settings that I pointed out you should do with ChatGPT. My thought is, even if you’re a Microsoft shop mostly using this, sometimes people are going to use ChatGPT. I’ll give you a reason here in a second. But number three, the source citing is just tremendously valuable. So much so, I think that ChatGPT is going to have to significantly improve how they do this.

And then number four, and this is maybe one more, in fact, this is probably the number one reason I come over to Edge. We were just using ChatGPT a minute ago. ChatGPT’s knowledge is guardrailed or cut off to not interact with you about anything that’s happened this year. So, you sometimes ask it something, and it’ll say, “I can’t tell you because my knowledge cut off is December 2023. Go check a current source.” Microsoft is playing a different game. They’re running their code completely live with no knowledge cut off. So, you can ask Edge Copilot in Edge about something that happened yesterday, and it can answer those questions. That’s often a reason that even if I’m using ChatGPT, I may come over here to Copilot in Edge. It’s worth understanding the two different games that are being played here. OpenAI is a relatively small company. They’re working to protect their reputation. They don’t want to make any kind of mistakes, and so they’ve got all the data from 2024, but they’re still testing, making sure the chatbot doesn’t do anything bad as a result. They do periodically move the knowledge cut off forward. Microsoft’s game is different. Let’s see, Google has 92% market share, Bing has 3% market share. Caution to the wind. So, they’re running this thing completely live, which is a real benefit to us. It’s unlikely for you to do anything crazy, but across their user base, every once in a while, someone does, and you’ll get to read about that on the front page of the Wall Street Journal. But Microsoft already has proven its reputation.

By the way, version one, do you expect perfection? In fact, that would be a good expectation setting right here. These products are new. The oldest one isn’t even two. Edge has not been available for, I should say, Copilot has not been available for even one year to any company. So, keep your expectations in check. When something doesn’t quite work, try it a different way or ask it why it didn’t work that way and if there’s a different way to do it, or check back tomorrow because it will be improved.

Now, GPT can handle 32 times more input than what Microsoft permits with Copilot in Edge. Sometimes I use it for that reason. The thing I am most excited about with respect to Microsoft, I’m a longtime Office user, is the integration with all these other tools. Microsoft Word, well, you already saw it working in there and got some of the benefits of that. I’m not a great writer. I’m an engineer by training, so I often like to write myself, but I often just consult on the fly. I was writing the sentence, “There are several other great features too, including,” and I just said to Copilot, “Is this the correct use of commas?” This time I got it right. So, yeah, that is the correct use. But here’s the thing that might be more valuable to you if you’re already a good writer: it knows all the features of the tools. “Hey, how do I insert a table of contents into this document?” Or simply ask it, “Could you insert a table of contents into this document?” How cool is that?

Now, in Word, that might seem to be of limited use because you probably know how to use Word pretty well. But think about Excel. Excel has 500 functions. Most of us know how to do 10 or 20 of them. How good are you with pivot tables or conditional formatting or whatever? My son took a six-month course in college. He’s an expert in Excel, but most of us aren’t. “How do I insert a table of contents?” Well, similar questions in Excel get killer answers. Once you have the data in Excel, you can ask questions like, “Which five customers are the least profitable?” If the core data is in Excel, it may well develop some kind of pivot table and then some kind of chart to give you an answer. That’s pretty cool.

PowerPoint, literally, I asked it one day, “What are the top four trends in the AI industry? Please add a slide.” Boom, it went out and researched the trends. It put them in that nice bullet form format, and the image you see in the background, it created uniquely for me in two seconds. Why would you ever make your own PowerPoints when it can do a beautiful draft for you like that?

Outlook, amazing things you can do. “Hey, create a rule so all incoming messages from this particular customer go into this particular folder.” Rules have been in Outlook for a long time, but most people have never used them, and they’re incredibly valuable. The reason we didn’t use them is they’re so hard to set up. So, have Copilot set them up for you.

Teams, already mentioned one example. “Hey, give me all the @mentions for me while I was offsite all day.” And of course, all the features with respect to Microsoft storage.

Now, I do have a comment again just to set expectations. I think that Copilot works really well in Word, PowerPoint, Teams, and with the file management. If you think about it, Excel is really complex as an application, and Outlook is pretty complex as an application. I would just say keep your expectations in check with respect to Copilot in those two, Excel and Outlook. Those two I am the most excited about. That is where the most value is going to come from, but they’re just not quite there yet.

Last thought I have for you is this: there’s a lot going on. You see new announcements every single week about new products. I discovered a tools directory. It’s called AI Scout.net. It’s a tools directory that last time I looked, profiles 1,777 different AI tools. “I want AI to do what?” and it will start to point you to the tools in that space. For sure, we’re going to use the large language models. We’re going to use Copilot probably integrated with our tools, but there’s probably some other things out there that can help with sales or marketing or whatever you see as your top business challenge. The next 12 months is a time of curious exploring. Try some things out. Get some people, some early pioneers, to explore. You’re going to fail fast on a few things or find some things aren’t ready for prime time, but you’ll learn a lot by doing this. About a year from now, you can probably harvest some of those learnings in new processes for, well, potentially, according to Goldman Sachs, a 30% productivity gain.

Closing Comments on Unleashing Productivity with Generative AI

Alright, we are straight up one o’clock Pacific time here, Kelly. So, let’s declare victory here. But let me suggest this: I’m happy to stay around and answer any questions right now. But if you didn’t get a question answered today, please chat it to Kelly. Kelly and I are going to work together, and we will get you written answers. Of course, I’ll be using Copilot to generate these for you, but maybe, maybe not. But we will work together to follow up with a document of FAQ questions and answers on anything that we didn’t address today.

Anything else we should talk about before we sign off here, Kelly?

EXP: I’m going to throw one quick one at you. So, where do you see this heading? I have my ideas of where this is all headed, but if we’re a year from now, or two or five years from now, what are we talking about?

DAVE NELSON: Well, these models are going to evolve. We think they are going to get roughly 10 times better every two years in terms of their capabilities. Now, consider for a second, we’ve gone through this phenomenal exponential called Moore’s Law. Our processors have been doubling in capacity or processing power every two years. So, we’ve gone from a supercomputer weighing 2.75 tons and costing $17 million to a phone that fits in your pocket or a watch you can wear on your wrist that has two to four times more power. We’ve been on this 2x every two years journey. What we’re talking about is going to evolve at a 10x every two years journey. So, get on that curve now because it’s bending away. It’s going to be harder later to learn.

I think ultimately, we’re seeing some really interesting developments. There’s some research going on right now on how to marry large language models with machines so we can talk to the machines to get them to do what we want. Of course, there are many other flavors of AI coming along, like autonomous vehicles. I was in Scottsdale, Arizona, and Mayo, which is part of Google, is completely autonomous there. You can, with your app, call an autonomous vehicle, climb in, and they’re no longer restricted to just the city streets or surface streets. I was on 101 going around Phoenix, and boom, there’s a 65 mph Waymo next to me. So, we’re seeing AI come into our vehicles and so on. I guess I would just say prepare for AI everywhere and it getting a lot better a lot faster than what we maybe have experienced with previous technologies.

EXP: Awesome. Well, Dave Nelson, we’re running a little bit over, but I wanted to thank you so much for joining. From the comments, I can tell that our audience is very appreciative. To everyone in attendance, you can email or chat questions to me in the closing minutes here. We will respond and follow up with answers to questions that have been unanswered. I will be in touch soon with the recording. With that, Dave, again, thank you so much for joining us.

DAVE NELSON: And thanks, everybody, for joining us. Good luck with your new super productive summer intern!