Why CoPilot Cowork is a Big Deal

Announcer (00:04): Welcome to Raw Data with Rob Collie. Real talk about AI and data for business impact. And now CEO and founder of P3 Adaptive, your host, Rob Collie.

Rob Collie (00:20): Welcome back, good to see you again.

Justin Mannhardt (00:22): Likewise.

Rob Collie (00:23): I have left the house once since the last time we recorded.

Justin Mannhardt (00:28): You had something going on last weekend, that didn't happen, right?

Rob Collie (00:31): Yeah, no, I did not go on the trip with our couple's friends. Jocelyn did.

Justin Mannhardt (00:37): Okay, good.

Rob Collie (00:38): This worked out sort of accidentally great because it left me on the couch, writing a book. The book is now estimated at 118 pages of content, we're cranking along. We have a working title, I'm not going to share it yet.

Justin Mannhardt (00:52): Oooh. This is sensitive, you shouldn't put this on the tape.

Rob Collie (00:56): Backstage, I just showed a preview of a potential book title and cover. Yes, it is definitely much better than past book titles. The very first book that I wrote, I titled it, Dax Formulas for PowerPivot. It's awful.

Justin Mannhardt (01:10): Wasn't there one in the middle? I forget. It had some sort of chemistry oriented-

Rob Collie (01:13): Oh yeah. There was PowerPivot Alchemy.

Justin Mannhardt (01:15): Alchemy, that's right.

Rob Collie (01:16): Which is actually one that I really loved, but that was like the ugly duckling of the bunch. But the ugly duckling in that case did not grow into a beautiful swan, it sold not very well. But I still had a place in my heart. And then the second edition of the first book, the best selling of the three, just because the audience was bigger at that point compared to the first edition. I even forget what the title of that one was. That one was like SEO, Word Salad... Modern Excel, blah, blah, blah, blah, PowerPivot, Power BI and modern... I don't know. Anyway, so yeah, still sick, but that's not what we're here to talk about today. Today's episode, we're going to talk a lot about Cowork. And there's going to be two phases of us talking about Cowork today. The first one's going to be kind of a sad story of how hard it's been to get it working for me.

(02:02): But the important thing is that I have been struggling with technology to get it working, which is something I normally don't do. If something doesn't work, I just leave it alone. But I have been crawling over broken glass to get to Cowork, it is that good. It is that life changing. I've gone through some extraordinary measures, which I'm sure you'll agree, Justin, knowing me as you do. But then after that phase of the conversation, we're going to get to the second part where Microsoft is just coming in and just making this all super easy. And if I had just been able to wait a few weeks, I'm pretty sure that I would have never done anything other than use Microsoft's version of this. Also, to underline this, what Microsoft is doing with Cowork is going to be a major hit for them. Not just a major hit for them, but also an inflection point in the mainstreaming of AI.

(02:55): Okay. But before we get to the part where the version from Microsoft magically fixes all of my problems, first, let's talk about my struggles with, and triumphs with, the versions that's just straight from Claude, that you have to download and install on your own, Claude Cowork, that has nothing to do with Microsoft. And it's so lucky for me that it worked the first time I tried it on my laptop, because installing it on my Windows laptop, it worked the first time, and it allowed me to do some things and get me hooked. And then it started failing.

Justin Mannhardt (03:25): But you knew the promise-

Rob Collie (03:27): Yeah, I'd had the taste. It is so unreliable on Windows and your mileage will vary. Even my computers at home are different. The desktop didn't work at all, laptop worked just long enough to get me hooked. Then the laptop goes through, it just puts me through hell. Just all kinds of intermittent and inscrutable error messages, which ultimately all the technical problems boil down to Claude Cowork's relationship with Windows VMs, Windows virtual machines, is ... It's got to be Windows fault. We're in the era of these problems just get fixed. If there's a solvable software problem, it's going to get solved.

(04:08): Their on record as saying that Claude Code is written using Claude Code. I'm a couple weeks into this now, they've had plenty of time, and I'm getting new updates downloaded all the time. And I have this batch file, Claude Code is written for me to help me fix Claude Cowork when it dies. I run this thing like six or seven times an hour. Sometimes I run it six or seven times in a row, Justin, and it doesn't fix it. And then the eighth time-

Justin Mannhardt (04:31): Doesn't take, so just retry?

Rob Collie (04:34): It is the weirdest thing. And then you don't know how long it's going to stay working for. Sometimes it works for like four or five hours in a row, other times it's like you get five minutes. So, the desktop was dead, the laptop was intermittent, I even added a feature to the batch file, so it started writing records to a log file on the desktop, so that how many times that I could visualize later how much pain I was willing to put up with. This is not how I normally behave with computers, I walk away from stuff like this.

Justin Mannhardt (05:04): Most people don't go to these lengths.

Rob Collie (05:07): Me, as a technical person, I have a much lower tolerance for stuff than this than most technical people. I have nope, nope, nope. But Cowork was so helpful to me that I was willing to do the equivalent of crawling over broken glass.

Justin Mannhardt (05:23): Just to get it for whatever time you could get it.

Rob Collie (05:26): So, then I ordered a Mac, which is again, something I don't do.

Justin Mannhardt (05:30): Man.

Rob Collie (05:31): I have not changed an operating system that I use since DOS. In 1992, I went from DOS to Windows, and that's the last time I changed a computing operating system. I do not want to hassle with this kind of stuff.

Justin Mannhardt (05:45): Right.

Rob Collie (05:45): But buying a Mac is worth it.

Justin Mannhardt (05:47): You bought a Mac so you could use Cowork.

Rob Collie (05:50): Cowork. In the same way that the spreadsheet was the reason to get a PC for the first time in the 1980s.

Justin Mannhardt (05:57): I was a big Apple fanboy in college, because I went to music college, and so if you were in a creative space, you were always on Apple. But then I got in the workforce, it was like, we're all on PCs. I also have a Mac coming.

Rob Collie (06:12): What's the impetus for you? I don't think it was Cowork.

Justin Mannhardt (06:14): No. So, I have Cowork, and I just haven't applied myself to trying to use it because I realized it couldn't do exactly all the things I was doing in Claude Code, but I absolutely see what you're describing as this inflection point for the world in terms of what AI is really meant to be. I'm getting a Mac because I want to jump on the whole OpenClaw fiasco, and get a little more hands-on with that. And also, I'm on the wait list for if you saw Perplexity is releasing something in a similar vein as OpenClaw that we're interested in trying out.

Rob Collie (06:50): OpenClaw, formerly known as Claudebot. I didn't know the name had changed three times already.

Justin Mannhardt (06:56): And then that person was hired by OpenAI.

Rob Collie (06:59): Yeah, I saw that, which totally makes sense, right? OpenAI is lagging behind in all of this tool usage stuff, let's go grab any shortcut we can. There's no way they were going to be outbid for that person.

Justin Mannhardt (07:12): Apparently he was quite the free agent. Zuckerberg tried to come make him a big offer and-

Rob Collie (07:16): Yahtzee for him. Okay. So, I ordered a Mac, in the meantime, I had Claude Code working on fixing my desktop, he's looking at the logs of Claude Cowork.

Justin Mannhardt (07:26): Try and figure out what's wrong.

Rob Collie (07:28): I'd been peeling that onion with no success. But I had kicked it off to go try one more layer of the onion, and I walked away and didn't look at it for days. Came back to it the other day, it did do something, let me try and see if it works now, and it does.

Justin Mannhardt (07:43): Incredible.

Rob Collie (07:46): My desktop works now and doesn't die... It's died once and all I needed to do was kill Claude and start it over. Whereas the other one seriously was dying 11 times an hour. So, now I have a Mac that at home I don't really need yet, if I'm using my laptop, I can still remote desktop to the desktop computer upstairs, and that's fine. When I go to a coffee shop or something to write, I'm going to need the Mac. In the meantime, this week, and this was the thing I texted you about, Microsoft announces Copilot Cowork. And I think your first reaction to it was the same as mine, which was, wait, are they allowed to use the Cowork name?

(08:26): Because that's an Anthropic thing, right? And if you read the press releases and blog posts and stuff about this, in the first 90% of the article, the word Anthropic does not appear. It's near the end of the articles that they acknowledge we're working very closely with Anthropic on this, and blah, blah, blah. So, this is Cowork, we haven't seen it yet, doesn't release until the end of the month.

Justin Mannhardt (08:50): I think even then in a limited capacity, if I understood it all.

Rob Collie (08:53): I'm going to be watching this with great interest to see how it actually works. It's the Cowork experience that I've been willing to crawl over broken glass to get to, with batch files and patience and frustration and ordering Macs. And it's that big a deal. But it runs in the cloud, so there aren't any of these VM issues, there aren't even the install it on the Mac. You don't have to even do that. It's just there, which is very, very, very appliance-like, very, very, very accessible. And because it's running in the cloud, they've done some significant integration work between Cowork and O365, the whole Office tenant and platform and everything, right? Because it's running there, it has access to much more than what's on your computer. Cowork is all about point me to a folder on your local computer, and we'll work with the stuff in there.

Justin Mannhardt (09:50): Which is a big deal.

Rob Collie (09:52): Huge deal.

Justin Mannhardt (09:52): Just to kind of paint the picture of the progression, I think still the vast majority of people... And this is why this is an inflection point. Understand AI as something I can go ask questions to and get answers. Something I can thought partner with, something that can riff with me, something that can write with me. But once you step into something that can go complete work on your behalf, go grab things, change things, you have a very different experience, and Claude Code, and Codex, and some of these other AI coding tools, they came out and they're a bit gatekeepy, right? Because they have the word code in them. Even though knowledge workers could get going with them and do a lot of amazing things, I think Cowork and its variants are just, they slip right into that world where there's just mass appeal for sure.

Rob Collie (10:41): Cowork is named and designed from the start to be a knowledge worker tool. Even me, I've got Claude Code right now ripping through all of the zipped up PDFs of handwritten scoresheets from the Indie Inline Hockey League, the old seasons.

Justin Mannhardt (10:57): Do you remember when we first tried to do that with ChatGPT 3 or something?

Rob Collie (11:00): Yeah, it was awful, right?

Justin Mannhardt (11:01): It was so bad.

Rob Collie (11:02): But it's going to win, right?

Justin Mannhardt (11:04): Yeah.

Rob Collie (11:04): We're going to unite the two stats regimes. I have that power up my hands, I don't really know why I can't pin it down. I would never use Claude Code to do this book editing project that I've been doing with Claude Cowork. I think I could. Using VS code is awful, there's so much formatting that you want to be able to see and stuff like that. The Claude desktop experience is so much better. But it occurs to me that I've never used the Claude desktop version of Claude Code.

Justin Mannhardt (11:36): Yeah, it's a better UI, feel, it feels more natural unless-

Rob Collie (11:40): Yeah. And I don't need the VS environment, I'm not using it for anything else. Anyway, even for me, again, I've got Claude Code sitting there and yet I'm fighting my way through batch files and restarts in order to have Claude Code working. So, people are used to AI as a riff buddy, or a Q&A bot, or more sophisticated things like research, or make me a plan for this thing that I want to do. But again, it needs to be plans about stuff that's public knowledge. Plans about my business, now you're stuck. The off the shelf experience doesn't really do a great job of that, it never will. Here's a cool story, I think it underlines your point. I have a friend, his name's Kevin, very bright. He's got a master's degree in engineering and everything, right? Because the career path he's on right now isn't forcing him into AI like it's been forcing us.

(12:37): So, he gets the Claude subscription for $20 a month, right? That was a big change for him. And he texted me in the middle of the day yesterday and he says, "Claude put me in timeout." He was trying out Projects, he was trying out, I think he was trying out Cowork as well. He's like, "Put me in timeout." And I go, "You need the $100 plan." And he goes, "Oh, this didn't happen to me with ChatGPT, $100 a month is too rich for me. I'm out." This really highlights the points you're making. $100 a month for the experience he's used to having, of that riff bot or the Q&A bot, that's too much. Why would I pay $100 for that? But he hasn't yet seen it doing things for him. At that rate, $100 a month is the best $100 he'll ever spend, but he's still kind of on the other side of that wall for the moment.

Justin Mannhardt (13:33): You need a catalyst to come over that wall. Just in the past week, since we last talked, I don't even know what the ROI multiple is on... Rob, I'm on the $200 a month plan. I'm on the big... But it's like, I'm getting, what we're able to do with this stuff, you wouldn't blink an eye at that cost at this point.

Rob Collie (13:54): No. I called him... Actually, he called me because he wanted to ask me questions. He's like, "So, I don't really understand Projects, and I don't really understand Cowork. I need to help me with this." I'm talking him through it and I said, "You could ask Claude this, and it will answer all these questions incredibly well." He's like, "But I'm in timeout." And I go, right, I know. In the future, if you're asking me questions that you could have asked Claude, the Rob Collie token subscription plan start at $5,000 a month.

Justin Mannhardt (14:21): I'm going to put you on time out.

Rob Collie (14:23): I'm not putting you on timeout, I'm going to start charging you. $5,000, that's what my brain costs a month for you to be able to phone a friend, right? Obviously I'm joking with him, but you're right, what's going to happen is I'm going to put him in timeout. I'm not going to answer those calls, right? I don't have time for that.

Justin Mannhardt (14:36): Maybe it's still out there, but there was a website that was just, letmegooglethatforyou.com.

Rob Collie (14:40): Yeah, right, right.

Justin Mannhardt (14:42): And you'd get a question at work and you'd respond with the link so they'd have to watch the question get typed into Google.

Rob Collie (14:48): I see this guy's future, he's so sharp, and was always a tinkerer in college. I wasn't a tinkerer. He's going to be another species. The version... He doesn't know what four weeks from now Kevin is going to look like. He has no idea what the ride he's about to go on, and he's going to go on that ride. I got off the phone with him, and I was just laughing so hard, I couldn't stop laughing as I was telling Jocelyn. I'm like, "Oh, Kevin, $100 a month, my sweet summer child, you aren't going to be thinking 100 is nothing, you're going to be like, I need the 200 plan." I can see it, but I don't want to tell him that he's going to be on the $200 plan because that'll scare him off. In the future, he's going to be like, "God, I wish I could spend more." The ROI on it is like, oh man, I would spend more, but I run out of time to direct it.

Justin Mannhardt (15:43): When I was still back at P3, we even had clients that were like, "We don't care, we'll approve the meter going as high as it needs to go for this person to have..." I was like, "It doesn't work that way."

Rob Collie (15:54): Did you see what you did there, "When I was still back at P3"? You were talking about it like it was like six months ago.

Justin Mannhardt (16:00): It was a long time ago. Maybe not the best phrasing. I don't know.

Rob Collie (16:05): Dude, you were just here.

Justin Mannhardt (16:10): I was just there. Back in my day.

Rob Collie (16:12): And I am right here, you're saying that right-

Justin Mannhardt (16:16): I'm sorry. Unintentional.

Rob Collie (16:18): It was beautiful.

Justin Mannhardt (16:19): Unintentional. The point is people be like, "I don't care what it costs, it's worth it."

Rob Collie (16:25): Yeah. Nathan here is spending $100 a day.

Justin Mannhardt (16:29): I bet. I believe it.

Rob Collie (16:30): Right? And is this a problem? No. This is a sign of amazing things happen. Imagine what $100 a day of tokens of Claude in the hands of someone like ... Now, people listening don't know Nathan, but we do. It's more like, Nathan, is there a way you could spend more? I've had this dawning realization, to be clear, that I think Copilot Cowork from Microsoft will be the single most successful and compelling Microsoft AI offering, probably ever. I would say that even with regard to the original Copilot that pioneered the writing of code in Visual Studio, this is going to be a major win for them, and it makes sense both for Anthropic and for Microsoft. Anthropic wasn't going to get this kind of unfettered integration access with the Microsoft properties without a deal.

(17:34): We've been talking forever about Microsoft having a really, really strong home field advantage, and they get 30 at bat strikes instead of three. This is them, they've taken 20 pitches, there are 20 strikes, and then they just hit a grand slam. I think this is going to be just super, super, super significant for them. And it's something that only Microsoft is positioned to do, and amazing.

Justin Mannhardt (17:59): Well, I think the way that they're announcing the commercial terms for this also confirms that. For those that weren't really on the inside baseball, to use the strikes analogy again, when we would work with Microsoft on a Power BI project for a common client, a lot of that had to do with Microsoft moving that client from one license package to the other. And that new package included Power BI and Power Apps and all these things, so they needed someone like P3 Adaptive to come in and help them realize value from all these new tools that they've got. So, you can see this coming again, where they're pushing this value of Cowork as being the single greatest value lever that exists in that ecosystem. You know what I mean?

Rob Collie (18:48): It's the reason for the next SKU. Funny aside there, when I worked on Office at Microsoft, worked on Excel, up until a certain point, I think that this went away, but there was a Standard SKU and a Pro SKU, this was back before subscriptions. These are the disks that you would buy, or the right to deploy it on a network or whatever. So, you buy Pro versus Standard, and for many years, the difference between Pro and Standard was that Pro contained Access and Standard didn't.

Justin Mannhardt (19:19): The database product?

Rob Collie (19:20): Yes. So, that was basically Pro was the Access SKU, right? They could have called it Office with Access, but they called it Pro, and then there was Office Standard, which came without Access. When IT started to become a real thing in the world, the first thing that IT did because they were held liable for all of these desktop solutions, right? The first thing IT did in the mid '90s was declare war on Excel and Access, like, we got to get that stuff out of there. Well, Excel thrived under attack, it just multiplied, whereas IT was very effective at stamping out Access, especially in enterprise environments. And this eroded the reason to buy Pro.

(20:02): In working with office marketing, they started to... This is the thing that made it such a pain in the butt for us, was that they came to us and started asking us to remove features from Excel and other apps in the Standard SKU, to prop it up. They're like, "Look, we just lost Access as the reason people buy Pro."

Justin Mannhardt (20:22): We need a reason.

Rob Collie (20:24): Naive me at the time said, "So, what's wrong with them buying Standard?" And they say, "Look..."

Justin Mannhardt (20:32): Oh, Rob.

Rob Collie (20:32): "Listen to you little youngster." They said, "No one buys the Standard SKU ever. The Standard SKU exists so that we can have a Pro SKU. The Standard SKU exists so that we can charge more for the Pro SKU, that is the only purpose of it. If people start buying the Standard SKU, we would rather not have one." And a relatively misguided attempt to shore up that product, they had to start taking features out of Standard Excel. So, now if you used a feature in Pro Excel, and then sent the spreadsheet to someone with Standard Excel, it doesn't guarantee that it always worked. I was dumb to ask the question I asked, but the plan that we implemented was dumb in another way.

(21:12): And so, eventually they stopped doing this, but it's like every time they got a new crop of marketers and strategists in there, they would try the same idea because they hadn't learned the lessons from the last crowd had moved on. So, the one thing about Copilot Cowork that I'm concerned about is that I think whatever subscription rate they charge probably isn't going to be enough to cover their cost. Do you know off the top of your head like what the Copilot for 365 per user cost is? I forget.

Justin Mannhardt (21:45): I absolutely do. The Copilot license on its own was around $30 a month, and now there's a new all in license at Microsoft. So, you're familiar with the term E5, which would include Office, and email, and Power BI, and bunch of stuff, right?

Rob Collie (22:09): Yeah.

Justin Mannhardt (22:09): So, now there's this thing called an E7 that's $100 a month-

Rob Collie (22:14): Oooh. Okay,

Justin Mannhardt (22:14): ... that includes basically everything you could get for the individual user at Microsoft and also Copilot. It effectively comes off as a discount on the Copilot stuff, if you have all that other stuff. So, let's just say it's somewhere between 20 and $30 a month for the Copilot stuff.

Rob Collie (22:30): I'm not sure that's enough.

Justin Mannhardt (22:31): I wish I would've saved this so I could link to it. There's been some research going around, hypothesizing what the per user cost of these types of subscriptions actually is. And I don't know how much of it is based in fact or speculation, but it was something on the order of, a Claude Mac subscriber actually costs something like one to $2,000 a month.

Rob Collie (22:57): Oh.

Justin Mannhardt (22:58): This problem for sure applies.

Rob Collie (23:01): That is a dramatic reveal.

Justin Mannhardt (23:03): Yeah.

Rob Collie (23:04): So, I was just basing my that's not enough on existing pricing that I'm familiar with with other plans. $20 a month, Claude, my friend who still is just scratching the surface, gets put in time out on his first day at $20 a month. So, in the 20 to $30 range, again, if you just assume that that is accurate pricing for a moment, let's assume that it's not lost leader, which is what you're hinting at.

Justin Mannhardt (23:33): Right. That research might be a bit hyperbolic and people are disputing and arguing, but the core idea here I think is good. The worry about does this cover the cost is real.

Rob Collie (23:42): I think the only way it works financially for Microsoft is if most of the users on the tenant aren't really using it.

Justin Mannhardt (23:51): Again, the articles I read were unclear... These were third party articles that were hinting on the announcement that had come up from Microsoft. I think there is still some extra consumption cost. I think there's still per message, per token type things that people might need to be accounting for at the company level.

Rob Collie (24:12): Just like we have overages with our Claude subscription at our company.

Justin Mannhardt (24:16): Right.

Rob Collie (24:17): Well, that would make sense to me.

Justin Mannhardt (24:19): This feels we've not had hands on keyboards with this thing yet, we know enough about how things like Claude Code and Cowork work that we understand like this is the right next step. I think even Satya came out and admitted like, "Listen, Copilot isn't the way people should be working with AI." Or any chat-based on its own AI is not really where this is headed. There's just so many people out there that have been really critical of Microsoft because they're not competing with Anthropic or OpenAI, they're not building models, they're not coming out with LLMs. And I think that's intentional, if I were to guess, because if they wanted to, they could.

(24:58): They have enough money, power, and resources to go in that direction, but I think their ability to leverage the graph, the completest state of everything in Microsoft, and the way people work today, if they can get the way Cowork interfaces with all these things working really, really well, they have something that can cook for sure.

Rob Collie (25:23): And we haven't really talked about what Cowork does, and we haven't defined the graph. So, the graph is everything that's in your OneDrive, everything that's in your SharePoints, everything in your email, everything in your Teams.

Justin Mannhardt (25:37): Everything in your calendar.

Rob Collie (25:38): Right? And so, an agent, a buddy that has access to all of that, not just the stuff that you have access to.

Justin Mannhardt (25:45): That's another big value prop for Microsoft because the fact that that's inherited is huge.

Rob Collie (25:51): Yeah. And so, if you want to tell it, go look at all my meetings from the past two weeks and make me a spreadsheet of all the times that we talked about topic X, it should be able to do that. And that is a sea change for people. Can I tell you about my Cowork editor for the book?

Justin Mannhardt (26:14): This is the thing that glitches all the time that led you to buy a Mac.

Rob Collie (26:16): This is the thing that made me buy a Mac. Jokes on me in a way, because my first experience with Cowork was Copilot Cowork. I might not have ever wanted the version that runs on my desktop.

Justin Mannhardt (26:29): Well, if you have the energy for it when you get access to Copilot Cowork, it'd be a cool compare and contrast.

Rob Collie (26:36): Yeah. I'm 100% going to do that. Because the thing is, is that all the docs I'm working with are in OneDrive to begin with. The project folder for Cowork on my desktop is a OneDrive folder. So, I'm editing the Word docs, for instance, on different computers, using the Office apps in OneDrive, the auto sync between, the multi-editing and all that kind of stuff. And I do need to tell those folders to always keep on this device, always download the full files in OneDrive, so that when Cowork goes to look for them, they're actually there, and not... Through the VM, I don't think it can force the download of them, they have to actually be there. And all the previous books, I didn't really have an editor, I had proofreaders.

(27:22): I did have some technical editors, right? My friend Scott, technical edited the first one to keep me honest and all that kind of stuff. But I never had an editor editor, probably because we didn't want to pay one, it was kind of a bootstrapped book. But also because any editor that we hired would have been incompatible with my writing style. They would have over and over and over again said, Rob, this is too informal, you need to fix this. Or they would have just felt sidelined. If they'd gone along with my style, they would have felt... It would have been almost like over their objections.

Justin Mannhardt (28:01): They wouldn't have been in it.

Rob Collie (28:02): Yeah, you wouldn't have gotten someone's best. The voice of my book ended up being a really great differentiator for the books. I didn't plan that, but that did work out that way. Sitting down today, I thought, well, this book I want to be a potential bestseller on the business bookshelf. I'm not playing in the, I'm writing for formula writer's circle anymore. The standards are higher. But I still want my voice. So, Cowork, I've "trained it," and remember, training is just essentially like a file in the folder of instructions to Claude Opus 4.6. Every time Cowork wakes up for me, every time an editor wakes up for me, it gets hydrated with these instructions, and that's what turns Opus 4.6 into Rob's editor buddy. Rob's editor buddy is amazing. If I go and I write a bunch... I write let's say four pages in Word, which ends up being much more than four pages in printed book form.

(29:10): So, four pages in Word, I launch Claude Cowork, it's already pointed at the folder full of all my Word docs, and I say, hey, I just wrote a bunch in Chapter 08, go take a look. And Cowork, you can't help but watch it work sometimes, like watching the Roomba. You're supposed to go do something else, but I just sit there and I watch it do it's stuff.

Justin Mannhardt (29:31): [inaudible 00:29:31]. What are you doing?

Rob Collie (29:31): It unpacks the whole Word file, and it also looks at all of the styles that are used because the word styling in the book is, there are styles like headings, and bulleted list, and image placeholder, and all these different styles that are semantically meaningful, it's not just the words there. This editor understands the styles and what they intend, and uses that as part of its understanding. Like, oh, this is a note, this is a call-out note. This is a pull quote. If it was just reading the words, it would be missing something really important. So, it's got all the definitions. I didn't give it all those definitions, I told it to go examine them and figure it out. So, all these instructions are largely written by it, one of my own little quirks.

(30:23): I want to get advice on this book in terms of the competitive landscape, because I believe that it fills a niche that really no one's filling. And so, I have it periodically go and research, boil the ocean of all the books that are out there and reviews and everything, but I've told it, I don't want to know about specific books. I find other people's work that's already done and already good, for some reason, it's deflating to me. So, I've told my editor to coach me on making sure that I'm in the lane that I want to be in, vis-a-vis the competition, but shield me from too much detail about that so that my fragile motivation... It takes a lot of energy to do this.

(31:11): I need a clean room. I've been very careful to tell this thing, never write for me. But I have given it instructions on generating three to four bullet summaries at the end of each chapter, which is something that people really benefit from. But I've never been able to do because it's too exhausting after finishing a chapter, I don't have the energy to go back and extract all that. It would take me probably like an hour or two to do a good job of that. It's now got instructions on how short those should be, how many of them there should be, how they should be formatted, and it just goes and does it.

Justin Mannhardt (31:47): Do you tweak?

Rob Collie (31:49): Oh yeah, I paste it in and then I tweak.

Justin Mannhardt (31:51): Cool.

Rob Collie (31:52): And sometimes significantly. But it does get me off the blank page, crucially, it's already summarizing things I already said, it's not coming up with stuff.

Justin Mannhardt (32:00): It's that concept of you've done the first draft and it's...

Rob Collie (32:03): Like really done a first draft, right? Like a 3000 word first draft. I'm okay with it pulling out three bullet points. Other things, I have it set up where I can tell it, if we have an exchange, me and the editor, that I think is really funny or kind of mind-blowing, I tell it, okay, save that one. And it writes a transcript of that section of the chat to another file so I have it for later. So, maybe I'll do like a bloopers reel at the end of the book, or maybe I'll just use it in promotions for the book, or just on LinkedIn or whatever. But think about that. That idea would float through your head, like, oh, some of these would be really cool to save. If you didn't have AI to help you do it, you'd just pass on that idea. There's an extra value being created here that wasn't before.

Justin Mannhardt (32:55): This is a great example, Rob. What hopefully Copilot Cowork also starts to enable for people is AI that can do something that's a repeatable task, with very specific instruction, contextualized to you and your business, do it in this very productive, compelling, creative way is way different than rolling up into ChatGPT or even Claude and being like, hey, I got this email from Rob, and it's got a cool idea, I want to build a thing around it. You'll make progress. We talked about this I think a couple episodes ago, how I shifted from doing the first drafts myself, and I found that to be immensely helpful to just grind through an idea that then opens the floodgates of creating all kinds of other stuff that doesn't drift in a significant way. It does drift, but it's much easier to bring it back.

Rob Collie (33:52): And writing this book, I'm doing it the really hard way. I'm trying to write the final draft myself.

Justin Mannhardt (34:03): Now, we should applaud you for it because I don't think it would be good if you just had AI write it.

Rob Collie (34:08): I don't think I'd even stay engaged, to be perfectly honest. You'd have to have such a distance from the project, like an indifference to it, that you'd be lying to yourself, and I would just know that I'm not really into this. One of the blooper reels or whatever, the highlights that I want to capture in that file... And I need to circle back and find the conversation so I can do this. This thing is so impressive, it's given me all kinds of advice. It's keeping me honest, it's like, "This section here, Rob, you should tighten this up, man. You're burying the lead," or whatever. The quality of my writing is going to be so much higher on net than it used to be. I've always been a good writer, I think a really good writer, actually, I would even go that far.

(34:53): Now I'm better because of it. When you have this relationship with this thing that is doing that for you, you just sort of inevitably... I do. I start to kind of almost put it above me. It's the authority in a way that I'm going to, and I'm listening to its advice. I'm not listening to all of it, I push back and say, "No, I'm not going to change that, that's going to stay as is. That's okay," or, "That's better the way... And the objection you're raising, editor, is a lesser problem than if I fixed it." But you get this relationship where you're almost having a respect for this thing that's really almost feeling a little bit senior. So, I finally just said, "You know what? Could you write this book?"

Justin Mannhardt (35:35): You asked it?

Rob Collie (35:36): "Could you have written this book?" I'm not asking it to do it for me, but I'm like, "Could you have written this book?" It was honest, right? It's like, "Rob, yeah, I can explain all these topics." I could write a B+ book on this. But then it ran through this list of things that it couldn't do, that I have been doing. And I knew it couldn't write this book, but it went so far in this relationship, I just had to stop and ask.

Justin Mannhardt (36:08): Because you're writing for a human audience, you're not writing for an AI audience. And human audiences, they appreciate the style, and the take, and the opinion, and all the things that come from you that, if you were just trying to educate people about topics, of course.

Rob Collie (36:27): As usual, I'm doing the things that keep me interested, little jokes, little-

Justin Mannhardt (36:31): Movie quotes? Got to be a lot of movie quotes?

Rob Collie (36:35): There's a conversation in the book where I'm having with Claude where I say, "Hi." Claude says, "Okay, what's up?" And I say, "Just contemplating the ifs," Pulp Fiction line. And it doesn't know what movie. Opus 4.6 doesn't know that that's a Pulp Fiction quote. But it does know some things about me in its pre-training. It also knew not to go search the web, it wasn't important enough to go search the web. It's like, "No, just tell me what the movie is, Rob."

Justin Mannhardt (37:05): And don't make me look it up.

Rob Collie (37:07): There's definitely movie quotes. One of the really cool thing I have it doing for me is, when you're writing a book, you're building layer on layer on layer of concepts. And so, you inevitably stumble into things that would not be kind to the reader if you explained them right then. They need to wait for that layer of the cake later. You need to control the [inaudible 00:37:30] so that they can learn. But you find yourself, that something comes up and you have to say something about it. So, you make a promise to pay it off later. This happens over and over and over in the book. So, one of the things I have the editor doing is cataloging all of the promises, all the forward promises I've made, and making sure that I pay them off.

(37:51): And another thing I have it doing, I told you about the bullet points it writes at the end of each chapter, that process has access to the entire index of the book, sort of like a meta summary of the entire book, and I noticed it would occasionally leak things from future chapters into the summary of that chapter.

Justin Mannhardt (38:06): Okay. Yeah. Why wouldn't it? It's in the context.

Rob Collie (38:09): It's in the context, right? And so I told it, okay, from now on, you can leak things from prior chapters into the summaries, but you can't leak things from future chapters, problem solved.

Justin Mannhardt (38:21): That's really cool, it's a really cool story of how you're creating a system more than you're creating an agent, or more than you're just using a chat. Leading back to Cowork, there's these additional layers and levels on what we all remember when ChatGPT first rolled on the scene.

Rob Collie (38:42): I'm in the writing the tool usage chapter, introducing the concept of tool usage. And when you stop to explain it at a fundamental level, it really helps put perspective on it. I take a step back and go, yeah, now that is different. The LLM taking action. We took it for granted for so long, like web search, it was right there under our noses the whole time, right? Long before you discovered MCP servers and all that kind of stuff, you were already just accepting that it did web search, not really knowing at that moment that the world changed dramatically when it started searching the web in response to your question. It's like, ooh, yeah. So, I got a chance to kind of re-experience that retroactively writing this chapter. I'm like, oh wow.

Justin Mannhardt (39:31): It's come a long way in a very short period of time.

Rob Collie (39:34): It really has.

Justin Mannhardt (39:35): I hope Copilot Cowork does bring more people into the party. I hope it's done well.

Rob Collie (39:43): If it's not done well at launch, it's the age-old Microsoft thing. They will get it, they will figure out the pricing. All of these problems, even the things that I'm concerned about are things that will be solved. They might have theirselves in introductory pricing and long-term pricing problem, especially if what you're saying in that research report you're talking about, where if they're only covering a fraction of their costs in general-

Justin Mannhardt (40:08): I think there's truth that there's an economic issue with AI at the moment, but when I look at problems and I see, okay, there is massive incentive for this problem to be solved, this will correct itself.

Rob Collie (40:21): I don't worry about it either, I'm in the same action taking place as you. Really, there's not any practical daylight between your take and my take, because we both end up behaving exactly the same way. My take on it is that they might not get the costs down until some massive qualitative change in the way that they do AI. In researching the book, when you ask it one question, it runs hundreds of trillions of calculations. There's just a physics to this. And again, it kind of brings you back to marveling about the efficiency of the human brain to perform the same kind of thought. We are using electrical impulses, but we're really, really, really low voltage. We don't require gigantic cooling fans. Nature has designed a piece of hardware that we can't build yet. Doesn't mean it can't be built, the current trajectory that we're on, today's LLMs are still very much direct descendants of the original ChatGPT. It's down the same evolutionary tree.

Justin Mannhardt (41:24): Did you tell me about there's some experimentation with etching the transformer directly on the chip? It was something like that where the tokens per second is way higher.

Rob Collie (41:36): So, I don't think this was me, but there is something else that's related to this, which is Google has their own hardware.

Justin Mannhardt (41:44): That's right.

Rob Collie (41:45): So, GPUs were already laying around as a really good fit type of hardware to process a large volume of mathematical calculations in parallel, a large volume of identical also mathematical calculations in parallel. That's what GPUs do. So, they were a great fit for AI. CPUs could do AI, they just take much, much, much longer and cost much, much, much more. So, Google has their own thing called a TPU. It's like a GPU, but doesn't have to pass the data back into memory each time. Because the GPU, even the GPU running everything in parallel, it has to feed things through in multiple passes. You get the answers to one set of questions, and then you feed it back in, right? And the TPU has it so that it can do that, it's got its own daisy chain series of hardware within the chip. It just flows like a wave, bing, bing, bing, bing, bing, bing, bing, bing, bing, all the way through end to end, doesn't have to go back out to memory, which costs time, costs money, costs power.

(42:48): Rather than sell these chips on the marketplace, they've kept them for themselves, which is interesting. We're going to see hardware changes, but again, even that is just really just an incremental optimization on the same hardware, on the same basic approach. So, anyway, I was saying, they might not ever solve the cost problem, or at least not on this trajectory, but the thing is we're going to learn that it's worth it. If we end up having to pay a couple thousand dollars a month for a subscription to something this productive, you and I both know that like right now the pricing, we kind of feel like we're getting away with something.

Justin Mannhardt (43:23): I'm going to keep getting away with it then.

Rob Collie (43:26): Until further notice, right?

Justin Mannhardt (43:27): Go as far and as fast as possible.

Rob Collie (43:30): The breaking bad meme, "He can't just keep getting away with this." Yeah. It might be that we have to pay a lot more eventually, but again, at that point we will. It'll be a good deal. All right, well, this was the fun one, I was a lot higher energy than the last time. I'm still sick, but not nearly as sick.

Justin Mannhardt (43:49): I sense you being on the upward track here after today.