×
Evaluate Your Team's AI Readiness

Take our official assessment and explore training paths to build secure, compliant AI workflows.

Access Portal

AI at Work Podcast

Ep #3 Quit Typing. Start Talking: Voice, Vibe Coding, and the Real ROI of AI (No, 95% Don’t Fail)

AI at Work, AI in Business

Audio episode

Listen to this episode

Summary

The conversation explores the future of artificial general intelligence (AGI) and robotics, particularly focusing on the emergence of humanoid robots. The speakers discuss the likelihood of AGI being a part of the technological landscape in the next decade and the potential for humanoid robots to become commonplace in various sectors, moving beyond simple tasks to more complex functionalities.

Takeaways

It’s increasingly likely that the 10-year picture will include AGI.

Humanoid robots will likely be prevalent in everyday life.

Robotics will evolve to perform tasks beyond current capabilities.

The conversation highlights the distinction between basic automation and advanced humanoid robotics.

The speakers express confidence in the advancements of robotics in the next decade.

Humanoid robots may take on roles that require more than just lifting boxes.

The discussion emphasizes the functional aspects of robotics in various industries.

There is a growing expectation for humanoid robots to integrate into daily activities.

The future of robotics is seen as a blend of AGI and physical capabilities.

The timeline for these advancements is projected to be within the next ten years.

Chapters

00:00 Early mornings, parenting exhaustion & AI bedtime hack
04:19 What is “superintelligence”? (AGI vs. SI)
08:12 The singularity explained & unpredictability
10:33 OpenAI progress claims, safety & global coordination
12:05 Planning beyond today’s AI; 10-year robot horizon
18:50 Deloitte refund cautionary tale (fake citations)
22:49 Company guardrails: policies, manifesto & non-abdication
26:04 Free vs. paid AI models; data control & opt-outs
31:48 Building on frontier models: unpredictability & guardrails
33:46 Measuring drift, compliance, and production monitoring
35:59 ROI studies: MIT controversy vs. Wharton findings
41:14 Where ROI shows up: content, research, knowledge mgmt
44:14 AI dictation as a workflow unlock
50:01 “Year of the agent” & voice in field ops
54:07 “Vibe coding” overview (Replit/Lovable vs. dev tools)
56:27 Scope first: why non-dev builds stall at 85%
59:16 Google’s AI Studio/Opal strengths & ecosystem fit
01:02:43 Using one model to scope prompts for another
01:04:07 Prompt hygiene: “one question at a time”
01:07:14 Next episode teaser: share real workflows

 

 

 

Explore practical AI implementation: https://assessment.ascendlabs.ai/

Book a conversation with Kevin: tidycal.com/kevinwilliams

Show Notes & Links: https://www.spark6.com/podcast

Submit listener questions:

[email protected]

[email protected]

Check out Kevin’s stuff:

AscendAI: https://ascendlabs.ai/

Follow Kevin on LinkedIn: https://www.linkedin.com/in/kevinguywilliams/

Check out Eli’s Stuff:

SPARK6 Agency: https://www.spark6.com/

Sign up for FREE AI Framework Friday Newsletter: https://www.spark6.com/newsletter

Follow Elijah on LinkedIn: https://www.linkedin.com/in/elijahszasz/

Transcript and citation details

Transcript

00:00 Play from here Welcome to AI at Work.

00:06 Play from here I'm Elijah Soss.

00:07 Play from here I'm a serial entrepreneur with 28 years of experience in tech, and today I run an agency

00:12 Play from here that builds AI solutions.

00:13 Play from here Hey, I'm Kevin Williams.

00:15 Play from here I spend most of my time helping companies figure out how to actually use AI, not in

00:20 Play from here theory, but in the work itself.

00:23 Play from here What's up, Kevin?

00:25 Play from here Good to see you.

00:28 Play from here You were waking up at 4.45 today, were you?

00:30 Play from here I was.

00:31 Play from here I was.

00:32 Play from here It's my way these days.

00:33 Play from here Yeah, it's awful.

00:35 Play from here It's something that I've just come to accept with getting older, is I'm just going to keep

00:37 Play from here waking up earlier and earlier every year, except my kids don't want to go to sleep,

00:42 Play from here so I'm just perpetually exhausted.

00:43 Play from here I think there's a point of divorce, too, that you have to watch out for.

00:46 Play from here What?

00:47 Play from here Did you just throw the D-word in there?

00:49 Play from here What's going on?

00:52 Play from here My wife will tolerate a lot, but 4.45 is not something that she's too keen on.

00:56 Play from here Yeah, I hear that.

00:59 Play from here You know what my wife is not too keen on is after I've been up for 18 hours and I just

01:01 Play from here completely tap out of parenting and slowly slink into a corner and try to just lick my

01:06 Play from here wounds from the day.

01:09 Play from here Happened last night, and I used AI to fix it.

01:10 Play from here You want to hear how I used AI to fix it?

01:12 Play from here Does it involve a neo-humanoid robot?

01:15 Play from here I sure hope so.

01:17 Play from here I'm not sure I fixed it, and I'll preface it by saying it's probably not my proudest

01:21 Play from here parenting moment.

01:23 Play from here Your kids are older, so I don't know if these books were around where it is this tomb of

01:28 Play from here really intricate illustrations.

01:32 Play from here There's bits of text that are always low contrast, like light yellow on light blue text that

01:37 Play from here says find the 100 unicorns on this page, and then there'll be the next one.

01:43 Play from here Find 17 magic wands over here.

01:46 Play from here I can not only read it, but this book is about the size of War and Peace, so it just never

01:51 Play from here ever ends either, and I just can't even think anymore, let alone see.

01:56 Play from here So I took out my phone.

01:57 Play from here I opened up GPT.

01:58 Play from here I took a picture of the book, turned it on the voice mode, just put the phone on my stomach,

02:04 Play from here and first of all, kind of spooky, it knew my daughter's name from some other conversation

02:09 Play from here in the past.

02:10 Play from here I don't know how.

02:11 Play from here This was a completely new threat, and so it said, oh, hey, Luna, first we're going to

02:15 Play from here find the 17 unicorns, and it just walked her through the entire thing, just from taking

02:21 Play from here the picture of the book, but it also did it in a really weird way.

02:27 Play from here This is probably going to be a whole topic for us at some point, but you can go into

02:30 Play from here the system settings for your account and choose all sorts of things, like please don't be

02:34 Play from here so sycophantic.

02:36 Play from here Don't congratulate me all the time.

02:38 Play from here Even little sliders for how much flowery language do you want versus just giving it to you straight,

02:44 Play from here and of course, I put all my stuff to just please, please stop and just give me the information.

02:50 Play from here So it's taking her along this little journey through this unicorn land, but it starts everything

02:58 Play from here with, okay, we're going to have fun, but I'm not going to sugarcoat it.

03:01 Play from here We're just going to get right to it, and we're going to find these 15 hidden frogs.

03:07 Play from here So of course, my wife then walks back into the room while this whole scene is going on.

03:12 Play from here It just kind of looks at me and my daughter talking to an AI storyteller and just shakes

03:17 Play from here her head and walks out.

03:19 Play from here So there's my use for AI for the night.

03:21 Play from here It takes a village, and sometimes that village is an AI village.

03:24 Play from here Sometimes that village is a village of robots.

03:29 Play from here Another thing you can do, and this doesn't work as well because so far I've noted that

03:34 Play from here GPT is an awful storyteller, but I will say she's obsessed with My Little Pony, so I'll

03:38 Play from here say the setting is a quest area, whatever their land is called.

03:43 Play from here She's going to go on an adventure.

03:44 Play from here It's a choose your own adventure, so it has all these forks in the road.

03:47 Play from here Do you want to go through the magic portal, or do you want to go into the enchanted forest?

03:51 Play from here And it is awful.

03:52 Play from here I mean, just absolutely awful.

03:55 Play from here I have a difficult time listening to it, but I think the writing is on par with My Little

04:00 Play from here Pony, so for a five-year-old, she loves it.

04:03 Play from here Right.

04:04 Play from here You have provided.

04:05 Play from here Okay.

04:06 Play from here We're all done.

04:07 Play from here That's the value for this episode, how to be a bad father and let various subpar robots

04:13 Play from here just do all your parenting for you.

04:15 Play from here It's not about being the best father.

04:17 Play from here It's about being an adequate father.

04:18 Play from here Adequate.

04:19 Play from here I like it.

04:20 Play from here Adequate parenting.

04:21 Play from here The best way to reach your goals is to keep lowering your bar.

04:27 Play from here That's how we do it.

04:28 Play from here I'm going to further derail you, Kevin, because so we have this shared Google document, and

04:34 Play from here I saw your first comment this morning.

04:35 Play from here I hate this as a first topic.

04:37 Play from here Fine.

04:38 Play from here We won't talk about it.

04:39 Play from here We'll just talk about something that wasn't even on here at all, but I think it might

04:42 Play from here have been, and then our robots scrubbed it off of the document.

04:47 Play from here So OpenAI will publish these blog posts every so often, and the most recent one was about

04:53 Play from here superintelligence, and I was wondering if you caught that or if you caught wind of that

04:56 Play from here in the news.

04:57 Play from here You're going to have to elaborate for me.

04:59 Play from here Okay.

05:00 Play from here Okay.

05:01 Play from here So superintelligence as a topic is probably something that we could spend five or six

05:06 Play from here episodes talking about, starting with what is the definition of superintelligence.

05:11 Play from here Maybe that's a good place to start since it is not a really agreed-upon definition at

05:16 Play from here all right now.

05:17 Play from here Kevin, what's your definition of superintelligence?

05:20 Play from here So I love it.

05:21 Play from here I'm an AI trainer.

05:22 Play from here As a group, we've sort of just left beyond AGI, which is general intelligence, to this

05:31 Play from here idea of superintelligence, and I do think it's important to sort of calibrate a little

05:35 Play from here bit on what's a mutual understanding of general intelligence.

05:40 Play from here In my mind, it has to do with the autonomous completion of median tasks, so that if you

05:47 Play from here can interact with an agent of some kind and not just have it do one specific thing or

05:54 Play from here a group of chained things, but essentially tell it what the output is supposed to be

05:59 Play from here and have it figure out what that chain of other agents is such that it yields that output.

06:05 Play from here That puts it more on par with human behavior and the type of thing that you would look

06:10 Play from here to a team member to do or something like that.

06:14 Play from here Superintelligence is different, and it does go down a giant rabbit hole of the singularity

06:19 Play from here and the rapid increase of intelligence, and there are some great graphics out there as

06:25 Play from here far as how humans think about our relative intelligence in the grand scheme of things,

06:32 Play from here about the IQ of a chimp being right here, right in the bottom quarter, and the IQ of

06:40 Play from here an average human being a little bit higher, and then an absolutely genius human being

06:46 Play from here quite a bit higher than that.

06:48 Play from here And the reality is that all of those things are super compressed.

06:53 Play from here And in the case of superintelligence, one of the possibilities and one of the fears

06:59 Play from here is that when you get to self-replicating systems, you're going to tick over from genius

07:05 Play from here level and like eight and a half seconds later have an IQ of 1,200 or 1,500.

07:13 Play from here And what the heck is an IQ of 1,500?

07:17 Play from here What does that mean for us as far as how the systems are perceiving us?

07:23 Play from here How does that mean with all of our approaches that are out there right now?

07:28 Play from here But for me, the less dystopian view of superintelligence is it starts solving much, much, much bigger

07:36 Play from here problems.

07:37 Play from here So it's not about, hey, hey, cool, let's launch a new blog and I can turn to my AGI agent

07:44 Play from here and expect a reasonably good blog to come out the other end, but more like, hey, let's

07:51 Play from here go after cancer today.

07:53 Play from here And it's going to start sucking in all of the available data and in a relatively short

07:59 Play from here amount of time, be able to start processing that information such that the humans in the

08:05 Play from here loop have a really, really narrow focus on the things that they should be focusing on

08:09 Play from here with their centrifuges and test tubes and such.

08:12 Play from here How about you?

08:13 Play from here I think that was really good.

08:14 Play from here The only thing I'm going to poke at you for is you threw out the S word with no explanation.

08:20 Play from here I don't think very many people have even heard of the singularity.

08:24 Play from here They may not be Ray Kurzweil fans or Isaac Asimov fans or spend as much time thinking

08:32 Play from here about this stuff as we might.

08:34 Play from here So a quick primer on the singularity is some point in time, and there's naturally occurring

08:42 Play from here singularities in our universe today.

08:45 Play from here A black hole is a singularity, a waterfall is a singularity.

08:48 Play from here So you have this very predictable, approachable path to an event horizon, this line that you

08:55 Play from here cross.

08:56 Play from here And after you cross that line, it's completely unpredictable.

08:59 Play from here Nobody can discern what exactly is going to happen after that line.

09:03 Play from here So this idea of a singularity is now really being molded into what some people are calling

09:10 Play from here AGI or what some people call a superintelligence, that it's really unpredictable.

09:16 Play from here We don't know exactly what that future looks like.

09:19 Play from here You mentioned the silver lining versus the dystopian view, and there's a lot of both

09:24 Play from here out there right now.

09:26 Play from here But I think that dovetails nicely into what OpenAI posted, which was kind of a progress

09:33 Play from here report predicting that systems will soon become smart enough to make discoveries.

09:39 Play from here And they're beginning this rally or this cry for global coordination on safety or oversight.

09:48 Play from here Basically resilience as this tech nears the superintelligent territory because of that

09:55 Play from here unpredictability.

09:57 Play from here And the only reason why I bring this up is because I feel like we dip in and out of this

10:01 Play from here conversation as AI evolves.

10:05 Play from here You'll hear a lot about it in the news, and then you won't hear anything.

10:08 Play from here You'll hear about Elon Musk and Sam Altman fighting because one of them says, I'm the

10:14 Play from here steward of safety and responsible AI development.

10:17 Play from here You can't be trusted.

10:19 Play from here And we hear those conversations.

10:21 Play from here But I bring this up because I feel like in the last week, I've heard several different

10:25 Play from here angles, including this post where OpenAI said that their current AI systems already outperformed

10:31 Play from here top humans in very complex intellectual tasks and are 80% of the way to an AI researcher.

10:38 Play from here So when you ask me, well, what do you think superintelligence is or what do you think

10:42 Play from here AGI is?

10:43 Play from here I really think it's those recursive loops.

10:47 Play from here Its ability to do its own research better than one of these research scientists and

10:52 Play from here start to improve on itself, at which point, for all intents and purposes, that could be

10:57 Play from here a runaway train.

10:58 Play from here When you start to scale those loops.

11:02 Play from here So that came out.

11:05 Play from here And yeah, Elon Musk was granted this absolutely insane trillion dollar package, but it wasn't

11:12 Play from here about the money.

11:13 Play from here It was about control because he says that Tesla is really a humanoid robot company and

11:20 Play from here that XAI is going to be the brain for those robots.

11:24 Play from here And he predicts there to be way more robots on the planet than people in the not so distant

11:29 Play from here future.

11:30 Play from here And that also bridges that conversation to superintelligence and safety and guardrails

11:35 Play from here and whoever really controls that domain controls the future of humanity.

11:40 Play from here And so that's very, very heavy topics, I know.

11:44 Play from here But there's just been a strange convergence in the news of this coming out.

11:49 Play from here And so I thought it was worth mentioning, especially since OpenAI posted it on their

11:52 Play from here blog.

11:54 Play from here From a practical AI or AI at work perspective, the takeaway is that these conversations have

12:01 Play from here been happening since AI was conceived.

12:04 Play from here So some of the conversations about the singularity have been going on for 50, 60 years.

12:10 Play from here And there was this period of time in particularly 2023, where it was like the thing at gatherings

12:16 Play from here to talk about what we were calling the P-Doom at the time, which is the percentage between

12:22 Play from here zero and 100 of the odds that this is going to kill us all.

12:26 Play from here And what was shocking about that is pretty much every major AI researcher had a positive

12:32 Play from here P-Doom.

12:33 Play from here Some were quite high.

12:36 Play from here But other than Jan LeCun, who's really the most notable dismisser of this idea, all of

12:44 Play from here them had a P-Doom of 0.5, 1, 2, 5.

12:48 Play from here So it's a 5% chance of what we're going to do here is going to end us all.

12:52 Play from here That seems a little bit problematic, I think.

12:55 Play from here It's a significant percentage.

12:57 Play from here You have a 1 out of 20 chance.

13:02 Play from here In the all-we-could-die realm of things, it's significant.

13:06 Play from here Yes.

13:07 Play from here Yes.

13:08 Play from here It does feel that way.

13:09 Play from here But from a practical perspective, I think that we're all sort of boiling frogs in this,

13:16 Play from here particularly those of us who are really enmeshed in it.

13:19 Play from here And the capabilities keep getting better, and they keep getting more practical, and

13:23 Play from here there is more data that we'll probably get to today about ROI on a lot of these projects.

13:30 Play from here And what that implies under the hood is that the promise of a lot of the growth curves

13:35 Play from here that are happening are happening.

13:39 Play from here And it's not just creating more efficiency in existing technology, it's the technology

13:44 Play from here ramping up.

13:46 Play from here When I work with companies, one thing that I almost always insist on talking about is

13:52 Play from here looking over the hill.

13:53 Play from here Because it's one thing to plan for where the technology is today as far as agents and workflows

14:00 Play from here and chaining things together and doing cool stuff that usually almost works, versus what

14:07 Play from here it's going to be like in another year or another five years or another 10 years.

14:12 Play from here And as a dad with kids in high school, I really need to understand what is that 10-year picture.

14:20 Play from here And it's increasingly likely that that 10-year picture will at least include AGI from a functional

14:27 Play from here perspective, will almost certainly include robotics in functional ways that go way beyond

14:35 Play from here where they are today.

14:36 Play from here We're talking about humanoid robots.

14:38 Play from here Humanoid robots.

14:39 Play from here We'll see.

14:39 Play from here that are lifting boxes onto warehouse shelves? Yeah. Yeah. Well, it could be things that are

14:44 Play from here working boxes, but if they do a really good job of lifting boxes, we got that already.

14:48 Play from here We got that already. You're talking a 10 year, a 10 year horizon percent. We're going to have

14:53 Play from here humanoid robots all over the place. You don't think so? No, I know. I do think so. But I want

14:58 Play from here to be a little bit conservative about it because these things are all, they're all happening.

15:01 Play from here So when you're thinking about your business and you're thinking about what your business does

15:05 Play from here today and you're really pleased with yourself because you're listening to guys like us who

15:09 Play from here are trying to give you really practical things going forward, you can't be thinking about it

15:14 Play from here statically. You need to be thinking about the what if, and how you're going to change

15:22 Play from here your business based on the opportunities that are being created, as opposed to merely,

15:28 Play from here air quotes, merely reinventing the way you do business, because that's going to become

15:32 Play from here table stakes, that we are all going to be reinventing the way our business works.

15:36 Play from here But what is our business? What does that future state do to change the landscape of your customers

15:44 Play from here and your competitors and where you're operating and what you're doing? And some businesses I

15:50 Play from here interact with, I'm like, yeah, you're still going to be selling the same thing because there's a

15:55 Play from here need for that. But the way you sell it, the way you go to market, the way that your customers

16:02 Play from here perceive your product is going to be totally different. And if you don't start thinking about

16:07 Play from here that soon, somebody else will think about it for you. Yeah, totally agree on all that. And

16:14 Play from here it's so fun to go down the rabbit hole of the singularity and superintelligence, and I promise

16:18 Play from here we're not going to for the next hour. Kevin giving me the stop motion. But last thing to just put

16:27 Play from here this to rest is a while ago, we talked about this book that I think I was intentionally avoiding

16:34 Play from here reading, titled If Anyone Builds It, Everyone Dies, right? By, I'm going to butcher his name,

16:40 Play from here like Eliezer, Yudowski, and Nate Soros, right? Who, to be fair, also a fiction writer. And I

16:49 Play from here heard him on some interviews, and I think we talked a little bit about that. And I said, yeah,

16:52 Play from here he just doesn't sound very convincing. He doesn't have really great answers to a lot of the questions

16:56 Play from here the interviewers were giving him. I started to listen to the book, and I'll tell you what,

17:01 Play from here one thing to note, and this really goes across the board with almost everything we talk about,

17:07 Play from here there's so much nuance to this. And there's so much going on behind the scenes that most people,

17:13 Play from here including us, have no idea. To be fair, including a lot of the researchers at these Frontier Labs

17:19 Play from here don't quite understand how these models work and their complexity as they evolve. And one thing

17:25 Play from here he's doing a really good job in that book, I'm not done yet, but so far, is a lot of thought

17:30 Play from here experiments. And I love a good thought experiment. I love just this, oh, you're an alien, and you

17:35 Play from here came to Earth, that it was 2 million years ago. What predictions could you have made? And I think

17:39 Play from here the TLDR is that nobody really knows. Nobody really knows where it's heading. And that's just

17:48 Play from here the way it's going to be until we get there. Nobody's going to quite know everything.

17:53 Play from here So I think that this is a good segue into how it also plays into current law and regulatory

18:02 Play from here and how we do business, because some unpredictable things can happen, and you have to have the right

18:08 Play from here guardrails in place. And I think that's really what this OpenAI blog post was about, just

18:14 Play from here bringing up that discussion on guardrails again. And some other things have happened in the news

18:19 Play from here that are related to these guardrails. Do you know which one I'm talking about, Kevin?

18:24 Play from here It happened to do with a big refund.

18:26 Play from here With Deloitte?

18:29 Play from here Yeah.

18:31 Play from here Yeah, you want to talk a little bit about what happened there?

18:33 Play from here Sure, sure. So our friends at Deloitte, who is a totally small firm with no global reputation

18:39 Play from here or anything like that, just had to refund the Australian government something like $350,000

18:48 Play from here because their team just GPT slopped out a report that looked beautiful. It had all of the bits and

18:57 Play from here pieces. It had citations, but it didn't take much clicking on those citations to recognize that they

19:03 Play from here went nowhere. So the idea that a team that was getting paid this much money was clueless enough

19:13 Play from here about the limits of the technology. It's both embarrassing for Deloitte, but it's also a

19:20 Play from here cautionary tale that these are smart people, right? And they're not paying enough attention

19:26 Play from here to the reality of where the technology is. And obviously, we're both super bullish and we're

19:32 Play from here really into this, but oh my gosh, proceed with caution. We're still at a stage of the technology

19:39 Play from here where it's trust, but verify. And if Deloitte can do this, how many other smaller consultancies

19:45 Play from here out there are just phoning things in? Or we may have talked about this last week, but lawyers

19:50 Play from here who are, there's something like 200 lawyers have been sanctioned for-

19:54 Play from here I was just going to bring that up, yeah.

19:55 Play from here Yeah, that have been sanctioned for phoning in citations and things like that. And the irony is

20:02 Play from here there are good tools that are out there for law, for finance, a proper approach within Deloitte.

20:11 Play from here In fact, I know that Deloitte has these internal tools that are using what we would call RAG

20:18 Play from here or embeddings to lock in information and they can query that information and get good data.

20:24 Play from here But this was just like amateur hour entirely. And again, if it's amateur hour for Deloitte,

20:31 Play from here imagine what it is for the SMBs of the world.

20:34 Play from here Yeah, it's so wild. And this just keeps going back to what we always talk about with

20:40 Play from here whether you are running a division inside of a small, medium, or large company,

20:46 Play from here or you are running a company, be it small or large. There are so many of these tools that

20:52 Play from here could 10X many, many different parts of your business, but they still need to have safety,

20:59 Play from here guardrails, policies. All of this stuff has to be understood. Everybody has to be on the same page

21:05 Play from here or you end up in this situation where it could be on getting off easy, embarrassing,

21:13 Play from here or hundreds of thousands of dollars in refunds. And when you scale that outward,

21:18 Play from here if you don't have these policies in place, could be millions of dollars in refunds.

21:23 Play from here You can see where this could get away. What do you suggest, Kevin, as far as just

21:29 Play from here the first steps if you were to look at, and listen, this could have also been an isolated

21:35 Play from here incident inside of a larger organization where maybe there were some policies and they were

21:41 Play from here ignored, right? But if you were to tease this apart and say, okay, this is the beginning of

21:46 Play from here where this went sideways, what things could people do inside their organization to keep something

21:52 Play from here like this from happening, where they're going to use the tools, they're going to help get a better

21:57 Play from here output and help serve clients faster and better without just throwing a slop out there. Or maybe

22:06 Play from here you have an employee that is just trying to get to five o'clock and they hit generate and they

22:13 Play from here throw it in a document and send it off as a PDF. You think that would be the easy stuff to fix, but

22:21 Play from here maybe not. Where do you start? Where does somebody start?

22:24 Play from here This is so my soapbox that when I'm working with companies on enablement or even on projects,

22:30 Play from here I just can't resist. I always ask, do you have a real policy around acceptable use of AI?

22:39 Play from here And often they do. Okay. Do you have a manifesto that's coming straight from the top of the

22:46 Play from here organization that explains the vision of the organization around AI? The answer is almost

22:52 Play from here always no. In that case, that has been delegated to the chief legal officer or whatever it is,

22:58 Play from here we're going to come up with a policy. It's about the types of ways that we can use it.

23:05 Play from here And generally it covers things like data leakage and acknowledgement of errors.

23:13 Play from here What I don't see enough, and I would suggest that, well, first off, anybody listening to this,

23:18 Play from here if you are in a position in your organization where you can dictate policy, you need to have

23:24 Play from here that manifesto and you need to have a practical policy that people can understand. And the one

23:31 Play from here line that I almost always insist on is a non-abdication clause in your policy. Like the

23:39 Play from here GPT did it is never an excuse for turning out substandard work. And I assume that heads rolled

23:46 Play from here at Deloitte in this case, and they should have rolled, but people need to understand that in

23:51 Play from here big, bold type and triple underlined that the work product that comes out the door,

23:56 Play from here whether it's amplified by AI or not, is your responsibility. Because the second it is no

24:03 Play from here longer your responsibility, you need to be questioning your role in the organization.

24:08 Play from here Because if the organization can do without that human judgment, yeah, what are you actually doing

24:13 Play from here there? Right? So you probably want to be leading into this, even in terms of self-justification

24:21 Play from here day in, day out, but you don't see it very often. You really don't. It's all like buried

24:25 Play from here and kind of legalized. So, you know, developing a policy is pretty straightforward. You can do

24:29 Play from here it using an LLM and just sort of craft it around the right pieces and the things we do.

24:35 Play from here You know, we don't just read it before you publish it to your company wiki.

24:38 Play from here Yeah. Yeah. For example, put up the citations. I don't know. See if it's good. Click something

24:43 Play from here along those lines. Yeah. So it seems so obvious, but you have to have some sort of guardrails in

24:48 Play from here place so people know how to use it. People truly don't understand that it's a bad idea to be using

24:53 Play from here a free tool like the numbers are astonishing. It's like 72 percent of millennials are using free

25:02 Play from here AI products at work, even if they're prohibited and they're putting consumer data in there.

25:08 Play from here They're putting, you know, proprietary information and the models are not as good.

25:15 Play from here The free models are simply not as good. They're very good, but the paid models do offer more

25:22 Play from here context, more features, better output, et cetera, and some controllability around, you know,

25:28 Play from here whether or not your information is going to be oozing around the sides or not.

25:33 Play from here Pause. Pause there. And let's differentiate the difference between a free model and a paid model.

25:38 Play from here Some people think it's all the same thing, or does paid mean I have the $200 a month version

25:44 Play from here versus a $19 a month or $30 a month version? What are the modalities of control and safety

25:53 Play from here and guardrails that you get with these different choices?

25:56 Play from here Well, number one, you got to think from a business model perspective, like it's a truism

26:00 Play from here in marketing that if a product is free, you're the product. So you are being monetized in some

26:06 Play from here ways. Hello, meta. Yes. Hello, advertising. Yes, exactly. So how are you being productized?

26:14 Play from here And in some cases of growth for a while, it may be just about rapid absorption of as many eyes

26:20 Play from here as they possibly can get. And over time, cynical eyes will look at the value of those eyes. And it

26:26 Play from here might be something as straightforward as advertising, but it's also something like

26:31 Play from here data. So we talked a little bit about humanoid robots earlier. So Neo made a big storm this week

26:38 Play from here with selling 10,000 Neo robots for $20,000 single purchase or $4.99 a month. And the cost of these

26:46 Play from here things, I think, is going to be $150,000. So you scratch your head a little bit and you're like,

26:51 Play from here okay, well, why is this so cheap? And it's so cheap because what they're doing in the $4.99

26:57 Play from here model is they have basically teleoperators behind the scenes. So it's kind of like a puppet,

27:04 Play from here but the puppet is unloading your dishwasher, but it's absorbing all of that real world data

27:10 Play from here that they can monetize. Yeah, it's all training data. And for them, that has worked so much more

27:16 Play from here than $499 a month that they'll want to do it every day of the week. And you have other companies like

27:22 Play from here Tesla, Fathom, others that are building these entire warehouses that have kitchens and bedrooms

27:28 Play from here and things in them. And then humans scurry out and make a mess. And then the robot tries to figure

27:35 Play from here it out. Well, this is a brilliant approach. And frankly, it's what Tesla did too with self-driving

27:40 Play from here cars. With their cars, they've been absorbing all of it. Except we didn't get them for free,

27:43 Play from here did we? We should have. We should have. We should get them for training data. Come on,

27:46 Play from here Elon. $12,000 a year. Free model what? Okay. Anyway, so if the product is free,

27:52 Play from here you're the product in some way, shape, or form. So if you're paying for the product,

27:56 Play from here you might still be the product if you haven't enabled certain controls. The easiest one in

28:02 Play from here all of the mainstream models is to simply opt out of training data. And it's usually in your

28:08 Play from here settings somewhere. Share your experience with the model. Well, that's code for we're using your

28:15 Play from here information for training. So generally, we advise that people turn off those settings.

28:21 Play from here And we will store everything that you put into this context window. And it may or may not start

28:28 Play from here to surface in random places on the internet as well, because that's been known to happen, right?

28:32 Play from here Mm-hmm. Well, true. Yeah. So turning off training being used in the model is the primary defense.

28:40 Play from here Using higher tiers of models like GPT's business, for instance, their terms of service are at least

28:48 Play from here implying that your information isn't going to be used. Whereas there are certain circumstances,

28:54 Play from here I believe, in the plus model where it still can be used, etc. So there is a bit of a trust element

29:01 Play from here here. There are a lot of good reasons to suspect that even though they say that data can't be used

29:08 Play from here in training, it might be used in training. So we do encourage people to be judicious about the

29:14 Play from here information they put in there anyway. So first is actually paying for more than a

29:19 Play from here consumer-grade tool.

29:20 Play from here Paying for control.

29:21 Play from here Paying for control.

29:22 Play from here Yeah.

29:23 Play from here You're paying for control.

29:24 Play from here Second is opting out of the training systems that you've been opted into.

29:30 Play from here Anthropic Clawed, for instance, opts everyone into training, and you have to proactively

29:34 Play from here opt your way out.

29:36 Play from here And then making sure you're on a tier of service such that the commercial terms that the model

29:42 Play from here company is offering align with some level of protection.

29:46 Play from here Yeah.

29:47 Play from here I think that's all good.

29:48 Play from here There's another category in there as we see billions of dollars of market value being

29:53 Play from here created with companies that are building products on top of these Frontier Lab LLMs, right?

29:59 Play from here You made that really good point of, hey, you have to take responsibility of the actual

30:04 Play from here product that you're selling.

30:06 Play from here You can't say, oh, well, it was AI did this.

30:09 Play from here It's like, no, this person who was responsible for the report didn't click on any of the

30:16 Play from here and just took everything from that output and pasted it into the report.

30:20 Play from here That's not cool, right?

30:21 Play from here But if you have a service or product that is built on top of one of these models, and

30:26 Play from here we kind of go back to this unpredictability of exactly what you're going to get every

30:33 Play from here time, that's almost a completely different category.

30:36 Play from here And there have been a lot of stories that have come out with that as well.

30:40 Play from here So yes, it is your product.

30:42 Play from here But what is powering your product under the hood isn't really 100% predictable.

30:49 Play from here That also really folds into a lot of the ways that some of these medium and large companies

30:55 Play from here are using these models, which really just reinforces the need for guardrails, policy,

31:03 Play from here literacy, training, all of these things that we talk about all the time.

31:07 Play from here So one piece that I want to add to there that I find particularly interesting as I

31:12 Play from here speak to companies about regulatory risk or legal risk and legal liability is there is

31:22 Play from here a line that you cross between trying to be compliant in terms of risk, like are we accidentally

31:29 Play from here violating X, Y, and Z law or putting ourselves in a place where somebody is going to be able

31:35 Play from here to successfully sue us because we're just using an AI model to hire people and we're

31:40 Play from here injecting bias into the system to a point of measurement and efficacy.

31:46 Play from here And I don't really want this to get too wonky, but it's this idea that all of the

31:51 Play from here regulations globally that are coming together and on the state level have this idea of

31:56 Play from here measurement that are built into them.

31:58 Play from here How do you measure model drift?

32:01 Play from here How do you measure bias, et cetera?

32:04 Play from here Well, guess what?

32:05 Play from here You need to be doing that from a production perspective anyway.

32:09 Play from here And a lot of implementations that I see aren't getting to that step.

32:14 Play from here They get to this point where it's bells and whistles and it's doing something cool with

32:18 Play from here workflows and things are moving around and they're like, yay, we're saving all this

32:21 Play from here time and money. But there's no attention being paid to how the model is performing

32:27 Play from here over time as, say, you're building a bigger database of documents that it's pulling

32:31 Play from here from. Is it drifting away from its original intent or not?

32:36 Play from here And even if you don't care about sort of regulatory risk, you need to care about the

32:43 Play from here accuracy of the tools you're putting together.

32:46 Play from here And I know, I don't suspect, I know that most organizations don't understand the fact

32:53 Play from here that LLM development is dynamic such that you don't just get to finish it and it's

32:59 Play from here deterministic and every single time it's going to do the same thing.

33:02 Play from here The dang thing drifts over time.

33:04 Play from here Unpredictable.

33:05 Play from here Unpredictable.

33:06 Play from here And, you know, trying to explain that from an engagement perspective, yeah, you know,

33:10 Play from here we do actually need some maintenance dollars because we have to pay attention to what

33:15 Play from here this model is going to look like in four months or six months to make sure that you're

33:18 Play from here still getting the outputs that you need from it.

33:21 Play from here Because if you just let it go, it could well be pretty much useless.

33:25 Play from here And worse, it could be misinforming its users over time.

33:29 Play from here And I'm not trying to anthropomorphize it at all.

33:31 Play from here It's just the nature of how neural networks work that there can be drift over time.

33:37 Play from here And it is convenient that if you focus on performance of your systems in a reasonably

33:46 Play from here sophisticated way, you're also giving some level of accountability in terms of liability

33:52 Play from here and accountability and et cetera.

33:54 Play from here So these things go hand in hand.

33:58 Play from here And, you know, at the moment you'll see a CTO's office who does care about performance

34:03 Play from here and you'll see like a CISO's office or chief risk officer's office who cares about

34:09 Play from here liability and compliance.

34:11 Play from here And really they're the same thing.

34:13 Play from here It's just different vocabulary and different application of the outputs.

34:17 Play from here Yeah, it's I mean, listen, this is a lot of macro talk.

34:20 Play from here And if you want to see the micro, I didn't realize until it's one of those things that

34:27 Play from here catches your eye once and then you can't stop looking at it.

34:29 Play from here How many times during the day I'll be in a conversation thread with an AI chat?

34:33 Play from here And then I will notice in the light gray text below the chat window how this model has

34:39 Play from here been updated, start a new thread or refresh your browser.

34:42 Play from here And it happens constantly.

34:44 Play from here Right. So you are building all this regulatory and compliance and measurement

34:50 Play from here on top of something that is evolving faster than the weather.

34:53 Play from here It's just it's constantly, constantly evolving.

34:56 Play from here Every time you fire up a new thread, you might be dealing with something that is updated,

35:01 Play from here but you don't know what those updates are.

35:03 Play from here You have no idea. You have no idea if it's going to start acting completely differently,

35:07 Play from here which it's very likely not.

35:09 Play from here It's happened a couple of times in history with I think it was one called Grok.

35:14 Play from here Is that right? Something like that.

35:16 Play from here Some weird stuff started happening after they updated the system prompt.

35:20 Play from here But really good points, Kevin.

35:21 Play from here And I think that is a really good segue.

35:23 Play from here Speaking of measuring things, there have been a bunch of studies that have come out around

35:29 Play from here ROI with AI and they have vastly different outcomes.

35:34 Play from here And it's probably a safe place to start with the fact that some studies are just shitty

35:40 Play from here studies. The controls are awful.

35:43 Play from here Things were skewed.

35:44 Play from here There were underlying incentives.

35:46 Play from here But we've got a couple of big names on the roster here.

35:49 Play from here We've got Wharton.

35:51 Play from here We've got Atlassian, a software company, a software company.

35:55 Play from here And we've got MIT, some big brand recognition here.

35:58 Play from here You want to talk a little bit about why one would say that about 75 percent of companies

36:04 Play from here already see positive ROI from Gen AI, whereas another study might say it's only 4

36:10 Play from here percent. What's what's real here?

36:12 Play from here What is what would you say based on these studies and the outcomes would be a fair and

36:18 Play from here definitive measurement of the real ROI from AI at work?

36:23 Play from here OK, so first, the the red herring was the MIT study of about a couple of months ago or

36:29 Play from here so. And this is the one that caused like hundreds of billions of dollars of market cap

36:34 Play from here changes in the market because it became a bit of a running joke, right?

36:37 Play from here A bit of a running joke.

36:38 Play from here Uncovered how the study was done and the study was was quite weak.

36:42 Play from here And, you know, hopefully some people will be held accountable for it because that that

36:45 Play from here was a massive destruction of economic value because it said that ninety five percent

36:51 Play from here of of of AI projects fail.

36:54 Play from here Ninety five percent.

36:56 Play from here Oh, my goodness. Are you out of your mind?

36:57 Play from here What are we doing?

36:58 Play from here Ah, everybody, stop.

37:02 Play from here Can't run on the bank subscription.

37:04 Play from here MIT has got to be right.

37:05 Play from here And then you peel back the onion a little bit.

37:07 Play from here And it was like, OK, in fairness to them, like, I'm not sure if somebody grabbed on to

37:12 Play from here their study or not, but they were basically looking at public reports and they they were

37:16 Play from here looking for public reports of companies that were indicating that they had AI projects

37:21 Play from here going on that were semi identifiable in a profit and loss statement.

37:25 Play from here And then they looked at the profit and loss statement and they decided, oh, well, the

37:29 Play from here P&L didn't actually change.

37:32 Play from here So therefore, the projects might have failed.

37:35 Play from here You know, business one on one is that you have to either save costs as in fire some

37:43 Play from here people or eliminate some systems or some other platforms and have that show up in the

37:49 Play from here bottom line, or you have to have increased revenue in some way, shape or form that's

37:55 Play from here going to have it appear at the top of the P&L and translate all the way down.

38:00 Play from here But that isn't really the nature of the AI projects that we're seeing quite yet.

38:05 Play from here Don't get me wrong.

38:06 Play from here Like focusing on ROI is absolutely vital in these projects, but that's not the stage of

38:12 Play from here development that a lot of companies are in.

38:14 Play from here They're they're not going to the team that did this and saying, OK, cool, half of you

38:18 Play from here are gone. So now we're saving a million dollars a quarter.

38:22 Play from here And, yay, it's a win right there.

38:24 Play from here So just the premise of the way that it was constructed was it was it was really pretty

38:30 Play from here odd to those of us who were paying attention.

38:31 Play from here It was also a really small sample set.

38:33 Play from here The difference with the Wharton study is, A, it was a much more robust study.

38:38 Play from here It was looking at, I don't know, something like twenty three hundred executives instead

38:43 Play from here of like one hundred and fifty.

38:44 Play from here And it was going through their own internal ROI assessment on a much more granular level.

38:53 Play from here Now, there's still a trust but verify aspect.

38:56 Play from here One benefit of the MIT study is the 10K is going to have data in it and you can validate

39:03 Play from here that data. The Wharton study, you have to trust that the people you're interviewing are

39:08 Play from here accurate about their assumptions.

39:10 Play from here And some of them are defending their jobs.

39:13 Play from here But, yeah, we spent a million dollars and it totally worked.

39:16 Play from here Trust us. Right.

39:17 Play from here So, yeah, the reality is somewhere in between.

39:20 Play from here But 63 percent of respondents came back with the fact that they feel that their investments

39:30 Play from here are already delivering positive business value.

39:33 Play from here The other related fact is most of those same people said just a year ago or roundabouts

39:40 Play from here that they expected to see a positive ROI in the next two to five years.

39:46 Play from here And what's happening is they're seeing the ROI a heck of a lot faster than they anticipated

39:51 Play from here before. So other takeaways, where are the value, where's the value coming from?

39:58 Play from here And this all resonates with me based on the sort of projects that we that we work on.

40:03 Play from here Content creation, really pretty low bar there.

40:07 Play from here Internal knowledge management.

40:09 Play from here So the ability to use rag structures to really pull out precise information.

40:14 Play from here Again, what Deloitte should have been doing, data summarization, et cetera.

40:20 Play from here But research, research, research.

40:22 Play from here Yeah. Yeah.

40:23 Play from here She's up a lot of hours.

40:24 Play from here And then human productivity is up.

40:28 Play from here You know, how do you measure human productivity when people are absolutely booked all day

40:32 Play from here long? You know, it's a lot of it's self-reported.

40:35 Play from here Right. Yeah.

40:37 Play from here So it feels much more real to me.

40:41 Play from here And again, there is trust, but verify there is a lot of fluff that's out there.

40:47 Play from here There is a lot of sort of internal defense of the dollars that are being spent.

40:53 Play from here But being able to point to specific ways that people are improving productivity, even if

41:00 Play from here you're not capturing that dollar value in terms of just letting those people go, to me

41:05 Play from here makes sense as far as an approach.

41:07 Play from here Yeah. Yeah.

41:08 Play from here I totally I totally agree.

41:10 Play from here And I think that when you look at ROI, there's a few important things to consider.

41:18 Play from here Right. And how you how you measure that, whether it's going to be over the long arc of

41:24 Play from here something, if it's for a project, if it's anecdotal of I feel like I got this much more

41:30 Play from here done in my day.

41:31 Play from here If you're just an individual worker inside an organization that hasn't really adopted a

41:35 Play from here bunch of AI yet, I, you know, as somebody who I'm inside of this every day, like you,

41:42 Play from here am constantly amazed at not only how much more I can personally get done with some of

41:47 Play from here these tools, but when I build the tools and then share them with my team and enable them

41:53 Play from here to be more productive as well, you can just see much faster in a smaller organization

41:58 Play from here what that output looks like.

42:00 Play from here And I think that the larger the organization gets, the more difficult it might be to see

42:07 Play from here that very clear light at the end of the tunnel as you're making all these investments, it

42:12 Play from here becomes much more complex and difficult to measure as well than when you are a small

42:18 Play from here and nimble company.

42:20 Play from here I really want to get into some of these specific tools, Kevin, I know that we've been

42:26 Play from here talking a lot between us on workflows and automation and things that can just not only

42:33 Play from here personally save us a bunch of time in our day, but also help our teams out.

42:39 Play from here I think it was the last episode that we spoke on, you said that your high level definition

42:44 Play from here of ROI with these AI tools is the number of hours saved, right?

42:50 Play from here The number of hours that you get back that you're no longer having to do something that's

42:55 Play from here usually pretty mundane, pushing pixels around a screen, moving things from one cell in a

42:59 Play from here spreadsheet to another, to a PDF and all this stuff, and how so much of this can be

43:03 Play from here automated now. Before we dive into that, you know, because we do this all the time, I

43:10 Play from here feel like we start thinking, well, everybody's doing this.

43:14 Play from here But one thing that we always talk about, and this gets kind of funny depending on the work

43:18 Play from here setting that we're in, is this whole new world of AI dictation for the machines, right?

43:24 Play from here And how there was this promise inside the operating system of all these different devices

43:30 Play from here of dictation and how you could just talk to your computer and it was so much faster than

43:35 Play from here typing. And so every time we would try this, what you would get on the screen was very,

43:40 Play from here very disconnected from what you actually said, basically every time.

43:45 Play from here So now with these tools, it's actually taking what you say, running it through their

43:51 Play from here language model, and not only doing a much better job with accuracy, but learning the

43:56 Play from here words that you use, right?

43:58 Play from here So it's this thing.

43:59 Play from here is learning. And it understands the context and the syntax. If I'm rattling off a list,

44:07 Play from here very often when it gets into the document, it's bulleted, or it's numbered, or things

44:12 Play from here have parentheses around them, or quotation marks. And it works so well right now. I think,

44:17 Play from here what was the stat? Something, it's about 400% faster than most people type, if you can just

44:24 Play from here talk to your machine all day. Is that right, Kevin? This is just one of my favorite things.

44:29 Play from here For me, this is the biggest single unlock in my professional life for the smallest and cheapest

44:37 Play from here piece of technology. And we did talk about this last week, I think, that your brain's

44:46 Play from here working at 1,000 words a minute, and your mouth can move at maybe 175, 200 words a minute. And

44:53 Play from here at least my terrible typing fingers can only move at 50. So you're throttling all of this down.

44:59 Play from here And when you unlock that, it's just relevatory. It's also just more natural. We type because we

45:07 Play from here type, right? Not because it's particularly convenient in any sort of way. It's just a

45:13 Play from here standardized way of communicating with devices. And once your workflow gets beyond that,

45:19 Play from here and you're incorporating dictation. So I use WhisperFlow. Super Whisper is another one.

45:27 Play from here No relationship with either of them. They're like 10 bucks a month.

45:30 Play from here Yeah, I'm on WhisperFlow. Well, I tried a few. For me, that worked the best.

45:36 Play from here It seems to work well enough, right? And then the next level that I'm getting to,

45:41 Play from here I've complained to you about SuperHuman as an email tool. And it's like certain people have

45:48 Play from here this thing about hotkeys. Oh, I'm just going to fly through all my workflows with a hotkey.

45:52 Play from here And I'm like, are you kidding me? Just like, really? You do that many emails that you need

45:57 Play from here a hotkey? But for me, the unlock is adding little snippets where I'm like, paste Zoom.

46:05 Play from here And if I say paste Zoom, which is not something you're ever going to say,

46:09 Play from here like randomly, naturally, yeah, it's going to add my Zoom link or paste my calendar or my email.

46:18 Play from here So all of those text shortcuts that I had built into my Mac over the years,

46:25 Play from here I'm now doing verbally and able to paste them in.

46:27 Play from here Oh, that's so funny. I use some Chrome extension. I can't remember the name of it.

46:31 Play from here But you basically load in a series of those commands. And I never thought about using it

46:37 Play from here with voice. I'm just so used to hitting the forward slash, whatever it is, cal,

46:40 Play from here and then it puts in my Zoom link. But I like that even better. I'll get deeper into the voice world.

46:46 Play from here So there are some security vulnerabilities with being able to paste from an LLM directly

46:52 Play from here into an app. And you'll see that a lot of apps block shortcuts for this reason.

46:57 Play from here I'm relatively comfortable with the risks that this is presenting, but it is something to be

47:01 Play from here aware of. And if you're getting pushback from your organization, oh, you can't possibly use that.

47:06 Play from here Well, there are some actually decent reasons why that's the case. But I find it endlessly

47:11 Play from here frustrating that I couldn't use some of my underscore. For me, my Mac replacement was

47:18 Play from here always underscore, whatever it was. And they don't work in certain environments, like

47:22 Play from here Mac's Calendar, for example. But now, because of the dictation, it's actually-

47:27 Play from here Shame on you for using Mac's Calendar, by the way. Shame on you. It's awful.

47:33 Play from here It's not Outlook, man.

47:35 Play from here Go to Google Calendar like everyone else. Come on.

47:37 Play from here Hate it.

47:41 Play from here Sorry, go ahead. Your shortcuts.

47:42 Play from here Anyway, so shortcuts are great. Okay, this is extremely practical, right? To the audience,

47:48 Play from here like 10 bucks a month, download one of these tools. And it is a little annoying in a communal

47:54 Play from here environment. You look like a crazy person as you're talking to yourself all day. But I get

47:58 Play from here so annoyed when I can't use it. And then you can combine it either with something like my go-to

48:05 Play from here shortwave for my email composer and be able to just dictate a rough email and have it make sense

48:13 Play from here of it based on the context. Some of them allow you to do custom prompts. So if you're sitting

48:18 Play from here in Gmail and you're prompting, it's going to know that it's an email. So it's going to fill in the

48:23 Play from here gaps of, hey, Eli, great show today, et cetera, whatever it is. We still have a little ways to

48:28 Play from here go with that context. But being able to get more information in front of the LLMs and in front of

48:35 Play from here your other tools faster is a great unlock. And thinking about how voice can unlock things in

48:41 Play from here other places of your business, I honestly feel like 2026, it's going to be the year of the

48:50 Play from here workflow and the agent. Again, 2025 is the year of the agent. It's always the year of the agent.

48:56 Play from here The year of the agent, but your voice that just transcending the friction of having to type things

49:05 Play from here in all of your environments. And it might be just the system overlay, something like WhisperFlow,

49:10 Play from here but it might be being able to do more voice interrogatory. And we have some cool projects

49:16 Play from here that are going on internally that are trying to get to information that's at the pointy end of

49:22 Play from here the spear, such like, for example, big organizations that have a lot of service personnel that are out

49:29 Play from here there. Those people actually know more about market intelligence on the ground than you're

49:36 Play from here ever going to be able to get to them. And yeah, over the years, people have tried like, oh,

49:40 Play from here we're going to put surveys out to all of our ground techs and we're going to get this information back.

49:44 Play from here And they're like, yeah, I'm busy. No, thanks. Or when they get in trouble for not doing it,

49:48 Play from here they respond with like a couple of words. But we have this project where we're hitting them up with

49:54 Play from here a voice agent and the voice agent is interviewing them very quickly, very simply, very in a very

50:00 Play from here friendly way about some of the things that they're seeing on the ground. And we're seeing really high

50:05 Play from here response rates. And because it's oral and they're not having to type, they don't have to type.

50:11 Play from here You're getting this information and all of a sudden we're learning about competitors and we're

50:17 Play from here learning about sort of pushback from clients and we're learning about quality issues and all of

50:24 Play from here these things that, I mean, it's deeply unstructured data that now we can use. And it's so cool. And

50:30 Play from here the feedback goes up to marketing, it goes to customer service, it goes to client management,

50:37 Play from here you know, it goes to product or supply chain. You know, the left-handed doohickey never works,

50:41 Play from here never has been fixed. Somebody needs to fix this or I'm going to quit.

50:45 Play from here You know, like literally. Anyway, so oral is huge.

50:49 Play from here Yeah, it's huge. It's one of those unlocks that you can have today and it will eventually be

50:55 Play from here in every OS on every machine. It's just because it's one of those things that once you start using

51:00 Play from here it, you can't go back, right? We've all played with these bits of tech that feel more like toys

51:07 Play from here and you think, oh, this is so cool. Two weeks later, it feels more annoying than it does cool

51:11 Play from here or it crashes or it's buggy. This isn't one of those things to the point that Kevin and I always

51:16 Play from here have these text exchanges where we're super annoyed that we're in a place where it's either

51:21 Play from here really loud or it'd be just very publicly uncomfortable to be talking to yourself in

51:27 Play from here the coffee shop. But I am now at the point where I feel like not having is such an impediment to my

51:34 Play from here workflow that I will just take out a single AirPod and I will just whisper into the tip of the AirPod

51:39 Play from here on my laptop just to keep that cadence of work up. That's how much of a game changer it is.

51:45 Play from here And I will say, especially when you're using a lot of these AI tools where you're having

51:51 Play from here constant conversations of putting in prompts, trying to get feedback, iteratively working on

51:57 Play from here something, being able to talk through that versus type through that, you just fly through those work

52:02 Play from here cycles. It really is incredible. So I want to add sort of for a tool perspective, a new toy that I've

52:07 Play from here been experimenting with this week is Gemini's Opal vibe coder. And you and I both have strong

52:13 Play from here opinions about vibe coding. Always Google getting closer. The ability to use Opal is particularly

52:19 Play from here interesting because it ties right in with the various bits and pieces of the Google sort of

52:25 Play from here app universe, not necessarily from a productivity perspective, but from a tooling perspective.

52:30 Play from here So if you want to build a voice bot, you can go to use the Google voice engine and specifically

52:38 Play from here build the app around that in a way that's way more integrated than you're seeing in other places.

52:43 Play from here And anybody who's played with Google Notebook LM knows that Google voices are exceptional.

52:50 Play from here The voices that I'm hearing through Opal are so much better than I can get through 11Labs or

52:56 Play from here Haygen, or I have a voice tool I use called VAPI that I like, Bland, Assembly AI. There are a whole

53:04 Play from here bunch of these that are floating around out there. I just call it VAPI. Is that too close to VAPID?

53:09 Play from here It just drives me nuts to call it VAPI. I don't know, VAPI. If it's called VAPI,

53:14 Play from here that's a terrible name for a company. I don't know. I don't know. By all means, if the people at

53:20 Play from here VAPI are listening, you should opine and tell us how to pronounce.

53:26 Play from here Kevin, I just work in HR, and I'm sitting in a cubicle outside of my boss's corner office,

53:31 Play from here and you just said vibe coding. The hell is that? What's vibe coding?

53:34 Play from here Boy, so vibe coding is the ability to translate simple instructions, to be honest, oral

53:40 Play from here instructions, if you're using one of these tools, into an app or an applet without going through a

53:47 Play from here dev process. From conversation to coding, basically.

53:50 Play from here Conversation to coding. This is what I'm trying to accomplish. This is why you go figure it out,

53:57 Play from here because I'm not a dev. Big names in this space are Repl.it, which hit $100 million in revenue,

54:03 Play from here Lovable. I wouldn't necessarily put Cursor into this category of consumer-facing

54:11 Play from here vibe coding, because that is more of a development tool. Don't get me wrong.

54:14 Play from here That's why I brought it up.

54:17 Play from here Yeah. Well, if you have any sort of development skills-

54:18 Play from here For me, it's a lot more powerful than the others you mentioned.

54:21 Play from here But you're a developer. But for the hypothetical HR assistant in the corner, there's no way they're

54:28 Play from here going to really make much forward progress with Cursor. But the promise of being able to build

54:35 Play from here these apps is remarkable. The practice is really quite fraught for a lot of reasons. And I would

54:42 Play from here say that the biggest reason is that one of the skills of a developer is in scope. And anybody

54:51 Play from here who's been a client on a development project knows that scope will make or break the project.

54:58 Play from here And if you are lazy about scope when you're vibe coding, the results will be predictably horrible,

55:05 Play from here because adding features late in the game will break everything that you did previously.

55:11 Play from here Exactly. Unlike real agile development done by developers, it's really difficult to get

55:17 Play from here that toothpaste back in the tube once it's out. And you start ending up with this spaghetti

55:22 Play from here that unless you know what you're doing, and this really brings up that whole human in the loop,

55:27 Play from here experts in their domains are still needed. Because if somebody who has never done any

55:33 Play from here sort of software development in their life picks up one of these vibe coding tools and thinks that

55:37 Play from here I can, through a conversation, build QuickBooks, it doesn't really quite turn out like that.

55:45 Play from here Whereas a seasoned developer who wants to build a new QuickBooks can probably today do it 10 times

55:52 Play from here faster than they could have before generative AI with this suite of tools. It takes a lot of just

55:57 Play from here the mundane base level coding out of things, as well as some new creative things that some

56:03 Play from here developers or architects may not have thought of. A lot of that comes out of these generative AI

56:08 Play from here tools as well. From a very practical perspective, I'm actually totally fine with that hypothetical

56:13 Play from here HR assistant trying to figure this out. But you've got to figure the scope first. And scope is

56:19 Play from here essentially, what are the inputs and what are the outputs? And what do those outputs look like?

56:23 Play from here And without having a solid technological foundation, you can, with a good scope, allow a lot

56:31 Play from here of these tools to define their tech stack. You want to keep it relatively simple. You don't want

56:36 Play from here to build a full CRM. But building some sort of a voice agent that interviews people about

56:43 Play from here performance or whatever, it's very doable. But what you'll find where most people break,

56:49 Play from here and where I suspect that Repl.it's revenue curve might break, is that they get to 80%,

56:54 Play from here 85%. And they're like, wow, this is amazing. And then they just can't get over that last 15%.

57:00 Play from here Right.

57:01 Play from here And everything gets jammed up.

57:03 Play from here Or can't debug it, or can't even really figure out how to effectively deploy it in the way that

57:07 Play from here they want to. A number of things can happen. And this is really across the board with all these

57:14 Play from here tools. If you are going to build a product through a conversation, you have to know what that

57:20 Play from here conversation entails. You have to know what conversation to have, right? If you're going to

57:25 Play from here generate a piece of art from one of these LLMs, you also have to know what to look for in that art,

57:33 Play from here how to describe that prompt in order to get the result that you want. And it's really no different

57:39 Play from here across all these different domains, where the expert who sits on top of all of this, their

57:47 Play from here knowledge, their experience is still a very, very important part of that entire workflow. It doesn't

57:54 Play from here work without it.

57:56 Play from here I would suggest that if people haven't, you do experiment with the Google tools. They are

58:01 Play from here under-marketed. And Google always just amazes me. They are not going to-

58:06 Play from here Very stealthy.

58:07 Play from here Very stealthy.

58:07 Play from here Very stealthy. And they have all of these amazing tools in their AI studio.

58:11 Play from here AI studio is incredible.

58:12 Play from here This Opal tool. And I did have a very lame weekend where I ran the same vibe coding task with

58:20 Play from here Repl.it, Lovable, another one that somebody had recommended. I honestly don't remember what it is.

58:27 Play from here And Google. And Google was by far the fastest and closest on a single prompt. Now, I like to

58:34 Play from here think that I, A, have prompting skills, so I know how to interact with these. And I have a decent

58:39 Play from here of scope. So I'm kind of starting on second base by having a well-crafted prompt, which by the way,

58:45 Play from here you can use GPT or Cloud or whatever to help you get your scope down. But Google is blazingly fast

58:53 Play from here and it's really, really cheap. It's part of your Gemini subscription that's built into your

59:00 Play from here Google Workspace account. So even on a variable cost basis, it's super, super cheap.

59:07 Play from here And Google also offers the promise. I'm not a giant Google fan boy, but it offers the promise.

59:12 Play from here Yes, you are. Yes, you are. It's okay.

59:15 Play from here Did we mention the Google calendar? I can't stand Gmail either. But it'll be connected

59:24 Play from here to your universe. If you're already a Workspace customer, where are your documents? It's in Drive,

59:30 Play from here right? Well, you're using a Gmail-backed email system, so you can tap into that.

59:35 Play from here So there's less duct tape in using a Google-oriented system than there is in using an

59:41 Play from here external system because Google knows how to talk to Google. So over time, that will get better and

59:47 Play from here better and better. And I assume that that's a lot of Google's commercial angle here.

59:52 Play from here What's our mantra, Kevin? This is the worst that you're ever going to see AI perform at anything.

59:57 Play from here It never comes in. Long-winded. This is the worst it will ever be.

01:00:01 Play from here It's always going to keep getting better and more impressive. It just keeps happening. It

01:00:06 Play from here keeps happening, keeps happening. You said something interesting about using possibly

01:00:12 Play from here a different model in order to get the prompt for the model that you want to use. For example,

01:00:17 Play from here maybe you use GPT-5 in order to get the prompt that you use in Google's AI Studio to build your

01:00:23 Play from here products. This is something that we do all the time, and I think we probably assume everybody

01:00:29 Play from here does it. But if you think about this in the frame of that cold-start problem where I just

01:00:37 Play from here don't even know where to begin with this stuff. I don't have prompting skills. I don't know what

01:00:43 Play from here to put in there. I've no idea whatsoever. You could go to really any of the models that you

01:00:49 Play from here want to today and say, here's roughly what I want to do, just like you were talking to a friend.

01:00:54 Play from here Hey, I've got this idea for this little web app and it contains this form, but then it does this

01:01:00 Play from here and this and this, and you just describe it. Then at the end say, I want you to keep asking me

01:01:06 Play from here questions about this product until you have exhausted everything that could possibly be

01:01:12 Play from here in this scope. It actually does a pretty good job of asking you those questions.

01:01:17 Play from here You can then take all that and you can port it over and use it in the prompt for whatever thing

01:01:22 Play from here you want to build or just play with or mess around with. There's a life skill in there too,

01:01:26 Play from here which is that interrogatory part, like ask me questions to get a more full-fledged thing.

01:01:31 Play from here People like us have been talking about this for years, but there is no limit really in a consumer

01:01:37 Play from here chatbot of the amount of information. Why ask for five headlines when you can ask for 20 or 50 or

01:01:42 Play from here whatever it is? Then because we're using the speech, the dictation, then just running over

01:01:49 Play from here it verbally as opposed to saying, I like number two and I like number seven. Well, my MO is to

01:01:56 Play from here see that scope or to see that list of questions or whatever it is, and then just sort of hold

01:02:01 Play from here down function on my keyboard and run all the way through them and get all of this data back the

01:02:07 Play from here other direction. Yes, it will guide you forward. The other related one, probably we'll bookmark

01:02:13 Play from here this a little bit, are CLAWD skills. The ability to equip Anthropx CLAWD with specific platform

01:02:21 Play from here skills is a new feature that's really, really pretty powerful. It's ironically better than using

01:02:28 Play from here say Gemini to talk to Gemini about Gemini. You can use a CLAWD skill that has better context and a

01:02:35 Play from here better understanding of how Gemini works than Gemini knows. That's a little wild and that's

01:02:39 Play from here something that won't persist over time. We're going to have to, yeah, we've got to loop back to

01:02:43 Play from here that one because there's a lot to unpack there. One thing you did say about how you really do have

01:02:49 Play from here this infinite number of outputs you can get. Why ask for five when you can get 500, right? In these

01:02:56 Play from here things where you are having a conversation and it's saying, okay, wonderful, here's a hundred

01:03:02 Play from here questions about that website that you want to build, right? And then you are scrolling all the

01:03:06 Play from here way up and down through this thread. Another little hack is to finish that prompt with,

01:03:12 Play from here please only give me one question at a time. Yep. Maybe you have a hundred questions. Just give me

01:03:17 Play from here one question at a time. And it really helps keep that conversation from going off the rails.

01:03:24 Play from here Kind of like with people, especially Kevin and myself, I'm included in this. We'll go down little

01:03:30 Play from here rabbit holes and there's side conversations and then an hour goes by and you realize that you've

01:03:34 Play from here completely deviated from the task at hand. So that is one of the things, and I do forget to do

01:03:39 Play from here it sometimes, but I'm quickly reminded when I get any sort of output, then I then have to go five

01:03:46 Play from here times up the screen to see what that first question was and how do I respond and get to where I want

01:03:51 Play from here to go. Do you do that too, Kevin? Oh, absolutely. And same thing that I always forget to ask. And

01:03:56 Play from here then it gives me, we're going to do the entire business plan at once. Like, no, no, no, no, no,

01:04:03 Play from here Encyclopedia Britannica in one context window for you. One thing at a time and then string it

01:04:08 Play from here together the other way. So yeah, remembering when you're, you know, I've talked before about

01:04:13 Play from here setting context and entering your bookstore, like part of that environment around your bookstore is

01:04:19 Play from here of expectations. So like the expectation is ask one question at a time and I'll tell you when to

01:04:25 Play from here move on. Yeah. Yeah. Well, Kevin, that's our time and I'm going to end it with a promise.

01:04:32 Play from here Here's the promise. Next week, we are going to kick off with an actual workflow. I'm going to

01:04:39 Play from here have one that I love. You're going to bring one that you love. We may or may not have built them

01:04:45 Play from here on the same platform. I don't know. We'll see what happens, but we will walk through not only

01:04:52 Play from here why we love them, why we think there's ROI there, but how you can do it too. And we're going to

01:04:58 Play from here challenge you to take this out of your AirPods and onto your computer and actually do it.

01:05:05 Play from here And listen, even if it falls flat, I think the experience of getting your hands on this will

01:05:11 Play from here start to change the way that you see a lot of it. Very often, a lot of the workflows that I

01:05:19 Play from here end up building, they don't end up the way that I intended in the beginning. A lot of the times

01:05:24 Play from here when you start down that road, you start getting a bunch of other ideas or ways that it could be

01:05:29 Play from here more efficient or things that you could add. It's a real fun iterative thing to take on.

01:05:36 Play from here And so that's my promise. You're going to hold me to it, Kevin, and not let me talk about using AI

01:05:42 Play from here to be a poor father right in the beginning of the episode? Well, you got to have like

01:05:46 Play from here playful banter to start things off. So after the banter. You have to. Right. Right. Until AI is just

01:05:52 Play from here doing that playful banter for us, which super intelligence, that's what first use case is going

01:05:58 Play from here to be, right? All right, buddy. We'll talk to you next week. Awesome.

01:06:07 Play from here Thanks for listening. If you'd like to send us a question to answer on the show, email me at

01:06:11 Play from here elijah at spark six dot com. That is S.P.A.R.K. the number six dot com or Kevin at Kevin at Ascend

01:06:19 Play from here Labs dot A.I. to learn more about A.I. strategy and training for your business. Visit Kevin's

01:06:24 Play from here company, Ascend A.I. Labs at Ascend Labs dot A.I. and to learn more about design and development of

01:06:32 Play from here custom A.I. solutions. Visit my agency at Spark six dot com. You can also sign up for my free

01:06:38 Play from here newsletter, A.I. Framework Friday by visiting Spark six dot com forward slash newsletter.

How to cite this episode

Credit the named speaker and AI at Work by AscendAI. Link to the episode or the passage you quoted.

“Ep #3 Quit Typing. Start Talking: Voice, Vibe Coding, and the Real ROI of AI (No, 95% Don’t Fail).” AI at Work by AscendAI, 2025-11-11. https://ascendlabs.ai/podcast/quit-typing-start-talking-voice-vibe-coding-and-the-real-roi-of-ai-no-95-dont-fail/

Recent episodes

View all episodes
×
AI Solutions Graphic
Free Resource
Get Started Now