On .NET Live - AI offers benefits, but at what cost?
dotnet
0:02 I know.
0:06 [music] [music] [music] [music] [music] [music] [music] [music] [music] [music]
2:00 I
2:23 [music] [music] [music]
3:17 Welcome, .NET friends.
3:19 You know what time it is?
3:20 It's the On .NET Live Show.
3:22 Where our mission here, yes, raise the roof.
3:25 Our mission here is to empower you with the .NET community to achieve more.
3:30 And uh hopefully energize you.
3:32 Uh for most of you, this would be a Monday morning.
3:34 Uh before we get into the usual, I wanted to plug an event we have coming up.
3:39 You all may have heard of this, KDQ, the the uh bumper here.
3:44 Microsoft Build is coming up this year.
3:46 San Francisco is is where this will be held, June 2nd through the 3rd.
3:51 We will have interactive sessions, meetups, workshops,
3:55 very familiar faces from the .NET ecosystem, including Scott Hanselman,
4:01 and representation across not only .NET, but we have VS Code content,
4:06 Visual Studio, GitHub, the list goes on and on.
4:10 Check that out at aka.ms/build26,
4:15 and do be aware that registration is already open.
4:18 We're excited to see you there.
4:20 Now we can move on to the regular meet of the show.
4:24 I'm your very familiar host, Scott Addie,
4:26 and I'm joined by familiar co-host Katie Savage and Frank Boucher.
4:32 Hello, hello.
4:33 We are excited to have Steve Smith on the show today.
4:37 Steve is no stranger to this show if you've tuned in before.
4:41 If not, Steve, I'll let you briefly introduce yourself.
4:45 Everybody, thanks Scott.
4:47 Yeah, my name is Steve Smith and because there are so many Steve Smiths,
4:51 I also go by Our Dallas online and so you'll
4:53 find me on YouTube and various other places as Our Dallas.
4:58 And I've been on the show a few times, it's true.
5:01 I've been doing stuff with .NET for most of my career and these days I'm I'm
5:06 like most of us learning a lot about AI and how to apply that in our software.
5:13 Sounds good.
5:14 Yeah, it's a brave new world.
5:16 I think a lot of us are tinkering trying
5:18 to figure out how we optimize our workflows moving forward,
5:22 but uh why don't we get into it, Steve?
5:25 Let you kick kick off the conversation for today.
5:30 Sure.
5:30 So, recently I've been talking to various clients who have
5:35 obvious interest in AI and and there's at least two,
5:40 but let's let's just start with two ways in which
5:42 most clients I've been talking to are looking at integrating
5:46 AI into their businesses and those two would be
5:50 using AI to increase the productivity of their own employees,
5:54 their team members, their developers, whatever it might be,
5:57 and then also integrating AI features into the the software and the tools
6:02 and the things that they build or that they
6:04 use themselves internally for for their organizations.
6:07 So, in both of those cases, there's,
6:10 you know, a pretty compelling cost-benefit story today.
6:16 Uh but but the the thing that concerns me and the thing that I
6:19 have been sharing with others and and doing a lot of research on is,
6:23 like right now, we are still in the the middle
6:27 of the the venture capital-funded period of these AI companies.
6:32 And so, they're all, you know, using other people's money, essentially,
6:36 to try and and prove the market and and gain market share,
6:41 and not necessarily to try and be profitable, right?
6:44 None of the major AI companies out there are currently profitable.
6:47 But someday they all hope to be, right?
6:50 And so, the the question is, you know,
6:52 how soon is is that going to change to where they're going
6:56 to be more focused on profitability than just trying to gain market share,
7:00 and what will those costs be?
7:02 Because if if I'm a business and I'm betting on this particular service,
7:07 if I'm doing so with the math of 2025 or 2026 and and then all
7:13 of a sudden that changes a couple
7:15 years from now after I've deployed my solution,
7:17 that that could be very disruptive to me as as a business.
7:21 And so, I thought that would be an interesting
7:23 conversation to have with the three of you.
7:26 Wow, I love it getting a little econ injected
7:29 into our our morning stream and I'm just I'm
7:33 just trying to think of like the parallels of what
7:35 other technologies we've seen a similar kind of situation.
7:41 I don't think we Have we or is this totally new?
7:46 Well, I think it plays into SaaS a little bit.
7:49 Right?
7:49 Like, you know, way back in the dark ages,
7:51 you bought if you needed compute stuff, right?
7:54 You bought a computer.
7:56 You bought a server and you put it on a rack
7:58 and you knew what the electricity cost might be.
8:01 If if you were leasing stuff,
8:03 you probably signed a long-term agreement that said that your your building
8:07 would cost so much per month for our you know, 36 months or 5 years.
8:11 You might have service agreements that you know,
8:14 again have multi-year terms on them.
8:16 And for SaaS products, there's the possibility that you know,
8:21 if you go and you buy Salesforce let's say
8:25 or or Slack or or some other software as a service,
8:28 that maybe that you know, price you're paying per user per month or whatever,
8:32 they might turn around and and suddenly charge a lot more.
8:35 Right?
8:35 Or all of us are maybe understand like
8:37 Netflix or Amazon Prime or whatever subscription for movies,
8:43 those sometimes will you know,
8:44 suddenly increase the rates and we don't typically
8:47 have like any type of protection against that, right?
8:49 We have a We don't have a 3-year,
8:51 5-year lock-in that Netflix is only going to cost us you know,
8:54 $9 a month or whatever it might be.
8:57 You know, if they want to make it
8:58 $15 a month the following month, it's like, okay.
9:01 Well, your only recourse is to cancel our subscription.
9:04 Um if you're building your business on something
9:07 that you don't have that long-term lock-in with you know,
9:11 that that you should look at that as as a risk that you know,
9:14 they might just turn around tomorrow and change that price.
9:17 How how tied in are you to that vendor?
9:21 That's true.
9:22 And like again, even like you mentioned that the TV
9:24 stuff and immediately the first thing that popped
9:26 in my mind is like we moved to the streaming
9:29 platform to avoid ads and now ads are back.
9:33 Right.
9:33 And and now recently we were seeing some pretty cool uh video on social
9:40 media about like some platform injecting ads
9:44 and their response and stuff like that.
9:48 So, yeah.
9:49 There's there's a parallel deal there.
9:52 Yeah.
9:52 It's an interesting conversation that now now that you bring it up,
9:55 it does you can kind of see it across sectors even.
9:59 I I feel like as subscription services are being injected into, you know,
10:04 our cars software for example and things like that.
10:06 It's It's interesting on the individual consumer level.
10:10 But Steve, when you're consulting with companies around AI for example,
10:14 how do you how do you like contextualize that risk and measure it?
10:18 Cuz at this point it's kind of hard
10:19 to say exactly where we're going to end up, right?
10:23 Well, exactly.
10:24 And so, that's that's the thing is, you know,
10:26 you have to think about your risk tolerance and whether or not you're
10:30 able to simply pass along the cost to your consumer and and you know,
10:34 what their your consumer, your customer,
10:36 what their tolerance is for, you know, pricing adjustments.
10:40 So, depending on where you sit in in like
10:43 the value chain and and whether, you know,
10:45 you're going to be to see the consumer
10:48 or or be to be to to other businesses, right?
10:50 All the all the math here might change.
10:53 Um one of the simplest ones I would say
10:57 is let's say you're a startup and you are
10:59 trying to get to market as quickly as possible
11:01 and you have a cool idea that leverages AI.
11:04 And so, to get to market as quickly as possible,
11:08 of course you're not going to necessarily just go and invent your own
11:11 models and go buy a bunch of hardware and start training those models, right?
11:15 You'll get to market a lot faster if
11:17 you just leverage the models that already exist,
11:19 the public models that are there.
11:21 Uh and and then you can ship faster, right?
11:24 Uh like the challenge will be like, okay,
11:27 let's say 18 months from now is when you expect to, you know,
11:30 really be a viable, uh, you know,
11:33 product in the market, what will those prices be?
11:36 Cuz today, they're they're relatively cheap, right?
11:40 And and what cheap means, when you look at a a quick spreadsheet here,
11:44 if I bring this up, you know,
11:45 if you're just consuming tokens, you know, per per million tokens,
11:50 depending on the the latest models that you're using,
11:52 uh, you're paying somewhere in the, you know,
11:55 one to two low low two digits, uh, dollars uh,
11:59 for the output, which is the the more expensive side of it, right?
12:02 So, you know, GPT-54 is $15 per million, uh, tokens.
12:08 Opus 46, which right now in April 2026 is widely
12:12 considered one of the best models for for many purposes,
12:15 especially for coding, um, is $25 per per million tokens.
12:19 And there there are some that are more.
12:21 Uh, so let's say that you build your pricing
12:24 around this with the assumption that those are reasonable prices,
12:28 um, but we know that all the companies selling
12:30 things at this these rates are are losing money.
12:32 So, you know, you you expect that you can charge a certain amount,
12:36 and then your cost of goods sold,
12:38 your price that you're paying for this this output, uh,
12:41 maybe it's $250 per million tokens, you know, a year and a half from now.
12:45 What what does that do to your, you know, your pricing strategy?
12:50 Yeah, and then also, cuz it seems like you're saying with a startup,
12:55 the cot or the the benefits make sense in that initial period,
12:59 certainly as you're as you're, you know, speed is a primary objective.
13:04 Uh, then I'm sure the story gets more and more
13:05 complicated as you're either a more established or a bigger business.
13:11 Yeah.
13:11 Yeah, it does.
13:12 And and again, it depends entirely on on your market,
13:16 uh, and your risk tolerance, right?
13:18 So, if if you have uh, uh, not a lot of risk tolerance,
13:22 and you want to make sure you're hedging your bets, right?
13:25 Maybe you want to take advantage of the AI models that are out there now
13:29 at their artificially cheaper rates because they're being subsidized
13:34 by venture capital and loans and and you know, there's a lot of competition.
13:38 So all of the vendors are trying to keep
13:40 their rates as low as they can to gain market share.
13:43 Why wouldn't you, you know, jump on that right now because it's as cheap
13:46 as it's ever going to be and and then
13:49 make sure that you have a plan for how you can get off of that, right?
13:53 You need to avoid vendor lock-in as much
13:57 as possible and I'm sure in the next, you know,
14:00 couple of years, we're going to see increasing interest
14:03 in being able to run your own model, right?
14:05 Whether that's inside your business or even for individual developers,
14:09 you might want to have, you know,
14:11 a graphics card or or another machine that's just operating as a local
14:14 model so that you're not having to pay for a public cloud, you know,
14:19 co-pilot or cloud coder or whatever or because maybe your company
14:24 has decided that they don't want to share all their proprietary
14:27 secrets with one of those public models who are then
14:30 going to take that and integrate it into their training data.
14:33 Yeah.
14:34 And and like like my David mentioned on on YouTube,
14:37 he's there's the AI we use as we like, you know, when we code as we are building
14:43 our services and there's AI that our services are consuming.
14:47 Exactly.
14:48 So, you know, you could be double increasing your fees if,
14:54 you know, the service is is going up.
14:57 Yeah, and David also was asking to you Steve in particular,
15:03 do consult like which piece of that puzzle are you consulting on?
15:07 How devs are using it or how Yeah, the service is.
15:09 Yeah, I mean it's it's both.
15:11 You know, I'm I'm I'm not claiming to be, you know,
15:13 the the best expert on any of these but cuz it's moving so quickly.
15:18 But I'm I'm talking to, you know, business leaders both on how to make sure
15:22 that they're not missing the boat on maximizing
15:25 the productivity of their teams and in how they
15:27 use AI internally to build things and you know,
15:30 ways in which that they might consider applying AI
15:33 to their product offerings in in ways that that add value.
15:39 And how are you seeing those different leaders
15:42 kind of weighing these things that we're talking about?
15:45 Is there Is it super case by case or you kind of seeing a a general
15:49 pattern in how folks are thinking about how
15:52 they want to implement these tools or not?
15:55 I would say it's case by case for the the features for sure.
16:00 There's There's definitely you know,
16:02 depending on on the industry and and the maturity
16:04 of the business like if you are a startup,
16:08 there's some case to be made that adding AI enabled
16:11 features even if they're just on your road map is
16:13 likely to make you more attractive to investors because they're
16:16 all trying to not miss the the AI bandwagon right now.
16:20 Um, and so there's a little bit of that from a you know,
16:22 funding perspective that wouldn't necessarily
16:25 apply to a privately held business.
16:28 You might see some of the same in publicly traded funds where if
16:32 they can make announcements that involve AI being added to their product line,
16:36 that might interest Wall Street and increase their their stock price.
16:39 Um, but I think universally all companies that have
16:43 in-house developers are interested in maximizing how quickly those developers
16:49 can develop their their systems and if AI is
16:52 a way to achieve that, that they want to you know,
16:55 implement that in in the most effective way possible, right?
16:58 They want to limit how much time they spend doing the wrong things
17:01 and going down the the wrong paths and and have someone that knows
17:05 what they're doing explain like here is a a proven effective way
17:09 that you can apply AI and and your team can can go faster.
17:13 Um, and I think there's there's a few of those that are sort of proven
17:16 effective ways right now and and a lot of us are constantly finding new ones.
17:22 [clears throat] Yeah, and kind of going off of that, someone
17:24 in the chat is specifically asking about,
17:27 you know, the features that .NET offers
17:29 for the development of AI-enabled apps and agents.
17:32 I think anyone in this group could probably point to a few,
17:35 but Steve, you know, what what are you seeing interest in there?
17:41 Uh sorry, I'm trying to read that comment specifically.
17:44 popped it up on this uh our view as well.
17:47 offers for development of AI-enabled apps and agents.
17:50 Um for the most part,
17:52 I've been using public uh APIs for things like um that Azure offers,
17:58 you know, and Azure keeps rebranding what what that's called,
18:00 but they've got their foundry um that you can use.
18:03 And then it the as far as .NET, you know,
18:05 there's there's different NuGet packages that you can grab to interact
18:08 with any of these, um but they're not like built into .NET framework
18:12 so much as they're uh a gateway to uh a set
18:16 of APIs that you use that's based on whatever that public model is.
18:20 Does that make sense?
18:23 Yeah, that makes sense to me.
18:24 One thing I'll toss in here that I've been using is
18:27 uh the .NET team has created a repo full of skills.
18:31 It's a .NET/skills repo.
18:33 Um we we can drop a link to that in the chat,
18:36 but uh that would be another thing to check out.
18:39 Uh for example, if you're migrating from one version of .NET to another,
18:42 there's a skill to assist with that.
18:45 There's another one that is pretty cool so about diagnostic.
18:49 So like when you want to dig deeper and like diagnose an issue
18:52 and and examine the logs and stuff like that, there's the skills.
18:57 When I saw that, I was like, "Oh, yeah,
18:58 I should do that more." And like I thought that one was pretty cool.
19:02 There was one for a different platform.
19:06 Um Yeah, and you're reminding me that there's
19:09 uh some new AI-enabled features in Visual Studio.
19:12 Um so so not directly .NET, but in the, you know,
19:15 the the the tooling that does stuff uh specifically diagnostics specific
19:20 to using AI to help you figure out like in the profiler,
19:24 um, you know, where where a bottleneck is or or things of that nature.
19:27 That's cool.
19:28 And and yeah, the skills are are great.
19:31 Uh, there's there's I see you posted the link.
19:33 Um, those are nice cuz you can just drop them in uh
19:37 and and suddenly you've kind of given whatever agent you're using, you know,
19:40 a set of uh of guidelines on on how they
19:43 should behave and and uh what what types of, you know,
19:47 decisions they should make.
19:48 Uh, so that hopefully they're going to make the decisions that an experienced
19:51 .NET developer would make uh for whatever it is they're doing,
19:55 whether it's, you know, using link or using Entity Framework or using ASP.NET.
19:59 Um, you know, you can break these skills down
20:01 so that they're fairly small so that they don't blow
20:03 out the context window of the agent um when
20:06 it's trying to tackle a particular unit of of work.
20:10 That's cool.
20:11 Just popping in my mind, I'm curious to see a chat what what
20:15 what model what tools are you using these days?
20:19 Feel free to uh tell us in the chat.
20:23 Give give some work reading the the chat messages.
20:27 Curious because like you said, it it changed so fast.
20:30 Like I feel like every other day I'm like, okay,
20:34 what what I I didn't know how to do that.
20:36 Like what should I do now?
20:37 Like Yeah.
20:39 Yeah, especially in the last 6 months and even
20:40 the last 3 months especially like it's been accelerating.
20:43 Um, and and there was a a while there where like a year ago, you know,
20:47 some of the the tools were were useful
20:49 and and you could do stuff with, you know,
20:51 ChatGPT or um similar things where it involved a lot
20:56 of copy and pasting of code back and forth and then,
20:58 you know, you might get an error message
21:00 and if you just wanted to kind of vibe code,
21:02 which only became a term little over a year ago,
21:05 uh you were you were still like largely copy pasting things to the to the agent.
21:10 Um, but since we've gotten tools that, you know,
21:12 basically can see all of our code and and work with it directly,
21:16 um, things have just accelerated dramatically.
21:17 And now, you know, some of the more cutting edge uh things that folks are doing
21:22 is is defining multiple agents and or sub-agents
21:25 and and sets of skills and having, you know,
21:28 orchestrators that are delegating tasks to other agents that are
21:32 playing roles just as if it were a a team
21:35 of of humans working on this where you've you've
21:39 got a developer and you've got a uh you know,
21:41 someone that's doing requirements analysis.
21:43 You've got someone who's responsible for QA,
21:45 someone that's responsible for making sure everything is ready to deploy,
21:48 uh and even someone, you know,
21:50 someone but you know, an agent that is uh now, you know,
21:53 responsible for pushing it to production
21:55 and going through all the checklist steps.
21:57 Uh, so a lot of the things that used
21:58 to be DevOps uh and still are still are DevOps, right?
22:02 Are also in in this realm of you know,
22:04 things that uh agents can be playing a large part in.
22:09 Yeah.
22:10 Stephen, as you were talking,
22:11 Squad was one project that came to mind uh from Brady Gaster.
22:16 Uh, is is that kind of what you had in mind?
22:19 Yeah, uh Squad and and some of the things that uh Claude is doing
22:23 as well um with with having multi-agent set up uh is is very powerful.
22:28 Um, yeah, and then I know Brady's stuff is uh is open source
22:32 and and anyone can play with that, so it's definitely worth checking out.
22:35 I know he's done some uh things online with uh
22:38 Jeff Fritz uh that you could check out as well.
22:42 Um, I don't know if if there's links to that.
22:43 I'm sure it's on YouTube.
22:44 But, uh yeah, it involves I essentially what
22:47 I just described where you can, you know, set up a team or squad in this case
22:51 of of agents that are going to, you know, work on things.
22:55 And the the the big goal of this, the big uh thing
22:59 that this is doing for us is making it so that individual agents uh
23:05 can work on something concrete and specific because they tend to do a better
23:09 job when they have a a very specific concrete goal with, you know,
23:14 a a limited uh I don't know, you know, set set of scope, right?
23:19 And it's it's really no different than us as developers, right?
23:23 If if I give a developer a very uh constrained task
23:27 that that I know exactly what I want the output to be,
23:30 I can, you know, have them just do this one thing,
23:33 and if they need to interact with somebody else that, you know,
23:36 has to do some, you know, dependent or related work,
23:39 I can give that somebody else, you know, a similar, really really, you know,
23:43 refined set of requirements and ask them to do this exact thing.
23:45 Um they're probably both going to, you know, do those tasks reasonably well.
23:50 But if I just give a team or a dev like a vague,
23:53 uh I need you to build me a a chat app, like and that's it, right?
23:57 Well, the odds that they give me the thing
23:59 that that was really what I needed are pretty low,
24:02 uh and I'm going to have to iterate quite a bit
24:03 to to get that into something that is is the thing I really wanted.
24:07 Yeah.
24:08 One skill that I start being I used more
24:11 and more and more since I discovered it is grill me.
24:14 I I didn't create it.
24:15 I found it online.
24:17 And and the thing is when you're coding
24:19 or like when we were coding and like doing manually,
24:22 if it would it be like putting down the spec or like starting to code,
24:27 I feel as we are doing that our thoughts about
24:31 what we're trying to achieve was kind of like evolving.
24:34 But now with that prompt, we don't.
24:37 We just have a draft ideas and there's just like, I do this.
24:40 Do it What did you say?
24:41 A chat app?
24:45 [snorts] Then using the skill grill me, it force So like that skill,
24:47 what it does is like you're you're asking your agent to yeah
24:51 bombard you with questions again and again and like each angle and stuff.
24:54 So it force you to kind of like spec it out all of it.
25:00 And I really like that because then, as you said,
25:02 the more descriptive and complete is the task the description of the task,
25:10 then the better are are you chance to have good results.
25:13 Yeah, definitely.
25:14 And And related to that, like the whole plan
25:17 mode and and the ability to or not the ability,
25:20 but the the um the process of defining what the the the outcome
25:26 should be uh up-front is is something we've learned in in, you know,
25:31 last, I don't know, 6 or 12 months, uh is really effective, too.
25:35 Uh and something that I've been doing probably for the last 6 months
25:38 or so is taking that and ensuring that it always has uh an actual,
25:43 you know, output on disk, right?
25:45 So, some of the skills and agents that I've spent a lot of time on are planning
25:49 agents that uh some of their instruction set
25:52 is basically like where to put the plan documents,
25:55 how to name them, uh if you're told to implement a plan,
25:59 make sure that you're doing it step by step
26:01 and updating the plan with your progress as you go.
26:04 Uh some of the things like the Ralph loops
26:06 that were very uh cutting-edge and popular a few months ago,
26:09 um maybe maybe I'm sure some folks are still using them effectively,
26:13 but, you know, that was a task or a uh pattern, let's say,
26:16 that that it would use where it would make sure
26:18 that it always updated how far it had gotten before it stopped.
26:22 And then the next one could just pick up and, you know,
26:24 scan the document and say, "Okay, well, here's the next step I need to do.
26:27 I'll tackle that." Um and and that's that's helpful.
26:30 Whereas I I found in some of the agents when I
26:32 use just inner plan mode and they start churning on the thing,
26:36 uh if something happens, I run out of tokens, or I, you know,
26:39 computer needs to restart, or whatever, uh that that context is lost, right?
26:43 They didn't They didn't save it anywhere by default.
26:45 Uh so, I have to like make my own skill sets to tell it, "No, no.
26:49 Save these things.
26:49 Save Save as you go.
26:50 Save them frequently.
26:51 Uh I want to make sure that, you know,
26:53 that's always there for the next stage and to pick up.
26:58 Yeah, that's super interesting and I feel like in the chat
27:00 we're also seeing some some ways that different folks are,
27:05 you know, trying to make their agents work better for them and and make them,
27:09 you know, smarter and more efficient.
27:12 Um Findel, for example,
27:14 has this good chunk here uh about iterating with Copilot to make
27:19 a plan and then asking it to run simulations with different personas.
27:24 Uh so, it's interesting the ways that kind of as the AI,
27:29 you know, this technology is evolving,
27:30 we're also evolving to be efficient with it.
27:34 It's so new.
27:36 We need we need to learn how to use that tool.
27:41 Yeah.
27:41 And there's different ways to like I was at a user
27:45 group recently and like there was a question like, "Oh, what's the best?
27:48 Is it the instruction document or is it the skills or like I can do?" Like,
27:52 "Well, it will works.
27:54 It will work.
27:55 It you know, you will arrive to obviously
28:00 different output [snorts] because it's generative AI.
28:04 And then like it's perfect.
28:05 You have different ways, so you take what works better with you
28:09 or with your enterprise because, you know,
28:12 there's still plenty of enterprise [snorts] are pretty scared
28:16 of like using AI for tons of different reasons.
28:20 So, like having different approach, I think it's great.
28:25 So, there's there's so many different ways to do stuff.
28:27 So, that's always fun to read.
28:30 I'm I'm reading all the chat is in fire.
28:32 Yeah.
28:33 Like [snorts] all the different ways, the different angles.
28:38 So, one thing I'm seeing in the in the chat,
28:41 I'm scrolling back to find this and I'll pop it up once I do find it.
28:45 Um here we go.
28:47 It's says topic of where the responsibility lies
28:51 with these AI tools when they're being used.
28:54 Um, David followed up this question with saying essentially he he
28:58 believes the responsibility is at the the human the developer level.
29:03 I tend to agree with that.
29:04 Um, you know, in the past you would have been
29:08 the the the person crafting the code in the IDE or the editor.
29:12 Um, I see this as an evolution of that.
29:16 It's just a more powerful tool, but at the end of the day it's
29:19 the human that is defining where the guardrails exist.
29:24 And if the AI goes outside of those guardrails,
29:27 ultimately that accountability still has to come back
29:29 to the developer that was steering that AI.
29:33 Uh, but what do others think about that?
29:39 I think only only humans can be ultimately
29:41 responsible because you can't sue an AI agent, right?
29:45 You can't you know, there's no you can't fire an AI agent.
29:48 Uh, if if the AI agent screws up something like you know,
29:52 that that causes a great deal of harm.
29:55 Um, you know, you you go out in the woods and and you know,
29:59 ask your phone like, "Hey,
30:01 is this mushroom safe to eat?" And then it says, "Yeah,
30:04 sure." And you eat it and you you you're lying
30:06 there on your deathbed and you tell the thing, "Well,
30:08 actually it was poisonous." It's like, "Oh my gosh, you're right." You know,
30:11 like you know, there's no there's no recourse like You are totally right.
30:16 Like of course it'll tell you that, but like
30:18 at that point the damage is done, right?
30:20 Who who was responsible for that action?
30:22 Well, it was you.
30:23 You're the one that took the actual action from it.
30:25 Um, and and the there will always be somewhere in the chain a human that's
30:31 that's going to have to be held
30:32 responsible if nothing else from a legal standpoint.
30:36 Uh, and there's this idea, um, I think Cory Doctorow is the one
30:40 that made it popular of the accountability sync.
30:43 And and basically [snorts] you you can expect
30:46 in certain flows that there's going to be
30:48 someone some expert that's in that flow
30:52 for the sole pure purpose of being the accountability sync.
30:55 Like their their job is to review everything that the AI-driven system does.
31:00 Uh and if any of it is wrong,
31:02 you're the you're the one that we can hold responsible.
31:08 For software, that's going to be developers.
31:11 I agree 100%.
31:18 [laughter] We're all like looking at the chat.
31:23 Yeah, I think also that kind of takes me back
31:26 to Steve what you were talking about with you know,
31:29 having these defined tasks versus really broad swaths
31:33 of things that you want the AI to do.
31:35 I feel like also from a reviewer standpoint,
31:38 that makes it easier to be like, "What is actually happening here?
31:42 And how can I you know,
31:44 as the one who's accountable for this code that's being shipped,
31:46 how can I make sure that it's actually
31:48 doing what we want and doing that safely?" Right.
31:54 And and there's a bunch of things you can do to increase your confidence.
31:57 And they're the same things that we have always been able to do like,
32:01 you know, writing more tests.
32:03 And and for folks that never really bought into like, you know,
32:07 test-driven development or writing a lot of tests for their code,
32:10 like now you can just ask the agent to do it.
32:12 And if you define an agent and and tell it
32:15 in its definition that it's like really good at doing
32:19 proper testing and give it some skills that that help
32:22 it with that, regardless of whether it's unit tests,
32:25 integration tests, usability tests, front-end tests, etc.
32:28 Like all of those tests,
32:29 the the whole goal of them is to give you confidence that the the code
32:33 does what you think it is doing what what what you think it should be doing.
32:37 Um and and at some point, you know, humans are going to use this code
32:41 and hopefully some humans will will touch the code,
32:44 you know, before the users do if it's if it's something really mission critical.
32:48 Um but like that that's one way that we can,
32:51 you know, scale our skills as reviewers uh using these tools.
32:57 And like in the past I don't know.
33:02 Far far away long long time ago uh I had using tests and one test was failing.
33:08 I asked AI to fix it and obviously say,
33:10 "Yep, it's fixed now." Deleted in the test.
33:13 Uh-huh.
33:13 Like all your tests are passing.
33:16 So, I I got I'm assuming they they they are getting better at it, right?
33:21 I I wouldn't make that assumption.
33:23 Um but again, you can you can try and constrain them uh with with skills
33:27 and with other things like we are getting
33:29 better at how we use the agents, I think.
33:31 And and yes, the agents are are getting better, too.
33:33 But um it's a little like, you know,
33:35 when when you get three wishes and you you make the wishes
33:39 uh and and the genie interprets everything in the you know,
33:42 least nice way possible, right?
33:44 Like yeah, you can ask the agent to make
33:46 all the tests pass and they can just be like,
33:47 "Oh, yeah, I added return true to the first line of every test." Like,
33:51 yeah, they all pass now.
33:52 Like, okay, that wasn't really what I wanted, but that's what you asked for.
33:57 That reminds me of um somewhat tangentially,
34:00 have you all ever seen that video of the dad who tells
34:03 his kids to tell him how to make a peanut butter and jelly sandwich?
34:08 And then they're like, "Yeah, they're like,
34:09 get some peanut butter." And he just dumps
34:12 the peanut butter out or these ridiculous things, but it it does make a good,
34:16 I guess, metaphor for what we're talking
34:18 about here of you need to understand how
34:21 to communicate with these tools and make
34:22 sure that they're actually doing what you want.
34:25 Certainly to check, I think, at some point.
34:27 Yeah.
34:28 You know, it's it's good that you trust a lot.
34:31 You know, even now with the YOLO mode
34:34 or all permission and they're running like doing everything.
34:38 Uh but I feel at some point you need to check and and you know, validate.
34:42 Okay, I have 300 tests.
34:45 Do they test the business what like what they are testing and like Right.
34:49 And and we can use metrics that we've had.
34:52 Uh we can look at code quality, you know, static analysis.
34:54 We can look at uh cyclomatic complexity and code coverage and things like that.
34:59 That won't necessarily tell us if the tests are good.
35:03 Um but we can also enlist an agent to do
35:05 a review just like a code review of the tests.
35:08 Um and and come back and just tell us like, you know, what do you think?
35:12 Are these are these good tests?
35:13 And and you know, that can be helpful, right?
35:15 There's there's a lot of ways that we can
35:17 try and use the agents to uh double-check themselves.
35:21 And you know, some folks uh swear by using two different models.
35:25 So like the one model will write the code and a different model will,
35:28 you know, review the code.
35:30 Uh and with the kind of the notion being that like the person that's reviewing
35:34 the code shouldn't be the one that wrote
35:35 it cuz they have blind spots and things.
35:37 Um I don't know that that's necessary because these models generally uh have
35:41 no recollection from in from run to run as far as their context window.
35:45 So um but it but it doesn't it can't hurt really, right?
35:49 If you're using two capable models in that way.
35:51 Uh and and these are all just ways
35:53 to try and take something that is inherently randomized,
35:56 inherently not predictable uh and get a predictable
36:00 result out of the other end of it.
36:02 Um and so there's there's going to be some inefficiencies there.
36:05 Uh which can bring us back to the economic argument which is like for now,
36:10 we're just trying to figure out how to make
36:11 this stuff work correctly where correctly means we get
36:15 like software that works out the other end uh
36:18 and does the thing that we want it to do.
36:21 Um I mean, there's a basically the two failure modes for software developers.
36:24 We we build the thing wrong or we build the wrong thing uh and and AI
36:28 is good at doing both of those you
36:31 know incorrectly and and depending on the guard rails
36:33 that we can put in place the more closely
36:36 we can get them to to building the right
36:38 thing the right way and then the question
36:41 becomes well how much does that cost how many
36:43 agents do I have to spin up run in parallel put in a pipeline and hit
36:48 this code every step of the way to the point where when I make a one line change
36:53 it's it's now costing me a hundred dollars to to you know do all that process
36:58 and and we don't know yet like we don't
37:00 know what that's going to cost in a couple years.
37:03 Yeah I think you almost directly just addressed a couple comments I'm seeing
37:08 based off of our reviewing conversation about
37:11 well isn't this a bottleneck or somebody
37:13 said you know oh sorry I I just switched the comment that somebody
37:18 put up but um if we ask them to do all kinds of tests
37:21 the pipeline will take hours to run so you know finding that line
37:26 between ensuring that we're accurate and you
37:29 know everything's being run safely but also
37:32 you know cost and time and and where does that line fall
37:37 and it sounds like we're still figuring that out as we go a bit.
37:43 Yeah I think it's going to depend on each organization go ahead Scott.
37:47 I wanted to go back to this question because when I read
37:51 this question how do you solve the code review bottleneck the the one
37:55 thing that comes to mind immediately is you have a higher volume
37:58 of let's say pull requests coming through
38:01 now in this new agentic development world.
38:05 How does that how do you scale the code review effort
38:09 on the team when you're seeing a much higher volume than
38:11 you've seen in the past and you could potentially have fewer
38:15 engineers on the team now than you had in the past.
38:17 What are some strategies for making that scale?
38:23 Uh like the the biggest one is to leverage
38:26 these these tools or other tools in the code review, right?
38:30 So, you know, why why are we doing code reviews, right?
38:33 We're doing code reviews because we want to make sure that the code
38:37 that was written has a certain level of of quality, right?
38:41 And and what is what is quality?
38:42 Well, quality is a bunch of uh usually
38:45 like non-functional requirements like is is it secure?
38:49 Uh is it is it, you know, maintainable?
38:51 Uh is it performant?
38:53 Is it accessible?
38:54 And some of those things we can pick up by just looking at the code,
38:57 uh hence the code review.
38:58 Uh you know, is it following our standards?
39:00 Things like that.
39:01 Um but a lot of that stuff can
39:03 also be done through various automated tools, right?
39:07 If if you've got an editor config and you and you run .NET format, right?
39:10 It can enforce all the coding standards for the most part.
39:13 Um if you, you know, really want to know if you're following
39:16 a lot of security uh standards, you can run,
39:19 you know, security scan uh both uh static analysis
39:23 and actively uh in a sandbox with the app running.
39:27 Um accessibility, same way.
39:28 There's various tests you can run.
39:30 Um so, for for most of the types of things
39:32 that we would be reviewing for, we can use
39:35 tools to increase the the speed at which we
39:37 can do that and the likelihood that we'll catch something.
39:40 Um I think the bottleneck then moves from can we review this code
39:45 fast enough to how many of the things that we find are are actionable?
39:50 Are are things that we want to prioritize, right?
39:53 Cuz when you when you set up all these different tools,
39:55 they're going to come back with, you know,
39:56 things of varying levels of, you know,
39:58 information or warning or error or what have you.
40:01 Uh and you have to decide then which ones are worth doing anything about.
40:05 Um and at the end of the day, that's probably a human or at least it's
40:08 humans that come up with the bar of like
40:11 this is what we're okay leaving in and these are
40:14 the things we think we really need to fix.
40:16 Does Does that make sense?
40:18 Yeah, I think it does.
40:19 I I think you have to make use of the agents in some way so that this can scale.
40:25 Without that, you will be underwater forever.
40:29 And And I mean, yes,
40:30 you you certainly can and and I do use the agents to to to make it scale,
40:34 but I want to make the point um having access to agents makes
40:38 it so that it is super easy to do things that aren't random.
40:43 Uh and and and put those into your processes.
40:46 Right?
40:46 What I mean is like, you know,
40:48 we all could have been doing .NET format inside of our uh you know,
40:52 GitHub actions or or whatever Azure DevOps pipelines uh to make
40:57 sure that all the code was standardized every time we committed.
40:59 But, it it took some effort and and, you know,
41:02 testing out GitHub actions and pipeline things is
41:05 like one of the most tedious things that you
41:06 can do as a programmer because the only way to test it is to tweak it,
41:10 submit it, put it in another commit that says test,
41:13 uh and then wait 10 minutes or whatever for the thing to run,
41:16 and then realize your YAML wasn't quite right, and then do it again.
41:19 Like, and that is terrible, so we mostly don't do it as often as we should.
41:24 Agents are really good at getting that right the first time.
41:26 And so, you can just tell an agent to to set
41:29 this up uh and and maybe do it in, you know,
41:32 using a pull request or something with Copilot, you know, monitoring it,
41:36 and just let it keep spinning on that thing until it gets it right.
41:39 And once it's right, it's done.
41:41 Um but, the point is that we use
41:43 the agent to build the thing that is deterministic.
41:47 Right?
41:48 So, we use the non-deterministic thing to help
41:50 us build a deterministic code profiler, code scanner, whatever it might be,
41:54 a quality tool that now is not going to consume any more tokens,
41:58 is not going to randomly decide one day that to tell us
42:01 X when it's always been Y um because it's a random model, right?
42:06 We're going to have that deterministic process
42:08 to improve our code quality from from point forward.
42:13 Good point.
42:14 It's true.
42:14 Like you you need to build a tool to have something stable and always the same.
42:23 There was something else in here that popped out.
42:25 It was seemed like a natural evolution of this conversation
42:28 we're having and it's a sentiment that if
42:30 uh if software engineering is is evolving
42:34 into this uh full-time code reviewer job, hey, I'm out.
42:39 Uh and I'm I'm hearing this a lot.
42:41 Um what do people think about when they hear statements like this?
42:46 Um do you think this is truly the evolution of the software engineering role?
42:51 We will just be steering agents in the right
42:54 direction or how do you see things unfolding?
42:58 I I personally think that yes,
43:01 uh we'll we'll be spending a lot more time using natural language to uh
43:06 talk to the computer to get it to build the thing that we want.
43:10 Um you know, we we've been getting close to where we had,
43:14 you know, voice AI uh integration where uh you know,
43:18 you could you could just, you know,
43:19 especially folks with disabilities could just
43:20 talk to the computer to write code.
43:22 Um but in many cases,
43:23 they were still having to like say every keyword of the, you know,
43:26 Python or C# or whatever language.
43:28 Um and now we're getting to the point where
43:30 you could just tell the the computer, you know,
43:32 not just uh add some logging to this method and it will do that, but, you know,
43:36 write write a whole class, write a whole method, write a whole feature.
43:39 Um and it's a a natural language input uh that produces,
43:45 you know, running software as as the output.
43:47 And uh yes, you can still type all that code in yourself.
43:51 Uh and and yes, you could still type everything in machine code yourself, right?
43:55 Um it's just we we've gotten from uh 3GL, 4GL, you know,
43:59 languages uh programming languages to now natural languages
44:03 as kind of a first-class input to building software.
44:07 Uh and there will still be a place for folks
44:09 that know how everything works uh at that lower
44:12 level and can see it and uh debug it and and optimize it and things like that.
44:18 Uh but it'll be I think fewer and fewer developers will need to know
44:23 uh as much about how individual common
44:26 programming languages today work uh in the future.
44:30 Is is my is my guess.
44:31 And that's what I think most of the the big AI companies are betting on as well.
44:35 So, if model are trained on code and code is written by model,
44:43 will we not plateau at some point?
44:46 To where the you know,
44:48 the the code all all the code that there is is is what it sees and and Well,
44:52 I I don't I feel like it's not possible.
44:56 But at at the same time,
44:57 like if the code is train if model are train on our existing code,
45:02 then like we're we're not pushing the boundaries, I feel.
45:06 Or like if there's a new version that net 10 came out,
45:09 there's no model cannot be trained on that because
45:13 there's no that net 10 code available everywhere.
45:16 Right.
45:17 Right.
45:17 Yeah, it certainly won't have like, you know,
45:19 thousands of blog posts and stack overflow answers
45:21 to to suck in to to do its training, right?
45:24 Um what it will have probably is,
45:27 you know, model context protocol server they can hit.
45:30 Um probably some docs um that that it can hit and I I run
45:35 into this still now where the the models
45:37 were trained before that net 10 shipped.
45:40 And so, I'm trying to get it to build me like a single file C# app,
45:43 which is, you know, really cool feature in that net 10.
45:46 Uh but the models are still like, "Oh, what's this?
45:49 Let me let me try and use, you know,
45:51 C# script from uh this new get package from 5 years ago." Like, "No, no, no.
45:55 You got you got to know there's this new thing." Like, so yeah, yeah,
45:57 there's there's going to be points in time where the model won't be up
46:00 to date because it [clears throat] takes
46:01 a long time to train and it's expensive, so it might be 6 months out of date.
46:06 Yeah.
46:06 But I I also think that the the agents will get
46:10 better at something I don't think they've been doing as much yet.
46:13 Um like like in the initial period what we've seen is
46:16 they've just been getting bigger and bigger and bigger models, right?
46:19 Uh and then training those models to to be better
46:21 at a particular task like writing software or what have you.
46:25 Um and now at this point like all
46:27 the publicly available data is is in the models, right?
46:30 The largest language models have sucked
46:32 in everything that's out there on the internet.
46:34 Um and so the place where they can get more data is two places.
46:37 It's either private proprietary data that's held
46:41 by various businesses and governments and things like that.
46:44 Uh and it's the data that they get
46:46 every day from users interacting with their system.
46:49 And users aren't just people like asking ChatGPT a question,
46:52 it's also all of the agents that are
46:55 doing things um are using these models, too.
46:58 And so they they can if they aren't already uh start to learn which
47:03 things work and which things don't um from both of those use cases, right?
47:07 They you're you're going to start seeing more and more and I just saw one like
47:11 a couple days ago I'd never seen before
47:12 in I think it was ChatGPT where it'll ask explicitly,
47:15 did this solve your problem?
47:17 And and they want to know that cuz if you say yes, then that's you know,
47:20 really valuable training data um that they can then leverage because they
47:24 they can't get it from Stack Overflow anymore cuz their traffic is,
47:27 you know, dried up, but they can get it from their own users.
47:30 Um and the agents can give them the same feedback, right?
47:32 The agent keeps hitting its head against
47:34 the wall until finally it compiles and it runs,
47:37 there's no reason why the agent can't,
47:38 you know, put in a little bit of metadata and say,
47:40 this this right here this was the one that worked.
47:42 You know, and the next time that training data gets updated now it'll know.
47:48 That's cool.
47:48 I never had that that it solved your issue.
47:51 I got at some point two answers and it say, which one did you prefer?
47:55 Yeah, I've seen that, too.
47:56 That's I think that's more of a usability thing cuz usually the answers
47:58 are saying the same thing in a different tone or different way.
48:01 kind of like uh Yeah, but this is like like for the for the whole conversation,
48:06 it popped up and and was asking did this solve your issue?
48:10 That's cool.
48:14 So, Steve mentioned the chat's going wild here.
48:17 I [laughter] declare bankruptcy.
48:21 Um you know, Steve was talking earlier about
48:23 there there's still being a need for, you know,
48:26 these very experienced developers who understand what's
48:30 going on kind of at the lower level.
48:33 Do we think we'll we'll reach a state where um those developers
48:37 will become the mysterious COBOL developers
48:41 that we have today where they're coming
48:42 out of retirement and they're paid millions of dollars because they're the few
48:45 that understand what's truly going on um down at that metal layer?
48:51 Do we see something like that happening in the industry or no?
48:56 I don't know.
48:57 It's It's hard to think that that would happen again like that though.
49:00 Like I I lived through the the the Y2K bug issue
49:05 and that was probably the first time that the you know,
49:07 that that that happened where there a whole lot
49:09 of COBOL developers that suddenly had a lot of demand
49:11 to fix a bunch of mainframe code um because
49:13 of the the date issues with turning over the the the century.
49:17 Um the thing is like all of the AI models
49:21 also know C# really well and they also know Java
49:25 and they also know IL and and all the lower
49:27 level things all the way down the line, you know.
49:29 Um and so, you don't need to go
49:31 find that uh that gray-bearded developer that you know,
49:35 knew it back in the early days because they've got that knowledge
49:37 locked in their head uh as far as the syntax goes, right?
49:41 Um you know, all all the models know all the syntax, you know,
49:44 they could write all the COBOL now for Y2K
49:46 without you having to go tap um that COBOL developer.
49:49 What I think you'll have are uh domain experts that that know, you know,
49:56 how a particular company solved a particular problem
49:58 perhaps because that won't be in the training model.
50:01 You know, some of that will be I think maybe more valuable than, you know,
50:05 syntax of a of an esoteric, you know,
50:07 programming language or a lower-level programming language.
50:10 But, you know, what they say,
50:11 it's it's it's hard to make predictions especially about the future.
50:15 So, um I I I don't know what how it's actually going to play out,
50:20 but I I do think there's still going to be
50:21 a place for smart people that can solve problems.
50:24 And and that's what most software engineers I know are.
50:28 And that's assuming that in the future companies
50:31 are still consuming as much AI as today.
50:36 You know, that the prices sky sky rock sky rock, sorry.
50:40 I'm having an issue with my English today.
50:43 You know, or super expensive.
50:45 Right.
50:46 Oh, well, and that's the question.
50:47 And then coming back to my article that that prompted this topic,
50:50 um if you want to switch the the screen for a second,
50:53 like I think that AI compute costs are going to increase.
50:58 I think we all can probably see that needs
50:59 to happen because they they have massive market share right now.
51:04 Like all kinds of companies are jumping on AI.
51:07 It's not like this is something that only 1% of users are interested in.
51:11 Like there's huge interest.
51:12 Um but they're still losing money, right?
51:14 And so they're not going to make it up in volume.
51:16 They'll just lose money faster.
51:17 So, the only way they're going to make money
51:20 is by raising the the prices to cover their costs.
51:24 And we know they're spending billions and billions on fixed capital expenses.
51:28 It costs them billions to do training for a particular model.
51:31 And they have to keep coming up with more models because
51:33 the the models keep getting better and there there's competition, right?
51:37 So, none of that's going to change.
51:38 So, I think the question isn't, you know, is it going to go up in price?
51:42 The question is how much?
51:44 And I am I'm that that will be somewhere
51:46 between a 10-fold and 100-fold increase in in price.
51:51 All right, so like right now if if
51:53 the the price for Opus is 25 per million tokens, you know, it might be 250,
52:00 it might be 2500 for, you know, not necessarily Opus but for, you know,
52:03 the latest best model per million tokens in a couple of years.
52:08 Uh and that'll be, you know, the question of, you know,
52:12 value at that point is is that still giving us positive ROI or is
52:17 it going to be less expensive for many use cases to to just throw old-fashioned,
52:22 you know, humans at the problem?
52:25 So then we might have still developer
52:29 in the future because AI will be too expensive.
52:34 Yes.
52:35 Certainly.
52:36 I think it goes back and I'll I'll scroll back and I'll pop it up here.
52:41 It goes back to a question towards the beginning of the show.
52:44 Will a company of two to three people be able to carry an AI bill,
52:47 let's say, $6,000 per month?
52:49 I think the answer is no because if
52:51 things go on this trajectory that Steve is describing,
52:55 I think that bill will be much larger than
52:57 $12,000 per month for those two to three developers.
53:01 Um which is a shame in some ways cuz I feel like
53:03 right now we're at a place where the playing field has a leveled.
53:07 Um with this price shift, things will change, of course.
53:13 Yeah, I mean it's it's really hard to predict this but all
53:16 the signs right now are that the costs are not going down
53:21 um for for the the AI that that we're consuming and I
53:26 do think that things like local models will get better and, you know,
53:30 [snorts] we'll continue to see uh Moore's law bringing down the price
53:34 of uh you know compute uh to be able to run, you know, locally and, you know,
53:39 maybe if if one of the big AI vendors goes out of business, suddenly, you know,
53:43 all that RAM that's super expensive right now or all
53:45 those Nvidia chips that might be super expensive right now,
53:48 maybe there's a whole bunch of those on the market
53:49 all of a sudden and and, you know,
53:51 folks can start picking them up to not not just individuals, you know,
53:54 but actual businesses could pick up some of those really heavy-duty things
53:58 that are able to run these large language models with with, you know,
54:02 millions and billions and tokens, uh you know,
54:06 those might become more available than they are today where,
54:09 you know, you can't even get them, right?
54:11 There's there's huge waiting lists and stuff
54:12 to to try and get some of this stuff.
54:15 Yeah.
54:19 Yeah, I think it's such an interesting point, right?
54:23 Like, you can't predict the future, of course,
54:26 and what does it mean if if these prices start to skyrocket,
54:30 but I also think there's an interesting case to be made about,
54:34 you know, the innovation that comes from constraint
54:38 and how we can think about you know,
54:40 we discuss AI holistically, of course, in a lot of these conversations,
54:44 but you know, earlier in the conversation,
54:46 we were talking about using it as a reviewer,
54:49 and what are these, you know, smaller,
54:51 more specific scenarios that maybe we should be
54:55 emphasizing AI's usage more on a scenario-by-scenario basis
54:59 and and how we can get more efficient
55:00 there as opposed to just this company's injecting AI,
55:05 you know, on this massive scale.
55:07 Uh I think we'll see a lot of that, certainly,
55:10 if if what you're saying comes to fruition, Steve,
55:13 and we'll just have to get smarter about where it makes sense to actually
55:17 use it and where we're really going to see that ROI versus not.
55:21 Yeah, and raising the price is not the only way that they can make money better.
55:26 Uh the other one is something we're seeing right now,
55:29 which is uh Anthropic uh is currently, you know,
55:33 capping their their limits uh at a at a lower
55:37 level during their peak use usage time of like 5:00 a.m.
55:40 to 11:00 a.m.
55:41 Pacific time.
55:42 Um and so it it may also be that you've bet your business on this public
55:47 AI model that you paid a subscription for and then all of a sudden overnight
55:51 you suddenly find that you're hitting caps
55:53 and usage limits that weren't there the day
55:56 before and your your customers are complaining
55:59 that your service is unavailable because of that reliance.
56:02 Right, so again, because we don't
56:04 have those long-term lease agreements or support
56:08 agreements that we might have had in in other areas of you know,
56:12 our business computing for this this AI stuff because it's so new.
56:16 Um there's not a lot of protection from them doing you know,
56:20 pulling the rug out from under your business whether that's just you know,
56:23 overnight increases in cost or overnight reductions
56:26 in the the service levels that they're providing you.
56:30 That made me think of like AI will become like
56:33 a power grid or like a water grid where you know,
56:36 you cannot water your plant outside you know.
56:39 Yeah.
56:40 Only on those days or you know,
56:43 electricity is more here that depending on when you're consuming electricity,
56:46 it could be more expensive.
56:48 On your counter than other hours.
56:51 Yeah.
56:52 Right.
56:53 Yeah, and and I focused in this article
56:55 mainly on business costs and business risk,
56:58 but there's there's other non-dollar costs to all of the AI stuff too.
57:02 Like you know, you're talking about you know,
57:04 there's limited resource of water or electricity in a particular area.
57:09 Well, there's there's a lot of impact on on the area where
57:11 these data centers are going in on electricity and on water too.
57:16 And then there's other you know, softer costs like we've alluded to a little bit
57:20 in this conversation so far of like developer skills.
57:24 You know, is is AI just going to make
57:26 us all that dumber and we won't really know
57:28 how to do things because we'll just be so
57:30 used to asking an agent to do it for us.
57:32 Uh and then and then saying, "Yeah, that looks right." and then moving on.
57:36 Um you know, what happens when that tool's not there?
57:39 How how will we, you know, manage for ourselves?
57:42 And the next generation, like when we were talking about this earlier as well,
57:45 I I couldn't help but think about education and how we are going to be
57:50 training the next generation of developers
57:52 to work with these tools and, you know,
57:55 to what degree what kind of expertise are we going to start to need um
58:00 as we go forward and who's going to Are those COBOL developers going to exist?
58:06 Are we going to train folks to kind of take those roles?
58:09 Or is the majority going to be
58:10 really focused on interacting with AI specifically?
58:13 It's an interesting question.
58:16 Yeah, another way um that AI can can get worse
58:21 as a way to to save money is, of course, ads.
58:24 Uh and and also, you know, just taking our data and selling it.
58:28 Um which sometimes is for ads, but sometimes for other reasons.
58:31 Uh and so those are things that, you know,
58:33 in in desperation to to make a profit,
58:36 you know, we can expect these companies to do more and more of.
58:39 Uh is is, you know, either adding advertising or selling our data.
58:44 Uh Same thing we see in in various other
58:47 models that started out free and started out great,
58:50 but over time got worse and worse and worse for their consumers.
58:54 AI vendors are likely to do that as well.
58:56 So, I do think we're going to want to have a solution
59:00 that includes being able to run some of this stuff locally.
59:04 Uh right now, the hardware and and the, you know,
59:07 specs aren't there realistically for for most developers to be able to do
59:10 that anywhere near the same quality that we've
59:12 come to expect from these public models.
59:15 Um but I'm I'm hopeful that in a few years that that'll be possible.
59:20 And then even if it is, you know, you still have like,
59:22 "Okay, that's possible, but is it Is it going to, you know,
59:24 still chew up a ton of electricity uh and you know hurt the the local power grid
59:30 or is it still going to make it so that you know you and and your children
59:34 or your or your upcoming employees don't know
59:37 how to do anything for real because they're used to leaning on the the tool like
59:41 there's a lot of open questions here for sure.
59:45 Yeah.
59:46 Well, will we get the the moment where like it's in deals
59:50 video we saw on social doing the like sport and they say yeah,
59:55 but you know if you use this at discount blah blah blah.
1:00:00 I I they make me laugh so much.
1:00:02 They make me think back of the Apple versus PC ads long time ago.
1:00:08 Yep.
1:00:09 They were very well done I thought.
1:00:12 So we've got three minutes remaining,
1:00:14 but there's a topic here I wanted to pop up.
1:00:17 Maybe we'll end on this discussion cuz I see this a lot out
1:00:21 in the wild too where it's it's saying
1:00:23 it seems everybody forgot the joy of programming.
1:00:27 One thing that comes to mind when I hear something like
1:00:29 this is I I still think it's possible to enjoy programming.
1:00:34 There's this there's this notion of parallelizing work now with agents where
1:00:40 you could potentially hand off the more mundane tasks to an agent
1:00:43 to accomplish while you focus on the piece of the programming
1:00:48 project that actually interests you where you want to be more hands-on.
1:00:52 I think maybe that's the way that you can kind
1:00:55 of still enjoy what you're doing while still being more productive.
1:00:59 What do others think about this?
1:01:04 I certainly agree that AI can do a lot of the kind
1:01:07 of you know grunt work things that maybe you weren't as excited to do.
1:01:11 I know for many folks that might be writing tests or writing documentation.
1:01:16 AI can can spit out you know lots and lots of tests or docs.
1:01:21 Whether or not they're right, you know,
1:01:22 again, that that's that can be a question,
1:01:24 but those are certainly areas you could automate
1:01:27 so that you don't have to do them.
1:01:29 And then maybe you just have to check them ideally.
1:01:32 But yeah, I I can see that for sure.
1:01:36 I do hear from a lot of folks like um that I'm working with that, you know,
1:01:43 from a hobbyist point of view as well,
1:01:45 it is really exciting to feel like you now have the tools
1:01:48 and the time to at least start or take on these projects
1:01:52 that maybe you've been putting off for a long time because you
1:01:55 just didn't have the time or the energy to take on, you know,
1:01:59 a big thing all on your own.
1:02:01 So, I think there is an excitement to AI
1:02:03 in that realm as well where going from idea to, you know,
1:02:07 your first MVP for example of something is is much easier than ever before.
1:02:13 Yeah, I think that's definitely true and it is a way to to get some joy
1:02:16 out of the stuff and to kind of play with it in its own sandbox, right?
1:02:21 Like if you're trying to figure out what works for you,
1:02:24 you know, which model or which, you know, agents or skills or strategies,
1:02:27 like go build that little side thing that you wanted to build for a while
1:02:31 that you just didn't have time for and and just let the agents,
1:02:34 you know, do it and figure out what works for you in in that process.
1:02:38 I've I've done that with half a dozen things in the last
1:02:40 four or five months and it's it's been a lot of fun.
1:02:43 Uh and and it's fine that I'm the only user, right?
1:02:47 Like it's not something that I need to worry as much about, uh you know,
1:02:50 security and some other things um because it's it's
1:02:53 something that isn't open to the world to use.
1:02:58 All right, but unfortunately we're at time.
1:03:00 Steve, did you have anything you wanted to provide
1:03:02 to the audience to help wrap up for today?
1:03:05 Uh yeah, if you want to learn more,
1:03:07 you can certainly find the the articles on the screen.
1:03:10 Hopefully we can put the link in the description.
1:03:13 If you're interested in in talking to me about how to, you know,
1:03:17 look forward to using AI in your company.
1:03:20 Um you can reach me through uh NimblePros.
1:03:23 Uh and that's a company I co-founded and and that's that's what
1:03:27 we do is help uh developer teams get better and and go faster.
1:03:30 So, uh that's that's pretty much it for me.
1:03:34 All right.
1:03:34 Well, thanks again for coming on the show today, Steve.
1:03:37 Uh it was been enjoyable conversation.
1:03:39 Kind of a different format today, but I I really enjoyed it.
1:03:42 Um hopefully everyone else did.
1:03:44 Thanks to our viewers for tuning in again today.
1:03:46 This was the On.NET Live Show.
1:03:49 We hope you check out other recordings of this show.
1:03:52 Those are all posted out at dot.net/live.
1:03:56 And we hope to see everyone next time.
1:03:58 Until then, see you.
1:04:01 Bye-bye.