Vox

Leitura de 16 min

Scott & Mark Learn To… Are Apps Dead?

Summary

Mark and Scott debate the recurring claim that “apps are dead” (and by extension “SaaS is dead”), arguing that most real-world work still requires durable, reliable software with clear interfaces—whether that interface is a GUI, a CLI, or an API. They frame the hype around disembodied AI agents and chat-based interaction as overreaching, especially when people treat LLMs like a universal scripting layer to paper over bad product design.

They explore why deterministic workflows (expense reporting, travel booking, video trimming) shouldn’t be re-solved from scratch by prompting an agent each time. Instead, repetitive, security-sensitive tasks should be codified into dependable software, potentially “sprinkling” AI into specific steps (like OCR) rather than letting an ambiguous agent drive the entire process. They also discuss how dynamic UI generation (e.g., tools that create drop-downs and forms on the fly) may blur the line between “chat” and “app,” but doesn’t eliminate the need for stable systems.

Finally, they push back on the idea that AI will quickly “vibe code” replacements for complex enterprise SaaS (e.g., systems with global deployments, compliance, data sovereignty, and decades of workflow hardening). The conversation closes with a broader reflection on expertise—arguing that the goal shouldn’t be to “extract” and replace human knowledge, but to unlock and augment it—plus a call for viewer comments and topic suggestions.

Topic Timeline

  • [00:00] — Opening question: are we heading toward “no apps,” only AI agents you talk to?
  • [00:01] — Planning more topical episodes; TUIs (text UIs) and agent experimentation as a potential show theme
  • [00:02] — Defining what an “app” is: durable code with a UI, CLI, or API interface
  • [00:03] — Critique of the “just ask a chatbot” model for routine tasks like expense reporting
  • [00:04] — Deterministic workflows vs. AI ambiguity; where AI fits (e.g., OCR) and where it shouldn’t
  • [00:05] — Pushback on “UX is dead”; why forms and structured input still matter
  • [00:06] — Dynamic UI generation in agent tools (drop-downs, multiple-choice prompts) and blurred boundaries
  • [00:07] — “Just-in-time UI” vs. building reusable tools; don’t use LLMs as a poor-man’s scripting language
  • [00:08] — Root-cause critique: bad expense-report UX and broken business processes shouldn’t be “fixed” with agent automation
  • [00:09] — Travel booking as another example where errors are costly; repeating old RPC/API cycles with added ambiguity
  • [00:10] — Why SaaS isn’t dead: scale, security, compliance, and deep workflow complexity (Dynamics/SAP example)
  • [00:11] — Expertise and “knowledge extraction” vs. enabling people; skepticism about turning expertise into a simple markdown file
  • [00:12] — Closing thoughts, future-looking humility, and call for audience comments and episode suggestions

Episode Info

  • Podcast: Scott & Mark Learn To…
  • Author: Microsoft
  • Category: Technology / Education / How To / Business / Careers
  • Published: 2026-02-04
  • Duration: 0h13m

References

Podcast Info

Transcript

[00:00:00] Mark are we getting to the point where it’s gonna be there’s no apps

[00:00:03] It’s just talk to and disembodied agent and you know, you’re looking you know, you’re poking me because we’ve talked about this

[00:00:10] My head explodes when people say apps are dead or sass is dead. Okay, Mark and Scott learned that apps are not dead

[00:00:16] I’m interested in your your take

[00:00:25] One of the things that I think we need to think about we need to work on is that mark

[00:00:29] And I feel like once we’ve done a topic we’ve done a topic and the topic is now done. Yeah, but there’s the agentic

[00:00:35] fun

[00:00:37] Explorations and people really like these shows because it’s real world and it’s factual and it’s it’s got clarity

[00:00:44] So that means that it’s you know, Mark and Scott vibe code for the sixth time

[00:00:48] So it makes a little hard for naming, but it does not a bad show. It’s just not

[00:00:52] A topic I do think that we should start looking at hacker news and tech meme and finding

[00:00:59] Things that are somewhat topical but not necessarily something that’s gonna get us to spin a news cycle

[00:01:04] Mm-hmm that we can talk about that Frank will get mad about. Yeah

[00:01:08] Because like like like my like two ease like I’m really excited about two ease right now

[00:01:15] Text you eyes and I don’t I mean this yeah, you know what team?

[00:01:19] You don’t agree team Lux is becoming like the everybody’s awesome tool of choice now, have you seen that yeah, dude

[00:01:26] But that there’s also arguments

[00:01:27] So that’s a show like that we can talk about two ease

[00:01:30] Because Jay has asked me to do a video on two ease and I did a whole deep research thing and I spun up

[00:01:35] A bunch of agents to experiment on different to II

[00:01:39] Styles and then then MCP UI goose just released MCP UI, but they’re doing it with Chrome and canvas and

[00:01:46] Another guy did Claude

[00:01:48] canvas so

[00:01:50] Basically, you’re asking it to prototype stuff

[00:01:53] Should it prototype and react should it automate Chrome should it make a new to II should it?

[00:01:57] How do you manage agent swarms? What’s the cube cuddle for agents?

[00:02:01] Like there’s a lot of fun conversations happening there people are using two ease because they’re tired of the weight

[00:02:07] Yeah, web view to

[00:02:09] Huh? I don’t necessarily think that that means that the next actually is terminals again. Well

[00:02:15] That could ties into another

[00:02:18] Debate that I’ve even just had with some people yesterday of our apps dead somebody claimed well

[00:02:24] But see Sam Solace said apps are dead. Yeah, I like another not but see this is the thing, right?

[00:02:31] What is an app is a whole philosophical conversation that we can have but we said our apps dead when dcom came out

[00:02:38] And everything was going to be a dcom

[00:02:41] But I an app for me is something that has a user interface that lets something that is

[00:02:49] Durable a piece of code that’s durable and has a user interface

[00:02:52] Okay, so David Fowler vibe coded tally his money system. Yeah, it has no user interface

[00:03:00] That’s you open the github CLI or you open Claude code and then it uses his tool calls and it

[00:03:06] Transforms it into a money. Okay, let me say user or a API interface

[00:03:11] Okay, so the API interface in the case of tally is it’s a CLI

[00:03:15] Yeah, but it is I cut you know when people say an app is dead

[00:03:19] I think what they’re saying is you know what we’re just gonna you you want to do something

[00:03:24] You’re just gonna ask the chat bot to do it. Well, yeah, you’re just gonna say oh, you know what go here

[00:03:30] Here’s my expense reports go OCR them and then file it and you know what?

[00:03:35] I didn’t use an app. I didn’t use an expense report out. I didn’t use an app 40 api’s you call 20 rest api’s and the

[00:03:42] Being this is where I’m like no that makes no sense at all

[00:03:46] No, sir. Like why okay, so here’s why this is the whole Scotty and Star Trek computer

[00:03:52] Yeah, like that’s what they want. That’s fine. But so I’ll tell you filing expenses

[00:03:56] Yeah, everybody that’s got to file expenses going to their chat bot and saying figure out how to OCR this and figure out

[00:04:04] It’s just stupid you’re gonna have an app

[00:04:07] That is like the okay the deterministic

[00:04:10] Here’s you know upload your expenses. Mm-hmm. I’m gonna call OCR on them

[00:04:16] I’m not asking AI to figure out how to do this

[00:04:18] I just know like that’s a flow workflow that is very deterministic very yeah

[00:04:23] It’s it’s gonna work it requires execution over and over again. I need to make sure it’s reliable

[00:04:29] I need to make sure it’s secure

[00:04:31] Are you saying that that should be an app because that’s a problem that we’re throwing an AI app when we just should fix the

[00:04:36] Expense reporting that’s right

[00:04:37] And there’s so many cases like that where you don’t want to leave yourself to the whims of an AI that might screw things

[00:04:43] Up and it’s not just a one like if this is a one-off case of like, you know

[00:04:48] Nobody else in the world is ever gonna want to do this followed by that followed by that

[00:04:52] Right. Sure have AI do it

[00:04:54] but if it’s something that is repetitive and

[00:04:58] Something that requires high security and reliability

[00:05:02] You’re gonna want to codify that in a deterministic manner and you know

[00:05:05] Maybe there’s a I sprinkled in there something like, you know, AI is gonna do OCR it

[00:05:10] Yeah, but the rest of the workflow should just be deterministic

[00:05:15] Totally agree. This is one of my you got me on a rant here because no

[00:05:20] Because that’s also I hear SAS is dead. I’m like, no, it’s not for the same reasons SAS is not dead

[00:05:28] UX is dead chats the new interface. No, that’s not true either

[00:05:32] There’s so many cases where humans need to review something or they need to input something

[00:05:38] That chat and text is just incredibly inefficient means of doing it. Yeah, and in those cases, you’re gonna have a UX like

[00:05:47] Dropdown that’s yes or no, you’re gonna say, you know in the chat bots gonna say yes or no

[00:05:53] You know a form you’re gonna have a chat with a form

[00:05:57] Yeah, so the thing is though that this there’s there’s pushback because the the terminal folks are using ask

[00:06:05] The of the ask user question tool. Have you seen this? No

[00:06:09] Okay. So the way that people are getting around this is this tool for Claude called ask user question

[00:06:17] And you tell it about the tool and then it makes

[00:06:21] drop-downs and yes-noes and interviews and it’s basically dynamically generating a

[00:06:26] SAT multiple choice questions and stuff like that. So rather than interviewing you it’s like, okay

[00:06:32] What behavior do you want retry fail Q and then you get to do that?

[00:06:36] So their dynamic user interface creation is gonna blur the lines

[00:06:41] Where you ask for a thing you’re like, I want to go and do this problem on a resize these videos

[00:06:46] It could do that and scripted it could generate an entire app and that would be an app by your definition or somewhere

[00:06:53] In the middle it dynamically create. Okay. Well, let me ask you this

[00:06:57] Look, so let’s let’s just test this zoom it and video clipping. Mm-hmm

[00:07:02] So often so I do it

[00:07:04] Let’s say AI is so good now that it it could do that in one prompt create that clip. So

[00:07:11] What’s better?

[00:07:13] What I just did which is create, you know, have it create it so now everybody can use it or yeah

[00:07:18] Yeah, every time somebody wants to trim something they go. Hey

[00:07:21] Copilot create it I want to trim this video and then it’s like, okay one second

[00:07:26] Let me go create a UX for that for you. Mm-hmm. And you know, you’re totally right. Yeah

[00:07:30] We don’t need just-in-time you eyes for stuff that is like if something is done twice

[00:07:36] Then make it deterministic totally agree when people are using LLMs as a poor man’s that that’s a phrase that people use

[00:07:43] So as a simplistic scripting language, you’re wasting power. You’re wasting the power of this thing

[00:07:50] Make the non-deterministic

[00:07:52] Yeah, it’s also like we like it’s not a scripting language. It’s not a scripting and AI should be used for cases where you need reasoning

[00:08:00] Ambiguity, it’s a great places where a little bit of ambiguity happens. The expense reporting thing pisses me off

[00:08:06] I’ve seen people try. Oh, yeah

[00:08:08] Look Claude can automate my browser because but what you’re saying is your UI for your expense reporting system sucks

[00:08:15] So bad, yeah that you need the power of the world burning computer

[00:08:19] to fix their UI

[00:08:21] and we want to how many and and

[00:08:25] 50, you know hundreds of thousands of people using the expense report horrible UX

[00:08:30] Rolling AI at it versus the expense report company throwing AI at putting an API on it

[00:08:36] And they love to go and do the whole like look at me OCR in my receipts rather than the business process problem

[00:08:42] Which is why are receipts in that format format? Can I get the receipts from American Express?

[00:08:47] Why am I physically taking pictures of receipts?

[00:08:50] Why are the receipts entering my brain and that exiting my brain entering my phone going to iCloud coming back down?

[00:08:55] Why do I have expense reports at all? It’s kind of like filing taxes in the United States. That’s stupid

[00:09:01] The government should just file the taxes for me. The company should just do the expenses for me

[00:09:06] So it’s very much high on your own supply. You think that that’s the way to solve the problem

[00:09:11] It’s like saying the root cause is I can’t touch that

[00:09:15] Yeah, that’s just too hard to fix. So what I’ll do is throw a whole bunch of crap in, you know

[00:09:21] Bailing wire and yeah, another one. Another great example is using an AI to book travel

[00:09:27] Yeah, that’s that’s done. What it is. That’s done. You’re gonna end up going to the wrong place in the wrong dates

[00:09:33] Probably. Yeah. Well, where is my travel API? Yeah, this is the thing we had decom

[00:09:39] We had core where we had RMI. We had rest we had soap. We have we have it we have MCP

[00:09:43] It’s just like we’re inventing remote method invocation over and over again and then putting an ambiguity loop around it

[00:09:49] And then we’re surprised when things don’t work out. We agree on that

[00:09:53] Well, so going back is SAS dead

[00:09:56] No, and I’ll tell you another thing like if you take a look at dynamics, which is where people you would file expense reports

[00:10:02] That’s a service. It’s a cloud service. It has yeah hundreds of features

[00:10:07] Hundreds of workflows built into it and it’s dynamically, you know extensible as well

[00:10:12] But it’s running in cloud servers all over the world. So because it’s got to honor

[00:10:17] regionality and a data sovereignty and latency and

[00:10:20] it’s got security controls all over it because this is sensitive data and

[00:10:25] You know that people are like, oh, you know what SAS like that is dead people

[00:10:29] You know a startup is just gonna be able to go create a dynamics killer based on AI

[00:10:33] Like what are you?

[00:10:36] That doesn’t make it when people are gonna vibe code SAP. Yeah, like I love looking forward to that

[00:10:41] And I’m like, what about scale? What about regional deployments? What about security? What about compliance?

[00:10:47] What about all the features that it has?

[00:10:49] That are built on real-world workflows and experience with real customers over at this point decades

[00:10:56] You’re just gonna go vibe code that in a in a few months or even a year or two years

[00:11:01] like it just doesn’t make any sense to me that people extrapolate this and

[00:11:07] Not just that but we talk about when do you want to use AI?

[00:11:10] Again, a whole lot of what these SAS applications do is very deterministic

[00:11:17] You’re not gonna go

[00:11:19] Anyway, that’s so let me ask you this. That’s cool. This is a whole other show

[00:11:24] When people talk about like Oh knowledge extraction like that’s not it’s very anti-human like people know stuff

[00:11:30] What is the I don’t want to extract knowledge from humans? I want to unlock it

[00:11:34] I want people to be enabled by these things not replaced by these things

[00:11:38] If you were gonna take something that an expert knew and then people seem to think that they can turn that into a markdown file

[00:11:44] And then have like Virginia vich dot MD and it’s like, okay. I know everything mark knows it’s in markdown now. Everything’s cool

[00:11:52] Right. Yeah, so that what is the role of

[00:11:55] Expertise in this context. Well, and I can how are you using this to be good smarter?

[00:12:00] How could I put my expertise in a markdown file? I

[00:12:03] Don’t be very big markdown file

[00:12:06] And be monstrous

[00:12:09] Yeah, your ego could fit in there as well

[00:12:17] You know

[00:12:17] I think it was Quentin Tarantino who said he who is most likely to make declarative statements is most likely to be called

[00:12:22] a fool in retrospect

[00:12:24] so

[00:12:25] Let’s go and see three five ten years later return to this episode of mark and Scott and see if mark is wrong

[00:12:32] Did apps get replaced it’s 2030

[00:12:34] Where are we? Are we at the the hillside mall with Marty and doc? Okay, like and subscribe

[00:12:39] Make sure that you leave a comment. I cannot say enough that we would rather have a

[00:12:44] Excited group of people who watch the show and comment and let us reply to you than a whole bunch of views

[00:12:49] It’s much more fun to be engaged. So if you are a viewer

[00:12:52] I’d love to hear people who are first-time callers longtime listeners that have something to say make sure to offer

[00:12:59] Suggestions on future episodes of mark and Scott because we are reading the comments and we will put episodes together if we think that we can

[00:13:07] Answer your questions and share with somebody else. So we got some more beers. Yeah, we would appreciate that. All right. Bye everybody

Visão de gráfico