Casey Muratori Says AI Coding Must Outpace Corporate FOMO to Save Code

Is "Vibecoding" Bad for the Industry | Casey Muratori — Ryan Peterman

Ryan Peterman

Widespread executive mandates and unreviewed AI prompting have locked the tech industry in a race where models must improve faster than naive adoption or software quality will collapse.

AI-assisted summary of Is "Vibecoding" Bad for the Industry | Casey Muratori — original source. Check the original for full context.

Source published:

Key ideas

  • Casey Muratori clarifies his warning on 'vibecoding', arguing that catastrophic software quality declines stem from non-engineers shipping unchecked, AI-generated code directly to production. (0:00)
  • Software reliability now hinges on an existential foot race: AI autonomous reasoning must improve faster than the rate at which organizations unthinkingly adopt and ship unverified code. (0:00)
  • Executive fear of missing out (FOMO) is driving irrational company-wide AI mandates, exemplified by Muratori's 'Slopify' criticism of Shopify's internal directives. (4:10)
  • Muratori rejects the narrative that AI ruined engineering standards, pointing out that baseline software quality was already remarkably poor before generative tools arrived. (4:10)

Detailed notes

## Background Published on YouTube by Ryan Peterman on October 2, 2026, the video features a conversation with video game developer and programming creator Casey Muratori. The page description states that the discussion covers surprising parts found while digging through computer history, where bad code comes from, and Muratori's career story.

## TL;DR In this discussion, Casey Muratori and Ryan Peterman examine the phenomenon of 'vibecoding' and the rapid, uncritical adoption of AI tools in software development. Muratori argues that while experienced engineers with computer science backgrounds can effectively leverage AI, shipping unchecked code written by non-engineers straight to production threatens industry-wide software quality. He also critiques corporate FOMO and executive AI mandates—dubbed 'Slopify'—while noting that software quality was already low long before the rise of large language models.

## Atmosphere & style The atmosphere is candid, pragmatic, and conversational, featuring a sharp critique of current tech industry trends delivered with wry humor and straightforward realism. The dialogue moves between blunt warnings about software degradation and lighthearted banter over puns like 'Slopify,' creating an engaged and critical tone about the realities of modern engineering.

## Bulletin - [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) Casey Muratori clarifies his warning on 'vibecoding', arguing that catastrophic software quality declines stem from non-engineers shipping unchecked, AI-generated code directly to production. - [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) Software reliability now hinges on an existential foot race: AI autonomous reasoning must improve faster than the rate at which organizations unthinkingly adopt and ship unverified code. - [04:10](https://www.youtube.com/watch?v=KNuBSNffSUo&t=250s) Executive fear of missing out (FOMO) is driving irrational company-wide AI mandates, exemplified by Muratori's 'Slopify' criticism of Shopify's internal directives. - [04:10](https://www.youtube.com/watch?v=KNuBSNffSUo&t=250s) Muratori rejects the narrative that AI ruined engineering standards, pointing out that baseline software quality was already remarkably poor before generative tools arrived.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) The premise behind Muratori's warning that software quality would degrade was people with zero computer science knowledge typing prompts and shipping results straight to production without review. - [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) A strong computer science background creates a massive divide in AI output quality; experienced engineers direct and validate AI effectively, whereas non-technical users merely hope for the best. - [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) The industry faces a foot race between the speed of autonomous AI capability growth and the unchecked adoption rate by individuals lacking technical evaluation skills. - [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) If AI capabilities were to plateau at current levels, pervasive uncritical adoption would leave the tech landscape facing severe software instability. - [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) Even AI skeptics must now hope for rapid, substantial model improvements because the technology is already too deeply embedded in widespread workflows to roll back. - [04:10](https://www.youtube.com/watch?v=KNuBSNffSUo&t=250s) Executive FOMO pushes enterprises to mandate company-wide AI usage before tools have proven they can reliably address real production demands. - [04:10](https://www.youtube.com/watch?v=KNuBSNffSUo&t=250s) Rational engineering organizations evaluate new technologies through controlled, measured experiments rather than top-down universal mandates. - [04:10](https://www.youtube.com/watch?v=KNuBSNffSUo&t=250s) High-risk, highly autonomous agent workflows are appropriate for isolated R&D experimentation, not blanket operational deployment across an entire workforce. - [04:10](https://www.youtube.com/watch?v=KNuBSNffSUo&t=250s) The tech industry was never operating at 'five-nines' reliability prior to LLMs, making claims that AI single-handedly ruined previously excellent software baseless.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) Vibecoding and the Foot Race Between AI Capability and Adoption

The conversation opens with Ryan Peterman revisiting Andrej Karpathy's famous February 2025 tweet that originally coined the term "vibecoding." At the time, Casey Muratori offered a blunt warning in response: "if you thought software was bad today, buckle up because it's about to get a whole lot worse." Asked whether that grim prediction has played out roughly a year and a half later, Casey clarifies the premise behind his warning. The impending collapse in quality was always predicated on an operating model where people simply "type some stuff in the computer and ship it to prod"—asking an AI to generate a feature and immediately pushing it live without verification, particularly when driven by individuals with zero computer science background. Muratori stresses that he does not want to be overly dismissive of AI's utility, noting that the disparity between users is night and day: an engineer well-versed in computer science leveraging AI can produce substantially better code than someone with no technical foundation who merely prompts a model and hopes for the best.

```mermaid flowchart TD accTitle: The Software Quality Foot Race accDescr: Diagram illustrating the competition between AI capability growth and unchecked adoption without domain expertise.

subgraph Adoption ["Adoption Dynamics"] A["Non-CS Users & Vibe Coders"] --> B["Ship Directly to Production Without Review"] C["Trained Engineers"] --> D["Validate & Direct AI Output"] end

subgraph Capability ["The AI Trajectory"] E["AI Model Autonomy & Reasoning"] end

B -->|"Drives quality down"| Outcome{"Net Software Quality"} D -->|"Preserves quality"| Outcome E -->|"Must outpace naive usage"| Outcome ```

The real danger, Muratori explains, emerges when organizations skip human code review entirely. In the worst-case scenario, "some random person in marketing somewhere who has no idea what programming is just types some stuff in and crosses their fingers." What results is an existential foot race between two competing curves: the rate at which AI autonomously improves versus the rate and capability level of the people adopting it. If model intelligence rises fast enough to consistently match or exceed the tasks delegated by end-users, software quality will not plummet. However, if uncritical adoption outstrips the model's actual reasoning capabilities, the tech landscape is "in for a world of hurt." Muratori notes that the industry currently teeters on a knife's edge between these two trajectories. While he admits he is personally "no fan of AI" and understands why skeptics actively root for its failure, the reality of widespread deployment means everyone must now root for AI to get dramatically better:

> "At the end of the day, if you're talking about something that tons of people are using, you're going to kind of want it to be good... if you want software to not be terrible, you have to kind of still be hoping that 6 months from now the AIs are again significantly better than they were, right? Like that is the only way out of the current situation as I see it."

- **The Human Competence Divide**: A computer science background acts as a critical filter; experienced developers using AI tools produce dramatically higher-quality code than beginners or non-engineers relying solely on conversational prompting. - **The Knife's-Edge Race**: Software stability depends on AI autonomous decision-making improving faster than the speed at which unchecked, review-free code is pushed to production. - **The Stagnation Risk**: Given current adoption levels, Muratori argues that if AI capabilities were to plateau at their current level—or where they were six months prior—the net outcome for industry software quality would be disastrous.

---

[04:10](https://www.youtube.com/watch?v=KNuBSNffSUo&t=250s) "Slopify," Corporate FOMO, and the Reality of Software Quality

Peterman turns to another viral exchange where Casey responded to an internal Shopify executive memo mandating AI adoption with a single-word jab: "Slopify." Muratori laughs, admitting he liked the pun, but uses it to highlight what he sees as one of the most destructive trends in tech: executive fear of missing out (FOMO). Instead of waiting until tools prove they can reliably handle production demands, leadership teams panic, convinced that failing to dive in immediately spells disaster for their enterprise. They force broad adoption before anyone even knows if the outcomes are viable or beneficial.

Muratori contrasts this top-down corporate frenzy with how any rational engineering organization should introduce new technology. A sane adoption model involves running controlled experiments: measuring tool capabilities, verifying whether they resolve concrete engineering bottlenecks faster than existing workflows, and integrating them into production pipelines only when they demonstrate a clear net positive. Having an internal R&D team explore the bleeding edge—"out there yolo swagging it" with 3,000 autonomous agents communicating across a cluster—is entirely appropriate. Mandating that every employee across the entire organization burn vast sums of tokens on daily tasks, however, is not.

Yet Muratori is careful not to pin the blame for bad software entirely on generative AI. He tempers his critique by reminding listeners that developer standards were already abysmal before large language models arrived on the scene:

> "I would temper that statement with just the obvious fact that like we were not exactly a five-nines industry to start out with. Like software quality was really pretty low rolling into the AI era... it's not like software was amazing and the AI showed up and ruined everything. That is a completely ridiculous narrative that that is not true at all."

- **The Corporate Adoption Anti-Pattern**: Companies rush deployments out of panic that waiting for mature tools will leave them behind, reversing the standard engineering practice of validating efficacy before scaling workflow changes. - **Targeted R&D vs. Company-Wide Mandates**: Exploring high-agent clusters and automated workflows belongs with dedicated experimental teams rather than being enforced blanket-style across non-technical departments. - **Historical Software Baseline**: While AI-generated code introduces distinct failure modes, pre-AI software engineering was already plagued by low reliability and poor performance rather than running at "five-nines" quality.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) if you thought software was bad today, buckle up because it's about to get a whole lot worse. — Casey Muratori

> [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) type some stuff in the computer and ship it to prod — Casey Muratori

> [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) some random person in marketing somewhere who has no idea what programming is just types some stuff in and crosses their fingers. — Casey Muratori

> [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) in for a world of hurt. — Casey Muratori

> [00:00](https://www.youtube.com/watch?v=KNuBSNffSUo&t=0s) At the end of the day, if you're talking about something that tons of people are using, you're going to kind of want it to be good... if you want software to not be terrible, you have to kind of still be hoping that 6 months from now the AIs are again significantly better than they were, right? Like that is the only way out of the current situation as I see it. — Casey Muratori

> [04:10](https://www.youtube.com/watch?v=KNuBSNffSUo&t=250s) Slopify. — Casey Muratori

> [04:10](https://www.youtube.com/watch?v=KNuBSNffSUo&t=250s) out there yolo swagging it — Casey Muratori

> [04:10](https://www.youtube.com/watch?v=KNuBSNffSUo&t=250s) I would temper that statement with just the obvious fact that like we were not exactly a five-nines industry to start out with. Like software quality was really pretty low rolling into the AI era... it's not like software was amazing and the AI showed up and ruined everything. That is a completely ridiculous narrative that that is not true at all. — Casey Muratori

## Mentioned Andrej Karpathy, Casey Muratori, Ryan Peterman, Shopify

## Video description Casey Muratori is a video game developer and programming creator also known for his talks about the history of software and computing. We talked about surprising parts he found while digging through computer history, where bad code comes from, and his career story.

Watch the full conversation: https://www.youtube.com/watch?v=jHLbL1Eg4gM

[Source](https://www.youtube.com/watch?v=KNuBSNffSUo)

中文简报

Casey Muratori称AI编码进步须跑赢企业FOMO以拯救代码 高管层广泛下达的指令以及未经审查的AI提示词，已让科技行业陷入一场竞争：模型的进步速度必须超越盲目采用的步伐，否则软件质量将走向崩溃。 在本次讨论中，Casey Muratori与Ryan Peterman探讨了“氛围编码”（vibecoding）现象以及软件开发中对AI工具迅速且不加甄别的采用。Muratori认为，尽管具备计算机科学背景的资深工程师能有效利用AI，但非工程人员将未经检查的代码直接推向生产环境，正在威胁全行业的软件质量。他还批评了企业的FOMO情绪和被称为“Slopify”的高管AI指令，同时指出早在大型语言模型兴起之前，软件质量就已经处于较低水平。 Casey Muratori澄清了他对“氛围编码”的警告，认为软件质量灾难性下滑源于非工程人员直接将未经检查的AI生成代码推向生产环境。 (0:00) 软件可靠性目前取决于一场生死竞速：AI自主推理的提升速度，必须快于各机构盲目采用并发布未经核实代码的速度。 (0:00) 高管对错失良机的恐惧（FOMO）正在催生全公司范围内的非理性AI指令，Muratori针对Shopify内部指令提出的“Slopify”批评便是一个典型例子。 (4:10) Muratori驳斥了AI毁掉工程标准的说法，指出在生成式工具出现之前，基准软件质量就已经非常糟糕。 (4:10)

Reading summary

Software engineer Casey Muratori argues that the industry faces a precarious foot race where rapid, uncritical AI adoption risks severely degrading software quality unless autonomous model capabilities improve fast enough to compensate. The Competence Divide and the Capability Race (0:00) Clarifying his warning following Andrej Karpathy's February 2025 coining of "vibecoding," Muratori argues that the primary danger lies in an operating model where non-technical individuals prompt AI models and ship unverified code directly to production. While experienced software engineers with computer science backgrounds can use AI to generate substantially better results, prompt-reliant workflows lacking technical foundations and code review introduce severe risks to system stability. Muratori frames the current landscape as an existential foot race between the speed of AI capability growth and the rate of unchecked adoption. Despite describing himself as no fan of AI, Muratori claims that because widespread deployment has already occurred, the industry must now hope that models make dramatic leaps in reasoning every six months. If model intelligence plateaus while uncritical adoption continues, Muratori warns the industry will face a disastrous collapse in software quality. Executive FOMO and Top-Down Deployment Mandates (4:10) Addressing his viral "Slopify" characterization of an internal Shopify executive memo mandating company-wide AI usage, Muratori critiques executive fear of missing out (FOMO) as a destructive industry pattern. He contends that leadership teams frequently panic over being left behind, forcing sweeping tool adoption before validating whether the software can reliably handle production demands. Muratori contrasts these top-down mandates with sound engineering methodology. Rational adoption requires isolated research teams to conduct controlled experiments that determine whether new tools genuinely resolve concrete engineering bottlenecks. While Muratori views internal R&D experimentation—such as testing clusters of autonomous agents—as entirely appropriate, he rejects blanket directives that mandate indiscriminate token consumption across an entire organization. The Pre-AI Baseline of Deficient Software Quality (4:10) Muratori qualifies his critique by rejecting the narrative that software was exceptionally reliable before generative AI entered the market. He points out that the software industry was never operating at a "five-nines" standard of reliability, noting that baseline quality and performance standards were already noticeably poor prior to large language models. From Muratori's perspective, generative AI introduces unique risks when organizations eliminate human verification, but it does not represent the sudden corruption of a previously pristine discipline. Instead, unchecked automated generation amplifies structural quality deficits that have long plagued software engineering. Muratori concludes that the industry's widespread integration of AI leaves software stability dependent on continuous model advancements, leaving unresolved whether model capabilities can sustainably outpace naive implementation before structural fragility compounds.

Source quotes

[0:00] Casey Muratori: if you thought software was bad today, buckle up because it's about to get a whole lot worse. [0:00] Casey Muratori: type some stuff in the computer and ship it to prod [0:00] Casey Muratori: some random person in marketing somewhere who has no idea what programming is just types some stuff in and crosses their fingers. [0:00] Casey Muratori: in for a world of hurt. [0:00] Casey Muratori: At the end of the day, if you're talking about something that tons of people are using, you're going to kind of want it to be good... if you want software to not be terrible, you have to kind of still be hoping that 6 months from now the AIs are again significantly better than they were, right? Like that is the only way out of the current situation as I see it. [4:10] Casey Muratori: Slopify. [4:10] Casey Muratori: out there yolo swagging it [4:10] Casey Muratori: I would temper that statement with just the obvious fact that like we were not exactly a five-nines industry to start out with. Like software quality was really pretty low rolling into the AI era... it's not like software was amazing and the AI showed up and ruined everything. That is a completely ridiculous narrative that that is not true at all.

Original transcript

[0:00] Andre Karpathy had this famous tweet [0:03] that kind of coined the phrase five [0:04] coding. [0:05] >> Yes. [0:06] >> And then you said if you thought [0:08] software was bad today, buckle up [0:10] because it's about to get a whole lot [0:12] worse. [0:12] >> Yes. [0:13] >> It's been about a year and a half since [0:16] that tweet came out. The tweet came out [0:18] February of 2025. [0:20] Would you say that that was an accurate [0:22] prediction? [0:24] >> To be clear, I think [0:28] it the whole lot worse part is [0:30] predicated on something that may not [0:32] happen. [0:34] And that is that the idea that we just [0:37] kind of type some stuff in the computer [0:38] and ship it to prod, right? Basically [0:40] like, "Hey, could you make me a thing [0:42] and then publish it?" [0:44] becomes a common way of doing things, [0:46] right? For example, also done by people [0:49] who maybe don't have a computer science [0:50] background. I think it's fair to say and [0:54] that I wouldn't be [0:56] being sort of overly dismissive of AI at [0:58] this point to say that [1:01] a person who is well trained in computer [1:03] science using an AI to make code right [1:06] now [1:07] can make substantially better code than [1:10] someone who doesn't know anything about [1:11] computer science who is just given, you [1:13] know, Fable and type some thought stuff [1:16] in, right? The difference is rather [1:18] dramatic, I would say, from everything [1:20] that I've seen. [1:22] So, part part of my [1:25] concern that I was trying to express in [1:26] that tweet was like [1:28] if the idea is like we're just going to [1:30] type stuff in and we're not going to [1:31] really be checking the code, you know, [1:33] someone who knows computer science is [1:35] not really going to be looking at it. Um [1:37] worst case scenario, it's literally just [1:38] like some random person in marketing [1:41] somewhere who has no idea what [1:43] programming is just type some stuff in [1:44] and hit crosses their fingers, right? I [1:47] think we're in for a [1:48] world of hurt, right? Now, it's a race, [1:52] so it's hard to say because it's [1:54] basically a race of how good can you [1:57] make the AI versus how much adoption [2:00] does it get? Right? It's a curve, right? [2:01] It's like if you can make the AI good [2:03] enough that the people who are adopting [2:05] it at a particular rate are always using [2:08] an AI that's good enough for what [2:09] they're adopting it for, we wouldn't [2:11] expect software to get significantly [2:13] worse. If those curves go the other way, [2:15] we're in a lot of trouble, right? So, [2:17] we're I feel like right now we're we're [2:18] almost kind of teetering on this knife's [2:20] edge. It's it's really to me it feels [2:22] like a foot race of like [2:24] improving AI so it can be more [2:27] autonomous and make better decisions [2:29] without your without you needing to make [2:31] them for it [2:32] versus the capability level of people [2:35] who are using it and the degree to which [2:37] they're paying attention to its output. [2:38] It's like these two curves that are just [2:40] like [2:41] you know, like what's [2:42] >> [laughter] [2:42] >> what's going to happen? [2:44] I don't have a prediction. I don't know [2:46] where we'll be in a year. Obviously, for [2:49] all of our sake, I'm hoping that the AI [2:52] curve wins [2:54] because I agree like I understand [2:58] certainly the perspective of people who [3:01] maybe just don't like AI and don't want [3:03] there to be AI. I can understand the [3:07] wanting it to fail. I I understand that, [3:10] right? [3:11] Um and I'm no fan of AI myself, so it's [3:14] not like I [3:15] like I'm going to criticize someone for [3:17] taking that position. [3:19] But at the end of the day, if you're [3:21] talking about something that tons of [3:22] people are using, [3:24] you're going to kind of want it to be [3:26] good. Like I think at this point, given [3:28] the level of adoption of AI, I really [3:31] don't think it's would be great if it [3:33] stopped getting any better right now. [3:36] Like if this was as good as it was going [3:37] to get, I think that might be bad. [3:39] Um certainly 6 months ago, I think that [3:42] was true. [3:44] Uh and I think it's probably still true [3:45] today. So, I think ideally, if you want [3:49] software to not be terrible, [3:52] you have to kind of still be hoping that [3:54] 6 months from now the AIs are again [3:57] significantly better than they were, [3:58] right? Like that that is the only way [4:01] out of the current situation as I see [4:03] it, right? [4:05] Uh I don't know if that's fair, but [4:07] that's my that's my sort of feeling on [4:09] that. [4:10] >> One of your other top tweets, it was, [4:13] you know, Shopify put out this internal [4:15] memo and you just, you know, you [4:17] replied, you know, "Slopify." [4:19] >> Yes. [laughter] [4:19] >> I guess it's cuz in this tweet it's [4:22] leadership pushing the adoption curve [4:24] maybe harder than the capabilities of [4:26] the AI in this case. [4:28] >> Yeah, although I also just like the pun. [4:31] One of the things that I think is most [4:33] unfortunate about the AI adoption as [4:36] I've seen it [4:38] is just the [4:40] because people think that it's going to [4:42] be this major um [4:45] I guess if I had to categorize the way [4:47] it appears that companies are reasoning [4:48] about it, they're assuming that if they [4:50] don't get in early, it will be a big [4:53] disaster for them, [4:54] right? Like like there's a tremendous [4:56] like they don't just think, "Oh, well, [4:57] we can just wait until the AI does what [4:59] we need it to do and then start using [5:01] it." They're like, "No, we have to do it [5:02] now, like even before we know whether it [5:05] can really do the thing that we want it [5:06] to do or whether we know whether the [5:07] outcomes will be good. Everyone has to [5:08] do it right now. Let's do this, right?" [5:11] Um and I understand why they want why [5:13] why they're going about that way because [5:14] they think that that that is critical, [5:15] right? They obviously believe that's [5:17] very important. [5:18] Um and to me, that's just that's just [5:21] kind of terrifying because [5:23] as with any technology, the sane way to [5:26] do it is to [5:28] measure its capabilities, see how well [5:30] it is able to solve problems that you [5:32] have, see if it solves them faster than [5:34] the way that you were do doing it, and a [5:37] put it into a workflow at such a time as [5:40] you've determined that it is a net [5:41] positive. That's just the same like [5:43] that's the what you would do with any [5:44] technology, right? And you'd probably [5:46] have like your team of people whose job [5:49] it is to assess this thing and they're [5:50] out there yolo swagging it, right? [5:52] They've got [5:53] 3,000 agents working on this cluster [5:55] talking to each other doing, you know, [5:57] God knows what, right? And uh [6:00] So, there's going to be that and [6:02] someone's going to be doing that, but [6:03] that is should not be every org, right? [6:04] Like you wouldn't just be like everyone [6:06] needs to use a ton of tokens, right? Um [6:10] So, yeah, like I I do have concerns [6:12] about that. I don't think that that the [6:15] way AI adoption was done was [6:18] was the best way for quality in [6:20] software. [6:21] But, [6:22] I would temper that statement with just [6:25] the obvious fact that like we were not [6:27] exactly a five-nines uh industry to [6:31] start out with. Like software quality [6:33] was really pretty low rolling into the [6:36] AI era. [6:37] So, I always try to just [6:39] also caveat most of the things that I [6:41] have to say that might be critical of a [6:44] particular thing happening AI with just [6:45] the fact that like look, it wasn't [6:47] particularly great beforehand, either. [6:50] Uh a lot of the software was pretty low [6:52] quality and so you can't [6:55] some AI things may may make things [6:58] worse, but it's not like software was [7:00] amazing and the AI showed up and ruined [7:02] everything. That is a completely [7:03] ridiculous narrative that that is not [7:05] true at all.