Hejlsberg Warns Handing Keys to AI Leaves Developers Liable

Creator of TypeScript & C#: AI Software Engineering Predictions | Anders Hejlsberg — Ryan Peterman

Ryan Peterman

Generative models cannot invent novel software architectures because they only extrapolate from memorized training data, leaving human developers fully liable when unreviewed code fails.

AI-assisted summary of Creator of TypeScript & C#: AI Software Engineering Predictions | Anders Hejlsberg — original source. Check the original for full context.

Source published:

Key ideas

  • Anders Hejlsberg challenges predictions that AI will soon write 90% of all code, arguing that while gross boilerplate volume is exploding, AI repeatedly fails at novel architectures like the TypeScript compiler. (0:00)
  • Hejlsberg warns against 'handing AI the keys' and abandoning source code review, emphasizing that legal liability and catastrophic failures fall squarely on developers rather than the models. (2:32)
  • The junior developer hiring pipeline is narrowing rather than disappearing, because cutting off entry-level hiring eliminates future senior engineers. (4:27)
  • Software development is shifting from direct syntax authoring to agent supervision and code review, though Hejlsberg notes reviewing AI diffs is often more mentally exhausting than writing code directly. (4:27)

Detailed notes

## Background Published on YouTube by Ryan Peterman on September 9, 2026, this video features an interview with Anders Hejlsberg, the creator of TypeScript and C#. According to the video description, the conversation explores Hejlsberg's perspective on how AI has impacted software engineering, as well as the TypeScript compiler's performance improvements.

## TL;DR Anders Hejlsberg offers a grounded critique of common AI hype, distinguishing the massive gross volume of boilerplate code generated by AI from truly novel, bespoke software architectures that stochastic models cannot invent. He strongly cautions developers against abandoning IDEs and code comprehension, underscoring that human engineers retain full legal and professional liability when software fails. Hejlsberg also dismantles predictions that junior engineers will be phased out, noting that cutting entry-level roles eliminates the pipeline for future senior engineers. Ultimately, he portrays the future of software development not as human obsolescence, but as a shift from manual syntax authoring to supervising and reviewing AI output.

## Atmosphere & style The conversation has a calm, pragmatic, and deeply seasoned tone that stands in stark contrast to prevailing tech industry hyperbole. Peterman acts as an engaged interlocutor presenting prominent AI predictions, giving Hejlsberg room to dissect extreme claims with measured realism, veteran wisdom, and dry humor. The exchange feels like a grounded masterclass from an industry veteran who happily embraces tools that eliminate rote drudgery while steadfastly defending craftsmanship, critical thinking, and human accountability.

## Bulletin - [00:00](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=0s) Anders Hejlsberg challenges predictions that AI will soon write 90% of all code, arguing that while gross boilerplate volume is exploding, AI repeatedly fails at novel architectures like the TypeScript compiler. - [02:32](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=152s) Hejlsberg warns against 'handing AI the keys' and abandoning source code review, emphasizing that legal liability and catastrophic failures fall squarely on developers rather than the models. - [04:27](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=267s) The junior developer hiring pipeline is narrowing rather than disappearing, because cutting off entry-level hiring eliminates future senior engineers. - [04:27](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=267s) Software development is shifting from direct syntax authoring to agent supervision and code review, though Hejlsberg notes reviewing AI diffs is often more mentally exhausting than writing code directly.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=0s) High output volume from generative tools makes 'AI writing 90% of code' a misleading, self-fulfilling metric driven by boilerplate rather than novel engineering. - [00:00](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=0s) AI models are fundamentally stochastic systems extrapolating over memorized internet data, leaving them ill-equipped to invent unseen algorithmic concepts. - [00:00](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=0s) The TypeScript compiler cannot be written by AI because its custom architecture and specialized implementation paradigms lack training corpus precedents. - [02:32](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=152s) Abandoning IDEs and source code comprehension leaves developers stranded when AI refuses or fails to solve nuanced edge cases and bugs. - [02:32](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=152s) Legal accountability and business liability rest entirely on human engineers when software fails, meaning developers must fully understand what they vouch for. - [02:32](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=152s) Frontier AI lab predictions often err by linearly projecting productivity gains into a theoretical 100% replacement rate, ignoring business context and human judgment. - [04:27](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=267s) Eliminating junior engineering roles is an industry failure mode because organizations cannot produce senior engineers without an entry-level developmental pipeline. - [04:27](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=267s) The engineering pyramid is narrowing at the base, requiring fewer entry-level hires who must advance faster into architectural and supervisory responsibilities. - [04:27](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=267s) Reviewing external pull requests and AI-generated code diffs is often more mentally taxing for senior developers than authoring solutions from scratch. - [04:27](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=267s) Delegating repetitive tasks like boilerplate test suites and rigid framework mechanics to AI provides genuine, high-leverage productivity benefits. - [04:27](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=267s) The code review workflow needs AI integration to summarize intent and explain structural deltas rather than forcing engineers to parse raw file diffs.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=0s) AI Writing 90% of Code: Volume vs. Novelty

Ryan Peterman raises a common prediction from late 2023 and early 2024: that AI will write over 90% of all code within a year. Anders Hejlsberg nuances the claim by separating sheer volume from high-value engineering. He points out that for certain classes of software, such as apps entirely built using AI scaffolding tools, machines are already producing 100% of the code. Because AI generates immense amounts of text effortlessly, the gross volume of written code is skyrocketing, making the "90% of all code" metric somewhat of a self-fulfilling prophecy. However, Hejlsberg questions whether AI will ever write 90% of the truly novel, high-quality code across the industry.

When asked what defines code AI cannot write, Hejlsberg cites bespoke architectures and algorithms that have never been seen before. As a concrete example, he explains that AI could not write the TypeScript compiler. The team has tried, and AI repeatedly fails because the compiler relies on unique concepts and implementation paradigms absent from training corpora:

> "AI we can't forget that AI is at heart a big stochastic machine that has memorized the entire internet, right? And has an ability to extrapolate somewhat over what it's memorized... but if it hasn't seen it before, it's not necessarily going to be all that good at doing it."

While acknowledging that AI's capabilities are wildly impressive and continually improving, Hejlsberg maintains that there is a vast gulf between assisting with existing patterns and independently creating unprecedented software architectures. He adds with a laugh that he hopes AI never completely takes over, asking, "then what's the point of humans anymore?"

- AI code generation skews gross output statistics because generative tools emit massive volumes of boilerplate code. - Unique problem domains—such as the TypeScript compiler's specific architecture and design choices—lack sufficient representation in training datasets, preventing AI from generating working implementations. - Fundamental limitation: LLMs extrapolate over memorized internet data rather than inventing entirely novel algorithmic paradigms from scratch.

---

[02:32](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=152s) "Handing AI the Keys": Liability, IDEs, and Responsibility

Peterman notes another prevailing hot take: that developers will no longer need IDEs or even need to read source code within a year. Hejlsberg offers a blunt warning to anyone willing to abandon the code level: "Well, good luck to you. I mean, if you're going to hand AI the keys... all bets are off." When an AI model refuses or fails to implement a nuanced feature or fix an elusive bug, an engineer who does not understand the underlying system is completely stranded.

Beyond developer ergonomics, Hejlsberg highlights the critical problem of legal and organizational accountability. If an application malfunctions—such as inadvertently compromising a user's bank account—the liability lands entirely on the human creators, not the model:

> "Well, they're going to sue you, not the AI. I mean, it's like at the end of the day, someone has to take responsibility for what AI is doing. And if you're giving it the keys, well, it's on you... I personally would not feel comfortable. I wouldn't be able to go to sleep at night if I didn't understand what it is that I am vouching for."

When Peterman asks why Hejlsberg's grounded perspective differs so sharply from the breathless projections coming out of frontier AI research labs like Anthropic, Hejlsberg explains that people tend to project trends to a hypothetical 100% completion rate. While AI can take on higher percentages of rote work, software engineering inherently demands human judgment to contextualize business logic, organizational constraints, and end-to-end responsibility.

```mermaid flowchart TD accTitle: Accountability and Understanding in AI-Assisted Engineering accDescr: Diagram contrasting uncritical AI delegation with grounded engineering oversight and legal responsibility.

subgraph Abdication["Handing AI the Keys (High Risk)"] A["Blind AI Code Generation"] --> B["Unreviewed Architecture / Obscured Internals"] B --> C{"Bug or Catastrophic Failure"} C --> D["Engineering Paralysis: Cannot Prompt a Fix"] C --> E["Full Legal & Business Liability on Developer"] end

subgraph Engineering["Grounded Engineering (Human in the Loop)"] F["AI as Productivity Tool"] --> G["Human Inspection & Business Context"] G --> H["Clear Understanding of System Invariants"] H --> I["Accountable, Verifiable Releases"] end ```

---

[04:27](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=267s) The Junior Engineer Dilemma and the Evolution of the Craft

Addressing the claim that AI will replace junior software engineers within three years, Hejlsberg exposes the obvious flaw in that logic: without junior engineers, the industry cuts off the pipeline required to create future senior engineers. Unless an organization believes it can operate indefinitely without senior staff, junior engineers must continue to be hired and trained.

Hejlsberg notes that the hiring pipeline is already morphing. The traditional employment pyramid has narrowed at the base, meaning fewer entry-level engineers are hired, but those who are must advance much faster into supervisory capacities. The fundamental craft of programming is shifting from manually typing lines of syntax to directing and reviewing AI agents:

- **From Authoring to Reviewing**: The day-to-day role moves toward reviewing agentic output, verifying logic, and ensuring architectural cohesion. - **The Toll of Code Review**: Hejlsberg admits that after 40 to 50 years of writing software, he still genuinely loves typing out core code and finds reviewing other people's—or an AI's—diffs significantly more mentally taxing than authoring solutions directly. - **Offloading the Tedium**: He embraces delegating mundane chores to AI, noting he is thrilled to avoid writing boilerplate test suites and conforming to rigid testing framework mechanics: "Heck no, if I can farm that out to AI, happy me." - **Improving Review Ergonomics**: Modern code reviews often consist of dumping raw file diffs for a human to decipher. Hejlsberg envisions AI transforming this workflow by explaining architectural deltas, synthesizing intent, and making oversight far more ergonomic.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=0s) AI we can't forget that AI is at heart a big stochastic machine that has memorized the entire internet, right? — Anders Hejlsberg

> [00:00](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=0s) And has an ability to extrapolate somewhat over what it's memorized... but if it hasn't seen it before, it's not necessarily going to be all that good at doing it. — Anders Hejlsberg

> [00:00](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=0s) then what's the point of humans anymore? — Anders Hejlsberg

> [02:32](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=152s) Well, good luck to you. I mean, if you're going to hand AI the keys... all bets are off. — Anders Hejlsberg

> [02:32](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=152s) Well, they're going to sue you, not the AI. — Anders Hejlsberg

> [02:32](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=152s) I mean, it's like at the end of the day, someone has to take responsibility for what AI is doing. And if you're giving it the keys, well, it's on you... — Anders Hejlsberg

> [02:32](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=152s) I personally would not feel comfortable. I wouldn't be able to go to sleep at night if I didn't understand what it is that I am vouching for. — Anders Hejlsberg

> [04:27](https://www.youtube.com/watch?v=qEDMAPpl_O4&t=267s) Heck no, if I can farm that out to AI, happy me. — Anders Hejlsberg

## Mentioned Anders Hejlsberg, Ryan Peterman, TypeScript, C#, Go, Anthropic, YouTube

## Video description Anders Hejlsberg is the creator of TypeScript and C#, and I asked him about how the TypeScript compiler got 10x faster through a rewrite in Go and his thoughts on how AI has impacted software engineering.

Watch the full episode: https://www.youtube.com/watch?v=cywK3XYYJ2o

[Source](https://www.youtube.com/watch?v=qEDMAPpl_O4)

中文简报

海尔斯伯格警告：将控制权交给AI将使开发者承担法律责任 生成式模型无法创造新颖的软件架构，因为它们仅基于记忆的训练数据进行推断，这使得在未经审查的代码出现故障时，人类开发者需承担全部责任。 安德斯·海尔斯伯格（Anders Hejlsberg）对常见的AI炒作提出了切合实际的批评，将AI生成的庞大样板代码量与随机模型无法创造的真正新颖、定制化的软件架构区分开来。他强烈告诫开发者不要放弃IDE和对代码的理解，并强调当软件出现故障时，人类工程师仍承担全部的法律与专业责任。海尔斯伯格还驳斥了初级工程师将被淘汰的预测，指出削减入门级岗位将阻断未来高级工程师的人才培养渠道。最终，他指出软件开发的未来并非人类被淘汰，而是从手动编写语法转变为监督和审查AI的输出。 安德斯·海尔斯伯格质疑了AI很快将编写90%代码的预测，他认为尽管样板代码总量激增，但AI在面对类似TypeScript编译器等新颖架构时仍屡屡受挫。 (0:00) 海尔斯伯格警告不要“把控制权交给AI”并放弃审查源代码，他强调法律责任和灾难性故障完全由开发者而非模型承担。 (2:32) 初级开发者的招聘渠道正在变窄而非消失，因为切断入门级招聘会消除未来的高级工程师储备。 (4:27) 软件开发正在从直接编写语法转向智能体监督和代码审查，不过海尔斯伯格指出，审查AI的代码差异（diffs）往往比直接编写代码更耗费脑力。 (4:27)

Reading summary

Anders Hejlsberg argues that while generative AI is dramatically expanding the gross volume of boilerplate code, it cannot replace human judgment, legal responsibility, or the design of unprecedented software architectures. He cautions against relinquishing code-level understanding and warns that eliminating junior developer roles would destroy the industry's future engineering pipeline. Boilerplate Volume Versus Genuine Architectural Innovation (0:00) Addressing industry forecasts from late 2023 and early 2024 that AI will write over 90% of all code within a year, Hejlsberg separates gross generation volume from high-value engineering. AI scaffolding tools already write 100% of the code for certain standard applications, causing the aggregate volume of machine-generated text to skyrocket. However, Hejlsberg questions whether AI will ever produce 90% of genuinely novel, high-quality code across the industry. Hejlsberg defines the limit of AI capability around bespoke algorithms and architectures unseen in training corpora. Characterizing large language models as stochastic systems that extrapolate from memorized internet data, he cites the TypeScript compiler as a concrete example of software AI cannot write. Despite internal attempts by his team, AI repeatedly failed because the compiler relies on unique conceptual paradigms that do not exist in the training data. Legal Liability and Preserving Code Understanding (2:32) Hejlsberg dismisses claims that developers will soon abandon IDEs and stop reading source code, warning that engineers who blindly hand control to AI will be stranded when a model fails to resolve subtle bugs or nuanced features. If developers do not comprehend the underlying codebase, they cannot intervene when the system behaves unexpectedly. Beyond technical ergonomics, Hejlsberg stresses that legal and corporate liability remains entirely with human engineers. If software malfunctions—such as accidentally compromising a user's bank account—legal action targets the human developers rather than the model. He argues that breathless projections from frontier research labs mistakenly extrapolate incremental automation into total delegation, ignoring that software engineering demands human judgment to contextualize business logic and enforce accountability. Talent Pipelines and the Shift to Code Review (4:27) Responding to predictions that AI will replace junior engineers within three years, Hejlsberg identifies an inherent contradiction: cutting junior positions eliminates the talent pipeline required to cultivate future senior engineers. While the hiring base has narrowed and entry-level engineers must now advance into supervisory duties much faster, the training of junior staff remains essential for long-term organizational continuity. Hejlsberg notes that the engineering craft is shifting from manual code authoring to guiding and reviewing agentic outputs. While he welcomes using AI to eliminate tedious tasks like boilerplate test suites and rigid framework mechanics, he observes that reviewing code diffs is far more mentally taxing than writing solutions directly. He envisions AI tooling evolving to improve review ergonomics by synthesizing intent and explaining architectural deltas, rather than merely presenting raw file diffs. Hejlsberg views AI as an impressive productivity tool for routine patterns rather than an autonomous replacement for software engineers. The enduring challenge for the industry will be balancing automated scaffolding with human oversight, ensuring developers retain foundational system comprehension while training the next generation of senior talent.

Source quotes

[0:00] Anders Hejlsberg: AI we can't forget that AI is at heart a big stochastic machine that has memorized the entire internet, right? [0:00] Anders Hejlsberg: And has an ability to extrapolate somewhat over what it's memorized... but if it hasn't seen it before, it's not necessarily going to be all that good at doing it. [0:00] Anders Hejlsberg: then what's the point of humans anymore? [2:32] Anders Hejlsberg: Well, good luck to you. I mean, if you're going to hand AI the keys... all bets are off. [2:32] Anders Hejlsberg: Well, they're going to sue you, not the AI. [2:32] Anders Hejlsberg: I mean, it's like at the end of the day, someone has to take responsibility for what AI is doing. And if you're giving it the keys, well, it's on you... [2:32] Anders Hejlsberg: I personally would not feel comfortable. I wouldn't be able to go to sleep at night if I didn't understand what it is that I am vouching for. [4:27] Anders Hejlsberg: Heck no, if I can farm that out to AI, happy me.

Original transcript

[0:00] Okay, so the first one that I've heard, [0:02] um, this is more common at the end of [0:04] last year, is that, you know, AI is [0:06] going to write 90 plus percent of the [0:09] code, you know, within a year. [0:11] Um, curious what you think on that. [0:14] >> Well, for certain classes of apps, that [0:16] that might be like all of I code coded [0:19] apps they're writing 100% of the code, [0:20] right? And and there's an awful lot of [0:22] that being written. [0:24] Are they going to write 90% of the high [0:26] quality code on the internet? I don't [0:28] know about that. Um, so [0:32] the the thing is like 90% of what? The [0:34] volume of written code is just like [0:36] going like that right now, right? [0:38] Because AI writes so much code. [0:41] >> [laughter] [0:41] >> So, so in a sense that's a [0:43] self-fulfilling prophecy, right? [0:45] >> [laughter] [0:46] >> But but but when it comes to like [0:50] the super high quality code that no one [0:52] has ever written before, [0:54] um, it's not clear to me that AI is [0:56] going to write 90% of that. [0:58] >> So, when you think about cases that AI [1:01] is not writing all the code, I mean, [1:03] maybe some examples of the super high [1:05] quality stuff. What is that? [1:07] >> [sighs] [1:07] >> Well, [1:09] I mean, it's it's algorithms that no one [1:11] has ever seen before, for example, or or [1:13] particular solutions like I mean, I I [1:15] know that AI could not write our [1:18] compiler, the TypeScript compiler. It it [1:20] just can't. [1:21] I mean, we've tried. It it it can't do [1:23] it. Um, [1:25] and I'm not asking it to, either. I [1:27] mean, I'm not expecting it to, either. [1:29] Um, [1:30] but [1:31] but our compiler is also not a typical [1:34] piece of code. I mean, that's why it's [1:36] bad at writing it, because it's seen [1:38] none of it in the training set. Well, [1:39] I've seen some compilers, but not like [1:41] this one, not with these concepts, not [1:43] with done the way that that we're doing [1:45] it, right? And so, [1:47] it's just not a good fit there. AI we we [1:50] we can't forget that AI is at heart a [1:53] big stochastic machine that has [1:56] memorized the entire internet, right? [1:58] And and his ability and has an ability [2:00] to extrapolate somewhat over what it's [2:02] memorized, right? Um but if it hasn't [2:05] seen it before, it's it's it's not [2:08] necessarily going to be all that good at [2:10] at doing it. Um it's getting to be [2:12] better. I mean, I'm not saying it's it's [2:14] not impressive. I did it's wildly [2:16] impressive what it can do. [2:18] But [2:19] I just think there's a long way still to [2:22] it doing all of it. [2:25] I don't and I don't know that we ever [2:26] get there. I hope we don't because then [2:29] what's the point of humans anymore then? [2:30] I mean, it's [2:31] >> [laughter] [2:32] >> What I see this take often too is [2:34] um you know, [2:36] you won't need to use an IDE within the [2:39] year or you don't need to read the code [2:41] anymore. [2:42] >> Well, good luck to you. I mean, if [2:44] you're going to hand AI the keys, [2:47] I mean, then all bets are off, right? I [2:48] mean, the minute AI you can't convince [2:50] AI to fix this issue or this new feature [2:54] that you want built, [2:55] >> what are you going to do? [2:57] >> I mean, [2:58] >> [laughter] [2:58] >> if you don't understand what what's [3:00] below, right? [3:01] There's a that's one problem. The other [3:03] problem is let's say a user comes to you [3:04] and goes, "Gee, your app just stole my [3:07] bank account. I'm [3:09] Well, it's going to sue you. It's not [3:11] they're going to sue you, not the AI. I [3:13] mean, it's like at the end of the day, [3:15] someone has to take responsibility for [3:17] what AI is doing. [3:19] And if you're giving it the keys, well, [3:21] it's on you. [3:22] You know, but but [3:24] I don't know. [3:25] >> [laughter] [3:26] >> I personally would not feel comfortable. [3:28] I wouldn't be able to go to sleep at [3:30] night. I mean, the if I didn't [3:32] understand what it is that I am vouching [3:34] for. [3:36] >> Why is there such a big difference [3:37] between this opinion and what I see from [3:42] maybe people who work at Anthropic? [3:45] >> Well, the difference is all I'm saying [3:47] here is not that AI can't do amazing [3:50] things. It does. It can. it does that [3:53] every day. But [3:55] but people love these like take it all [3:58] the way to 100%, right? And that's where [4:00] I get off the bus a little bit, right? I [4:03] think oh yeah, okay, fine. We can get we [4:04] can get higher and higher percentage. [4:07] But there still has to be someone [4:09] understanding what's going on and how [4:11] it's relevant to our business problem [4:13] and how it relates to what our [4:15] organization is doing and like and [4:17] there's so many other things that have [4:18] to be connected that it's I think it's [4:20] wrong to just view it as AI did it all [4:23] and we were completely unnecessary, you [4:25] know, and [4:27] >> Yeah, the next one I was going to ask [4:28] you was, you know, 3 years from now will [4:31] AI replace junior software engineers or [4:34] what junior software engineers do today? [4:36] >> Well, if it does then how do you ever [4:38] get senior software engineers? [4:41] Right? I mean it it's just like I don't [4:43] think so. Well, I'm not unless you also [4:45] believe that you will be able to do your [4:47] business [4:48] 3 years later than that without any [4:50] senior software engineers. [4:53] >> Right. [4:53] >> Because where are they going to come [4:54] from? [4:55] You got to train them. [4:57] I mean and but to me [5:00] I think maybe one thing that to some [5:03] extent has happened is that for certain [5:04] classes of programmers the the the [5:06] pyramid has narrowed at the bottom, [5:08] right? There are fewer people coming in. [5:11] Um [5:11] and we are wanting them to advance [5:13] higher up faster. [5:16] Um [5:17] so that they can get to this more [5:18] supervisory role, right? I think the the [5:22] craft of being a software engineer is [5:25] definitely changing from your your [5:27] typing in lines of code [5:29] to you're having agents type in lines of [5:32] code and you are reviewing the work of [5:34] the agents, right? And so [5:38] so you do more reviewing, less typing. [5:41] Um and some people love that. You know, [5:44] I mean and and and and really are [5:46] fantastic at at that workflow, right? [5:49] But it's it's a change in nature of of [5:51] what what the job looks like as a [5:53] programmer and this tool [5:55] that you now have in a toolbox is having [5:57] that impact, right? But it is but a [5:59] tool. [6:00] And there are still programmers involved [6:03] in the in the process. That I firmly [6:05] believe. [6:05] >> Sounds like you enjoy typing out the [6:08] code. Would you be less happy if 5 years [6:11] from now [6:11] >> I I I I for for 40 50 years I have [6:15] enjoyed typing out code. Um [6:18] I'm actually there there [6:20] some and so I will continue to type [6:23] certain parts of the code, you know, but [6:25] there are other parts of the code that I [6:26] don't particularly care to type. I don't [6:28] care to type in tests and follow all the [6:31] rigor of the particular testing [6:33] framework. Heck no, if I can farm that [6:34] out to AI, happy me. [6:37] You know. [6:37] Um [6:39] I do think reviewing has always been [6:41] harder for me. It's harder for me to [6:42] like spend time on reviewing other [6:44] people's code than it is, you know, [6:45] writing the code. [6:47] But I also think that over time we will [6:49] we can make that process of reviewing [6:52] code much more ergonomic than it is [6:54] today, right? I mean, often today you [6:56] just get here's like the list of the [6:58] files that were changed and here are the [6:59] deltas. You figure it out. Right? I [7:02] mean, AI could help us more there by [7:04] explaining what the changes are and and [7:06] and whatever. And so I I think we're [7:07] going to get better. This is going to [7:09] get more interesting, but it will change [7:11] the nature of of the craft.