Has AI Conquered Coding (It’s Not So Simple…)

Has AI Conquered Coding? (It’s Not So Simple…) ![rw-book-cover](https://readwise-assets.s3.amazonaws.com/static/images/article0.00998d930354.png) 🔗 **Source:** [Has AI Conquered Coding? (It’s Not So Simple…)](https://www.youtube.com/watch?v=SOKfVVOq-Ck) > [!note]+ Document Notes > tags: [ai-development, programming-skills, developer-productivity, artificial-intelligence, software-engineering, cognitive-atrophy] > author: Cal Newport > source: Deep Questions > reference: https://www.youtube.com/watch?v=SOKfVVOq-Ck > date created: {{date:YYYY-MM-DD}} {{time:HH:mm}} > > # 🏷️ Title: Has AI Conquered Coding? (It's Not So Simple…) >  ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ## 📝 Summary - **What happened:** Cal Newport examines Lars Fay's essay "Agentic Coding is a Trap" which argues that AI-driven coding is creating skill atrophy among developers. - **Why it matters:** Current AI coding workflows may undermine the very expertise needed to use AI tools effectively, creating a paradox in software development. - **Key stakeholders:** Senior developers experiencing skill loss, junior developers struggling without AI, software teams facing productivity measurement challenges. - **Risks and tradeoffs:** Speed gains vs. cognitive skill degradation; short-term productivity vs. long-term capability. - **Bottom line:** AI coding tools are powerful but require careful integration that preserves human programming skills rather than replacing them entirely. ## 📖 Main Content The software development industry is experiencing a fundamental shift with AI-powered coding agents. These tools allow developers to describe requirements in plain English and have AI generate code automatically. Some developers report feeling like "10x" or even "100x" developers when using tools like Claude Code. However, programmer Lars Fay argues this approach is unsustainable in his essay "Agentic Coding is a Trap." The core problem: effective use of AI coding tools requires skilled developers who can review thousands of lines of generated code, but AI tools are actively degrading the critical thinking skills needed for this review. **The Skill Atrophy Problem** Evidence from Reddit posts shows both established and junior developers struggling: - Experienced developers report losing their ability to code without AI - Junior developers who learned primarily through AI cannot debug code they didn't write - Computer science students hit a "junior year wall" when AI-assisted early coursework leaves them unprepared for advanced challenges **Industry Reality Check** A veteran developer with 30 years of experience confirmed these observations while noting AI tools do increase speed when used properly. However, he identified additional problems: - Constant context switching due to 1-2 minute AI wait times - Loss of deep work opportunities in coding - Management using problematic metrics like token counts for productivity **Proposed Solution** Rather than abandoning AI tools, Fay advocates for "demoting AI's role" by: - Using LLMs primarily for specs and planning rather than direct coding - Writing 20-100% of code manually depending on task importance - Providing pseudo-code to AI rather than natural language descriptions - Never asking AI to implement something the developer couldn't do themselves This approach maintains a world where AI makes development faster and cheaper while preserving human programming capabilities rather than creating an unrecognizable landscape where coding skills become obsolete. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=0) Introduction to Lars Fay's Essay - Describes current AI-driven development workflow as "slot machine" approach - Growing distance between orchestrator and generated code - Industry experiencing "frenzy of excitement" about AI coding #### ⏱️ [00:01:30](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=90) Developer Enthusiasm Examples - Quote about becoming "10x developer" with Claude Code - Comparison to "Double Rainbow" viral video enthusiasm - Vision of English as new programming abstraction layer #### ⏱️ [00:03:00](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=180) The Core Problem - Only skilled developers can spot issues in AI-generated code - AI tooling negatively impacts critical thinking skills needed for review - Ironic twist: AI undermines skills needed to use AI effectively #### ⏱️ [00:04:00](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=240) Reddit Evidence of Skill Loss - "Losing my ability to code due to AI" - syntax understanding problems - "Handling feeling dumber" - professionals feeling skill degradation - "Letting AI do 100% coding fried my brain" - cognitive skill struggles #### ⏱️ [00:05:30](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=330) Junior Developer Crisis - "Juniors ship fast but can't debug anything they didn't write" - "Hired a junior who learned to code AI, cannot debug without it" - Junior year wall phenomenon spreading industry-wide #### ⏱️ [00:07:00](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=420) Veteran Developer Perspective - Tools work faster for highly skilled developers with good guidance - Context switching and attention fragmentation major problems - 1-2 minute waits lead to multitasking and mental exhaustion - Loss of deep work opportunities in coding #### ⏱️ [00:09:00](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=540) Management and Productivity Issues - Token counts becoming misguided KPI like lines of code - Engineers burning out while producing worse code - Reducing friction in wrong places creates new problems #### ⏱️ [00:10:30](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=630) Proposed Solutions - Demote AI's role rather than eliminate it - Use LLMs for specs and planning, less for direct coding - Write 20-100% of code manually based on task importance - Provide pseudo-code instead of natural language to AI - Never ask AI to implement unfamiliar concepts #### ⏱️ [00:12:00](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=720) Conclusion and Reality Check - AI makes development faster but doesn't eliminate need for coding skills - Vision of evolution rather than revolution in software development - Reality more complicated than enthusiast predictions ## 📚 References - [Lars Fay's essay "Agentic Coding is a Trap"](https://example.com) - Reddit developer community discussions - Cal Newport's newsletter at calnewport.com/ideas ## 🗣️ Paraphrase The software industry is grappling with a paradox around AI coding tools. While these tools can dramatically accelerate development when developers describe requirements in plain English, they're creating a dangerous dependency. Skilled developers are losing their ability to write and understand code—the very skills needed to effectively direct and review AI output. Junior developers who learn primarily through AI never develop debugging skills, creating a generation unable to work independently. The solution isn't abandoning AI but using it strategically: leveraging AI for planning and specifications while maintaining hands-on coding skills through selective manual implementation, especially for complex or critical features. ## 📰 Newsletter **The AI Coding Reality Check You Need to Read** The WordPress community is buzzing about AI coding tools, and for good reason—they can genuinely accelerate development. But before you hand over your PHP skills to Claude or ChatGPT, there's a critical conversation happening in the developer community that affects all of us building themes, plugins, and custom solutions. Cal Newport's latest deep dive explores why experienced developers are warning that "agentic coding is a trap." The issue? AI tools are creating skill atrophy at exactly the moment we need human expertise most. Junior developers who learned primarily through AI can't debug their own code, while senior developers report losing their ability to think architecturally about complex problems. For WordPress developers, this hits close to home. Whether you're building custom blocks, optimizing database queries, or debugging theme conflicts, you need that deep understanding of how things work under the hood. The sweet spot seems to be using AI for planning and boilerplate while keeping your hands on the keyboard for the architecture and problem-solving. **Resources worth your time:** - [WordPress Coding Standards](https://developer.wordpress.org/coding-standards/) - Keep these skills sharp - [WordPress Developer Blog](https://developer.wordpress.org/news/) - Latest on Block Editor development - [Make WordPress Core](https://make.wordpress.org/core/) - Where the real architectural discussions happen The takeaway? AI is a powerful assistant, not a replacement for understanding your craft. Use it wisely, and keep coding. ## ✅ Action Items #### DevRel - **Developer education:** Create workshops on "AI-assisted development best practices" showing when to code manually vs. use AI - **Documentation:** Develop guidelines for using AI tools while maintaining skill development - **Community engagement:** Host discussions on sustainable AI integration in development workflows #### OSPO - **Policy & Governance:** Establish guidelines for AI-assisted contributions to open source projects - **Developer Support:** Create training programs that balance AI efficiency with skill preservation - **Security Management:** Develop review processes for AI-generated code contributions #### WordPress - **Core:** Establish best practices for AI-assisted WordPress development that maintain code quality standards - **Training:** Update learning paths to include AI tool integration while preserving fundamental PHP and JavaScript skills - **Documentation:** Create DevHub guidance on when and how to use AI coding tools in WordPress development - **Community:** Facilitate discussions about sustainable AI adoption in WordPress development #### The WPCC - **Programs:** Fund research into best practices for AI-assisted development education - **Projects:** Support initiatives that balance AI efficiency with developer skill preservation ## 🧪 Hypotheses 1. **Skill atrophy hypothesis:** Developers who rely heavily on AI for code generation will show measurable decline in debugging and architectural thinking abilities within 6-12 months. 2. **Junior developer gap hypothesis:** Entry-level developers trained primarily with AI assistance will struggle significantly more with complex debugging tasks compared to traditionally trained developers. 3. **Context switching degradation hypothesis:** The 1-2 minute wait times inherent in AI coding tools will reduce developers' ability to maintain deep focus and produce lower quality architectural decisions. 4. **Review capability paradox hypothesis:** The cognitive skills required to effectively review AI-generated code will be undermined by the very act of frequently using AI for code generation. 5. **Productivity measurement distortion hypothesis:** Organizations using token counts or AI usage metrics as productivity measures will see decreased code quality and increased developer burnout over time. ### 📊 Methodology - **Longitudinal studies:** Track developer skills over 12 months comparing AI-heavy vs. balanced approaches - **Code quality analysis:** Measure bug density and architectural quality in AI-assisted vs. traditional development - **Cognitive assessment:** Test debugging and problem-solving abilities in developers with varying AI exposure - **Productivity metrics:** Compare multiple measures (features shipped, bug resolution time, code maintainability) across different AI adoption levels ### ❓ Questions - What percentage of coding tasks should remain manual to maintain skill development? - How can we measure the "critical thinking" skills that AI supposedly degrades? - What are the long-term consequences for software quality if junior developers never learn to code manually? - How do different AI integration approaches affect team collaboration and knowledge transfer? ### 🔮 Future Research - Optimal ratios of manual vs. AI-assisted coding for skill preservation - Cognitive load studies on context switching during AI-assisted development - Long-term career trajectory analysis for AI-native vs. traditionally trained developers - Development of assessment tools for evaluating architectural thinking skills ### 💡 Implications - **Education:** Programming curricula may need to explicitly balance AI tools with foundational skill development - **Hiring:** Companies may need new assessment methods that test actual coding ability beyond AI orchestration - **Career development:** Senior developers may need to actively preserve skills through deliberate manual coding practice - **Software quality:** Industry may need new code review standards for AI-generated code - **Team dynamics:** Organizations may require hybrid teams mixing AI-native and traditionally skilled developers ### 🧠 Learning Styles Assessment Reading level estimate: Grade 12-13 (Flesch-Kincaid). Uses technical terminology but explains concepts clearly. #### 👀 Visual learners - **Present:** Clear quote formatting, structured bullet points, video timestamps - **Improve:** Add flowcharts showing AI vs. manual coding decision trees, skill atrophy progression diagrams #### 🔊 Aural learners - **Present:** Video format with clear narration, quoted examples - **Improve:** Audio summary of key points, podcast-style discussion format #### 🗣️ Verbal learners - **Present:** Extensive quotes, clear definitions of terms like "pseudo-code" and "agentic coding" - **Improve:** More analogies comparing to historical technology adoption patterns #### 🧩 Kinesthetic learners - **Present:** Actionable recommendations from Fay, concrete examples from Reddit - **Improve:** Step-by-step guide for implementing balanced AI/manual approach, coding exercises #### 🧠 Logical learners - **Present:** Clear problem-solution structure, evidence-based arguments - **Improve:** Decision matrices for when to use AI vs. manual coding, quantified skill assessment rubrics #### 🧑‍🤝‍🧑 Social learners - **Present:** Community examples from Reddit, developer discussions - **Improve:** Discussion prompts for development teams, peer review frameworks #### 🧘‍♀️ Solitary learners - **Present:** Self-reflection prompts about personal AI usage - **Improve:** Self-assessment questionnaires for skill atrophy, individual practice schedules #### 🌿 Naturalistic learners - **Present:** Evolution metaphors for technology adoption - **Improve:** Ecosystem analogies for balanced AI/human development, sustainability frameworks ### 📊 Accessibility Assessment #### Screen reader semantics - **Present:** Clear heading structure, descriptive quote formatting - **Improve:** More descriptive alt text for any future visual elements, ARIA labels for complex sections #### Keyboard navigation - **Present:** Standard navigation structure - **Improve:** Skip links for long transcript sections, logical tab order #### Contrast and readability - **Present:** Standard text formatting meets basic contrast requirements - **Improve:** Ensure quote formatting maintains sufficient contrast, check timestamp link visibility #### Text resizing and reflow - **Present:** Responsive text structure - **Improve:** Test at 200% zoom for timestamp formatting, ensure no horizontal scrolling #### Media and captions - **Present:** Full transcript provided for video content - **Improve:** Synchronized captions with video, downloadable transcript format #### Forms and interactions - **Present:** No complex interactive elements - **Improve:** N/A for this content type ### 🌍 DEIB Assessment #### Cultural breadth - **Present:** US-centric tech industry perspective - **Improve:** Include global developer perspectives, non-English speaking developer experiences with AI tools #### Inclusive language - **Present:** Gender-neutral technical language, defined jargon - **Improve:** Avoid assumptions about educational backgrounds, explain CS-specific terms more broadly #### Neurodiversity support - **Present:** Well-structured content with clear headings - **Improve:** More frequent headings in transcript section, consistent formatting patterns #### Representation - **Present:** Focuses on developer experiences regardless of demographics - **Improve:** Include diverse voices beyond Reddit posts, acknowledge different learning and working styles #### Psychological safety - **Present:** Acknowledges legitimate concerns about skill loss without shaming AI users - **Improve:** Emphasize that using AI tools isn't "wrong," just needs balance ### ✅ Logical Validity #### Strong reasoning - **Evidence-based argumentation:** Newport cites specific Reddit posts and developer testimonials rather than relying solely on theory - **Balanced perspective:** Acknowledges AI tools' benefits while highlighting concerns, avoiding false dichotomy - **Expert consultation:** Includes 30-year veteran developer perspective with specific, nuanced points #### Fallacies - **Hasty generalization:** Reddit posts may not represent broader developer population - needs larger sample size - **Appeal to authority:** While the veteran developer's experience is valuable, it's still anecdotal - location: veteran developer quotes section - **Post hoc reasoning:** Assumes AI usage directly causes skill atrophy without controlling for other factors - improvement: longitudinal studies needed #### Tone of voice - **Measured and analytical:** Uses phrases like "it's hard to predict the future" and "I don't know if Fay is correct" - **Conversational but informative:** Includes humor (Double Rainbow reference) while maintaining serious analysis - **Cautiously critical:** Questions AI enthusiasm without dismissing benefits entirely ### 🛡️ DARVO Analysis No DARVO patterns detected. The content presents a balanced critique of AI coding practices without deflecting criticism, attacking opponents, or reversing victim/offender roles. Newport acknowledges the legitimacy of both AI enthusiasm and concerns about skill atrophy. ## 🎯 Conclusion The integration of AI into software development represents a complex challenge rather than a simple productivity win. While AI coding tools demonstrably increase development speed for skilled practitioners, they create a paradox: effective use requires precisely the skills that heavy AI reliance may erode. The solution lies not in abandoning these powerful tools but in thoughtful integration that preserves human expertise. Developers must maintain hands-on coding skills while strategically leveraging AI for appropriate tasks, ensuring they remain capable of the architectural thinking and code review that AI cannot replace. This balanced approach promises a future where AI enhances rather than replaces human programming capability. ## 🧱 Refactor Appendix ### 📚 Literature Split Pad ### AI coding tools create skill atrophy paradox > [!quote] [Has AI Conquered Coding? (It's Not So Simple…)](https://www.youtube.com/watch?v=SOKfVVOq-Ck) by [[Cal Newport]] via [[Deep Questions]]. Published 2024. Accessed {{date:YYYY-MM-DD}}. > "Only a skilled developer who's thinking critically and comfortable operating at the architectural level can spot issues in the thousands of lines of generated code before they become a problem." AI effectiveness depends on skills that AI usage undermines. Critical for understanding sustainable AI integration. [view highlight](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=180) #xref #articleseed ID:rw_001 ### Junior developers cannot debug AI-generated code > [!quote] [Has AI Conquered Coding? (It's Not So Simple…)](https://www.youtube.com/watch?v=SOKfVVOq-Ck) > "Juniors on my team ship fast but can't debug anything they didn't actually write." Fundamental gap in problem-solving skills when learning relies primarily on AI assistance. [view highlight](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=330) #xref ID:rw_002 ### Context switching degrades AI coding effectiveness > [!quote] [Has AI Conquered Coding? (It's Not So Simple…)](https://www.youtube.com/watch?v=SOKfVVOq-Ck) > "A gentic system shouldn't go unsupervised for large chunks of time, and they aren't immediate. So, you're inherently stuck with one to two minute waits." Built-in delays force multitasking that reduces quality and increases cognitive load. [view highlight](https://www.youtube.com/watch?v=SOKfVVOq-Ck&t=420) #xref ID:rw_003 --- ### 🔖 Atomic Split Pad (Zettel Candidates) #### AI effectiveness requires skills AI usage erodes - Applies to any cognitive domain where AI assistance reduces practice of underlying skills - Evidence from programming context but principle likely generalizable - Links: [[Skill-Maintenance]] [[AI-Integration-Paradox]] #### Manual practice necessary even with AI assistance - Strategic coding by hand prevents skill atrophy while maintaining AI productivity gains - 20-100% manual work based on task criticality provides skill preservation - Links: [[Deliberate-Practice]] [[Tool-Integration-Strategy]] #### Context switching reduces AI tool effectiveness - 1-2 minute delays inherent in AI systems create multitasking pressure - Deep work capacity compromised by frequent AI interaction cycles - Links: [[Attention-Management]] [[Tool-Design-Consequences]] --- ### ✏️ Fleeting Queue (Extract by Selection) - [F1] Test AI coding skill atrophy in controlled study — context/https://www.youtube.com/watch?v=SOKfVVOq-Ck #fleeting #inbox - [F2] Develop WordPress AI coding guidelines balancing speed and skill — context/https://www.youtube.com/watch?v=SOKfVVOq-Ck #fleeting #inbox - [F3] Create debugging skill assessment for AI-assisted developers — context/https://www.youtube.com/watch?v=SOKfVVOq-Ck #fleeting #inbox --- ## Full Transcript [00:00] Within technology circles, there's a lot [00:02] of buzz right now about a recent essay [00:04] written by a professional programmer and [00:06] entrepreneur named Lars Fay. It opens [00:09] with the following description of the [00:11] current state of affairs of AIdriven [00:13] software development. I'm going to read [00:15] here. [00:16] This workflow takes many shapes at this [00:18] point, but in general, it is a process [00:20] where someone defines the project's [00:22] requirements, generates a plan, and then [00:25] pulls the slot machine lever over and [00:27] over, iterating and reiterating with [00:29] often multiple agent instances until [00:32] it's done. All the while putting a [00:34] growing distance between the [00:36] orchestrator and the code that is being [00:38] generated and committed. [00:41] Now, this new approach to computer [00:42] programming has thrown the industry into [00:45] a frenzy of excitement. And I really do [00:48] mean excitement. Let me read you a real [00:50] quote from an essay that a developer [00:53] posted just a couple months ago. Now, [00:54] I'm reading here. What a fantastic time [00:57] to be alive. With Claude Code, I have [01:00] become, if I do say so myself, a 10x [01:02] developer. Sometimes it feels like 100x. [01:05] I find it all thrilling and amazing. [01:08] It's all intoxicating. To watch Claude [01:10] code work, to ask it to do something [01:12] that I know would take a week or to have [01:15] it figure out some complex bug that I [01:17] would have taken three days to debug is [01:19] almost too much to believe. I don't have [01:21] the superlatives to describe it. H. Now, [01:25] what does that rhetoric remind me of? [01:29] Whoa, that's a full rainbow all the way. [01:33] Double rainbow. Oh my god, it's a double [01:36] rainbow all the way. [01:39] Whoa, that's so intense. [01:41] >> All right, joking aside, this vision [01:45] points to a massive change [01:48] for the world of software development. [01:50] In this new world, no one would need to [01:53] learn to code again. English will be the [01:55] new abstraction layer. Just explain what [01:56] you want and the AI will create it. And [01:58] as a consequence, [02:01] millions of well-paid and highly skilled [02:03] software developers will will be [02:04] replaced by a small number of [02:07] hypercaffinated orchestrators who manage [02:09] hordes of tireless coding agents. [02:12] But is this vision accurate? [02:15] Well, it's Thursday, which means it's [02:17] time for an AI reality check episode of [02:19] this show, which is a good opportunity [02:21] to take a closer look. And indeed, if we [02:24] return to Lars Fay's essay, [02:27] we quickly find that he's not a believer [02:30] in the idea that we're on the path to a [02:32] world without code. If anything, [02:35] Fay thinks that this new style of [02:37] software development isn't sustainable. [02:40] A belief he captures bluntly in the [02:42] title of his essay, agentic coding is a [02:46] trap. Now, this seems like an argument [02:48] that gets at the heart of so many issues [02:50] about AI and knowledge work and hype and [02:52] reality. So clearly we need to take a [02:55] closer look and that's exactly what [02:56] we're going to do. As always, I'm Kell [02:58] Newport and this is Deep Questions, the [03:02] show for people seeking depth in a [03:04] distracted world. [03:12] So what's Lars Fay's issue with the [03:15] current vision of agents taking over all [03:17] coding? I want to read you another quote [03:19] from his essay. [03:21] Being successful with a with this [03:23] approach to coding agents hinges on a [03:25] rather crucial element. Only a skilled [03:28] developer who's thinking critically and [03:30] comfortable operating at the [03:32] architectural level can spot issues in [03:34] the thousands of lines of generated code [03:36] before they become a problem. Yet, in an [03:39] ironic twist of fate, it's the [03:41] individual's critical thinking skills [03:43] and cognitive clarity that AI tooling [03:46] has now been proven to impact [03:48] negatively. [03:50] As Fay elaborates, this idea that AI is [03:53] undermining the very skills needed to [03:55] use AI effectively is not an abstract [03:58] fear. It's actually happening right now [04:00] and at wide scale. And it's something [04:01] that programmers are really starting to [04:03] talk a lot about. [04:06] To support this claim, Fay points to a [04:08] bunch of posts from Reddit where people [04:12] are worried about exactly this skill. I [04:14] want to read you a few headlines and [04:16] quotes from uh that Fay points out here. [04:18] Okay, so here's one. The headline is [04:20] losing my ability to code due to AI. And [04:23] the post goes on to say, I feel like my [04:25] skills are really deteriorating. I think [04:28] that inherently not understanding the [04:30] syntax is a problem in itself. I might [04:32] be missing something, but uh a lot of it [04:36] takes a lot of time to check the AI [04:37] hasn't made a mess or created subtle [04:39] bugs. And you have to understand the [04:40] language and how to write in it. All [04:41] right, here's another headline from [04:42] Reddit. handling feeling dumber or like [04:46] losing skills due to the need of using [04:48] AI. That post includes the following [04:51] quote. Every time I use AI to do [04:53] anything, I feel like I'm losing my [04:55] skills that I and that I become a worse [04:57] professional. Another Reddit post is [04:59] titled, "Letting AI do 100% coding fried [05:03] my brain. Help." It goes on to say, [05:06] "Lately, I found myself struggling with [05:08] my general cognitive skills and problem [05:10] solving skills." [05:12] Now, these examples are about [05:14] established developers who are losing [05:15] their ability to think and code. But [05:17] what about junior developers? As FY [05:19] points out, the problem here becomes [05:21] even worse. Struggling to write code [05:23] from scratch is how you get better at [05:26] understanding and reviewing code. If you [05:28] skip the struggle step and jump straight [05:30] to the code review, [05:32] you don't have any skills to fall back [05:34] on. All right, here are two recent uh [05:37] Reddit headlight headlines about this [05:39] phenomenon that Fay highlights. First [05:40] is, "Juniors on my team ship fast but [05:44] can't debug anything they didn't [05:45] actually write." Another headline says, [05:47] "Hired a junior who learned to code AI, [05:50] cannot debug without it, don't know how [05:51] to help them." Now, in my world of [05:54] computer science, this is a well-known [05:55] effect called the junior year wall, [05:58] right? which describes what happens in a [06:00] CS curriculum when a CS student uses AI [06:02] to do most of their basic coding related [06:04] classes but then when they face [06:06] challenging classes during their junior [06:07] year their lack of understanding catches [06:09] up to them and they struggle. [06:12] Fay is saying that this is basically uh [06:15] the junior the junior year wall effect [06:18] is basically happening in the whole [06:19] coding industry writ large and that this [06:21] is a problem. [06:24] So after I read this essay, I asked a [06:26] veteran developer I know, someone with [06:28] 30 years of experience, whether or not [06:30] it matched his current experience, and [06:33] he said it largely did. He then offered [06:36] four additional points about the current [06:38] state of affairs and programming. [06:39] They're nuanced. I'm going to read them [06:41] to you because I think it's important. [06:42] All right, here's point number one. He [06:44] said, "These tools do get the job done [06:48] faster [06:50] if you are a highly skilled developer [06:52] and have lots of good rules and guidance [06:54] in place. If you use the tools [06:56] correctly, it is in fact much faster [06:58] than typing code out by hand while still [07:00] producing quality code. This is why I'm [07:02] still using them despite all the [07:04] problems I'm about to state." He then [07:05] goes on, "Point number two, context [07:08] switching and attention fragmentation [07:10] are a huge problem. A gentic system [07:12] shouldn't go unsupervised for large [07:14] chunks of time, and they aren't [07:15] immediate. So, you're inherently stuck [07:17] with one to two minute waits. Your [07:19] choices are to sit around or multitask, [07:21] and this leads to out of control context [07:23] switching, lower quality output, and [07:24] mental exhaustion. All right, point [07:26] number three. He says, on a personal [07:29] level, I absolutely feel the effects of [07:31] skill atrophy. This used to be my source [07:34] for deep work. It no longer is. His [07:37] fourth point, agentic coding is going to [07:40] become a huge problem for software teams [07:42] in the next few years. I'm already [07:43] seeing management at several companies I [07:45] work with turning to token counts as a [07:48] productivity measure. Just like lines of [07:50] code was a misguided KPI, this will be [07:52] another measure of how fast you're [07:54] spinning on the hamster wheel. Engineers [07:56] will burn out and produce worse and [07:58] worse code as time goes on. It has the [08:00] same problem of reducing friction in the [08:01] wrong places, leading to the same [08:02] problems you've outlined with Slack and [08:04] email. [08:06] All right. So, having outlined these [08:08] complaints and again, let me just [08:10] summarize them real quick, right? [08:12] If we rely almost entirely on agents to [08:16] generate code and we just describe what [08:18] we want, [08:20] the senior developers will lose their [08:22] ability to actually understand code, [08:23] which is needed to ask for the right [08:25] things and to check the code the agents [08:26] produce. And junior developers will [08:28] never learn coding skills in the first [08:29] place, which means they won't really be [08:31] able to work usefully with agents at [08:34] all. [08:35] All right. So, what is the solution [08:36] here? Is it to throw AI tools out the [08:38] window? No. [08:40] Fay still uses them. The developer I [08:42] still talk to use them. But we're going [08:44] to have to think much more carefully [08:46] about how we integrate AI coding tools [08:48] into software development. Now, Lars Fay [08:50] has a suggestion which he summarizes as [08:53] demote AI's role. All right. He says, [08:57] I'm quoting here, I'm certainly not [08:59] advocating for typing out all code [09:01] manually. He then soon adds, "What I am [09:05] advocating for though is leveraging LLMs [09:07] and coding agents as secondary [09:09] processes, a way that doesn't sacrifice [09:12] the individual skills at the altar of [09:14] productivity." [09:16] Fay then goes on to give more details [09:18] about what he means here. He he [09:19] clarifies that he uses LLM [09:23] LLM more for specs and planning and less [09:25] for coding, which he does on 20% to [09:28] 100%. He'll he'll write 20% to 100% of [09:31] the code himself depending on the [09:32] importance of the task. More important [09:34] task, more code he writes. [09:37] When he does have the LLM write code, he [09:39] often provides it pseudo code. So [09:41] basically pseudo code, if you don't [09:43] know, is like um it's like a programming [09:46] language but less syntactically [09:48] demanding. So it's like you're actually [09:49] giving com uh programming language style [09:52] demands but you don't have to be super [09:54] exact about every semicolon or [09:55] parenthesy but you're basically saying [09:57] this is what I want you to do this loop [10:00] this check and then the LLM can take [10:02] that and put it into the proper [10:03] programming language that's being used. [10:07] He also says I never ask an LLM or agent [10:10] to implement something that I've never [10:11] done before or couldn't do on my own. [10:14] All right let's step back here. What [10:16] Lars Feay and others are beginning to [10:18] realize is that AI being good at [10:20] something doesn't necessarily mean that [10:22] it can easily replace humans. [10:25] They are seeing a world in which people [10:26] still need to learn to code. At first, [10:29] they might code things AI could easily [10:30] do just to build skill. And then as they [10:32] get more senior, they can use AI more [10:35] often and leave their coding time for [10:37] more complicated or missionritical [10:38] features and spend more of their time [10:40] reviewing code. and they can do this [10:42] review because they've learned and [10:44] maintain their own programming [10:45] abilities. Now, to be clear, this is a [10:49] vision of a world in which AI coding [10:51] tools are ubiquitous and they really do [10:53] make software development much faster [10:54] and cheaper. As I mentioned, Fay uses [10:56] coding agents, as do most of the [10:59] developers who agree with his warnings. [11:02] But this vision is of a world that is [11:04] not unrecognizable from the one we live [11:06] in today. It's not a world where [11:08] computer code doesn't exist. It's not a [11:09] world in which the ranks of software [11:12] developers catastrophically contracts. [11:15] Now look, I don't know if Fay is [11:17] correct. It's hard to predict the future [11:20] and what's going on within computer [11:21] programming and software development is [11:23] fastm moving. But the fact that he's [11:26] attracting a lot of respectful attention [11:28] in the software industry at the moment [11:30] is an important reminder of something [11:31] that we've learned time and again in the [11:33] history of technology. [11:36] reality is almost always more [11:38] complicated and uneven than enthusiasts [11:42] would like or predict. And there's no [11:45] reason yet to believe that AI is an [11:48] exception to this longstanding rule. All [11:51] right, that's all I have for today. [11:55] Remember, you should care about AI but [11:56] not trust everything you read about it. [11:58] All right, thanks. Hey, if you made it [11:59] this far, you must be ready to join my [12:03] fight for depth in a distracted world. [12:05] Now, the best way to do this is to join [12:07] over 125,000 people who receive my email [12:12] newsletter each Monday. You can sign up [12:15] at calupport.com/idas. [12:18] And when you do, I will send you a free [12:20] guide to my seven best ideas about [12:23] cultivating a deep life. Sign up today, [12:25] calubber.com.  ## Highlights - Test AI coding skill atrophy in controlled study — context/https://www.youtube.com/watch?v=SOKfVVOq-Ck - Note: ✏️ Fleeting capture #atom/fleeting/f1 - Develop WordPress AI coding guidelines balancing speed and skill — context/https://www.youtube.com/watch?v=SOKfVVOq-Ck - Note: ✏️ Fleeting capture #atom/fleeting/f2 - Create debugging skill assessment for AI-assisted developers — context/https://www.youtube.com/watch?v=SOKfVVOq-Ck - Note: ✏️ Fleeting capture #atom/fleeting/f3 - ** Reddit posts may not represent broader developer population - needs larger sample size - Note: 🔍 Fallacy: Hasty Generalization #fallacy/hasty-generalization - ** While the veteran developer's experience is valuable, it's still anecdotal - location: veteran developer quotes section - Note: 🔍 Fallacy: Appeal to Authority #fallacy/appeal-to-authority - ** Assumes AI usage directly causes skill atrophy without controlling for other factors - improvement: longitudinal studies needed - Note: 🔍 Fallacy: Post Hoc #fallacy/post-hoc