AI Killed the Coding Interviews. Here’s What Meta Built Instead  🔗 **Source:** [AI Killed the Coding Interviews. Here’s What Meta Built Instead](https://www.youtube.com/watch?v=3u97MCRA7mY) > [!note]+ Document Notes > tags: [hiring, technical-interviews, ai-native-process, calibration, psychological-safety, engineering-leadership, meta] > author: Denitz (spelling per auto-transcript) — software engineering manager at Meta > source: https://www.youtube.com/watch?v=3u97MCRA7mY > reference: "AI killed the coding interviews. Here's what Meta built instead" — conference talk > date created: 2026-08-26 19:50 > > > [!info] Generated analysis — not the speaker's words. Sections below are machine-written > > from the captured transcript. Only text inside `> [!quote]` blocks is verbatim from the speaker. > > Do not quote or attribute any other line to the named speaker. > > # 🏷️ Title: AI killed the coding interviews. Here's what Meta built instead > > **Attribution.** Reader's `author` field reads `youtube.com`. The speaker introduces himself: *"hi, I'm Denitz. I'm a software engineer manager at Meta."* **The name is an auto-caption rendering and the spelling should not be trusted** — it is plausibly Denitz, Denitsa, Dennis, or similar. **Verify before citing.** He states he had been at Meta three years as of the story two years prior, and refers to *"my 20 years of experience."* > > **Register.** A conference talk delivered as a personal-failure-to-rebuild narrative, describing a **real, deployed, large-scale process change** at a named company. **This is the most operationally concrete source in the batch** — 9,000 interviews, 5,000 interviewers trained, a stated launch date, and an honest account of what it cost to run. > > **Why it matters here.** It **directly contradicts Elliott Minns's hiring criterion from earlier in this batch** — and the contradiction resolves into something more useful than either position alone. It also supplies the **resourcing answer** to Cunningham's transmission problem that nobody else in the batch offered: Meta paid for the learning time explicitly. > <!-- deep-read:sections start --> ## 🧭 Table of Contents 1. The contradiction with Minns, and how it resolves 2. Summary 3. Main Content and Notes 4. References 5. Paraphrase 6. Newsletter 7. Action Items 8. Hypotheses 9. Methodology 10. Questions 11. Future Research 12. Implications 13. Learning Styles Assessment 14. Accessibility Assessment 15. DEIB Assessment 16. Logical Validity 17. DARVO Analysis 18. Conclusion 19. Refactor Appendix 20. Template Library ## ⚖️ The contradiction with Minns, and how it resolves **Two practitioners in this batch, opposite hiring philosophies.** **Elliott Minns** rejected every script-editor candidate: *"I want to hire somebody who can do the work without AI, because I know that even if you do use AI in your role, you are more capable without having to rely on a tool that I could just then go ahead and ask."* **Meta built the opposite.** Each interview question ships with a codebase that *"cannot be explored without AI"* — and the speaker is explicit that his first instinct, in-person interviews to prevent tool use, *"was just a plaster. It was me trying to run into my comfort zone again."* **But they are testing the same thing by opposite means, and that's the finding.** Meta's assessment criteria, in his own words, *"didn't really change. The only thing that changed is the way we look at it."* They still watch code quality, maintainability, and testing — and the criterion they added is judgment about the tool's output: > *"we don't want engineers who assume the code is working. And we also don't want engineers who assume the AI assistant gave them a working code or a code that does what they asked them to… How do they validate the AI output? Do they know to recognize AI slops and get rid of that?"* **Minns's candidates failed because they added nothing beyond the tool** — they ran his outline through a model and returned it. **Meta's candidates fail if they add nothing beyond the tool** — if they accept output without validating it. **Same target: does this person supply judgment the model doesn't?** Minns tests it by **removing** the tool. Meta tests it by **making the tool insufficient alone.** **And Meta's method is the better one**, for a reason Minns's own framing exposes: he conceded that using AI to speed up work you could do unaided is *"absolutely fine"*, which means his criterion tests a **proxy** for judgment rather than judgment itself. Meta tests the thing directly. **Minns's approach also carries the equity risk I flagged in that analysis — screening for when someone learned rather than what they can do. Meta's does not.** ## 📝 Summary **The opening story, and it's a good one.** Two years ago, three years into Meta, his team was doubling. He was so pleased about a new hire that he attended new-hire orientation with them *"to see how his onboarding experience looks like."* On stage was **a senior manager he admired** — *"I've used almost all the tools his team built on a daily basis. tools which before joining Meta I didn't even believe existed. I look up to him."* The talk was on AI adoption, since his org built Meta's internal AI integrations. **Then it went wrong.** He was *"explaining what LLM means, large language model"*, teaching *"how to write a prompt, saying things like, 'You need to give a role to your agent or context is important. Let me explain to you what tokens are.'"* In a room of new hires in their twenties. > *"It felt like he was reading a well-memorized script which any of the new hires around the room could have written."* **And the moment the talk is built on:** > *"One of the new hire raises a hand and asks a question about vibe coding. The speaker freezes, mumbles something and moves on. He has no idea what vibe coding is and he is afraid to ask."* **His own position, stated plainly:** *"I am not an early adopter. I'm a traditionalist. I work with a notebook and pen and I refuse to replace my phone unless I really, really have to."* And then: *"when I saw the manager's embarrassment, I couldn't help thinking what it would have been like for me to be up there. I got scared."* > *"This is not just a generational gap. This is an era where all of us leaders are required to adapt to a technology that didn't even exist for the majority of our career."* **Three things he found broken.** The **front door** — coding interviews defeated by cheating tools. **Calibration and growth** — *"The bar wasn't clear. Engineers career expectations were a mystery."* And **leadership** — *"How can I teach a team to use a tool I don't know how to use myself?"* > *"My 20 years of experience were worthless. even worse. Those years hold me back."* **Fixing the front door.** The cheating tool — **interview coder** — produced *"a non-detectable overlay of the coded solution and the thoughts of the candidate on top of the screen. Even with screen sharing, no one would have been able to see it but the candidate… My grandma could have passed this interview."* **His first instinct was in-person interviews**, and he rejects it himself: *"that was just a plaster. It was me trying to run into my comfort zone again."* Because: > *"those engineers who has a problem in front of their face and then they find the right tool to solve that problem and learn how to use that tool in a way that will help them to solve their problem. Those are the engineers we need."* **The process.** A cross-functional group — *"senior engineers, leaders, juniors, recruiters, analysts"* — questioning how coding interviews were originally designed, *"how we pulled out the pure skills that needed for an engineer to code well"*, then what coding would look like in future and which skills to measure now. **This was early 2025, and the resistance was real.** *"I am here over a decade. I know how things works. This AI thing it will pass."* And a list of complaints he does not dismiss: AI *"waste our time… reduces the code quality, keeps the engineers average and causes outages in production."* > *"They weren't fully wrong."* **The design move that resolves the whole debate.** Rather than argue about whether AI use should be mandatory, they made the question moot: > *"each question full code base it cannot be explored without AI. That's how we bypass this question of are we making it mandatory or not. This is the bar. If you can reach it without AI, great."* **And the surprise:** *"the assessment criteria didn't really change. The only thing that changed is the way we look at it."* - **Problem exploration** — formerly back-and-forth with the interviewer over a deliberately vague problem. Now *"the back and forth is with AI assistant… the interviewer at that point is just observing this communication."* - **Coding** — the volume bar *"moved higher"*, but still requiring quality, maintainability, validatability, and *"we want to avoid an overengineering."* - **Validation** — the crux, quoted above. - **Communication** — unchanged in substance, moved to the front for two stated reasons: it is now *"measure[d] twice"* (with the interviewer and with the assistant), and *"People have stopped writing themselves… if your communication skills were originally poor, now it will be even worse."* He adds, unprompted: *"AI helped me grammar proof this whole slide."* **Scale and cost, stated honestly.** Launched **September 2025**; he believes Meta was first and is still the only big tech company interviewing this way. **9,000 interviews, 5,000 interviewers trained.** As of **last April**, Meta *"moved to fully AI native hiring process and abandoned completely the classic SW coding interviews."* **What it costs to run** is the part most talks would omit: - Questions must be heavy enough that an hour with an assistant is still work — *"and then a new model of cloud is out and suddenly solves this question in 10 minutes."* - So they needed *"a big team of engineers who kept chasing the state-of-the-art models"*, and **every model upgrade required upgrading every question.** - A large question bank, because *"we are the only company interviewing like that questions will leak."* - Training 5,000 interviewers, including *"a full system of mock interviews."* **Calibration.** He describes how he originally built a sense of the bar — years of comparing engineers across levels. Then: *"everyone keeps telling me that with AI the bar should be higher. So I thought, great, I will just raise the bar. Now juniors can be mid-level, mid-level will be seniors. That's it, right?"* > *"I thought I know what I'm doing, but that was just my comfort zone trying to push in again."* **The actual answer was continuous recalibration** — peer-manager sessions *"sometimes even on a weekly basis"*, because *"the real bar kept moving up."* **And the ending, which is the best part.** Returning to the frozen manager: > *"He froze because for his entire career, his value was tied to his knowledge. But this approach today can make your seniority a liability. You are anchored to traditional ways of doing things while the juniors flying past you. In the age of AI, your value is your ability to ask questions."* The prescriptions: *"Give the microphone to the power users. Let the junior engineers who adapted faster to lead and teach the rest of the team to teach you."* AI learning hours, hackathons, team demos. *"Talk about the failures that you created."* **And the resourcing statement nobody else in this batch made:** > *"Invest time and resources. Learning doesn't come free. It cost tokens. It costs time. It cost deadline which have to be pushed forward but it's worth it."* > *"We don't need leaders who can vibe code better than the 20 years old. We need leaders who can create the psychological safety for the teams to learn."* > *"You are in a senior position not because you know everything. You are in a senior position because you know what to ask."* ## 📖 Main Content ### 🗒️ Notes **The setup.** - Two years ago, three years into Meta; his team doubling in size. *"I had been waiting for the opportunity to hire like that for so long."* - He attended new-hire orientation with his first new joiner *"to see how his onboarding experience looks like. I just wanted to make sure that he will have the best onboarding ever."* - The room: *"young 20s something years old, many of them landing their first tech job."* - On stage: a senior manager whose org built Meta's internal AI tooling. *"I look up to him. I quietly hope my own career will grow to a similar state as his."* **The failure.** - The talk taught *"what LLM means, large language model"*, then prompt basics — *"You need to give a role to your agent or context is important. Let me explain to you what tokens are."* - *"I'm thinking to myself, I wonder how many in here already knew that… it would be weird if someone here who just hired into meta didn't already know that."* - *"It felt like he was reading a well-memorized script which any of the new hires around the room could have written."* - **The Q&A:** *"One of the new hire raises a hand and asks a question about vibe coding. The speaker freezes, mumbles something and moves on. He has no idea what vibe coding is and he is afraid to ask."* **His reaction.** - Self-description: *"I am not an early adopter. I'm a traditionalist. I work with a notebook and pen and I refuse to replace my phone unless I really, really have to."* - *"when I saw the manager's embarrassment, I couldn't help thinking what it would have been like for me to be up there. I got scared. That is when I realized I have to do something."* - *"This is not just a generational gap. This is an era where all of us leaders are required to adapt to a technology that didn't even exist for the majority of our career."* - *"Walking out of that orientation room, I started debugging it. Was that moment on stage just one manager's failure or was it our whole system breaking?"* **Three breakages.** 1. **The front door.** *"Traditional coding interviews were failing because candidates were using tools like interview coder to cheat their way into the company."* 2. **Calibration and growth.** *"The bar wasn't clear. Engineers career expectations were a mystery. And I'm not even talking about career goals."* 3. **Leadership.** *"I had no idea how to lead. How can I teach a team to use a tool I don't know how to use myself? Or how to learn what exists? How to predict what will be in the future?"* - *"My 20 years of experience were worthless. even worse. Those years hold me back."* - **The choice:** *"I could panic. Try to bend AI and claim back to my comfort zone, hoping to find a new workplace which will do the same. Or I could rewrite the playbook."* **Interview coder.** - *"the AI have been able to create a non-detectable overlay of the coded solution and the thoughts of the candidate on top of the screen. Even with screen sharing, no one would have been able to see it but the candidate."* - *"All they needed to do is just type it in. I mean my grandma could have passed this interview."* **The rejected first instinct.** - *"My first instinct was an in-person interviews. If they have a tool, let's put them in the office where they cannot use the tool and that's it. Problem solved, right? So no, that was just a plaster. It was me trying to run into my comfort zone again."* - **The reframe:** *"those engineers who has a problem in front of their face and then they find the right tool to solve that problem and learn how to use that tool in a way that will help them to solve their problem. Those are the engineers we need."* - *"less the cheating part, but we definitely want to see those engineers who know how to find the different tools in this sea of AI tools that are attacking us in a way and learn how to use them. Those are the engineers who will thrive in an AI native environment."* **The working group.** - Composition: *"senior engineers, leaders, juniors, recruiters, analysts."* - Method: questioned *"How coding used to look and how it looks now"*; examined *"the original coding interviews and how that was developed. How we pulled out the pure skills that needed for an engineer to code well and how we going to measure that in an interview"*; then *"try to assess how coding will look like in the future"* and *"which skills we will need now."* - **Timing:** *"remember it was early 2025. So obvious principle that we see now, back then that was a huge struggle."* **The resistance, reported fairly.** - *"Senior engineer pushed back. I heard sentences like I am here over a decade. I know how things works. This AI thing it will pass."* - *"Engineers saw AI as a problem that waste our time, that reduces the code quality, keeps the engineers average and causes outages in production."* - **"They weren't fully wrong."** - **The junior pattern:** *"Junior engineer just come into the company. Oh, I have this great tool that can write the code for me. They don't even bother to write that code, try to land it, and then some of the senior engineers need to catch it a moment before it hit production."* - The task became balancing: *"make sure that the engineers will know how to validate the code, but also not ignore the tool completely."* **The design.** - *"we developed a full coding environment. Engineers were able to see all the files in the coding interview and they had an AI assistant which they can leverage in order to learn about the code base in front of them. Sort of like simulate how real world coding looks like."* - **The resolving move:** *"each question full code base it cannot be explored without AI. That's how we bypass this question of are we making it mandatory or not. This is the bar. If you can reach it without AI, great."* - Honest aside: *"That wasn't as easy as it seems. We saw a lot of back and forth with it."* **Assessment criteria — and the surprise.** - *"what really surprised me is that the assessment criteria didn't really change. The only thing that changed is the way we look at it."* - **Problem exploration:** formerly *"the interviewer used to present a vague problem and then the candidates were asking question and the vague problem became clearer and clearer in time."* Now: *"the back and forth is with AI assistant. The question the candidate asks the AI assistant, their prompting skills, the interviewer at that point is just observing this communication."* - **Coding:** formerly *"how much you can write on a whiteboard."* Now *"the bar of how much code you can produce and we expect you to produce moved higher. But we still expect high quality code… that solves the problem that is easy to maintain and easy to validate. We want to avoid an overengineering."* - **Validation — the crux:** *"we don't want engineers who assume the code is working. And we also don't want engineers who assume the AI assistant gave them a working code or a code that does what they asked them to."* What they watch: *"the questions they ask back and forth with the AI. How do they validate the AI output? Do they know to recognize AI slops and get rid of that? And we also want to see their traditional quality levels. Which tests do they run? Which follow-up questions do they have?"* - **Communication:** *"pretty much stay the same. The only thing that changed is that we are placing it in the forefront of what we evaluate."* Two reasons: 1. *"communication now measure twice. It is the communication of the candidate with the interviewer how it used to be. But this is also the communication the candidate has with the AI assistant. So we have two instances of it, the verbal communication but also the written one."* 2. *"People have stopped writing themselves. AI helps all of us with grammar, scripting, almost everything we do around communication. **AI helped me grammar proof this whole slide.** So if your communication skills were originally poor, now it will be even worse."* **Scale and operational cost.** - **Launched September 2025.** *"Meta was the first company to launch AI interviews and correct me if I'm wrong but I think that's still from the big tech company Meta is still the only one interviewing like that today."* - **9,000 interviews conducted, 5,000 interviewers trained.** - **The model-chasing problem:** *"we had a lot of tradeoff between state-of-the-art models and the process of writing questions to produce a real codebase that will be that heavy… with an hour of an interview and an AI assistant we still will be able to make progress. That takes a lot of time. And then a new model of cloud is out and suddenly solves this question in 10 minutes."* - *"we needed to create a big team of engineers who kept chasing the state-of-the-art models. And every time we decided to upgrade the model we allow the engineers to interview with, we needed to upgrade all the questions."* - **Leakage:** *"we are the only company interviewing like that, questions will leak. So we cannot only have three or four questions."* - **Interviewer training:** *"5,000 interviewers is a lot and to make sure that all of them has the right mindset, the right bar, that they even understand this is not the regular interview you used to do… We had to create a full system of mock interviews versus regular interviews."* - **Completion:** *"last April, Meta moved to fully AI native hiring process and abandoned completely the classic SW coding interviews."* - *"This is one of the biggest win of Meta in the past two years and to be honest, this is the biggest win of my career by now."* **Calibration.** - How he built his original sense of the bar: *"it was my experience. It was different engineers in different levels and different skills and me exploring the impact that they were able to achieve. Me comparing engineers in different levels and learn what to expect from an IC4 versus IC6, 7 and so on. As I saw more and more engineers, I developed my own sense of expectations, my own bar."* - **The failed shortcut:** *"everyone keeps telling me that with AI the bar should be higher. So I thought to myself, great, I will just raise the bar. Now juniors can be mid-level, mid-level will be seniors. That's it, right? I thought I know what I'm doing, **but that was just my comfort zone trying to push in again.**"* - **The actual answer:** *"The truth is that the real bar kept moving up. We saw new tools, new capabilities. The only way to catch up was to keep calibrate. We had countless sessions with peer managers and I discussed about what their engineers were able to achieve and it was sometimes even on a weekly basis. We needed to add more and more dots to this graph and we needed to do that fast and continuously through time."* - *"The goals became current goals and I credited the ambitious engineers who continuously kept hitting the goals month after month."* **Leadership — the payoff.** - *"Remember that senior manager from orientation? He froze because for his entire career, his value was tied to his knowledge. But this approach today can make your seniority a liability. You are anchored to traditional ways of doing things while the juniors flying past you."* - *"In the age of AI, your value is your ability to ask questions."* - *"Give the microphone to the power users. Let the junior engineers who adapted faster to lead and teach the rest of the team to teach you. Make this a team effort. Run AI learning hours, AI hackathons, team demos."* - *"Create a safe environment. Talk about the failures that you created. Make sure your team feels comfortable to learn."* - **The resourcing statement:** *"Invest time and resources. Learning doesn't come free. It cost tokens. It costs time. It cost deadline which have to be pushed forward but it's worth it."* - *"We don't need leaders who can vibe code better than the 20 years old. We need leaders who can create the psychological safety for the teams to learn."* - *"when you are that senior manager and a kid asks you about something that you don't understand, don't freeze. You just have to be willing to ask them what have they meant."* - *"The future isn't about knowledge. You are in a senior position not because you know everything. You are in a senior position because you know what to ask."* - *"Those who will thrive in the AI era are those who know how to adapt. I have no idea how the future going to look like, but from what I've seen, I'm excited for it."* ## 📚 References - **Interview coder** — the cheating tool that forced the redesign. Named, described mechanically, not linked. - **Meta's AI-native interview process** — launched **September 2025**, fully adopted **last April**. **Externally checkable**, and the most verifiable claim in the talk. - **Cross-batch pairings, and this source connects to more of them than any other:** - **Elliott Minns** — **direct contradiction on hiring**, resolving into a shared target. - **The LeadDev manager** — the junior-ships-unvalidated-output pattern, near-verbatim, independently observed at Meta. - **Zoe Cunningham** — *"be more useful than ChatGPT"* meets *"your value is your ability to ask questions."* **Both demand-side reframes.** - **The Monzo staff engineer** — the frozen manager is the **exact inverse** of Hanselman's ask-the-acronym technique. - **DX** — the senior complaints (reduced quality, outages) are DX's measured J-curve dip, reported from inside. - **What's missing:** any outcome data on the new process — no hire quality, no retention, no false-positive rate, and no comparison against the old interviews. ## 🗣️ Paraphrase **On the design move that deserves to be copied.** The debate everyone has — *should candidates be allowed to use AI in interviews?* — was dissolved rather than answered: > *"each question full code base it cannot be explored without AI. That's how we bypass this question of are we making it mandatory or not. This is the bar. If you can reach it without AI, great."* **Scale the problem past what's feasible unaided, and the permission question evaporates.** No policing, no detection arms race, no honour system. **The task itself enforces the constraint** — which is the seventh instance in this batch of the enforcement-boundaries pattern, and one of the cleanest: **a constraint built into the work needs no rule.** **On the finding that surprised him and should surprise the reader less than it did.** *"the assessment criteria didn't really change. The only thing that changed is the way we look at it."* Problem exploration, code quality, testing discipline, and communication were always what mattered. **What changed is the medium in which they're observed** — prompting instead of clarifying questions to a human, and validation of model output instead of validation of your own. **That's a genuinely reassuring result** for anyone worried these tools invalidate accumulated hiring wisdom. **They mostly don't. They relocate where the same evidence shows up.** **On the contradiction with Minns, which is the batch's most interesting disagreement.** Minns rejected candidates who ran his work through a model. Meta requires candidates to run the work through a model. **Both are testing whether the person adds judgment the tool doesn't** — Minns by removing the tool, Meta by making the tool insufficient. **Meta's method is better, and Minns's own words show why.** He allowed that using AI to speed up work you could do unaided is *"absolutely fine"* — which concedes that unaided competence is a **proxy** for what he actually wants. Meta measures the thing itself: *"How do they validate the AI output? Do they know to recognize AI slops?"* **And Meta's version avoids the equity problem I flagged in that analysis** — testing unaided competence screens partly for *when* someone learned; testing validation doesn't. **On the opening story, which is doing more work than it appears.** A senior manager, expert enough that this speaker uses his team's tools daily, stands in front of new hires **teaching them what a token is** — and then freezes at a question about vibe coding because *"he is afraid to ask."* **Two failures stacked.** He mistook his org chart position for domain currency. And when currency failed, **the thing that stopped him was embarrassment.** Which is exactly what the Monzo staff engineer's Hanselman technique is designed to prevent — asking what the acronym means when you already know, so that others can ask when they don't. **One source describes the intervention; this one opens with the failure it prevents.** **On the resourcing line, which nobody else in this batch supplied.** Cunningham diagnosed the transmission failure as structural: *"there's just not the time in system to help train other people."* **Everyone else in the batch prescribed attitudes.** Meta paid: > *"Learning doesn't come free. It cost tokens. It costs time. It cost deadline which have to be pushed forward but it's worth it."* **That is the structural remedy for a structural problem** — and naming pushed deadlines specifically, rather than gesturing at "investing in people," is what makes it credible. **On the inversion at the end.** *"We don't need leaders who can vibe code better than the 20 years old. We need leaders who can create the psychological safety for the teams to learn."* Plus: *"Give the microphone to the power users. Let the junior engineers who adapted faster to lead and teach."* **This is Cunningham's demand-side reframe arriving independently, and pushed further.** She said be worth asking. He says **your value is the asking** — and the corollary is that the people with the knowledge may be the most junior people in the room, and your job is to make it safe for them to hold the floor. **What to hold loosely.** **No outcome data whatsoever.** 9,000 interviews and 5,000 trained interviewers is impressive throughput and says nothing about whether the hires are better. **No false-positive rate, no retention comparison, no quality measure against the old process.** He calls it *"the biggest win of my career"* on the strength of it shipping. **And the costs suggest this may not generalise.** A dedicated team chasing frontier models, every question rewritten on every model upgrade, a large question bank against leakage, and 5,000 people trained. **This is a Meta-scale solution to a Meta-scale problem**, and a company without that budget cannot copy it — though **the design principle can be copied for free.** ## 📰 Newsletter **How Meta dissolved the "can candidates use AI in interviews?" debate.** Not by allowing it. Not by banning it. **By making the question meaningless:** > *"each question full code base it cannot be explored without AI. That's how we bypass this question of are we making it mandatory or not. This is the bar. If you can reach it without AI, great."* Scale the problem beyond what anyone can do unaided in an hour, and permission stops mattering. **No detection, no arms race, no honour system — the task enforces the constraint.** The backstory is better than the solution. A Meta engineering manager sat in new-hire orientation next to someone he'd just hired, watching **a senior manager he genuinely admired** — whose team built tools he used daily — explain to a room of twenty-somethings *"what LLM means"* and *"let me explain to you what tokens are."* > *"It felt like he was reading a well-memorized script which any of the new hires around the room could have written."* Then a new hire asked about vibe coding. *"The speaker freezes, mumbles something and moves on. He has no idea what vibe coding is and **he is afraid to ask**."* The speaker's own description of himself: *"I am not an early adopter. I'm a traditionalist. I work with a notebook and pen."* And: *"I couldn't help thinking what it would have been like for me to be up there. I got scared."* **What broke first was hiring.** A tool called *interview coder* put *"a non-detectable overlay of the coded solution… on top of the screen. Even with screen sharing, no one would have been able to see it but the candidate. My grandma could have passed this interview."* His first instinct — in-person interviews — he rejects himself: *"that was just a plaster. It was me trying to run into my comfort zone again."* **And here's the result that should reassure people:** > *"what really surprised me is that the assessment criteria didn't really change. The only thing that changed is the way we look at it."* Problem exploration still matters — but the back-and-forth is with the assistant, and the interviewer observes. Code quality still matters. And the new criterion is the one worth stealing: *"we don't want engineers who assume the AI assistant gave them a working code… **Do they know to recognize AI slops and get rid of that?**"* **This directly contradicts something else I read this week.** Elliott Minns rejected every candidate for a role because they'd secretly run his work through a model: *"I want to hire somebody who can do the work without AI."* **They're testing the same thing by opposite means** — does this person add judgment the tool doesn't? Minns removes the tool. Meta makes the tool insufficient. **Meta's version is better**, because Minns himself allows that using AI to speed up work you could do unaided is fine — which means his test is a proxy. Meta measures the thing. **The numbers:** launched September 2025, **9,000 interviews, 5,000 interviewers trained**, fully replaced classic coding interviews last April. And the costs are stated honestly: a dedicated team chasing frontier models, because *"a new model of cloud is out and suddenly solves this question in 10 minutes"* — and **every model upgrade means rewriting every question.** **Two more.** He tried to shortcut calibration — *"I will just raise the bar. Now juniors can be mid-level"* — and caught himself: *"that was just my comfort zone trying to push in again."* The real answer was recalibrating with peer managers **weekly**. And the ending, which is the best statement of the leadership problem I've seen: > *"He froze because for his entire career, his value was tied to his knowledge. But this approach today can make your seniority a liability… In the age of AI, your value is your ability to ask questions."* Plus the line nobody else this week was willing to say: **"Learning doesn't come free. It cost tokens. It costs time. It cost deadline which have to be pushed forward."** Everyone prescribes attitudes. **Meta paid.** **Caveat worth keeping:** no outcome data at all. 9,000 interviews tells you about throughput, not about whether the hires are better. And the operating cost — a team permanently chasing model releases — means most companies can copy the principle and not the process. ## ✅ Action Items **DevRel** - `#action/h1` **"Make the constraint part of the task"** is a portable design principle well beyond interviews. **Scale the problem past unaided feasibility and the permission debate disappears.** - `#action/h1` **The frozen-manager story is the best opening anecdote in this batch** for any talk about leadership and expertise obsolescence. - `#action/h2` *"Your value is your ability to ask questions"* pairs directly with Cunningham's "be worth asking" as a matched pair of demand-side reframes. **OSPO** - `#action/h1` **"Learning doesn't come free. It cost tokens. It costs time. It cost deadline which have to be pushed forward."** — **the resourcing statement**, and the structural answer to a structural problem. - `#action/h2` **The assessment criteria didn't change** — accumulated evaluation wisdom mostly survives; what changes is where the evidence appears. **Community** - `#action/h1` **"Give the microphone to the power users. Let the junior engineers who adapted faster to lead and teach."** — directly applicable to contributor onboarding, and it inverts the usual seniority assumption. - `#action/h2` *"Talk about the failures that you created"* as the mechanism for psychological safety — same practice the Monzo speaker and WP Rocket both used. **WordPress subteams** - **Community** — `#action/h1` **The most applicable subteam.** Reverse mentoring, learning hours, and the explicit statement that learning costs deadline time are all directly usable for contributor programmes. - **Core AI** — `#action/h1` The validation criteria are a ready-made rubric for reviewing AI-assisted contributions: *"How do they validate the AI output? Do they know to recognize AI slops?"* - **Training** — `#action/h1` **The orientation failure is a teaching case**: content pitched below the audience, delivered by someone whose authority was positional rather than current. **And the fix is asking, not knowing.** - **Core Program Team** — `#action/h2` Continuous recalibration with peers, weekly, beats trying to reason your way to a new bar. Applicable to release-squad expectations. - **Plugins, Themes** — `#action/h2` *"Do they know to recognize AI slops and get rid of that?"* is the right review question for directory submissions. - **Core, Meta, Documentation, Design, Accessibility, Hosting, Support, CLI, Playground, Polyglots, Openverse, Photos, TV, Mobile, Marketing, Core Performance, Tide** — **Not applicable to this source.** **The WPOCC** - `#action/h1` **Strong fit and unusually direct.** The audience includes people whose accumulated experience feels devalued, and this talk's message is that **seniority becomes a liability only when it's anchored to knowledge rather than to asking.** Delivered by a self-described traditionalist who was scared, which makes it land. **FAIR** - Not applicable to this source. **Personal / vault** - `#action/h1` **RESOLVES THE MINNS CONTRADICTION.** Two opposite hiring philosophies, same target: does this person add judgment beyond the tool? **Minns removes the tool; Meta makes it insufficient. Meta's is better** — Minns's own concession that speeding up work you could do unaided is fine shows his test is a proxy — **and Meta's avoids the cohort-screening equity risk.** - `#action/h1` **SUPPLIES THE RESOURCING ANSWER TO CUNNINGHAM.** She diagnosed the squeeze as structural; everyone else prescribed attitudes. **Meta named the price: tokens, time, pushed deadlines.** - `#action/h1` **The frozen manager is the inverse of the Hanselman technique** from the Monzo talk. **One source gives the intervention, this one gives the failure it prevents.** Write them together. - `#action/h2` **"Make the constraint part of the task"** — seventh enforcement-boundaries instance, and the cleanest: a constraint built into the work needs no rule. - `#action/h2` **The junior-ships-unvalidated-output pattern** is now independently observed at Meta and by the LeadDev manager. **Two large-company observations, same behaviour.** - `#action/h3` **Verify the speaker's name** before citing — "Denitz" is an auto-caption rendering. - `#action/h3` **Check whether Meta has published outcome data** on the new process. The talk has none. ## 🧪 Hypotheses 1. **The constraint-in-the-task hypothesis.** Scaling a task past unaided feasibility removes the need to police tool use. **The most elegant idea in the talk**, and it generalises. *Deployed at scale; no comparison against a policed alternative.* 2. **The criteria-survive hypothesis.** What to evaluate barely changes; where the evidence appears does. *Reported as a surprise, which suggests it wasn't assumed.* 3. **The validation hypothesis.** The discriminating skill is recognising and rejecting bad model output. **The same construct DX measured as staff token efficiency and the LeadDev model calls accelerator behaviour.** *Now three independent formulations.* 4. **The proxy-versus-direct hypothesis (reviewer).** Testing unaided competence is a proxy for judgment; testing validation measures it directly. *Follows from comparing this with Minns; neither source makes it.* 5. **The seniority-as-liability hypothesis.** Expertise anchored to knowledge rather than to inquiry becomes a handicap when the knowledge turns over. *Illustrated by one vivid case.* 6. **The reverse-mentoring hypothesis.** The fastest adapters are often the most junior, and leadership means making room for them to teach. *Prescribed, not measured.* 7. **The paid-learning hypothesis.** Learning requires budgeted tokens, time, and moved deadlines. **The structural remedy nobody else in the batch offered.** *Asserted; no cost figures.* 8. **The continuous-calibration hypothesis.** A moving bar can't be reasoned to, only re-observed — weekly, with peers. *Reported practice.* 9. **The generalisability counter-hypothesis (reviewer).** The process requires a dedicated model-chasing team, per-upgrade question rewrites, and 5,000 trained interviewers. **Most organisations can copy the principle, not the implementation.** *Not addressed.* ## 📊 Methodology **Design: a practitioner talk describing a deployed system.** **The most operationally verifiable source in the batch** — named company, dated launch, stated volumes — **and it contains no outcome measurement whatsoever.** | Claim | Basis | Strength | |---|---|---| | Interview coder defeats screen-shared interviews | **Described mechanically** | **Credible**; the tool is real and the mechanism is specific. | | The AI-native format launched Sept 2025 | **First-hand** | **Strong and externally checkable.** | | 9,000 interviews, 5,000 interviewers | **Stated** | **Credible**, unaudited. **Measures throughput, not quality.** | | Meta first and still only big tech doing this | **Hedged** — *"correct me if I'm wrong"* | **Appropriately uncertain.** | | Assessment criteria barely changed | **First-hand, reported as a surprise** | **Moderate-to-strong.** The surprise makes it less likely to be post-hoc. | | Model upgrades invalidate questions | **Operational experience** | **Strong and specific** — *"a new model of cloud… solves this question in 10 minutes."* | | The process is better | **Asserted** | **Unevidenced. The central gap.** No hire quality, retention, false-positive rate, or comparison. | | Continuous calibration is the only way | **Reported practice** | **Plausible**, unmeasured. | | Seniority can become a liability | One vivid anecdote | **Weak as evidence, strong as illustration.** | **Methodological strengths.** - **He reports his own wrong instincts twice** — in-person interviews as *"a plaster"*, and raising the bar arithmetically as *"my comfort zone trying to push in again."* **Reporting your own rejected reasoning is rare and it's the best feature of the talk.** - **He concedes the skeptics' case:** AI wasting time, reducing quality, causing outages — *"They weren't fully wrong."* - **Operational costs are stated**, including the ongoing ones nobody would volunteer: model-chasing, per-upgrade rewrites, leakage, and training burden. - **The group composition is named** — engineers, leaders, juniors, recruiters, analysts — which is a real methodological choice, not a gesture. - **He discloses using AI on his own slides** while making a point about communication skills degrading. - **The claim about being first is hedged.** **What is not established.** - **No outcome data of any kind.** This is the central weakness in a talk whose thesis is that the new process is better. - **No comparison to the old process** on any dimension. - **No false-positive analysis** — the whole premise was that the old interviews let unqualified candidates through; nothing shows the new ones don't. - **Self-assessed success** — *"the biggest win of my career"* rests on shipping. - **Generalisability unaddressed.** The costs described are prohibitive outside a company of Meta's size. - **The junior-ships-bad-code claim is anecdotal**, though independently corroborated elsewhere in this batch. **Bias checks.** *Employer interest* — **present**: a Meta manager presenting a Meta system he led, and calling it his career's biggest win. **Not hidden, and the costs disclosed cut against pure promotion.** *Survivorship* — the process that shipped is described; abandoned designs appear only as his own rejected instincts, which is better than nothing. *Self-reported* — everything. *False precision* — absent; 9,000 and 5,000 are round and plainly approximate. *Anecdote generalised* — the orientation story carries a large thesis. *Newer-is-better* — **partly resisted**, via *"they weren't fully wrong"* and the retained assessment criteria. ## ❓ Questions 1. **Are the hires better?** **The question the talk exists to answer and doesn't.** No quality, retention, or performance comparison. 2. **What is the false-positive rate now?** The old process's failure was letting cheaters through. Nothing shows the new one doesn't. 3. **What does it cost annually?** A dedicated model-chasing team plus per-upgrade question rewrites plus 5,000 trained interviewers is a large standing expense, never quantified. 4. **How do candidates experience it?** No candidate feedback anywhere — and this is a process 9,000 people have been through. 5. **Does it disadvantage anyone?** **Unexamined.** Candidates with less exposure to frontier tooling, or on slower connections, or unfamiliar with the specific assistant, may be systematically penalised. 6. **Which assistant, and does the choice matter?** If the interview assistant differs from what a candidate uses daily, that is a confound. 7. **What is the speaker's name?** The auto-caption rendering is unreliable. 8. **Has anyone else adopted this since?** He believes Meta is still alone. ## 🔮 Future Research - **Look for published Meta data** on the new process — quality, retention, false positives. - **Verify the speaker's name and role** before citing. - **Watch whether other large companies adopt this**, which would test whether it's viable outside Meta's scale. - **Investigate the equity question** — who does an AI-native interview disadvantage? **Nobody in the talk asks.** - **Compare with Minns deliberately** — two hiring philosophies, one target, and the comparison is more useful than either. - **Test the constraint-in-the-task principle elsewhere** — it should apply to take-homes, assessments, and coursework. ## 💡 Implications **Practice.** The transferable idea is free: **build the constraint into the task rather than into a rule.** If a task genuinely requires a tool, permission stops being a question. And the assessment insight is reassuring — **what you evaluate barely changes; where you observe it does.** **Policy.** Two items. **"Learning doesn't come free… it cost deadline which have to be pushed forward"** is the honest form of investing in capability, and it is a budget decision rather than a values statement. And **reverse mentoring as policy** — *"Give the microphone to the power users"* — inverts the assumption that seniority implies currency. **Security.** Adjacent and real. **The old interview was defeated by an undetectable overlay**, and the response was not better detection but **a task where cheating is not the winning move.** That is the correct general answer to an arms race, and it applies well beyond hiring. **Education.** The orientation failure is a complete teaching case: **content pitched below the audience, authority that was positional rather than current, and a question that couldn't be answered because asking felt worse than not knowing.** And the remedy is small — *"You just have to be willing to ask them what have they meant."* **For Courtney's practice.** Four things. **It resolves the Minns contradiction** into a shared target with a better method. **It supplies the resourcing answer to Cunningham** that the rest of the batch left as an attitude. **It's the inverse of the Hanselman technique** — the failure that intervention prevents. And **"make the constraint part of the task"** is the seventh and cleanest enforcement-boundaries instance. ## 🧠 Learning Styles Assessment - **Visual** — Slides referenced (*"AI helped me grammar proof this whole slide"*, *"more dots to this graph"*) and not described. **The calibration graph in particular is invoked and invisible.** - **Aural** — **Strong.** Well-paced narrative with a clear emotional arc: admiration → embarrassment → fear → rebuild. - **Verbal** — **Good, with non-native-English constructions throughout** (*"I didn't had"*, *"engineers was as were able"*, *"those engineers who has a problem"*). **These never obscure meaning**, and several formulations land hard regardless. - **Kinesthetic** — **Strong.** A deployed system with a stated design principle, named criteria, and concrete practices (learning hours, hackathons, demos, mock interviews). - **Logical** — **Strong.** Story → diagnosis of three breakages → fix each in turn → return to the opening image. **The frozen manager bookending the talk is well-constructed.** - **Social** — **Strong.** The whole final section is about safety, reverse mentoring, and making room for others to teach. - **Solitary** — Good. Self-contained, no prior context needed. - **Naturalistic** — Not applicable to this source. **Reading level estimate.** Roughly grade 11–12. Vocabulary is plain; the assumed context is large-company engineering (IC4/IC6/IC7 levels, calibration, orientation, interview loops). **The core argument survives without that context; the specifics don't.** ## 📊 Accessibility Assessment Assessed as a **recorded conference talk with an auto-generated transcript**. 1. **Perceivable — captions (1.2.2).** Present. **Two categories of error.** Grammatical artifacts from non-native English are faithfully transcribed and mostly harmless. **But "a new model of cloud" is clearly Claude**, and **the speaker's own name renders as "Denitz"**, which may or may not be correct. **A caption that misrenders a product name and possibly the speaker's name is a real defect**, and it makes the talk hard to find by search. 2. **Perceivable — audio description (1.2.3, 1.2.5).** **Fails in the usual way.** Slides referenced but never described — most consequentially the calibration graph (*"more and more dots to this graph"*), which is the visual anchor of that section. 3. **Perceivable — info and relationships (1.3.1).** N/A for spoken delivery. 4. **Operable — timing (2.2).** Recorded, pausable. **Pass.** 5. **Understandable — language (3.1).** **Good.** Plain vocabulary; jargon (IC levels, calibration) mostly inferable from context. 6. **Understandable — predictable (3.2).** **Strong.** Three named problems addressed in order, with the opening image returned to at the close. 7. **Robust (4.1).** Extracted cleanly. **Content-level note, and it's the substantive one.** **The talk's central prescription is an accessibility-adjacent intervention:** *"when you are that senior manager and a kid asks you about something that you don't understand, don't freeze. You just have to be willing to ask them what have they meant."* **Lowering the cost of admitting you don't know is an inclusion practice**, and it is the exact behaviour the Monzo talk's Hanselman technique is designed to make safe. **Highest-value remediation:** correct the speaker's name and "Claude" in the captions, and describe the calibration graph. ## 🌍 DEIB Assessment - **The reverse-mentoring prescription is the substantive content, and it's a genuine structural inversion.** *"Give the microphone to the power users. Let the junior engineers who adapted faster to lead and teach the rest of the team to teach you."* **This asks senior people to be taught publicly by junior people** — which redistributes both authority and visibility, and is much harder than it sounds in a levelled organisation like Meta's. - **The framing of the frozen manager is generous rather than mocking.** He is described with admiration first — *"I look up to him"* — and his failure diagnosed structurally: *"his value was tied to his knowledge."* **The speaker explicitly puts himself in the same position** (*"what it would have been like for me to be up there. I got scared"*), which converts a story that could have been contemptuous into a shared predicament. - **Psychological safety is named as the leadership deliverable**, not as a nice-to-have: *"We don't need leaders who can vibe code better than the 20 years old. We need leaders who can create the psychological safety for the teams to learn."* **And the mechanism offered is self-disclosure** — *"Talk about the failures that you created."* - **The unexamined equity gap, and it's significant.** **An AI-native interview may systematically disadvantage candidates with less prior exposure to frontier tooling** — people from organisations that block these tools, from regions with restricted access, on constrained connections, or unfamiliar with the specific assistant Meta provides. **The talk never asks who this format disadvantages.** Given that the previous format's failure was fairness (cheating), the absence of a fairness analysis on the replacement is conspicuous. - **The communication criterion has a real equity edge, both ways.** Moving communication *"to the forefront"* and measuring it twice — verbally and in writing to the assistant — **may disadvantage non-native English speakers**, and it's worth noting the speaker is himself apparently a non-native English speaker who discloses using AI to proof his slides. **His argument is that AI makes weak communication weaker, which is plausible; the countervailing point — that AI also levels written communication for non-native speakers — goes unmade.** - **Age is handled carefully.** *"the 20 years old"* and *"a kid asks you"* could read dismissively; in context they don't, because the argument is that those people are ahead. - **Inclusive language.** Neutral throughout. No individuals named. ## ✅ Logical Validity **Strong reasoning.** 1. **He reports his own rejected instincts, twice** — in-person interviews as *"a plaster"*, and arithmetic bar-raising as *"my comfort zone trying to push in again."* **Showing the reasoning you discarded is the strongest available signal of genuine inquiry.** 2. **He concedes the opposing case:** *"They weren't fully wrong"* about wasted time, quality loss, and outages. 3. **The design move dissolves the dispute rather than settling it** — a genuinely better outcome than winning the argument about mandatory tool use. 4. **The surprise is reported as a surprise.** *"what really surprised me is that the assessment criteria didn't really change"* — an expectation-violating finding is more credible than a confirmed one. 5. **Ongoing costs are disclosed**, including the ones that make the system look burdensome. 6. **The claim to primacy is hedged:** *"correct me if I'm wrong."* 7. **The talk's structure is honest about sequence** — he fixed hiring, then discovered calibration was also broken, then leadership. **Problems surfacing in order rather than being presented as a designed programme.** **Fallacies and weaknesses.** 1. **No outcome evidence** — `#fallacy/unfalsifiable-claim`. **The central weakness.** The thesis is that the process is better; the support is that it exists and is large. 2. **Success asserted by the person who built it** — `#fallacy/conflict-of-interest`. *"the biggest win of my career"*, from the programme's owner. 3. **Throughput mistaken for validation** — 9,000 interviews and 5,000 interviewers demonstrate adoption, not quality. 4. **One anecdote carrying a large thesis** — `#fallacy/hasty-generalization`. The frozen manager supports claims about seniority in general. 5. **Generalisability unaddressed** — `#fallacy/suppressed-evidence`. The stated costs make this uncopyable at most companies, and the talk recommends it without noting that. 6. **Fairness analysis absent** — the old process failed on fairness; the new one's fairness is never examined. 7. **"My 20 years of experience were worthless"** is rhetorically strong and analytically false — **and the talk's own conclusion contradicts it**, since experience is what lets you know what to ask. **Tone of voice.** Warm, self-deprecating, and structurally disciplined. **The distinguishing feature is how much of his own wrongness he volunteers** — the traditionalist self-description, the fear, the two rejected instincts, the admission that his experience felt like a handicap. **That makes the confident conclusion land better than it otherwise would**, because it has been earned rather than asserted. The non-native English constructions give it an unpolished quality that reads as authenticity rather than carelessness. **The one place confidence outruns evidence** is the assessment of the programme itself, which is stated as settled success on the strength of having shipped. ## 🛡️ DARVO Analysis **Not applicable to this source.** DARVO describes a pattern in an accused party's account of alleged wrongdoing. **This is a talk about redesigning a hiring process.** No accusation, no accused party, no contested harm. **Two adjacent observations, and both are creditable.** **First, the blame framings available here are numerous and all declined.** Candidates cheating with interview coder could have been a story about declining integrity — instead the response is *"those are the engineers we need"* minus the deception, and the fix targets the process rather than the candidates. Senior engineers resisting could have been a story about dinosaurs — instead: *"They weren't fully wrong."* Juniors shipping unvalidated code could have been a story about entitlement — instead it becomes a training and criteria problem. **And the frozen manager, the most obvious target in the talk, is introduced with admiration and diagnosed structurally.** **Second, he locates the failure in himself repeatedly.** *"My 20 years of experience were worthless"*, *"I got scared"*, *"that was just my comfort zone trying to push in again"* — twice. **Every prescription at the end is an obligation on leaders, meaning himself.** **Recorded as: not applicable; a talk with unusually many available blame framings, all declined — the eighth source in this batch to do so.** ## 🎯 Conclusion **Bottom line.** The most operationally concrete source in the batch, with a design principle worth copying for free — and no evidence that the thing it built actually works better. **The principle:** *"each question full code base it cannot be explored without AI. That's how we bypass this question of are we making it mandatory or not."* **Scale the task past unaided feasibility and the permission question dissolves.** No detection, no arms race, no honour system — **the work itself enforces the constraint.** That is the seventh enforcement-boundaries instance in this batch and the cleanest: **a constraint built into the task needs no rule.** **The reassuring result:** *"the assessment criteria didn't really change. The only thing that changed is the way we look at it."* Problem exploration, code quality, testing, communication — all still the targets. **What moved is where the evidence appears**: prompting instead of clarifying questions, and validating a model's output instead of your own. And the discriminating criterion is now explicit: *"Do they know to recognize AI slops and get rid of that?"* **The batch's most interesting disagreement, resolved.** Elliott Minns rejected every candidate who ran his work through a model: *"I want to hire somebody who can do the work without AI."* Meta requires it. **Both test whether the person adds judgment the tool doesn't — Minns by removing the tool, Meta by making it insufficient.** Meta's method is better, and **Minns's own concession shows why**: he allows that using AI to speed up work you could do unaided is *"absolutely fine"*, which means unaided competence is a **proxy**. Meta measures the thing directly. **And Meta's version avoids the cohort-screening equity risk** I flagged in that analysis. **The line nobody else in this batch would say.** Cunningham diagnosed the transmission failure as structural — *"there's just not the time in system."* Everyone else prescribed attitudes. **Meta named the price:** > *"Learning doesn't come free. It cost tokens. It costs time. It cost deadline which have to be pushed forward but it's worth it."* **A structural problem with a budget answer**, and naming pushed deadlines specifically is what makes it credible rather than aspirational. **And the ending inverts the usual assumption.** *"He froze because for his entire career, his value was tied to his knowledge… In the age of AI, your value is your ability to ask questions."* Plus: *"Give the microphone to the power users."* **This is Cunningham's demand-side reframe arriving independently and pushed further** — not just *be worth asking*, but *the asking is the job*, and the people worth listening to may be the most junior in the room. **What to hold against it.** **No outcome data whatsoever** — no hire quality, no retention, no false-positive rate, no comparison to the old process. The support for *"the biggest win of my career"* is that it shipped at scale. **Throughput is not validation.** The **fairness question is never asked**, which is conspicuous given that the old process failed *on fairness*. And the operating costs — a standing team chasing frontier models, every question rewritten on every upgrade, a leak-resistant question bank, 5,000 trained interviewers — **mean most organisations can copy the principle and not the process**, which the talk recommends without noting. **What earns it credit:** he reports his own rejected reasoning twice, concedes the skeptics were *"not fully wrong"*, discloses the ongoing costs, hedges his claim to primacy, reports the criteria-unchanged finding as a genuine surprise, and declines every available blame framing — cheating candidates, resistant seniors, careless juniors, and the frozen manager he opens with and treats generously throughout. **Next steps.** 1. **Record the Minns resolution** — same target, opposite methods, and Meta's is the better test. 2. **Add the resourcing line to the delegation note** as the structural remedy the batch was missing. 3. **Verify the speaker's name** before citing anywhere. ## 🧱 Refactor Appendix ### Literature Split Pad (H3) > [!quote] `#atom/literature/h1` > "each question full code base it cannot be explored without AI. That's how we bypass this question of are we making it mandatory or not. This is the bar. If you can reach it without AI, great." **The most elegant design move in the batch.** Dissolve the permission debate by making the constraint part of the task. **Seventh enforcement-boundaries instance and the cleanest.** > [!quote] `#atom/literature/h1` > "we don't want engineers who assume the code is working. And we also don't want engineers who assume the AI assistant gave them a working code or a code that does what they asked them to... How do they validate the AI output? Do they know to recognize AI slops and get rid of that?" **The direct test for judgment** — and the resolution to the Minns contradiction. Meta measures what Minns proxies for. > [!quote] `#atom/literature/h1` > "Invest time and resources. Learning doesn't come free. It cost tokens. It costs time. It cost deadline which have to be pushed forward but it's worth it." **The resourcing statement the rest of the batch lacked.** Cunningham diagnosed a structural squeeze; this is a structural answer. > [!quote] `#atom/literature/h1` > "He froze because for his entire career, his value was tied to his knowledge. But this approach today can make your seniority a liability... In the age of AI, your value is your ability to ask questions." **Cunningham's demand-side reframe, arrived at independently and pushed further.** Not just *be worth asking* — the asking is the job. > [!quote] `#atom/literature/h2` > "what really surprised me is that the assessment criteria didn't really change. The only thing that changed is the way we look at it." **Reassuring, and reported as a surprise** — which makes it more credible. Evaluation wisdom mostly survives; the medium of observation changes. > [!quote] `#atom/literature/h2` > "One of the new hire raises a hand and asks a question about vibe coding. The speaker freezes, mumbles something and moves on. He has no idea what vibe coding is and he is afraid to ask." **The exact inverse of the Hanselman technique** from the Monzo talk. One source gives the intervention; this gives the failure it prevents. > [!quote] `#atom/literature/h2` > "My first instinct was an in-person interviews... So no, that was just a plaster. It was me trying to run into my comfort zone again." **Reporting your own rejected reasoning.** He does this twice, and it's the best methodological feature of the talk. > [!quote] `#atom/literature/h2` > "We don't need leaders who can vibe code better than the 20 years old. We need leaders who can create the psychological safety for the teams to learn." The leadership deliverable, named. **And the mechanism offered is self-disclosure: "Talk about the failures that you created."** > [!quote] `#atom/literature/h3` > "Engineers saw AI as a problem that waste our time, that reduces the code quality, keeps the engineers average and causes outages in production. They weren't fully wrong." **Conceding the skeptics' case** — and note these complaints are DX's measured J-curve dip, reported from inside. > [!quote] `#atom/literature/h3` > "everyone keeps telling me that with AI the bar should be higher. So I thought, great, I will just raise the bar... but that was just my comfort zone trying to push in again." **A moving bar can't be reasoned to, only re-observed.** Weekly peer calibration was the actual answer. > [!quote] `#atom/literature/h3` > "Junior engineer just come into the company. Oh, I have this great tool that can write the code for me. They don't even bother to write that code, try to land it, and then some of the senior engineers need to catch it a moment before it hit production." **Near-verbatim the LeadDev manager's observation**, independently made at Meta. Two large-company sightings of the same pattern. ### Atomic Split Pad (H4) - `#atom/zettel/z1` **Build the constraint into the task, not into a rule** — If a task genuinely requires a tool, permission stops being a question and enforcement stops being necessary. Policing a rule invites an arms race; scaling the work past unaided feasibility ends one. - `#atom/zettel/z1` **Test the thing, not its proxy** — Removing a tool tests whether someone can work without it, which is a stand-in for whether they add judgment. Requiring the tool and watching how they check its output tests the judgment directly. - `#atom/zettel/z1` **Evaluation criteria survive; the medium of observation changes** — What made someone good is largely unchanged. Where the evidence appears moves, and mistaking a relocated signal for a lost one is how accumulated wisdom gets discarded prematurely. - `#atom/zettel/z1` **Learning has a price and it is paid in deadlines** — Capability investment is not a values statement; it is budgeted tokens, budgeted hours, and shipped dates that move. A structural squeeze cannot be answered with an attitude. - `#atom/zettel/z1` **Seniority anchored to knowledge becomes a liability when knowledge turns over** — Positional authority does not confer currency. What survives turnover is knowing what to ask, which is why the senior move is inquiry rather than answer. - `#atom/zettel/z1` **The fastest adapter may be the most junior person present** — Making room for them to teach redistributes authority and visibility, and it is harder than it sounds precisely because it inverts the hierarchy it operates inside. - `#atom/zettel/z1` **A moving bar can only be re-observed, not reasoned to** — Adjusting expectations arithmetically is a way of avoiding recalibration. The bar is found by repeatedly comparing real work, frequently, with peers. - `#atom/zettel/z1` **Report the reasoning you rejected** — Showing the instinct you had and discarded is the strongest available evidence that inquiry actually happened, and it is nearly absent from the genre. ### Fleeting Queue - `#atom/fleeting/f1` **RESOLVES THE MINNS CONTRADICTION.** Same target — does this person add judgment beyond the tool — opposite methods. **Meta's is better and avoids the equity risk.** Write it up. - `#atom/fleeting/f1` **THE RESOURCING ANSWER TO CUNNINGHAM.** *"It cost deadline which have to be pushed forward."* Structural problem, structural remedy. **The batch was missing this.** - `#atom/fleeting/f1` **The frozen manager is the inverse of the Hanselman technique.** Monzo gives the intervention; Meta gives the failure. **Write them as one note.** - `#atom/fleeting/f1` **"Make the constraint part of the task"** — seventh enforcement-boundaries instance, and the cleanest. - `#atom/fleeting/f1` **Junior-ships-unvalidated-output now observed at Meta and by the LeadDev manager** — two independent large-company sightings. - `#atom/fleeting/f1` **The senior complaints are DX's J-curve dip reported from inside** — wasted time, reduced quality, outages. *"They weren't fully wrong."* - `#atom/fleeting/f1` **Verify the speaker's name** — "Denitz" is an auto-caption rendering. **And "a new model of cloud" is Claude.** - `#atom/fleeting/f1` **No outcome data at all** — carry this caveat wherever the process is cited. - `#atom/fleeting/f1` **The fairness question is never asked**, which is conspicuous given the old process failed on fairness. **Who does an AI-native interview disadvantage?** - `#atom/fleeting/f1` **Eighth source in this batch to decline every available blame framing.** Firmly a pattern now. <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/AI Killed the Coding Interviews. Here’s What Meta Built Instead.md|AI Killed the Coding Interviews. Here’s What Meta Built Instead by Meta engineering manager (name per auto-transcript: Denitz - verify)]] ## Highlights - coding looks like. Now each question full code base it cannot be explored without AI. That's how we bypass this question of are we making it mandatory or not. This is the bar. If you can reach it without AI, great. That wasn't as easy as it seems. We had we saw a lot of um back and forth with it. I will talk about it a little bit later. Um so we we needed then to define ([View Highlight](https://read.readwise.io/read/01m0z8mm3vvhm58vhmq2gj9h2b)) - Note: THE MOST ELEGANT DESIGN MOVE IN THE BATCH. They didn't allow AI or ban it — they made the question meaningless by scaling the task past unaided feasibility. No detection, no arms race, no honour system: THE WORK ITSELF ENFORCES THE CONSTRAINT. Seventh enforcement-boundaries instance this week and the cleanest — a constraint built into the task needs no rule. Generalises to take-homes, assessments, coursework, and any policy currently policed. Note also the security lesson: the old interview was defeated by an undetectable overlay, and the answer wasn't better detection, it was a task where cheating stops being the winning move. #atom/zettel/z1 - created. Make sure you feel your team feels comfortable to learn. Invest time and resources. Learning doesn't come free. It cost tokens. It costs time. It cost deadline which have to be pushed forward but it's worth it. We don't need leaders who can vibe code ([View Highlight](https://read.readwise.io/read/01m0z8n1syvaz31d04380td3pm)) - Note: THE RESOURCING ANSWER THE WHOLE BATCH WAS MISSING. Cunningham diagnosed the transmission failure as STRUCTURAL — "there's just not the time in system to help train other people." Lachance, the LeadDev manager, and Minns all prescribed attitudes or criteria. META NAMED THE PRICE: tokens, time, and deadlines that move. A structural problem with a budget answer, and naming pushed deadlines specifically is what makes it credible rather than aspirational. Continues: "We don't need leaders who can vibe code better than the 20 years old. We need leaders who can create the psychological safety for the teams to learn." #atom/zettel/z1