<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Brett Martin on Medium]]></title>
        <description><![CDATA[Stories by Brett Martin on Medium]]></description>
        <link>https://medium.com/@fonziai?source=rss-687975a4ecf8------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*66pp2ojCAqofpmUcB3PerA.jpeg</url>
            <title>Stories by Brett Martin on Medium</title>
            <link>https://medium.com/@fonziai?source=rss-687975a4ecf8------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 23:41:45 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@fonziai/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[How Do You Know Your Remote Team Is Actually Working?]]></title>
            <link>https://medium.com/kumospace/how-do-you-know-your-remote-team-is-actually-working-492534bd97e6?source=rss-687975a4ecf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/492534bd97e6</guid>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[management]]></category>
            <category><![CDATA[employee-monitoring]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[remote-work]]></category>
            <dc:creator><![CDATA[Brett Martin]]></dc:creator>
            <pubDate>Thu, 22 Oct 2026 15:31:01 GMT</pubDate>
            <atom:updated>2026-10-22T15:31:01.759Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*GjOOe4e76sFXALQJnRbzQg.jpeg" /></figure><p><em>California just banned AI tools that try to read employees’ emotions at work. Time and activity tracking isn’t covered. Our product can show managers how long people spend in it, so I’ve had to think about this one.</em></p><p>If you manage a remote team, you’ve asked yourself this. Probably not out loud.</p><p>I’ve asked it too, though it’s never kept me up at night. Our teams keep a regular workday with a lot of flexibility inside it, and people are spread across time zones, so there are plenty of stretches when I can’t see who’s doing what. Honestly, that doesn’t worry me much. What I care about is that nothing important falls through the cracks. On the things that matter, I’d much rather overcommunicate than have people confused, or miss a chance to work on something together because nobody knew it was happening.</p><p>In September, California signed <a href="https://leginfo.legislature.ca.gov/faces/billTextClient.xhtml?bill_id=202520260AB1883">AB 1883 into law</a>. Starting January 1, 2027, California employers generally can’t use AI-powered surveillance tools to infer an employee’s emotional state or collect neural data. That sounds like science fiction. It isn’t. There are products that analyze faces, voices, and other signals and then label people as engaged, stressed, focused, or checked out.</p><p>It’s easy to read about that and feel superior to whoever bought it. I’d rather understand why they did.</p><h3>The worry is real</h3><p>Managers buy monitoring software because they’re worried, and the worry isn’t stupid. I’ve heard it straight from a buyer. A contact-center manager told me his team stopped using a virtual office because “it became a distraction.” What I took from it: the most important thing for these people is control.</p><p>When your team sat around you, you got information without trying. Who came in early. Who looked stuck. Who had been on a tense call for an hour. Remote work turned that off, and nothing replaced it by default.</p><p>So people reach for whatever turns it back on. Activity trackers, keystroke counts, random screenshots, and now software that guesses moods. Different tools, same limitation. Each one is a proxy for the thing you actually care about, which is whether the work is good. A proxy is a poor stand-in for that.</p><h3>Two bad outcomes</h3><p>The tradeoff looks like this.</p><p>Measure too little, and you don’t know what’s happening. Problems surface late. Someone who’s struggling doesn’t get help until it’s a crisis. That’s a real failure, and it happens on remote teams all the time.</p><p>To be clear, I’m not against measuring things. Inside our own company, I push to track almost everything we build. People sometimes don’t use a feature simply because they don’t know it exists, and without data we’d never find that out. I’ve asked this about notifications nobody had turned on: “But what’s the point of launching the feature but not doing the work necessary to get it used?” If something isn’t worth tracking, it probably isn’t worth building. But measuring whether a feature gets used is a very different thing from measuring whether a person is doing good work.</p><p>Measure the wrong things, and people learn to produce the measurement instead of the work. Mouse jigglers are the punchline, but the real cost is quieter. You lose the people with the most options first, because they don’t have to work somewhere that watches them like that. And everyone else learns exactly what you think of them.</p><p>I’d rather risk the first mistake and work to fix it than lock in the second. Not knowing is fixable. Not being trusted is much harder to undo.</p><h3>The uncomfortable part</h3><p>I should be honest about something.</p><p><a href="https://www.kumospace.com/signup?ajs_event=Sign%20Up%20Clicked&amp;ajs_prop_nav_src=top_nav">Kumospace</a>, the virtual office company I co-founded, gives teams a persistent online space where they can see who’s around and talk without scheduling a call. It also shows the people who run a space how much time each person spent in it. Some of our customers use that a lot.</p><p>I’ve gone back and forth on whether that’s a feature or a problem. Here’s where I’ve landed.</p><p>Hours in a shared space tell you someone was around. That’s useful. It tells you when you can find people, and whether your team overlaps enough to work together at all. It tells you nothing about whether anyone did good work. Once hours become the scoreboard, you’ve built a fancier mouse jiggler.</p><p>So if you use any tool that counts time, ours included, use it to coordinate, not to grade. And tell your team exactly what you can see. If you’d be uncomfortable explaining a metric to the person it measures, that’s a good sign you shouldn’t be using it on them.</p><h3>What I’d look at instead</h3><p>So how do you know your team is working? Mostly through boring things.</p><p>Start with outcomes. If I can’t say what good looks like for someone this week, another source of employee data isn’t going to fix that.</p><p>Then look for evidence that work is moving: features shipped, decisions made, customers answered, problems raised before they became emergencies. Real work leaves evidence. Being around mostly leaves logs.</p><p>Availability matters too, but because it helps people work together, not because it proves anyone is working. On a good remote team, people can find each other when they need to and disappear when they need to focus. I do this myself. When I’m at my computer working, I turn sound and video off but stay available. The person controls their own availability. That’s the line between visibility and surveillance, as far as I’m concerned.</p><h3>Where I land</h3><p>AB 1883 draws one line: AI shouldn’t try to read how employees feel or collect neural data from them. It’s a narrow law. It doesn’t answer the far more ordinary question of what managers should do with everything else software can tell them.</p><p>We can show a manager that someone spent seven hours in our virtual office. That’s a fact. Whether those seven hours say anything about the quality of the work is a judgment, and I don’t want software making it for me.</p><p>Use visibility to help people find each other. Use the work to evaluate the work.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=492534bd97e6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/kumospace/how-do-you-know-your-remote-team-is-actually-working-492534bd97e6">How Do You Know Your Remote Team Is Actually Working?</a> was originally published in <a href="https://medium.com/kumospace">Kumospace</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What RSA-260 Teaches You About Getting Hired]]></title>
            <link>https://medium.com/fonzi-ai/what-rsa-260-teaches-you-about-getting-hired-f4369d8a974c?source=rss-687975a4ecf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/f4369d8a974c</guid>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[hiring]]></category>
            <category><![CDATA[software-engineering]]></category>
            <dc:creator><![CDATA[Brett Martin]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 19:26:47 GMT</pubDate>
            <atom:updated>2026-10-15T23:45:32.051Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*eqovm7hUsHvXeaaZiVC2GA.jpeg" /></figure><p><em>The agents built the GPU pipeline. The engineer sent them more than 80,000 words of direction, questions, and course corrections. I think that second part is worth paying attention to.</em></p><p>On September 3, an engineer at <a href="https://cognition.com/blog/factoring-rsa-260">Cognition named Eric Lu</a> posted a 130-digit number on X with two words underneath it: divides RSA-260.</p><p>If you don’t follow this stuff, here’s the short version. RSA-260 is one of a set of very large numbers published in 1991 as a public challenge. Factoring these challenge numbers has long been a way to benchmark progress on the math problem RSA security relies on. Nobody had publicly factored a larger one before. The previous record had stood since 2020.</p><p>A joke quickly spread that he’d somehow done it by hand. Six days later, he published a write-up explaining that he’d directed a fleet of Devin agents, which built a GPU version of an existing factoring method and then ran it on spare compute. About three weeks from first prompt to answer. Essentially no new math.</p><p>Most of the coverage focused on the record, but <em>something else stuck with me.</em></p><h3>The part I keep rereading</h3><p>Lu described his own role in one sentence. His job was to “<em>set priorities, establish benchmarks, and recognize when work was going off-track.</em>”</p><p>Then there’s the number. By Devin’s own count, Lu sent 3,328 messages across 192 sessions, about 82,700 words in total. That’s roughly the length of a novel written as instructions.</p><p>So Lu’s main contribution wasn’t writing GPU kernels. It was deciding what mattered, defining what success looked like, and noticing when the agents went somewhere unproductive.</p><p>It would be easy to read that and conclude that engineering is turning into management. I don’t think that’s right, and the companies hiring through Fonzi don’t either.</p><h3>What companies are telling us</h3><p>I run a hiring marketplace, so I read a lot of engineer profiles. I also hear what companies actually want once you get past the job description.</p><p>“Agentic” and “AI-native” now show up on almost every candidate profile. They spread so fast they stopped telling anyone anything.</p><p>The companies noticed. When they describe who they’re really looking for, a few things keep coming up.</p><p>They discount the label. Interest in AI, or having worked near an AI team, counts for much less than having personally built something with it that real users touched. One company told us that a résumé where the only AI work is a chatbot wrapper is a reason to pass.</p><p>They don’t want people who can only direct. The same companies asking for AI fluency still want engineers who can get deep enough into the implementation to debug it, challenge it, and own what ships. A profile that reads as pure coordination gets pushed down.</p><p>And at least one has started asking a much better question than “do you use AI?” Theirs is closer to: <strong>how has it changed the way you build, and do you know when not to trust it?</strong> They also pointed out something most hiring processes miss. How someone works with AI is mostly invisible on a résumé, so a missing mention should be a question for the interview, not a reason to pass.</p><p>That last point is the whole problem in miniature.</p><h3>Two ways to get this wrong</h3><p>There are two bad outcomes here, and we only get to pick which one we risk.</p><p>The first is over-crediting the label. Someone writes “agentic systems” in their headline, and we treat it as a skill when it might just mean they’ve used a coding agent. That mistake costs a company an interview slot.</p><p>The second is the one I worry about more: under-crediting someone like Lu. His résumé bullet would say “factored RSA-260.” That’s an outcome, and it hides almost everything that made him good at it. If you judged him by his commit history, you’d find the agents wrote a lot of the commits. On paper, the most important engineer in the story at times looks like the least productive one. That mistake costs a company the person it was looking for.</p><p>I’d rather risk the first one. I’m not going to die on that hill, though, because the better fix isn’t a screening rule. It’s engineers making the real work visible.</p><h3>What the real work looked like</h3><p>Read his write-up and you can see the three things he described actually happening.</p><p>He set priorities. He picked a target that was hard but tractable, and he chose to build on an existing open-source tool rather than invent new math.</p><p>He set benchmarks. Two hours after his first prompt, he added that the new siever had to handle the parameters used for RSA-250, the previous record. He later wrote that his biggest contribution was probably getting one consistent set of measurements and benchmarks in place, something the agents were not going to assemble on their own. And before committing to RSA-260, he ran the system up a ladder of progressively larger numbers to see whether it held.</p><p>He knew when things were going off-track. His write-up lists the moments he stepped in: telling an agent to skip a measurement it didn’t need, pointing out an inefficiency it kept repeating, catching it when it gave up on a direction too early. That is what answers the management question. Lu wasn’t coordinating people from a distance. He was close enough to the work to tell good output from output that only looked good.</p><p>The mistake is treating “hands-on” as a measure of how many lines you personally typed. <strong>Lu was hands-on because he knew when the work was wrong.</strong></p><p>Where did that judgment come from? He’d been factoring numbers as a hobby for about a decade. The agents gave him hands. The decade gave him enough context to know where to point them and when something looked wrong.</p><p>That lines up with something else companies on Fonzi weigh heavily: serious work people did before anyone paid them to. <em>Side projects, research, open source, competitions, a business that actually made money.</em> Not as résumé decoration, but as <strong>evidence that someone goes deep on hard problems because they want to.</strong> Lu’s hobby is a perfect example. He was also candid that his understanding of some components was closer to a hobbyist’s than an expert’s, and I trust the rest of his account more because he said so.</p><p>Sixteen days later, Stephen Weis at Anthropic announced an even larger factorization using a strikingly similar recipe: an AI agent, the same open-source factoring software, GPUs, and idle compute. Weis isn’t a newcomer either. He did his PhD at MIT under Ron Rivest, the R in RSA. Both records paired a capable agent with a person who had unusually deep context. The implementation pattern repeated quickly. The judgment still came from the humans.</p><h3>If you’re an engineer</h3><p>Write your own version of Lu’s sentence.</p><p>When you describe a project in a résumé or an interview, the outcome is only the first line. The more interesting part is how you got there. Say what you decided to build and what you decided not to. Say how you knew it worked, meaning the specific test, baseline, or benchmark. And say when you caught something going wrong and what tipped you off.</p><p>Don’t lead with the label. “Agentic” in a headline doesn’t do anything for you anymore. A specific system you built and shipped does.</p><p>Expect the question about how you work with AI, because the good companies know it doesn’t show up on paper and they’re going to ask. Have a concrete answer about what changed in how you build. Have a specific example of output you didn’t trust, and why.</p><p>And if you have something like Lu’s decade of factoring, a domain you know deeply enough to tell when output is wrong, put it in front of people. It’s becoming one of the more valuable things an engineer can have, because it’s the part an agent can’t supply.</p><h3>Where I land</h3><p>I don’t think engineering moved toward management. It moved toward judgment. The engineers I’d most want to hire can define what good looks like, measure it, notice when they’re not getting it, and stay close enough to the work to know the difference. The companies we work with keep asking for exactly that, in one form or another. Lu’s 82,700 words are one of the clearest public examples I’ve seen.</p><p>At <a href="https://talent.fonzi.ai/?utm_source=medium&amp;utm_medium=blog&amp;utm_campaign=medium-rsa260">Fonzi</a>, we’re trying to make this kind of signal easier for great companies to find. If you work this way, we built it for you.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f4369d8a974c" width="1" height="1" alt=""><hr><p><a href="https://medium.com/fonzi-ai/what-rsa-260-teaches-you-about-getting-hired-f4369d8a974c">What RSA-260 Teaches You About Getting Hired</a> was originally published in <a href="https://medium.com/fonzi-ai">Fonzi AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[We Stopped Making Engineers Wait for Monday]]></title>
            <link>https://medium.com/fonzi-ai/we-stopped-making-engineers-wait-for-monday-2ab6327eb6e9?source=rss-687975a4ecf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/2ab6327eb6e9</guid>
            <category><![CDATA[job-hunting]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[careers]]></category>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[hiring]]></category>
            <dc:creator><![CDATA[Brett Martin]]></dc:creator>
            <pubDate>Fri, 09 Oct 2026 17:01:03 GMT</pubDate>
            <atom:updated>2026-10-09T17:01:03.042Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8kcgHpnPQtFQ-LMkBorGuQ.jpeg" /></figure><p><em>There were two bad outcomes on the table. As we grew, one of them got a lot more expensive.</em></p><p>At <a href="https://talent.fonzi.ai/">Fonzi</a>, we connect vetted software engineers with AI startups and high-growth companies that want to hire them. For most of our history, we did that through something we called Match Day, a weekly cycle that set the rhythm for how companies and engineers met.</p><p>Recently, a founder we had just onboarded asked our team whether Match Day is actually a day.</p><p>Fair question. It used to be. It isn’t anymore, and I want to explain why, because it was one of the harder calls we made this year.</p><p>Here’s how it worked. Every Monday, a new group of engineers who had passed our vetting went live. Companies reviewed them and sent interview requests. Candidates had until the end of the week to respond. Then the cycle reset.</p><p>The cycle had real benefits. Everyone knew when things would happen. Companies knew when to expect new people, and candidates knew when they’d be visible. For a young marketplace, that kind of predictability is worth a lot.</p><p>The problem was what it did to candidates.</p><p>If you finished vetting on a Tuesday, you waited six days before any company could see you. If a company sent you a request on Saturday, it didn’t reach you until Monday. Because every request expired when the cycle ended, how long you had to decide depended on which day a company sent it. A Monday request gave you most of a week. A Saturday request gave you about a day.</p><p>None of that was intentional. It’s just how a weekly schedule works.</p><h3>Two bad outcomes</h3><p>So the decision looked like this.</p><p>Keep the cycle, and we keep the predictability that made it easy for everyone to plan around. Candidates keep waiting on a schedule that has nothing to do with them.</p><p>Go continuous, and candidates stop waiting. But we lose the one moment each week that did a lot of work for us. Without it, we have to make the experience worth coming back to every day, not just on Mondays.</p><p>Neither option is free. What tipped it was scale. The weekly rhythm made a lot of sense when the marketplace was smaller. As more engineers and companies joined, the waiting started to cost more than the rhythm was worth. So we went continuous.</p><h3>What we changed</h3><p>Now, engineers go live as soon as they’ve been vetted, companies can send interview requests whenever they find someone they want to meet, and every request gets the same full week to respond. The calendar no longer decides how much time a candidate gets.</p><p>And if a company sends you a request at 3 am, you hear about it at a reasonable hour where you live.</p><h3>What we didn’t expect</h3><p>Removing a weekly cycle sounds like deleting a cron job. It wasn’t.</p><p>A lot of things quietly assumed Monday existed. Our rules, our emails, our product copy, the way we talked about ourselves. Even simple questions, like what “this week” means when there’s no <a href="https://talent.fonzi.ai/#match-day">Match Day</a> to anchor it, had to be answered again.</p><p>I think this is the most underrated part of changing any system. The change itself is usually the easy part. The hard part is everything built on top of the old assumption, because nobody wrote that stuff down. Nobody expected it to change.</p><h3>What this means if you’re looking</h3><p>Every request gives you the same amount of time. You can actually think about it, talk to people, and compare it against whatever else you have going on.</p><p>I also want candidates to respond, even when the answer is no. A decline gives the company information it can act on. A timeout gives nobody anything and holds up the next person they might have contacted. I’d rather reward responsiveness than punish silence, but the goal is the same: keep the marketplace moving.</p><p>And pay attention to how fast companies move with you. I’ve watched plenty of companies lose great engineers without ever rejecting them. They just moved too slowly, and somebody else made the offer first. How a team handles your time during the process tells you a lot about how it makes decisions once you’re inside.</p><h3>Where I land</h3><p>I don’t think we’ve seen every second-order effect yet. We’ll keep measuring and keep adjusting.</p><p>But I’m confident about the principle. The person with the most to lose from waiting shouldn’t be the one doing the waiting.</p><p>So, to answer that founder’s question: Match Day isn’t a day anymore. It happens whenever a company finds someone worth meeting.</p><p>At <a href="https://talent.fonzi.ai/?utm_source=medium&amp;utm_medium=blog&amp;utm_campaign=medium-matchday">Fonzi</a>, vetted engineers receive interview requests from AI startups and high-growth companies, with the role and compensation visible upfront, and a full week to decide on each one.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2ab6327eb6e9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/fonzi-ai/we-stopped-making-engineers-wait-for-monday-2ab6327eb6e9">We Stopped Making Engineers Wait for Monday</a> was originally published in <a href="https://medium.com/fonzi-ai">Fonzi AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Before You Take the AI Job, Interview the Company]]></title>
            <link>https://medium.com/fonzi-ai/before-you-take-the-ai-job-interview-the-company-767bccb934d6?source=rss-687975a4ecf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/767bccb934d6</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-safety]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[job-interview]]></category>
            <category><![CDATA[careers]]></category>
            <dc:creator><![CDATA[Brett Martin]]></dc:creator>
            <pubDate>Mon, 05 Oct 2026 17:19:13 GMT</pubDate>
            <atom:updated>2026-10-05T22:52:23.500Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iYK_l7p7nwwCFyRZ9dge1A.jpeg" /></figure><p><em>Jacob Coxon left Anthropic before his equity vested. You don’t have to agree with his AI-risk forecast to take the career lesson seriously.</em></p><p>Jacob Coxon had the kind of résumé most AI engineers would love.</p><p>Three years doing pretraining research across OpenAI and Anthropic. A seat close to the frontier. Access to the models, people, and problems everyone else is trying to get closer to.</p><p>Then he quit.</p><p><a href="https://time.com/article/2026/09/15/ai-anthropic-researcher-quits-coxon-slowdown/">Coxon resigned from Anthropic on September 8</a> and said publicly that he believed both Anthropic and OpenAI were racing toward self-improving superintelligence without acting responsibly enough. Axios later reported that he left roughly two months before his Anthropic equity would have vested.</p><p>You can think his forecast is reasonable. You can think it’s wildly alarmist. There is no consensus on how likely catastrophic AI outcomes are or when they might occur.</p><p>But if you’re an engineer deciding where to spend the next few years, I think there’s a much less speculative lesson here:</p><blockquote><strong>Getting into the room is only half the decision. You also have to decide whether you want to be in it.</strong></blockquote><h3>The prestigious job can still be the wrong job</h3><p>Most engineers know how to compare offers.</p><p>Compensation. Equity. Scope. Technical challenge. Team quality. Company trajectory. If you’re doing frontier AI work, add compute and access to models.</p><p>Those variables matter. But Coxon’s resignation puts another one on the table: <strong>Do you actually believe in what this company is trying to accomplish, and how it intends to get there?</strong></p><p>That’s different from liking a mission statement.</p><p>Anthropic is an especially interesting case because it isn’t a company that publicly dismisses AI risk. Its Responsible Scaling Policy is explicitly designed around catastrophic-risk scenarios, and its safety roadmap covers alignment, security, safeguards, and government oversight.</p><p>Four days after Coxon resigned, Anthropic CEO Dario Amodei went further. He argued publicly that frontier AI development should be paced so safety work has time to keep up, citing accelerating AI-assisted AI development and recent alignment incidents among his reasons.</p><p>That makes this more interesting than “employee cares about safety, company doesn’t.”</p><p>The disagreement is happening <strong>inside the safety conversation itself</strong>.</p><p>How much uncertainty is acceptable? How fast is too fast? When should a company stop? What happens when slowing down means a competitor keeps going?</p><p>If you’re joining an AI company, those aren’t abstract questions anymore.</p><p>They’re part of the job.</p><h3>Candidates spend too much time preparing answers</h3><p>We train candidates to expect scrutiny.</p><p>Be ready to explain the system design. Defend the architecture. Walk through the failure. Know your complexity. Have good <a href="https://fonzi.ai/blog/unique-interview-questions-to-ask-employer">questions for the hiring manager</a>.</p><p>But most of the “good questions” candidates are told to ask are remarkably safe.</p><p>What does success look like in 90 days?</p><p>How does the team collaborate?</p><p>What’s the roadmap?</p><p>Fine.</p><p>If you’re going to help build increasingly capable AI systems, I’d ask harder questions.</p><p><em>What happens when an evaluation produces a result leadership doesn’t like?</em></p><p><em>Has the company ever delayed a launch because it wasn’t comfortable with the risk?</em></p><p><em>Who has the authority to stop one?</em></p><p><em>How are serious technical disagreements resolved?</em></p><p><em>What happens when the safe decision is also the expensive one?</em></p><p>That’s the actual trade-off you’re joining, whether anyone says it out loud in the interview or not.</p><p>You don’t need to conduct a Senate hearing during your onsite. You do need to understand the institution you’re joining. An interview is due diligence in both directions.</p><h3>Don’t outsource your career judgment to prestige</h3><p>This is where engineers can be surprisingly irrational.</p><p>We’ll spend a week benchmarking databases before adopting one, then accept a job because the company name is impressive and the offer expires Friday.</p><p>Prestige is information. It isn’t a decision. Honestly, it’s the same mistake as picking a vendor because everyone else uses it.</p><p>Your employer will take thousands of hours of your life. Its reputation becomes part of yours. You’ll contribute to decisions you won’t fully control, and if the company succeeds, you helped build whatever it becomes.</p><p>That’s a bet.</p><p>Treat it like one.</p><p>And this isn’t specific to AI safety. An engineer might leave because a company’s privacy practices change, because the product starts harming users, because leadership keeps shipping around security concerns, or simply because the incentives inside the business turn out to be different from the ones advertised during recruiting.</p><p>The point isn’t that Coxon made the “correct” decision. I’m not going to die on that hill either way.</p><blockquote>The point is that he had a line.</blockquote><blockquote>You should know yours too.</blockquote><h3>“Doomer” tells me less than you think</h3><p>AI discourse has become very good at producing labels.</p><p>Doomer. Accelerationist. Safetyist. Optimist.</p><p>They’re useful shorthand and terrible substitutes for reasoning.</p><p>If I’m evaluating an engineer, I’m much more interested in what happens underneath the label.</p><p>Can you explain why you believe something?</p><p>What evidence would change your mind?</p><p>Can you separate a plausible failure mode from a dramatic story?</p><p>Can you work with smart people who assign radically different probabilities to the same risk?</p><p>And when your beliefs finally become inconvenient, were they actually beliefs?</p><p>Coxon gave up something tangible to act on his. You don’t have to agree with where he drew the line to recognize that.</p><h3>The goal isn’t to get picked</h3><p>This is the part candidates sometimes forget.</p><p>You are not trying to get every great company to hire you.</p><p>You are trying to become good enough that great companies want you, then build enough options that <strong>you get to choose among them</strong>.</p><p>That changes the interview.</p><p>Ask about the architecture, but ask about the incentives too. Meet the brilliant researcher, but figure out who makes the final call. <a href="https://fonzi.ai/blog/equity-compensation">Compare the equity</a>, but compare the things each company expects you to help create.</p><p>The best offer on paper can still be the wrong place to spend four years.</p><p>Getting picked feels good.</p><p>Choosing well matters more.</p><p>At <a href="https://talent.fonzi.ai/?utm_source=medium&amp;utm_medium=blog&amp;utm_campaign=medium-ai-doomerism">Fonzi</a>, exceptional engineers receive interview requests from AI startups and high-growth companies that want to meet them, with the role and compensation visible upfront. We built it that way on purpose; you shouldn’t have to choose blind.</p><p>The point isn’t to get you into every interview. It’s to give you better information about which interviews are worth taking. Because the company gets to evaluate you. <strong>You should be evaluating them just as hard.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=767bccb934d6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/fonzi-ai/before-you-take-the-ai-job-interview-the-company-767bccb934d6">Before You Take the AI Job, Interview the Company</a> was originally published in <a href="https://medium.com/fonzi-ai">Fonzi AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Fly Brain Doom Demo Wasn’t the Story]]></title>
            <link>https://medium.com/fonzi-ai/the-fly-brain-doom-demo-wasnt-the-story-da9363d3a8c5?source=rss-687975a4ecf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/da9363d3a8c5</guid>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[engineering-management]]></category>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[hiring]]></category>
            <dc:creator><![CDATA[Brett Martin]]></dc:creator>
            <pubDate>Thu, 01 Oct 2026 00:25:21 GMT</pubDate>
            <atom:updated>2026-10-09T12:22:45.923Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Fly brain Doom validation: a halftone fruit fly whose second wing is built from the same dots shuffled, a control that looks just like the real result." src="https://cdn-images-1.medium.com/max/1024/1*RBQduRFKbOqD3TarV1zmUA.jpeg" /></figure><p><em>The experiment failed its own validation tests. That may be the strongest engineering signal in the entire project.</em></p><p>Founders love to say we hire for slope, not intercept. Then we open the résumé and screen for the intercept anyway.</p><p>Something happened recently that made the alternative unusually easy to point at.</p><p>In September, researchers from HHMI Janelia, the University of Cambridge, the MRC Laboratory of Molecular Biology, and <a href="https://male-cns.janelia.org/">Google Research published the MaleCNS paper</a>, describing a complete wiring map of an adult male fruit fly’s central nervous system. The dataset contains more than 166,000 neurons, and an earlier version had been public since 2025.</p><p>Once the paper put it in front of the internet, engineers started building.</p><p>Doom. Super Mario 64. Minecraft. Driving simulations. Escape rooms. Within days, people were tracking dozens of projects using fruit-fly connectomes in increasingly strange ways.</p><p>The Doom one went viral.</p><p>Alex Wormuth, an engineer at Coinbase, connected the MaleCNS graph to a Doom environment. Game pixels became modeled neural input. Neural activity became movement and shooting. Damage triggered an artificial reinforcement signal through two identified dopamine cells.</p><p>Worth being precise about what that is. The wiring came from the biological reconstruction. The neuron dynamics, the retinal interface, the reinforcement mechanism, and the game controller were all modeling choices layered on top. Nobody was running a fly.</p><p>The question was simple: <strong>could the system learn to survive?</strong></p><p>More importantly, Wormuth built tests capable of telling him no.</p><p>They did.</p><p>The project’s current status is blunt: learned survival has not been demonstrated. Its visual, conditioning, and survival validations failed.</p><p>That is the part I cannot stop thinking about.</p><h3>The experiment he didn’t have to pass</h3><p>One of the conditioning tests looked promising at first.</p><p>The model paired a visual cue with an imposed punishment signal and allowed thousands of modeled connections to change. Afterward, the response to the trained cue dropped by 93.8%.</p><p>Great result.</p><p>Then Wormuth ran the control that mattered: the same cue without the imposed punishment.</p><p>The response dropped by <a href="https://fly-brain-doom.awormuth.chatgpt.site/learning"><strong>93.8%</strong></a> <strong>again</strong>.</p><p>Whatever was causing the effect, it couldn’t be attributed to the punishment signal.</p><p>The counterbalanced test failed too. And when Wormuth moved from the conditioning assay into Doom itself, using a later version of the model, the learning-enabled system didn’t outperform the frozen control. In a pilot of two episodes per condition, the frozen version lasted longer on average. One of its runs was still alive when the 8-second limit cut it off.</p><p>He could have stopped much earlier.</p><p>Nobody was forcing a side project through peer review. A changing neural network connected to Doom already made for a great demo and an even better headline.</p><p>Instead, he ran the controls.</p><p>Then he put the failures where everyone could see them.</p><p>There’s a line in the project documentation I wish more engineers would internalize:</p><blockquote><a href="https://github.com/nftechie/doomfly/blob/main/docs/doom-live-training.md">“Passing software tests is not evidence of biological validity.”</a></blockquote><p>You could replace “biological validity” with a lot of things we’re building right now.</p><h3>What the résumé cannot show</h3><p>I’ve read a lot of résumés. I’ve also sat in plenty of debriefs where a team argues about a candidate for forty minutes without once discussing whether that person can recognize when their own work is wrong.</p><p>We screen heavily for the ability to build.</p><p>For a long time, that made sense. Implementation was expensive. Getting the thing to work was the bottleneck.</p><p>AI is changing that.</p><p>Several of the fly-connectome projects were built with heavy AI assistance. Suddenly an engineer can wander into an unfamiliar scientific domain, feed a coding agent the documentation, and get surprisingly far before dinner.</p><p>That’s incredible.</p><p>It also means getting something to run tells me less than it used to.</p><p>The interesting signal moves to what happens next.</p><p>Did you understand what you were measuring?</p><p>Did you build a control?</p><p>Did you test the assumption everybody else quietly accepted?</p><p>Did you know which result would prove you wrong?</p><p>And when you got that result, what did you do with it?</p><p>Some of the strongest projects in this little <a href="https://fonzi.ai/blog/fly-brain-doom">fly-brain</a> wave were interesting for exactly that reason. Builders added ablations, rewired-network comparisons, negative controls, provenance checks, and explicit limitations. The better repositories are noticeably less sensational than the coverage around them.</p><p>That gap matters.</p><p><strong>Engineering judgment isn’t just knowing how to build the thing. It’s knowing whether the thing you built actually did what you said it did.</strong></p><h3>This is hard to hire for</h3><p>The problem is that judgment is almost invisible in the formats we use to evaluate engineers.</p><p>A résumé mostly shows outcomes. Outcomes can hide a lot.</p><p>A shipped project looks the same in one bullet whether you validated the underlying mechanism or got lucky.</p><p>Take-homes can be even worse for this. They’re often scoped so tightly that there’s almost nothing consequential to be wrong about. You can demonstrate competence, but rarely the instinct to challenge your own model of the problem.</p><p>Interviews surface it occasionally, usually when someone volunteers a story about something that failed.</p><p>I’ve started paying more attention to those stories.</p><p>Where this quality really shows up is in artifacts.</p><p>A repository with negative results in it. A postmortem that names the decision instead of blaming the circumstances. A commit reverting your own clever idea. A benchmark where you included the dumb baseline and the dumb baseline won.</p><p>A status line saying the gates failed.</p><p>Those artifacts are cheap to publish and hard to fake convincingly because they expose the reasoning between the result and the claim.</p><h3>What I would tell an engineer</h3><p>Publish the run that killed your hypothesis.</p><p>Put it in the README, not in a file called LIMITATIONS that nobody opens.</p><p>That feels like it costs you something. On the first read, maybe it does. Your demo sounds slightly less magical.</p><p>But with anyone evaluating you carefully, it does the opposite.</p><p>When I see someone document the test that made their own claim weaker, I learn several things immediately. They know how to construct evidence. They aren’t emotionally attached to the result. And the claims they <em>didn’t</em> retract get more of my trust.</p><p>I learned some version of this the hard way myself. After my first company failed, I wrote publicly about the mistakes instead of trying to turn the whole thing into a story about market conditions.</p><p>There is a strange instinct in careers to hide the places where we were wrong.</p><p>I think that instinct is becoming increasingly expensive.</p><p>The tooling got cheap.</p><p>The demos got cheap.</p><p>A lot of what sits downstream of “can you build this?” is compressing very quickly.</p><p>Knowing whether your own work holds up did not get cheaper.</p><p>If anything, it became scarcer, because we now have dramatically more output to evaluate, and that output looks better than it used to.</p><p>So if the last thing you built contains the experiment that could have killed it, lead with that.</p><p>It may be the most differentiated thing on your résumé.</p><p>Most people will bury it.</p><p>We run <a href="https://talent.fonzi.ai/?utm_source=medium&amp;utm_medium=blog&amp;utm_campaign=medium-flybrain">Match Day</a> every week at Fonzi. If this is how you work, we’ll help you find work that matters.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=da9363d3a8c5" width="1" height="1" alt=""><hr><p><a href="https://medium.com/fonzi-ai/the-fly-brain-doom-demo-wasnt-the-story-da9363d3a8c5">The Fly Brain Doom Demo Wasn’t the Story</a> was originally published in <a href="https://medium.com/fonzi-ai">Fonzi AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Real Numbers on Working From Home Productivity]]></title>
            <link>https://medium.com/kumospace/the-real-numbers-on-working-from-home-productivity-ed4d45bf5e65?source=rss-687975a4ecf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/ed4d45bf5e65</guid>
            <category><![CDATA[business-template]]></category>
            <category><![CDATA[meeting-productivity]]></category>
            <category><![CDATA[meeting-minutes]]></category>
            <category><![CDATA[workplace-collaboration]]></category>
            <category><![CDATA[remote-work-tools]]></category>
            <dc:creator><![CDATA[Brett Martin]]></dc:creator>
            <pubDate>Thu, 06 Aug 2026 15:01:03 GMT</pubDate>
            <atom:updated>2026-08-06T15:01:03.432Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2N5SpdOoqLsBdB_LfqnGBQ.jpeg" /></figure><p><em>The same numbers get read in opposite directions depending on who’s holding them.</em></p><p>Working from home productivity is one of those topics where the same underlying data gets interpreted in opposite directions depending on who’s reading it. Employees see the research saying their output holds up fine at home. Managers see the 2022 Microsoft survey reporting that 85% of leaders struggle to feel confident their hybrid employees are actually getting things done. Both are looking at real numbers. The disagreement is about what those numbers actually prove.</p><p>Here’s what the research actually shows, and what it means for how managers should be designing work.</p><h3>What the Productivity Data Says</h3><p>The case for working from home is strongest on focused, independent tasks, though the best-known evidence is narrower than it sounds. A <a href="https://siepr.stanford.edu/publications/essay/bright-future-working-home">Stanford-run tria</a>l at a Chinese call center found home workers posted 13% better numbers. Most of that came from working more minutes, not from working better: fewer breaks, fewer sick days.</p><p>The picture changes when the work changes type. When Microsoft sent its US workforce home in early 2020, researchers tracked 61,182 employees and watched the collaboration network turn inward. Time spent working across group lines fell by about a quarter. Teams talked more to themselves and less to everyone else, and conversation shifted out of real time into inboxes and message threads.</p><p>Junior people pay the steepest price for that. Economists studying software engineers at a Fortune 500 firm found that sitting near your teammates raises the feedback you get on your code by about 18%, and nearly all of that gain lands on younger, less-tenured engineers. Their code gets better for it. Mentoring isn’t free, though. Senior engineers ship more of their own work when they aren’t sitting beside juniors. Proximity buys long-run development and bills you in short-run output.</p><p>Both findings come with limits worth naming. The Microsoft data captures a lockdown, not remote work anyone chose. The proximity study covers engineers at one company. Neither tests fully remote against hybrid. So the honest version is narrower than “hybrid wins.” Focused individual work travels home well. The work that runs on ambient contact does not: mentoring, cross-team ties, the correction that happens when someone overhears you.</p><h3>Why Managers and Employees See This Differently</h3><p>The manager-employee perception gap is documented and persistent. In Microsoft’s 2022 survey of 20,000 employees across 11 countries, 87% said they were productive at work. Only 12% of their leaders had full confidence that was true. Those two numbers can coexist because they’re measuring different things.</p><p>Employees are reporting on their own output. Managers are reporting on their visibility into it. Microsoft named this <em>‘productivity paranoia</em>.’ The gap isn’t about performance. It’s about what managers can see, and when activity is invisible, visibility gets mistaken for productivity. The employees producing excellent work from home and the managers who feel uncertain about whether they’re productive are often the same people working together.</p><p>The fix isn’t surveillance tools or <a href="https://www.kumospace.com/blog/return-to-the-office">return-to-office</a> mandates. It’s output-based measurement: goals, milestones, and deliverables rather than hours observed.</p><h3>The Commute Dividend</h3><p>Across 27 countries, people who work from home save about 72 minutes a day on commuting. They put roughly 40% of that back into their jobs, 11% into caregiving, and the rest into everything else.</p><p>Remote-only employees also log an average of 29 extra minutes of productive work per day compared to hybrid or in-office peers, and save an average of $42 daily in commuting, food, and related costs. That financial and time return is part of why the preference for remote and hybrid work remains so strong despite ongoing return-to-office pressure: the trade employees are being asked to make is real and measurable.</p><h3>What Actually Kills WFH Productivity</h3><p>The research is clear that working from home can work well, and equally clear about when it doesn’t. The common failure modes are worth naming specifically.</p><ol><li><strong>No structure on the workday.</strong> Remote workers who don’t have explicit start times, work blocks, and end signals tend to either overwork or underwork. The blurring of work and personal time is cited by 47% of remote workers as a top concern. A workday without edges tends to expand until it crowds out recovery, which accelerates burnout.</li><li><strong>Distractions without accountability.</strong> 39% of remote workers say home distractions actively hurt their productivity. A meaningful proportion admit to spending work hours on non-work activity: social media, online shopping, watching shows. The absence of social accountability that the office provides is real, and not everyone builds an equivalent structure independently.</li><li><strong>Isolation that limits creative feedback.</strong> 28% of remote workers report feeling less motivated than in an office setting, often tracing back to a lack of informal contact with colleagues rather than dissatisfaction with the work itself. Ideas that would have been stress-tested by a passing conversation or a hallway comment don’t get that friction, which can let weak thinking persist longer.</li><li><strong>Burnout from always being “on.”</strong> <a href="https://www.eaglehillconsulting.com/news/workforce-burnout-survey-2025/">61% of fully remote workers</a> experienced burnout in 2025. The same flexibility that increases productivity in the short run can erode it over time when workers never fully unplug.</li></ol><h3>What Good WFH Structure Looks Like</h3><p>The remote workers who sustain high productivity over time tend to share a few habits that aren’t particularly glamorous but show up consistently.</p><ol><li><strong>Defined time blocks for different kinds of work. </strong>Deep focus work, email and messages, and meetings occupy different windows of the day rather than competing for the same hours. Most people have a peak concentration period, usually in the morning, and scheduling the highest-stakes independent work there rather than during meetings or admin time produces better output for the same hours invested.</li><li><strong>A physical cue that the workday ends.</strong> Shutting the laptop, going for a walk, or changing out of work clothes are all proxy rituals for what a commute used to do automatically. Without some deliberate signal that the workday is over, work hours tend to stretch indefinitely, and the recovery time that makes tomorrow’s work possible gets consumed.</li><li><strong>Sync with the team’s rhythm, not just your own preferences.</strong> Remote productivity isn’t purely an individual equation. Engineers and managers who optimize entirely for their own schedules without accounting for when their team is available create coordination overhead for everyone else. Knowing when your team is most responsive and building enough overlap with those hours tends to reduce the latency that makes async work frustrating.</li><li><strong>Short </strong><a href="https://www.kumospace.com/blog/asynchronous-communication-tips"><strong>async updates</strong></a><strong> rather than long meetings. </strong>No study isolates this cleanly, so treat it as a judgment rather than a finding: teams that coordinate less synchronously tend to protect more of the unbroken time that focused work needs. Replacing status meetings with written updates is the cheapest way to test it.</li></ol><h3>The Manager’s Actual Lever</h3><p>For project and marketing managers specifically, the research suggests the manager’s highest-leverage move isn’t designing a productivity monitoring system. It’s designing a meeting structure that protects focus time.</p><p>The average employee spends <a href="https://www.notta.ai/en/blog/meeting-statistics">57% of their time in meetings, email, and chat, leaving 43% for focused work</a>. Most managers are consuming a large share of their team’s productive capacity through coordination overhead, not enabling it. Auditing which meetings can become async updates, which check-ins are actually necessary, and how much uninterrupted time each person on the team has in a given week tends to produce more output than any individual productivity hack.</p><p>Teams that want to stay connected without meeting overhead have found persistent virtual office environments useful for exactly this. <a href="https://kumospace.com/">Kumospace</a> lets team members be visibly available for quick questions and informal check-ins without scheduling anything, which handles the ambient coordination that would otherwise become another calendar block.</p><h3>The Bottom Line on the Productivity Question</h3><p>Working from home is more productive than offices for focused individual work. It’s less effective for mentoring junior people, complex real-time collaboration, and maintaining the informal social contact that sustains motivation over time. Whether hybrid splits the difference is still an open question, and the researchers studying it say so. What the data doesn’t support is the productivity case for forcing full-time office attendance.</p><p>The more productive question isn’t where people work. It’s how the work is structured, and whether the coordination costs of staying aligned have been kept low enough to leave room for actual work to happen.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ed4d45bf5e65" width="1" height="1" alt=""><hr><p><a href="https://medium.com/kumospace/the-real-numbers-on-working-from-home-productivity-ed4d45bf5e65">The Real Numbers on Working From Home Productivity</a> was originally published in <a href="https://medium.com/kumospace">Kumospace</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Quiet Bias Built Into Every Hybrid Meeting]]></title>
            <link>https://medium.com/kumospace/the-quiet-bias-built-into-every-hybrid-meeting-450597817d09?source=rss-687975a4ecf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/450597817d09</guid>
            <category><![CDATA[remote-work-tools]]></category>
            <category><![CDATA[workplace-collaboration]]></category>
            <category><![CDATA[meeting-minutes]]></category>
            <category><![CDATA[business-template]]></category>
            <category><![CDATA[meeting-productivity]]></category>
            <dc:creator><![CDATA[Brett Martin]]></dc:creator>
            <pubDate>Tue, 04 Aug 2026 15:01:02 GMT</pubDate>
            <atom:updated>2026-08-04T15:01:02.761Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Hybrid meeting with in-person team and remote participants on video call." src="https://cdn-images-1.medium.com/max/1024/1*7crH-DET2ogLOkTncBMW-Q.jpeg" /></figure><p>A hybrid meeting, where some participants are physically in a room and others join by video, has a structural problem that pure in-person or pure remote meetings don’t: it tends to produce two different experiences inside the same meeting, and one is consistently worse. The people in the room get full presence, easy side conversations, and natural turn-taking. The people on screen get a wide shot of a room, muffled audio if they’re lucky, and a much harder time getting a word in.</p><p>Fixing this requires treating hybrid meetings as their own format, not as an in-person meeting with a video call bolted on.</p><h3>Why Hybrid Meetings Default to Favoring the Room</h3><p>The core issue is that <a href="https://ronspotflexwork.com/blog/inclusive-hybrid-meetings/">hybrid meetings</a>, run without deliberate design, default to optimizing for whoever’s physically present. Side conversations happen in the room and don’t reach remote participants. Body language and informal signals that the room reads naturally are invisible on a wide-angle camera. And the social dynamics of in-person turn-taking, where someone leans forward or makes eye contact to signal they want to speak, simply don’t translate to a video feed.</p><p>This produces a predictable outcome: remote participants speak less, contribute less, and over time, disengage from meetings where they consistently get less airtime and worse audio than their in-room colleagues. If your hybrid meetings have a pattern where the same handful of in-room voices dominate discussion, that’s not a personality issue. It’s usually a structural one.</p><h3>The “Everyone Remote” Principle</h3><p>The single most effective mental model for hybrid meetings: design the meeting as if everyone were remote, even when some people happen to be in a room together.</p><p>This means defaulting to individual video tiles rather than one wide shot of the conference room, even for people sitting next to each other. It means using a shared digital document or <a href="https://www.kumospace.com/online-whiteboard">whiteboard</a> rather than a physical whiteboard that only the in-room group can see clearly. It means running discussion and decisions through the same channel for everyone, rather than letting in-room participants hash something out verbally while remote attendees watch and wait.</p><p>Teams that adopt this principle consistently report better hybrid meeting outcomes, because it removes the structural advantage of being in the room rather than trying to patch around it after the fact.</p><h3>Technical Setup That Actually Matters</h3><p>Individual cameras, not a single room camera. A single wide-angle camera capturing the whole room makes every in-room participant small and hard to read for remote attendees, and makes it hard to tell who’s currently speaking. Individual laptops or dedicated camera setups per person, even for people in the same room, produce a meaningfully better remote experience.</p><p>Microphone placement that actually picks up the room. Built-in laptop mics or a single ceiling mic often fail to capture people speaking from different parts of a room clearly. A dedicated conference microphone, positioned centrally, reduces the all-too-common experience of remote participants hearing only fragments of in-room conversation.</p><p>A <a href="https://www.kumospace.com/screen-sharing">shared screen</a> that’s actually shared, not a physical whiteboard photographed badly through a webcam. If visual collaboration is part of the meeting, use a digital tool that both in-room and remote participants can see and edit with equal fidelity.</p><p>Consistent audio quality testing before, not during, the meeting. A meeting that starts with five minutes of “can you hear me?” troubleshooting sets a bad tone and burns goodwill, particularly for remote participants who already feel like an afterthought in hybrid setups.</p><h3>Facilitation Techniques That Level the Playing Field</h3><p>Call on remote participants by name, proactively. Don’t wait for remote attendees to find an opening in a fast-moving in-room conversation. A facilitator who periodically checks in directly (“Sarah, what’s your take on this?”) gives remote voices an entry point they wouldn’t get by competing for airtime naturally.</p><p>Use the chat as a parallel channel, and actually read it. Remote participants often type a thought into chat because jumping into verbal conversation feels harder over video. If the facilitator or a designated note-taker isn’t monitoring chat actively, those contributions get lost entirely.</p><p>Rotate who participates from where. If the same people are consistently in the room and the same people are consistently remote, the power imbalance compounds over time. Mixing it up, so everyone experiences both sides of the hybrid dynamic, builds empathy for the remote experience among people who’d otherwise never feel it.</p><p>Set explicit norms around side conversations. In-room side conversations during a hybrid meeting actively exclude remote participants who can’t hear or participate in them. Either bring side discussions into the main conversation or save them for after the meeting.</p><h3>Structuring Hybrid Meetings for Decisions vs. Discussion</h3><p>Not every hybrid meeting needs the same level of intervention. A meeting that’s purely informational, a status update or an announcement, doesn’t suffer much from the in-room/remote imbalance because there’s limited back-and-forth to begin with.</p><p>Meetings built around genuine discussion or decision-making are where the imbalance does the most damage, because the in-room group’s natural advantage in spontaneous back-and-forth directly shapes outcomes. For these higher-stakes meetings, more deliberate facilitation (structured turn-taking, explicit prompts to remote attendees, written input collected in advance) is worth the extra effort that a routine status meeting doesn’t require.</p><h3>What to Do When the Room Has More People Than Remote Attendees</h3><p>A <a href="https://us.neat.no/resources/a-complete-guide-to-seamless-hybrid-meetings-for-large-teams-in-2026/">common hybrid pattern</a> is a small number of remote attendees joining a meeting with a much larger in-room group. This is the hardest version of the problem to solve, because the social gravity of the room pulls discussion toward itself almost regardless of facilitation effort.</p><p>If this pattern happens regularly with the same remote attendees, it’s worth asking whether the meeting genuinely needs to be hybrid, or whether it would serve everyone better as either fully remote (leveling the field entirely) or fully in-person with a recorded summary and async follow-up for whoever couldn’t attend. Forcing a permanently outnumbered remote minority into the same meeting repeatedly tends to produce quiet disengagement over time, even with good facilitation.</p><h3>Documentation Matters More in Hybrid Settings</h3><p>Because hybrid meetings are more prone to information asymmetry between in-room and remote participants, clear written documentation afterward matters even more than in a fully remote or fully in-person setting. A meeting summary that captures decisions, context, and action items protects against the scenario where in-room participants walk away with shared context that never made it to remote attendees.</p><p>This is also where <a href="https://www.kumospace.com/blog/recording-online">recording</a> matters. If a hybrid meeting is recorded, remote participants who had connection issues or who couldn’t fully follow an audio-degraded segment have a way to fill the gap rather than just missing whatever happened.</p><h3>Rethinking the Office-Centric Default</h3><p>Many hybrid meeting problems trace back to an unstated assumption: that the meeting is fundamentally an in-person meeting that remote people are tuning into, rather than a meeting that happens to have participants in different locations. Shifting that assumption changes design choices throughout: where the camera points, how the agenda gets shared, whether decisions get made verbally in the room or documented in real time for everyone.</p><p>Some teams address this by removing the office-centric default entirely for certain meeting types, using a shared virtual environment instead of a physical conference room as the default meeting space. Platforms like <a href="https://kumospace.com/">Kumospace</a> create a digital space where everyone, regardless of physical location, joins on equal footing, removing the structural asymmetry that a physical-room-plus-video-call setup inherently creates. For teams that consistently struggle with hybrid meeting imbalance, treating certain meetings as fully virtual, even when some participants happen to be in the same building, can be simpler than trying to perfectly engineer a hybrid setup every time.</p><h3>The Core Test for Any Hybrid Meeting</h3><p>After a hybrid meeting, ask whether a remote participant would describe the experience as equivalent to an in-room participant’s. If the honest answer is no, something in the setup, facilitation, or format needs to change before the next one. Hybrid meetings that pass this test consistently tend to retain remote participant engagement over time. The ones that don’t tend to quietly lose it, meeting by meeting, until remote team members stop bothering to speak up at all.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=450597817d09" width="1" height="1" alt=""><hr><p><a href="https://medium.com/kumospace/the-quiet-bias-built-into-every-hybrid-meeting-450597817d09">The Quiet Bias Built Into Every Hybrid Meeting</a> was originally published in <a href="https://medium.com/kumospace">Kumospace</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Behavioral Interview Questions for Software Engineers]]></title>
            <link>https://medium.com/fonzi-ai/behavioral-interview-questions-for-software-engineers-3ea6573da3f9?source=rss-687975a4ecf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/3ea6573da3f9</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Brett Martin]]></dc:creator>
            <pubDate>Tue, 28 Jul 2026 22:08:10 GMT</pubDate>
            <atom:updated>2026-07-28T22:08:10.859Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NwjNO3pEAbkaaCKRtWka3g.jpeg" /></figure><p>Behavioral interview questions are the round most software engineers under-prepare for, and the round most likely to be the actual deciding factor when two candidates have similar technical scores. Companies don’t want to hire brilliant jerks. A talented engineer who can’t collaborate, handle feedback, or navigate ambiguity well is, in aggregate, a net cost to a team, regardless of raw coding ability.</p><p>This covers how to actually prepare, not just which questions to expect.</p><h3>What Behavioral Interviews Are Actually Testing</h3><p><a href="https://fonzi.ai/blog/interview-questions-to-ask-candidates">Behavioral interview questions</a> evaluate how a candidate has handled real situations in the past as a predictor of how they’ll behave on the job. The premise is that specific past behavior predicts future behavior more reliably than hypothetical answers or stated values.</p><p>As engineers become more senior, the weight of behavioral evaluation increases. Senior engineers need to lead and influence without direct authority, resolve conflicts, anticipate risks, and communicate technical trade-offs to non-technical stakeholders. Technical interviews test whether you can do the work. Behavioral interviews test whether people will want to keep working with you while you do it.</p><h3>The STAR Method: What It Actually Means</h3><p><a href="https://fonzi.ai/blog/master-star-method-cheat-sheet">STAR</a> stands for Situation, Task, Action, Result. It’s a structure for organizing your answer, not a script to memorize word for word.</p><p><strong>Situation (roughly 20% of your answer):</strong> Set the context in two to three sentences. What team, what project, what was happening. Keep this brief; the interviewer doesn’t need every detail, just enough to understand the stakes.</p><p><strong>Task (roughly 10%):</strong> What was your specific responsibility, not the team’s broader goal. This is where candidates often drift into describing what the team needed to do rather than what they personally owned.</p><p><strong>Action (roughly 60%):</strong> What you actually did. This should be the longest part of your answer by a wide margin, and it should be specific. <em>“I held a meeting”</em> is vague. <em>“I scheduled a 30-minute sync with the three engineers who owned the conflicting services, proposed a shared interface contract, and wrote the first draft myself”</em> is concrete and answers the question the interviewer is actually asking: what did you personally do?</p><p><strong>Result (roughly 10%):</strong> What happened. Quantify wherever you can. <em>“We improved system performance”</em> is weaker than <em>“we saw a 40% reduction in latency, which reduced support tickets by half.”</em> Numbers make the result verifiable and memorable.</p><p>Some frameworks add a fifth element, Reflection, where you briefly note what you learned or would do differently. This extra step signals self-awareness and growth, which is a strong signal at senior levels specifically.</p><h3>Why “Tell Me About Yourself” and “Tell Me About a Conflict” Matter Most</h3><p>Two questions appear in nearly every behavioral loop regardless of company, and they’re worth specifically preparing rather than improvising.</p><p><a href="https://fonzi.ai/blog/common-mistakes-answering-tell-me-about-yourself"><strong>“Tell me about yourself</strong></a><strong>.”</strong> This usually opens the interview and sets the tone for everything after. Keep it under two minutes, focus on your trajectory and what you’re looking for now, and avoid simply reciting your resume line by line. The interviewer has your resume. They want the narrative connecting it.</p><p><strong>“Tell me about a time you resolved a conflict.”</strong> This is the most universally asked soft-skill question across companies and levels. Have a specific story ready that shows you handling a genuine disagreement professionally, not a manufactured story where you were obviously right and the other person obviously wrong. The strongest conflict stories show real tension and a resolution that required give on both sides.</p><h3>Building Your Story Bank</h3><p>Rather than trying to prepare a unique answer for every possible question, the more effective approach is preparing 5 to 8 strong stories from your career that each demonstrate multiple behaviors, then mapping them to whichever question comes up.</p><p>To build this bank, look back across your career for projects with high impact, high complexity, and high personal involvement: major launches, significant refactors, system migrations, or architecture decisions that affected multiple teams. For each one, identify the different “behaviors” embedded in the story: did it involve ambiguity? Conflict? Mentoring someone? Failure and recovery? A single well-chosen project often contains material for four or five different behavioral questions.</p><p><strong>A reasonable target story bank:</strong></p><figure><img alt="Table titled ‘Story type’ and ‘What it demonstrates,’ listing six interview story categories — technically complex problems, disagreements with colleagues or managers, shifting requirements, mistakes made, mentoring struggling teammates, and influencing without authority — each paired with the skills they demonstrate, such as problem-solving, conflict resolution, adaptability, self-awareness, leadership, and persuasion." src="https://cdn-images-1.medium.com/max/1024/1*Se57cdYgU9WIoYVTGBTK5A.jpeg" /></figure><h3>How to Talk About Failure Without Sabotaging Yourself</h3><p>Questions about failure are not traps, and treating them as traps is itself a mistake. Interviewers ask about failure specifically because dodging it (claiming you’ve never really failed, or deflecting blame entirely onto others) reads worse than genuine accountability.</p><p>The strongest failure stories follow a clear pattern: what went wrong, your specific role in it, what you learned, and concrete evidence that you applied the lesson afterward. A story that ends with “and I learned to communicate more” without a specific example of changed behavior afterward feels unfinished. A story that ends with “and on my next project, I implemented X specifically because of what I learned here” lands much harder.</p><h3>Company-Specific Patterns Worth Knowing</h3><p>Some companies structure behavioral interviews around explicit frameworks, and preparing generically without accounting for this can cost you.</p><p><a href="https://thebehavioral.substack.com/p/whats-different-about-amazon-behaviorals">Amazon</a> evaluates behavioral answers against 16 specific Leadership Principles (Customer Obsession, Ownership, Bias for Action, Disagree and Commit, and others), and these show up even in technical rounds, not just dedicated behavioral interviews. If you’re interviewing at Amazon specifically, prepare at least two stories mapped to each principle rather than relying on generic preparation.</p><p>Other companies are less formalized but still consistently probe for the same underlying behaviors: ownership, communication under ambiguity, and how you handle disagreement. Research the specific company’s stated values before the interview; many publish them, and tailoring your story selection to echo that language (without being obviously performative about it) signals genuine fit.</p><h3>Common Mistakes That Undercut Strong Candidates</h3><ol><li>Spending too long on the Situation. Two or three sentences is enough context. Interviewers get impatient when candidates over-explain the setup before getting to what they actually did.</li><li>Describing what the team did, not what you did. “We decided to refactor the service” doesn’t tell the interviewer about your individual contribution. Be explicit: “I proposed the refactor, built the initial prototype, and convinced the team to adopt it.”</li><li>Only telling success stories. A candidate who has no failure stories, or who dodges every question about mistakes, raises a flag rather than looking impressive. Self-awareness about your own limitations is itself a positive signal.</li><li>Rambling without structure. Behavioral questions without a clear answer structure tend to wander, lose the interviewer’s attention, and never land on a clear result. This is exactly what preparing a story bank in advance prevents.</li><li>Memorizing answers so rigidly they sound scripted. There’s a balance between being prepared and sounding rehearsed. Know the key facts and structure of each story cold, but allow some natural variation in how you tell it depending on the specific question asked.</li></ol><h3>Practicing Without Sounding Robotic</h3><p>The goal of preparation is internalizing the structure and key facts of your stories well enough that you can adapt them naturally to whatever specific angle the interviewer asks about, not memorizing a fixed script.</p><p>Practice out loud, not just by writing notes. The cadence and pacing of spoken delivery is different from written prep, and stories that read well on paper sometimes ramble when spoken without practice. Recording yourself or running a mock interview with a friend or mentor surfaces problems that silent review misses entirely.</p><h3>Where Behavioral Strength Matters Most</h3><p>At small AI startups, behavioral fit often matters proportionally more than at large companies, simply because there’s less organizational structure to absorb friction between people who don’t work well together. A 12-person engineering team has nowhere to hide a difficult collaborator, and a strong individual contributor who communicates poorly under ambiguity can meaningfully slow down a team that size.</p><p><a href="https://talent.fonzi.ai/?utm_source=medium&amp;utm_medium=organic&amp;utm_campaign=june&amp;utm_content=software_eng_blog">Fonzi</a> matches vetted software engineers with VC-backed AI startups where this kind of fit is evaluated carefully, connecting candidates with teams where both the technical and behavioral signals genuinely align.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3ea6573da3f9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/fonzi-ai/behavioral-interview-questions-for-software-engineers-3ea6573da3f9">Behavioral Interview Questions for Software Engineers</a> was originally published in <a href="https://medium.com/fonzi-ai">Fonzi AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Remote Engagement Number Nobody Expects]]></title>
            <link>https://medium.com/kumospace/the-remote-engagement-number-nobody-expects-82e243343e91?source=rss-687975a4ecf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/82e243343e91</guid>
            <category><![CDATA[remote-work-tools]]></category>
            <category><![CDATA[business-template]]></category>
            <category><![CDATA[meeting-productivity]]></category>
            <category><![CDATA[workplace-collaboration]]></category>
            <category><![CDATA[meeting-minutes]]></category>
            <dc:creator><![CDATA[Brett Martin]]></dc:creator>
            <pubDate>Tue, 28 Jul 2026 21:38:37 GMT</pubDate>
            <atom:updated>2026-07-28T21:38:37.049Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_lC5xT46LdbVLL4mhoJmuw.jpeg" /></figure><p>Remote employee engagement is the degree to which a distributed employee feels emotionally committed to their work and connected to their team, and the data on it cuts against the common assumption that remote work inherently damages engagement. Remote workers actually report higher engagement than hybrid or fully on-site employees in most current research. The challenge isn’t remote work itself. It’s that engagement at any location is increasingly hard to sustain, and remote settings remove some of the passive mechanisms that used to mask disengagement until it became a real problem.</p><h3>The Engagement Numbers Are Worse Than Most Managers Realize</h3><p>Only <a href="https://www.gallup.com/workplace/349484/state-of-the-global-workplace.aspx">20% of employees worldwide</a> report feeling actively engaged with their work, leaving the large majority disengaged or actively disengaged. In the US specifically, the ratio of engaged to actively disengaged employees fell to 1.8-to-1 in 2024, down from 2.1-to-1 the year before, continuing a downward trend that’s been called “the Great Detachment” by some researchers tracking it.</p><p>This matters financially, not just culturally. <a href="https://www.gallup.com/workplace/708071/global-employee-engagement-continues-decline.aspx">Disengaged employees</a> cost the global economy an estimated $10 trillion in lost productivity, roughly 9% of global GDP. Companies with genuinely engaged employees see 23% higher profitability than those without, which makes engagement one of the more directly measurable levers available to a manager, not just a soft cultural goal.</p><h3>Remote Workers Are More Engaged, Not Less</h3><p>This finding surprises people who assume remote work erodes connection. <a href="https://www.gallup.com/workplace/660236/remote-work-paradox-engaged-distressed.aspx">Gallup’s research</a> shows remote employees report 29% engagement, compared to 21% for hybrid workers and just 20% for fully on-site employees. The flexibility and autonomy that remote work provides appear to outweigh whatever connection loss comes from not sharing a physical office.</p><p>This doesn’t mean remote engagement happens automatically. It means the baseline starting point is better than assumed, and the gap that does exist is addressable through specific management practices rather than requiring a return to in-person work to fix.</p><h3>Why Managers Matter More Than Location</h3><p>Managers account for roughly 70% of the variance in team engagement levels. This is the single most important fact for anyone trying to improve remote engagement: the manager’s behavior matters more than whether the team is remote, hybrid, or in-office.</p><p>What specifically does an engaged manager do differently? Regular, substantive check-ins rather than purely transactional status updates. Clear communication of expectations and how individual work connects to broader goals. Genuine recognition of good work, delivered specifically rather than generically. And perhaps most importantly, demonstrated investment in each person’s growth and development, not just their current output.</p><p>A manager who treats 1:1s as a box to check rather than a real conversation, who only reaches out when something’s wrong, or who never discusses career growth beyond the current role, will struggle to produce an engaged team regardless of how good the company’s broader perks and policies are.</p><h3>The Isolation Problem and How to Address It</h3><p>Remote workers do face a specific engagement risk that in-office workers don’t: loneliness. 20% of all workers report <a href="https://www.gallup.com/workplace/645566/employees-worldwide-feel-lonely.aspx">daily loneliness</a>, and remote workers are disproportionately affected. This is a real risk even within a population that reports higher overall engagement, which suggests connection and engagement aren’t quite the same thing; someone can be professionally engaged with their work while still feeling personally isolated from their colleagues.</p><p>Addressing this requires deliberate social infrastructure, not just task-focused management. Daily or near-daily virtual touchpoints, even brief ones, help. Regular check-ins that ask about more than just task status give people room to surface how they’re actually doing, not just what they’ve shipped.</p><h3>Building Engagement Through Recognition</h3><p>Recognition is one of the more reliably effective and cheapest engagement levers available, and it’s also one of the easiest to under-deliver in remote settings, where the casual “nice work on that” that happens organically in an office doesn’t occur unless someone makes a deliberate channel for it.</p><p>A dedicated space for public recognition, whether a Slack channel, a section of a regular team meeting, or a more structured program, gives visibility to good work that would otherwise go unnoticed by anyone beyond the immediate collaborators. The specificity of recognition matters more than the format: “great job this week” lands far weaker than naming the specific thing someone did and why it mattered.</p><h3>Accountability and Engagement Are Connected</h3><p>A <a href="https://www.ajg.com/employeeexperience/-/media/files/gallaghercomms/gcommssite/employee-communications-report-2025.pdf">rise in joint accountability</a> appears to correlate with stronger engagement in recent research, with one report finding a 74% increase in measured accountability, the large majority of it framed as shared rather than individual. This connects to a broader pattern: employees who understand how their work fits into a larger shared goal, and who feel real ownership over outcomes rather than just tasks, report stronger engagement than those who experience their work as a disconnected stream of individual assignments.</p><p>Practically, this means structuring work so people understand not just what they’re doing but why it matters and how it connects to what their teammates are doing. Project kickoffs that explain the broader goal, not just the task list, and retrospectives that connect individual contributions to team outcomes both reinforce this sense of shared purpose.</p><h3>Learning and Development as an Engagement Driver</h3><p>A lack of growth opportunity is one of the more consistently cited drivers of disengagement and eventual attrition. Employees who feel they’re developing new skills and have a visible path forward report meaningfully higher engagement than those who feel stuck in a static role.</p><p>For remote teams specifically, development opportunities require more deliberate design, since the informal mentorship and stretch assignments that happen organically through physical proximity in an office don’t occur by default. Structured growth conversations, explicit stretch project assignments, and access to learning resources all need to be intentionally built into a remote team’s operating rhythm rather than assumed to emerge naturally.</p><h3>What Actually Drives Engagement, Ranked by What the Data Shows</h3><p>Based on the available research, the highest-leverage levers for remote engagement, roughly in order of impact, are: manager quality and consistency of 1:1 engagement, clarity of role expectations and how individual work connects to broader goals, genuine and specific recognition, visible development and growth opportunities, and deliberate social connection that addresses isolation risk directly.</p><p>Perks and benefits matter, but they show up lower on this list than most companies assume. A team with strong manager relationships and clear purpose will generally out-engage a team with better perks and weaker management, which is worth remembering before defaulting to a wellness stipend or a new tool as the fix for an engagement problem that’s actually rooted in management quality.</p><h3>Where Persistent Connection Tools Fit</h3><p>Part of addressing the isolation specific to remote engagement is giving people low-friction ways to be present with colleagues outside of formally scheduled interactions. Persistent virtual spaces, like <a href="https://kumospace.com/">Kumospace</a>, let remote team members have the kind of casual, ambient interaction that builds the personal connection underlying genuine engagement, the equivalent of being visible and approachable rather than only existing as a name on a calendar invite.</p><p>This doesn’t replace the manager relationship as the primary engagement driver, but it does address a specific gap: the informal social texture that makes a team feel like a team, rather than a collection of people who happen to work on the same projects.</p><h3>Telling If Any of This Is Actually Working</h3><p>Regular engagement surveys, ideally short and frequent rather than long and annual, give you a trend line rather than a single snapshot. Track engagement specifically alongside retention and performance data, and pay particular attention to manager-level variance: if one manager’s team consistently reports lower engagement than peers, that’s a coaching opportunity, not a location or remote-work problem.</p><p>The data is clear that engagement is solvable, and that it’s solvable primarily through management practice rather than through location policy or perks. Teams that treat it that way tend to see real, measurable improvement rather than chasing initiatives that don’t address the actual driver.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=82e243343e91" width="1" height="1" alt=""><hr><p><a href="https://medium.com/kumospace/the-remote-engagement-number-nobody-expects-82e243343e91">The Remote Engagement Number Nobody Expects</a> was originally published in <a href="https://medium.com/kumospace">Kumospace</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Prepare for a System Design Interview]]></title>
            <link>https://medium.com/fonzi-ai/how-to-prepare-for-a-system-design-interview-a0a22c146616?source=rss-687975a4ecf8------2</link>
            <guid isPermaLink="false">https://medium.com/p/a0a22c146616</guid>
            <category><![CDATA[software-engineering]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[data-science]]></category>
            <dc:creator><![CDATA[Brett Martin]]></dc:creator>
            <pubDate>Sat, 25 Jul 2026 23:23:10 GMT</pubDate>
            <atom:updated>2026-07-25T23:23:10.482Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jkDidJJl2Hm-Xeyk2UdmfA.jpeg" /></figure><p>A system design interview tests something different from a coding interview, and most candidates prepare for it the wrong way. They memorize architecture diagrams for “design Twitter” and “design a URL shortener” without understanding that the diagram was never the point. The interview is testing whether you can scope an ambiguous problem, reason about trade-offs out loud, and make defensible decisions under uncertainty.</p><p>Here’s what actually matters in preparation and execution.</p><h3>What a System Design Interview Is Actually Testing</h3><p>A <a href="https://fonzi.ai/blog/system-design-interview">system design interview</a> evaluates a candidate’s ability to scope ambiguous problems, reason through architectural trade-offs, and communicate technical decisions clearly under time pressure. It’s not primarily a test of memorized architecture patterns.</p><p>This is the part most candidates get backward. Being a good programmer has a surprisingly small role in passing system design interviews. Many talented distributed systems engineers struggle with the format itself, while engineers without deep production experience with the specific technologies sometimes perform well because they’ve practiced the meta-skill: structuring an ambiguous conversation under pressure.</p><p>That distinction matters for how you prepare. Reading “Designing Data-Intensive Applications” gives you vocabulary. It doesn’t teach you how to run a 45-minute conversation.</p><h3>The Five-Step Structure That Works Across Companies</h3><p>Regardless of company or specific prompt, a reliable structure carries you through almost any system design interview.</p><ol><li>Clarify functional requirements. Don’t start designing before you understand what the system actually needs to do. If the prompt is “design TikTok,” ask which features matter: video upload, feed generation, search, live streaming. The prompt is intentionally vague. Narrowing scope is part of the test.</li><li>Establish non-functional requirements and constraints. Ask about scale: how many users, what request volume, what latency targets, what consistency requirements. Concrete numbers (queries per second, data volume, read/write ratio) signal that you think about systems quantitatively, not just conceptually.</li><li>Propose a high-level architecture. Sketch the major components and how data flows between them before diving into any single piece. Use generic component names rather than specific brand names unless you’re prepared to defend that specific choice. Say “a message queue,” not “Kafka,” unless you can explain why Kafka specifically over alternatives.</li><li>Go deep on two or three components. You won’t have time to detail every part of the system. Pick the components that are most interesting or most central to the problem, and go deep there. This is where trade-off reasoning happens: discuss the options, name the trade-offs, and commit to a choice.</li><li>Address failure modes and operational concerns. A design that works on paper isn’t enough. Discuss what happens when a component fails, how you’d monitor the system, and how you’d handle scaling beyond the initial estimate. This is the part junior candidates skip and senior interviewers specifically probe for.</li></ol><h3>How to Handle Trade-Off Discussions</h3><p>The single most common failure pattern in system design interviews: presenting options without committing to one.</p><p>A candidate says: “We could use this type of database, or this other one, and these are some pros and cons…” and then moves to the next component without deciding. This reads as indecisive even when the analysis itself is correct. In the real world, you have to make decisions. The interview is testing whether you can do that under uncertainty, not whether you can list every possible option.</p><p>The fix is simple in structure: state the options, name the trade-offs briefly, then explicitly commit. “We could use a relational database for stronger consistency or a key-value store for better write throughput. Given that this system prioritizes consistency for financial transactions, I’ll use a relational database.” That’s a complete trade-off statement. It’s fine to revise the decision later if new information changes the calculus, but always land on a choice.</p><h3>Common Pitfalls That Cost Offers</h3><ol><li>Jumping straight to a solution. Skipping requirement-gathering to start drawing boxes signals you don’t know how to scope ambiguous problems, which is the core thing being tested.</li><li>Over-engineering for the stated scale. Designing a globally-distributed, multi-region system for a problem that specifies 10,000 users wastes time and signals poor judgment about appropriate complexity. Start simple. Add complexity only when the requirements justify it.</li><li>Under-engineering and missing obvious failure modes. The opposite failure: a design that works for the happy path but falls over the moment a component fails or load spikes. Senior-level interviews specifically test for this.</li><li>Memorized templates that don’t fit the actual prompt. Interviewers increasingly introduce mid-conversation changes specifically to see if a candidate can adapt, rather than executing a rehearsed answer regardless of fit. If your design doesn’t change at all when the interviewer adds a new constraint, that’s a signal you weren’t actually listening.</li><li>Silence while thinking. Verbalize your reasoning as you go. The interviewer is evaluating your thought process, not just your final answer, and silence gives them nothing to evaluate.</li></ol><h3>What Changes at AI-Focused Companies</h3><p>System design interviews at AI-first companies are increasingly framed around AI workloads, but the underlying patterns are classic distributed systems problems wearing a different costume. A common example: design a batch inference API for a GPU cluster. The patterns being tested (queuing, batching under resource constraints, mapping async work to synchronous responses, managing a scarce and expensive compute resource) are general infrastructure patterns that show up regardless of whether the workload is AI inference or anything else.</p><p>Interviewers at AI companies explicitly look for candidates who can abstract the AI-specific framing into general patterns. “Batch inference on a GPU” becomes “batched processing on a constrained compute resource.” If you can make that translation cleanly, you demonstrate that your systems thinking transfers across domains rather than being narrowly tied to one type of workload.</p><h3>How to Actually Practice</h3><ol><li>Solo whiteboard practice. Pick a common prompt (design a URL shortener, design a chat system, design a rate limiter) and talk through the full five-step structure out loud, by yourself, with a timer. This builds the structural muscle memory that matters more than memorizing specific solutions.</li><li>Read real design docs and post-mortems. If your current job involves any system design work, study your own company’s internal docs and incident retrospectives. Understanding how real systems actually fail under load is more valuable than abstract textbook knowledge.</li><li><a href="https://fonzi.ai/blog/ai-interview-practice">Mock interviews</a> with feedback. Practicing alone teaches structure. Practicing with a partner who can interrupt, challenge your choices, and change requirements mid-conversation teaches adaptability, which is what real interviews actually demand.</li><li>Study trade-offs, not just components. Instead of memorizing what a message queue is, study why you’d choose one queue technology over another in a specific context. The interview rewards reasoning about trade-offs far more than reciting facts about individual technologies.</li></ol><h3>What to Expect at Different Seniority Levels</h3><p>System design rounds calibrate differently by level, and the same prompt produces different expectations depending on the seniority being evaluated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BKLlDosE8-2_r60DLv5IDg.jpeg" /></figure><p>Fumbling a design interview can result in being down-leveled even if you pass, since design interviews are often a key factor in determining where a candidate lands on the seniority ladder. This is worth knowing going in: the interview isn’t pass/fail in a simple sense. It’s calibrating where you land.</p><h3>Getting Into Rooms Where Preparation Pays Off</h3><p>System design interviews matter most at companies building genuinely complex systems, which describes most VC-backed AI startups working at meaningful scale. The preparation above applies regardless of company size, but the specific patterns tested (constrained compute, async processing, data pipelines) show up disproportionately at AI-focused companies.</p><p><a href="https://talent.fonzi.ai/?utm_source=medium&amp;utm_medium=organic&amp;utm_campaign=june&amp;utm_content=software_eng_blog">Fonzi</a> matches vetted software engineers with VC-backed AI startups in NYC and SF where these systems problems are central to the actual work, not just the interview process.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a0a22c146616" width="1" height="1" alt=""><hr><p><a href="https://medium.com/fonzi-ai/how-to-prepare-for-a-system-design-interview-a0a22c146616">How to Prepare for a System Design Interview</a> was originally published in <a href="https://medium.com/fonzi-ai">Fonzi AI</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>