<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>Constraint Lab</title>
    <description>System exploration</description>
    <link>https://constraintlab.com/</link>
    <atom:link href="https://constraintlab.com/feed.xml" rel="self" type="application/rss+xml"/>
    <pubDate>Sat, 08 Aug 2026 05:04:17 +0000</pubDate>
    <lastBuildDate>Sat, 08 Aug 2026 05:04:17 +0000</lastBuildDate>
    <generator>Jekyll v4.4.1</generator>
    
      <item>
        <title>How AI changes our system of work</title>
        <description>&lt;blockquote&gt;
  &lt;p&gt;So your company is making thirty-six percent more money from your plant just from installing some robots? Incredible. - Jonah from &lt;em&gt;The Goal&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;AI is impressive technology, but more interesting to me is how we apply the technology. The fast rollout has some good and some bad. In my experience, no one has done a sufficient job of explaining the rules on determining what’s good and what’s bad. And I’m not talking about experience reports, vague mandates, or arguments that hand-wave the important details.&lt;/p&gt;

&lt;p&gt;It seems obvious to me that our current plan of tokenmaxxing is not just bad from a Goodhart’s Law perspective. It’s also bad because simply maxing out running of LLMs doesn’t help even if everyone was fully bought into the mission.&lt;/p&gt;

&lt;p&gt;Eli Goldratt’s four questions of new tech seem like the way to get the rules I’m looking for.&lt;/p&gt;

&lt;ol&gt;
  &lt;li&gt;What is the power of the new technology?&lt;/li&gt;
  &lt;li&gt;What limitations does the new technology diminish?&lt;/li&gt;
  &lt;li&gt;What rules helped us bypass the limitation?&lt;/li&gt;
  &lt;li&gt;What are the new rules?&lt;/li&gt;
&lt;/ol&gt;

&lt;h2 id=&quot;what-is-the-power-of-ai&quot;&gt;What is the power of AI?&lt;/h2&gt;

&lt;p&gt;There are a few different ways I’d describe it, depending on how abstract I need to be:&lt;/p&gt;

&lt;ol&gt;
  &lt;li&gt;AI is good at taking loosely structured data and transforming it into differently structured data.&lt;/li&gt;
  &lt;li&gt;AI is good at cheaply generating plausible hypotheses.&lt;/li&gt;
  &lt;li&gt;AI is good at cheaply testing hypotheses.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The connection is, the transformation in step 1 is not guaranteed correct. It has to be validated against reality. That process is just a small, scientific experiment. Therefore, the output of the transformation is an hypothesis.&lt;/p&gt;

&lt;p&gt;AI is very good at generating a reasonable hypothesis, testing it against whatever reality it can access, and then creating a new hypothesis based on the results.&lt;/p&gt;

&lt;p&gt;Put succinctly: AI can, by the process of generating and testing hypotheses, cheaply answer any question with a hypothesis that has a ~good chance of being proven correct. (It’s a bit like Jeopardy where every answer to the question must be a question. Any statement can be easily transformed into a question, it just takes a bit to get used to thinking about that way.)&lt;/p&gt;

&lt;h2 id=&quot;what-is-the-limitation-ai-diminishes&quot;&gt;What is the limitation AI diminishes?&lt;/h2&gt;

&lt;p&gt;All software is a hypothesis: “I predict that this set of instructions will result in this behavior.” If you tried to randomly generate instruction sets, you’d be surprised if you could even get something that had the proper syntax, let alone behavior. Every plausible set of instructions required human effort and understanding.&lt;/p&gt;

&lt;p&gt;Restated in our AI language, generating a reasonable hypothesis is expensive because it requires human effort and understanding. Validating a hypothesis is expensive because it is potentially risky and requires human effort and understanding. Depending on the context, AI has made those statements mostly false or completely false.&lt;/p&gt;

&lt;h2 id=&quot;what-rules-helped-us-bypass-the-limitation&quot;&gt;What rules helped us bypass the limitation?&lt;/h2&gt;

&lt;h4 id=&quot;1-we-route-work-to-get-the-most-efficient-use-of-human-effort&quot;&gt;1. We route work to get the most efficient use of human effort.&lt;/h4&gt;

&lt;p&gt;Take code reviews. If it’s a simple, safe change, ~anyone can sign off. If it’s risky or complex, get a principal involved. Don’t submit half completed code for review, unless there’s a specific reason you need help at that point. Prefer reviews from someone who has context, so we don’t require the extra human effort of gaining context.&lt;/p&gt;

&lt;h4 id=&quot;2-we-use-linters&quot;&gt;2. We use linters.&lt;/h4&gt;

&lt;p&gt;Here go a bunch of rules to amortize the cost of human effort that goes into deciding things like coding style. Let’s answer this question once so we never have to spend human effort on it again. Let’s make it automatic so no one ever has to think about it again.&lt;/p&gt;

&lt;h4 id=&quot;3-we-have-upfront-design&quot;&gt;3. We have upfront design.&lt;/h4&gt;

&lt;p&gt;It’s expensive to build the wrong thing. The further a design gets into production, the more human effort required to pull it back.&lt;/p&gt;

&lt;h4 id=&quot;4-we-reject-upfront-design&quot;&gt;4. We reject upfront design.&lt;/h4&gt;

&lt;p&gt;It’s expensive to design a bunch of things on paper that don’t translate to code and/or end up never getting built.&lt;/p&gt;

&lt;h4 id=&quot;5-we-use-automated-testing-canaries-and-phased-rollouts&quot;&gt;5. We use automated testing, canaries, and phased rollouts.&lt;/h4&gt;

&lt;p&gt;We’re not sure our code is correct, so we verify it against reality. How we verify is determined by cost and value. Unit tests are low on the cost scale but also low on the value scale. Production is the opposite.&lt;/p&gt;

&lt;h4 id=&quot;6-risk-separated-code&quot;&gt;6. Risk separated code&lt;/h4&gt;

&lt;p&gt;Config changes require approvals from multiple principals. A typo on the careers page gets a “LGTM”. Credit card processing is a completely separate org with access strictly gated.&lt;/p&gt;

&lt;p&gt;This isn’t an exhaustive list of rules nor does it explain every reason the rule exists (e.g. code reviews share accountability). It does show how reducing human effort, specifically of the “human thinking” variety, along with risk management, underlines so much of our system of work.&lt;/p&gt;

&lt;h3 id=&quot;analyze-the-rules&quot;&gt;Analyze the rules&lt;/h3&gt;

&lt;p&gt;Before figuring out the new rules, we should see that our old rules are built on an important assumption. Because every reasonable hypothesis required human effort/understanding, you could assume human understanding + reasonable hypothesis always traveled together. It was safe to use “this is reasonable” interchangeably with “some human understands this.”&lt;/p&gt;

&lt;p&gt;&lt;em&gt;That assumption is &lt;strong&gt;totally broken&lt;/strong&gt;. It is not true anymore at all.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Code review approval? Could be Alice, could be Claude.&lt;/p&gt;

&lt;p&gt;PR that passes all the tests? Could be Bob, could be Codex.&lt;/p&gt;

&lt;p&gt;Ten page design doc? Definitely an LLM.&lt;/p&gt;

&lt;p&gt;Code reviews assumed there was a human on the other side we could jump on a call with and explain how it works. We assumed that the twenty commits going out in the next release were backed by ~20 people who understood each commit. It was safe for all of our rules to assume this, so every rule potentially needs to be changed. This assumption was load-bearing (a human person typed that, relax, it’s just a cheap joke).&lt;/p&gt;

&lt;h2 id=&quot;what-are-the-new-rules&quot;&gt;What are the new rules?&lt;/h2&gt;

&lt;p&gt;Answering this is a lot of work. The new rules aren’t determined by blindly negating the old rules. Getting rid of code reviews and just shoving all the commits on main &lt;em&gt;might&lt;/em&gt; work, but other parts of the system would need to change. And code reviews are not just about ensuring correctness, they’re also about security/accountability. So if we get rid of code reviews, we need to satisfy all of the reasons that rule was there. We’re changing a system of work, so any local changes have to be weighed against their impact on the entire system.&lt;/p&gt;

&lt;p&gt;As I mentioned, it’s a lot of work, so that will be the work of the next post.&lt;/p&gt;
</description>
        <pubDate>Fri, 07 Aug 2026 00:00:00 +0000</pubDate>
        <link>https://constraintlab.com/blog/how-ai-changes-our-system-of-work.html</link>
        <guid isPermaLink="true">https://constraintlab.com/blog/how-ai-changes-our-system-of-work.html</guid>
        
        <category>ai</category>
        
        
      </item>
    
      <item>
        <title>AI is a bubble, just like dot-com</title>
        <description>&lt;blockquote class=&quot;twitter-tweet&quot;&gt;&lt;a href=&quot;https://twitter.com/bostonou/status/2085097056159318275&quot;&gt;August 5, 2026&lt;/a&gt;&lt;/blockquote&gt;

&lt;p class=&quot;post-disclaimer&quot;&gt;I tested out having Claude generate the final draft. Results were quite clear that people would not read it lol.&lt;/p&gt;

&lt;p&gt;Andrej Karpathy has written large language models the way most of us have written CRUD apps. He was a founding member of OpenAI, ran AI at Tesla, and joined Anthropic this spring to train Claude. He wrote nanoGPT and llm.c, the small readable codebases people use to learn how a language model works end to end.&lt;/p&gt;

&lt;p&gt;When Anthropic put Claude inside Slack, he called it &lt;a href=&quot;https://x.com/karpathy/status/2069547676849557725&quot;&gt;“a new paradigm”&lt;/a&gt; — the third major redesign of how humans use these systems. Not a website you visit anymore, not an app you install, but “a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans.”&lt;/p&gt;

&lt;p&gt;He closed with: “it works and it is awesome.”&lt;/p&gt;

&lt;p&gt;The replies had a different read. One of them, in full:&lt;/p&gt;

&lt;blockquote class=&quot;twitter-tweet&quot;&gt;&lt;p lang=&quot;en&quot; dir=&quot;ltr&quot;&gt;&amp;gt; This is a new paradigm&lt;br /&gt;&lt;br /&gt;brother it&amp;#39;s a slack integration&lt;br /&gt;&lt;br /&gt;I miss the old andrej&lt;/p&gt;&amp;mdash; @MatthewFoxAF &lt;a href=&quot;https://twitter.com/MatthewFoxAF/status/2069709303234146384&quot;&gt;June 24, 2026&lt;/a&gt;&lt;/blockquote&gt;
&lt;script async=&quot;&quot; src=&quot;https://platform.twitter.com/widgets.js&quot; charset=&quot;utf-8&quot;&gt;&lt;/script&gt;

&lt;p&gt;Hold that exchange still for a second. Nobody in it is stupid. That’s what makes it worth understanding.&lt;/p&gt;

&lt;p&gt;The same split runs through everything right now, and nothing around it holds still.&lt;/p&gt;

&lt;p&gt;Leadership says “use more AI,” and never defines responsible use. Engineers hear it through an older instinct: don’t ship what you don’t understand. Nobody explains how both survive.&lt;/p&gt;

&lt;p&gt;Layoff announcements credit AI in the same breath as the earnings beat. A colder fear sits underneath: people who own the AI on one side, people whose work it does on the other, no visible way across.&lt;/p&gt;

&lt;p&gt;The ground keeps moving. Prompt engineering was a job title. OpenClaw picked up a hundred thousand GitHub stars and caused a Mac Mini shortage in the same week. Ralph was the future for about a weekend. Context engineering, then harness engineering, then graph engineering. Each was the obvious way to work, right up until the next one was. The gaps keep getting shorter.&lt;/p&gt;

&lt;p&gt;And everywhere, teams ship code faster than they can understand it. Ask whoever’s on call.&lt;/p&gt;

&lt;p&gt;I’d love to stand cleanly on one side of all this. I can’t. I haven’t written a line of code in about a year. Claude writes it all. These posts too. I read all the code before it ships, and I still feel the difference. Understanding used to be a byproduct of writing the code myself. Now understanding is a separate job.&lt;/p&gt;

&lt;p&gt;Both perspectives, one person.&lt;/p&gt;

&lt;p&gt;There’s a default way to resolve contradictions like these: decide who’s deluded. “I miss the old andrej” does it in five words: if Karpathy calls a Slack bot a paradigm, something must have happened to Karpathy. Sorting people into right and wrong passes for rigor. It assumes that when two perspectives contradict, one of them must be false.&lt;/p&gt;

&lt;p&gt;Our industry tested that assumption once before, at scale.&lt;/p&gt;

&lt;p&gt;In March 2000 the NASDAQ peaked above 5,000. By October 2002 it had lost 78 percent of its value. Pets.com IPO’d in February and was gone by November. The sock puppet outlived the company. Betting against it was exactly right.&lt;/p&gt;

&lt;p&gt;Amazon lost more than 90 percent of its stock price in the same crash and kept shipping. Within a decade the internet had rearranged retail, news, music, and how you found a job. Betting on Amazon was exactly right too.&lt;/p&gt;

&lt;p&gt;Same year. Same technology. “Wildly overhyped” and “changes everything” were both correct at the extreme.&lt;/p&gt;

&lt;p&gt;Nobody had to be deluded.&lt;/p&gt;

&lt;p&gt;And there were four ways to stand, not two. Short everything: it’s all a bubble. Long everything: it’s all the future. Long Pets.com, short Amazon: backwards twice. Long Amazon, short Pets.com: right twice. The first two are moods about a technology. The last two are judgments about cases. The only thing separating right twice from backwards twice was understanding.&lt;/p&gt;

&lt;p&gt;None of it was magic. The information separating a Pets.com from an Amazon was public the whole time. The hard part was knowing what question to ask of it. What could the internet do that nothing before it could? A bookstore with every book in print, yes. Free shipping on forty-pound bags of dog food, no. Most people, watching the same screens, never asked. They picked a mood and waited for the grade.&lt;/p&gt;

&lt;p&gt;We’re inside the next one now. That may be the only point the amazed and the unimpressed already agree on: whatever this is, it’s big.&lt;/p&gt;

&lt;p&gt;Transformational technology comes along maybe twice in a working life. I was in school for the last one, watching different screens: South Park, Perfect Dark, Conker’s Bad Fur Day. This time I want to understand it while it’s moving: test hypotheses against what happens, and decide how I work while the decision still matters.&lt;/p&gt;

&lt;p&gt;You’re inside it too. Your feed has already assigned you a position: be amazed, or be afraid. Both are moods. Understanding isn’t assigned. You have to go get it.&lt;/p&gt;

&lt;p&gt;So start where understanding starts. What is the disagreement about?&lt;/p&gt;

&lt;p&gt;Mostly not models. There are arguments over which model is best, and over whether open weights beat closed ones. They stay small and almost polite. The real fights are somewhere else. Do you read every line the model writes? Can you ship code you don’t understand? Do passing tests mean it works? Does an AI review count as a review? Every team I know is re-deciding these rules on the fly. Some decide with conviction. More sound like: we don’t know what to do, but we’d better do something.&lt;/p&gt;

&lt;p&gt;Model quality enters at one point only: “the models will get so good that none of this will matter.”&lt;/p&gt;

&lt;p&gt;Maybe.&lt;/p&gt;

&lt;p&gt;So here’s my first hypothesis, stated so you can hold me to it. Models will not become dependable enough, soon enough, to settle these questions for us. If you’re sure that’s wrong, sure that near-flawless output arrives in a year or two, you can stop reading. The questions will answer themselves, and this series with them. I don’t think that’s where we are.&lt;/p&gt;

&lt;p&gt;Now set the pieces next to each other. Intelligent people look at one technology and see wildly different things. History says conclusions that far apart can both be correct. And the rule debates run on, team after team, without resolving. Arguments that run this long between capable people are built on assumptions nobody has checked. We’ve already seen one: somebody has to be wrong. The rule debates lean on another that nobody has said out loud. Behind that one is the question that sorted Amazon from Pets.com, aimed at our own work this time.&lt;/p&gt;

&lt;p&gt;What did this enable that we couldn’t do before?&lt;/p&gt;

&lt;p&gt;Each of those rule debates assumes an answer to it. That’s the question this series works out. One small, checkable step at a time. I’m writing to answer it for myself. You’re welcome to come along and check the work. By the end you won’t need my conclusions. You’ll have your own read on any AI claim that crosses your feed.&lt;/p&gt;

&lt;p&gt;Including this one: AI is a bubble, just like dot-com.&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;&lt;em&gt;Thanks to &lt;a href=&quot;https://www.linkedin.com/in/emma-bukacek-%F0%9F%8D%89-69822249/&quot;&gt;Emma Bukacek&lt;/a&gt; and &lt;a href=&quot;https://www.linkedin.com/in/timothypote/&quot;&gt;Tim Pote&lt;/a&gt;, who read early drafts and told me what wasn’t working.&lt;/em&gt;&lt;/p&gt;
</description>
        <pubDate>Wed, 05 Aug 2026 00:00:00 +0000</pubDate>
        <link>https://constraintlab.com/blog/ai-is-a-bubble-just-like-dot-com.html</link>
        <guid isPermaLink="true">https://constraintlab.com/blog/ai-is-a-bubble-just-like-dot-com.html</guid>
        
        <category>ai</category>
        
        
      </item>
    
      <item>
        <title>It&apos;s Supposed To Be Hot</title>
        <description>&lt;p&gt;Several NFL players were sitting in a meeting room, complaining about the heat. It was training camp for the Tennessee Titans in Nashville, TN.&lt;/p&gt;

&lt;p&gt;Summers in Nashville are brutal. It’s hot. It’s humid. There’s no cloud cover.&lt;/p&gt;

&lt;p&gt;&lt;a href=&quot;https://bleacherreport.com/articles/2134876-what-does-an-nfl-training-camp-schedule-look-like&quot;&gt;It’s miserable.&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;As the players complained about the rigors of training in the stifling heat, future Hall of Famer Eddie George spoke up.&lt;/p&gt;

&lt;p&gt;“It’s training camp. It’s supposed to be hot.”&lt;/p&gt;

&lt;hr /&gt;

&lt;p&gt;I consider myself to be a software engineer of average talent with one very useful focus: I hate hard things. Software engineering is full of complexity that is a nightmare to work with. In my experience, most engineers fight this by trying harder to wrangle the complexity. My brain, however, has gotten to the point where it just stops. It refuses to spend any energy trying to figure out some convoluted hairball of complexity.&lt;/p&gt;

&lt;p&gt;Instead, the vast majority of my effort goes to making things easy.&lt;/p&gt;

&lt;p&gt;The correctness of this focus has come under stress recently as I’ve listened to &lt;a href=&quot;https://www.amazon.com/Shoe-Dog-Memoir-Creator-Nike-ebook/dp/B0176M1A44&quot;&gt;&lt;em&gt;Shoe Dog&lt;/em&gt;&lt;/a&gt;, the memoir of Phil Knight, founder of Nike. The way he talked about running was jarring to me. He was an accomplished long distance runner in college and spent significant time with the world’s best runners.&lt;/p&gt;

&lt;p&gt;The best runners in the world never expected running to be easy. There was always a point where their lungs burned and their legs begged them to stop.&lt;/p&gt;

&lt;p&gt;I hate running. Every time I’ve started jogging, I’ve assumed I would train enough to where that desire to quit would go away. At some point, running would become easy.&lt;/p&gt;

&lt;p&gt;Wrong.&lt;/p&gt;

&lt;p&gt;It never gets easy to run a 4 minute mile.&lt;/p&gt;

&lt;p&gt;Training my mind to expect difficulty and beat it into submission is significantly different than training through difficulty with the hope of some future ease.&lt;sup id=&quot;fnref:everyone-else&quot;&gt;&lt;a href=&quot;#fn:everyone-else&quot; class=&quot;footnote&quot; rel=&quot;footnote&quot; role=&quot;doc-noteref&quot;&gt;1&lt;/a&gt;&lt;/sup&gt;&lt;/p&gt;

&lt;p&gt;Understanding this has helped me refine my understanding of good software engineering. Thinking is hard, and there is no tool that will do your thinking for you. My focus has never been about making thinking easy. Instead, I focus on removing anything that makes thinking harder than necessary.&lt;/p&gt;

&lt;blockquote class=&quot;twitter-tweet&quot;&gt;&lt;p lang=&quot;en&quot; dir=&quot;ltr&quot;&gt;I write code for free. I get paid to think.&lt;/p&gt;&amp;mdash; Jonathan Boston (@bostonou) &lt;a href=&quot;https://twitter.com/bostonou/status/842568360550436864?ref_src=twsrc%5Etfw&quot;&gt;March 17, 2017&lt;/a&gt;&lt;/blockquote&gt;
&lt;script async=&quot;&quot; src=&quot;https://platform.twitter.com/widgets.js&quot; charset=&quot;utf-8&quot;&gt;&lt;/script&gt;

&lt;p&gt;I suspect this is true about anything worthwhile. It’s supposed to hurt. It’s supposed to be hot.&lt;/p&gt;

&lt;div class=&quot;footnotes&quot; role=&quot;doc-endnotes&quot;&gt;
  &lt;ol&gt;
    &lt;li id=&quot;fn:everyone-else&quot;&gt;
      &lt;p&gt;It’s also motivating to know that &lt;em&gt;everyone else&lt;/em&gt; is experiencing that same burning of the lungs. Then, it’s just about being willing to hurt more than the person you’re competing against. &lt;a href=&quot;#fnref:everyone-else&quot; class=&quot;reversefootnote&quot; role=&quot;doc-backlink&quot;&gt;&amp;#8617;&lt;/a&gt;&lt;/p&gt;
    &lt;/li&gt;
  &lt;/ol&gt;
&lt;/div&gt;
</description>
        <pubDate>Mon, 21 Aug 2023 00:00:00 +0000</pubDate>
        <link>https://constraintlab.com/blog/supposed-to-be-hot.html</link>
        <guid isPermaLink="true">https://constraintlab.com/blog/supposed-to-be-hot.html</guid>
        
        <category>discipline</category>
        
        
      </item>
    
      <item>
        <title>Is HashiCorp a Broken Brand?</title>
        <description>&lt;blockquote&gt;
  &lt;p&gt;What it means to have a brand is you’ve made a promise to people. - Seth Godin&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;In 1983, Halston signed a one billion dollar deal with JCPenney.&lt;/p&gt;

&lt;p&gt;&lt;a href=&quot;https://www.fastcompany.com/90637111/what-happened-to-halston-the-rise-and-fall-of-an-american-fashion-icon&quot;&gt;That deal was the beginning of the end for Halston.&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Halston started as a fashion designer in the 1960s, eventually being described as “the top of the fashion show-biz heap.” He designed the pillbox hat that first lady Jacqueline Kennedy wore to her husband’s inauguration, made uniforms for the U.S. Olympic teams, and created costumes for a famous ballet. He was seen with a group of models so often that they became known as “the Halstonettes.”
In short, he was the picture of glamor and exclusivity.&lt;/p&gt;

&lt;p&gt;His deal with JCPenney was self-described as being “for the American people.”&lt;/p&gt;

&lt;p&gt;And, while noble perhaps, it was not what he had promised.&lt;/p&gt;

&lt;p&gt;&lt;a href=&quot;https://www.hashicorp.com/blog/hashicorp-adopts-business-source-license&quot;&gt;HashiCorp’s decision to change their license&lt;/a&gt; looks like a similar breaking of a promise.&lt;/p&gt;

&lt;p&gt;HashiCorp changed their license from Mozilla Public License v2.0 (MPL 2.0) to Business Source License (BSL). The key result is their software is no longer free if competitors use it to compete with HashiCorp. Regardless of what business sense this makes (or not), the open source community takes the principles of open source software (OSS) seriously. A vocal group has come out in opposition to the change, believing that HashiCorp has profited off OSS only to abandon it once inconvenient.&lt;/p&gt;

&lt;p&gt;Is this a significant misstep from HashiCorp in the vein of Halston?&lt;/p&gt;

&lt;p&gt;Even if this choice is a breaking of the brand promise, it might not matter because of several potential HashiCorp strengths.&lt;/p&gt;

&lt;p&gt;If the cost of switching from HashiCorp’s offerings is too high, it doesn’t matter in the short term if trust is broken. And given the current high cost of capital, the appetite for paying that switching cost is lower than recent past. The high capital cost also means fewer new startups that can choose a different product from the start.&lt;/p&gt;

&lt;p&gt;There is also no 800 lb. gorilla competitor with the resources to easily fork the project into a similar offering. Making the change now makes such an event more unlikely as time goes on.&lt;/p&gt;

&lt;p&gt;Current direct competitors immediately face pressure to either work out a licensing deal or create a fork (HashiCorp’s advantages with a licensing deal are obvious). Not only does creating a fork create an immediate drain on competitors’ resources, but they are all competing with each other and at a significant disadvantage to HashiCorp owning the main branch. Even if HashiCorp’s customers are turned off by the broken brand promise, they’d still need to believe a) another company will figure out a “true” OSS business model, and b) they are able to figure out which company that is up front.&lt;/p&gt;

&lt;p&gt;If we look at past history to determine the wisdom of the decision, Mongodb and Elastic are two companies that are especially interesting comparisons. Both companies share relevant similarities:&lt;/p&gt;

&lt;ul&gt;
  &lt;li&gt;Changed licenses to add restrictions on commercial usage after going public.&lt;/li&gt;
  &lt;li&gt;AWS released a similar product (&lt;a href=&quot;https://techcrunch.com/2019/01/09/aws-gives-open-source-the-middle-finger/&quot;&gt;DocumentDB for Mongodb&lt;/a&gt;, &lt;a href=&quot;https://aws.amazon.com/blogs/opensource/introducing-opensearch/&quot;&gt;OpenSearch for ElasticSearch&lt;/a&gt;).&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;However, their trajectories are significantly different. Mongodb is up ~464% since their license change (Oct 18 - Aug 23) while Elastic is down ~63% since theirs (Jan 21 - Aug 23).&lt;/p&gt;

&lt;figure&gt;
&lt;img src=&quot;/resources/elastic-stock.png&quot; alt=&quot;Elastic Stock&quot; /&gt;
&lt;figcaption&gt;Elastic Stock Price Jan 21 - Aug 23&lt;/figcaption&gt;
&lt;/figure&gt;
&lt;figure&gt;
&lt;img src=&quot;/resources/mongo-stock.png&quot; alt=&quot;Mongodb Stock&quot; /&gt;
&lt;figcaption&gt;Mongodb Stock Price Oct 18 - Aug 23&lt;/figcaption&gt;
&lt;/figure&gt;

&lt;p&gt;If we use stock price as a way to judge the results of their decision, we can say that simply breaking the OSS brand promise is not a significantly bad decision, even when there is a readily available and trusted alternative.&lt;/p&gt;

&lt;p&gt;Given those considerations, I think HashiCorp has made a good business decision. What it means to be OSS is a regular debate, so it’s not even clear that a significant portion of the customer base feels that HashiCorp broke a promise (or that they had any other options). And even if that group is significant, HashiCorp’s positioning appears strong enough to weather the criticism.&lt;/p&gt;

&lt;p&gt;If I were to further explore the decision, here are a few threads I’d follow:&lt;/p&gt;
&lt;ul&gt;
  &lt;li&gt;What were the competitive environments for Mongodb and Elastic, and how similar were they to HashiCorp’s?&lt;/li&gt;
  &lt;li&gt;What other strategic decisions might explain the difference between Mongodb and Elastic, and what similar options are available for HashiCorp?&lt;/li&gt;
  &lt;li&gt;HashiCorp lists other companies that have made similar license changes. Apply similar analysis to those companies to help understand the best HashiCorp strategy moving forward.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;&lt;a href=&quot;https://www.amazon.com/7-Powers-Foundations-Business-Strategy/dp/0998116319&quot;&gt;Powers at play&lt;/a&gt;:&lt;/p&gt;
&lt;ul&gt;
  &lt;li&gt;&lt;strong&gt;Switching Costs&lt;/strong&gt;: Changing infrastructure &lt;em&gt;that is working adequately&lt;/em&gt; is a hard sell, especially given the current high cost of capital.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;Scale Economies&lt;/strong&gt;: HashiCorp has the cost of OSS built into their model. Competitors that were simply building &lt;em&gt;on&lt;/em&gt; OSS now have to add that cost. This is enhanced by the current high cost of capital.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;Branding&lt;/strong&gt;: HashiCorp is negatively affected as mentioned previously, but they also enjoy the benefit of owning the main fork. Customers that choose a branch are introducing uncertainty. HashiCorp might also get a bump in trustworthiness if customers are satisfied that they now have a sustainable business model.&lt;/li&gt;
  &lt;li&gt;&lt;strong&gt;Network Economies&lt;/strong&gt;: The network exists between companies and engineers. The more engineers that use HashiCorp, the easier it is to hire. And the more companies that use HashiCorp, the easier it is to get hired.&lt;/li&gt;
&lt;/ul&gt;

</description>
        <pubDate>Wed, 16 Aug 2023 00:00:00 +0000</pubDate>
        <link>https://constraintlab.com/blog/hashicorp.html</link>
        <guid isPermaLink="true">https://constraintlab.com/blog/hashicorp.html</guid>
        
        <category>hashicorp</category>
        
        <category>branding</category>
        
        
      </item>
    
  </channel>
</rss>
