2,505 karma · joined July 18, 2019
Now the sharpening of a coherent point, challenging one’s assumptions, and editorial decisions of what (not) to include are super important because they’re no longer a byproduct of the writing process.
I also think the reason this works is because it forces exactly what AI makes it convenient to skip:
The hard part of thinking about our assumptions and intentions, challenging them, and resolving them.
I think that stuff is what's often lost with AI. I work in growth nowadays, and it's easy to say "Write me copy for this landing page based on our docs" and it'll write something that contains real features, integrations, and potential benefits.
The hard part isn't finding features and how they could be useful, but to decide which are the 3-4 to spotlight based on what users care about most and how they actually use the product.
Or to realize that yes, there's a set of general "best practices", but we need to overrule some because we're in a niche category, so people will care about some niche details that other products don't need to surface on their home page.
I found the same is true in engineering. We recently rebuilt our Kubernetes operator because one CRD replicated connection and auth for each resource, which led to scalability issues:
-Each resource authenticated independently on restart, and the burst of API calls led to rate limits, which led to latency in getting to steady state.
-Each resource held its own client in memory, which led to OOM issues, which led to restarts.
If you ask AI on how to fix this, it did realize the issue was burst API calls and replication of connection/auth. But its suggestions were that we needed to build event watchers/dither/queues, etc.
This would've solved that specific issue, but not the underlying problem that we had an overloaded CRD.
The real solution was to turn auth/connection into separate CRDs that operate as a lazy, shared cache. This solved the underlying architectural problem, not just the symptoms of it.
Doing this kind of thinking is the important part, and you can't skip it and expect good results. Something like grill-me is a forcing function to do this.
But:
1. As companies grow, customers continue to want more integrations.
If you imagine a massive matrix of companies and all the tools they use, you need to expand the number of rows where your product is a viable solution.
This matters for implementation now (any tool must work with your stack) and for your long-term prospect of using a product since companies expect to migrate, add, or retire tooling.
2. I think the problem with a vibe-codeable plugin platform is “what is a good data model” because that depends principally on what a tool is for.
Let's take project/task management stuff. You can vibe-code a kanban board, task tracker with different statuses, and project folders quite easily.
For my personal use, I can probably build a "better" Jira. I could have an agreeable interface, a data model molded to the types of projects I work on and how I work on them, useful constraints that assuage my weaknesses (e.g. you can't have more than 3 active projects), and a persistent list of things I've shipped to combat imposter syndrome.
That solution would be far too opinionated to scale across even a small team, let alone an enterprise.
Companies buy Jira because it integrates with just about everything, because it's the de-facto standard so you don't need to teach people how to use it, and because you have all the SSO, workflows, approval policies, RBAC, etc. anyone could want.
SaaS vendors don't compete to offer the core features, but the genuinely hard stuff.
I think the issue is not disagreeing with one's boss, but what relationship you have with your boss.
There's a big difference between hostility and wanting to win against the other person vs. trying to find the best possible solution for the team, and clashing on what that should look like.
Some of this is about communication (continually centering on what goal one is trying to reach and validating the other party's perspective) other stuff is basic manners (don't make a public spectacle of your disagreements, especially in front of your boss's boss), but much of it boils down to relationship skills.
I think more people in tech should understand basic relationship skills and how they apply at work. Work is more transactional than a friendship or a marriage, but the core parts still matter: everyone wants to feel seen and heard, not invalidated and attacked.
The rise of knowledge work made many people far less physically active because moving one's body was no longer a given part of one's job. This led to a lot of people (who assumed sports was exercise on top of one's work, not the only source of exercise) moving very little. This meant we needed to rediscover the importance of exercise as a pillar of health.
I think something similar will happen with knowledge work, where we have to do a lot less cognitive exercise due to AI (as well as the decline of reading and rise of short-form video), which will likely lead to eventual issues and subsequently, a rise in activities designed to replicate the cognitive exercise work used to provide.
It’s interesting how ruling out those options of spending my time made me default to reading more.
I also noticed I had a much better attention span for long-form movies (2+ hours used to be hard for paying attention to one thing).
This made a big contribution because vertical short-form video feeds require extreme stimuli to get anyone’s attention - but they add nothing to the actual experience and often detract from it.
This has also led to the absolutely horrific acoustics where even in non-nightclub bars and normal restaurants, you have to yell to understand each other because the decor is made of tile, tables and chairs are at odd angles that increase distance, etc.
Everything now is subordinate to the visual environment because that’s what gets shared on Instagram.
Not saying interior design doesn’t matter, but its point should be to create a great overall experience, not to be visually stimulating at the expense of the rest.
My kitchen knives are decent knives, but no hand-forged Japanese masterpieces. Using them is a joy though, because I have an ingrained understanding of their ergonomics and how they cut.
I remember how clumsily I handled them at first. I take them to a sharpener regularly.
That experience of built-up familiarity puts me into a state of mind where I feel competent and joyful.
This is also true for my mechanical keyboard and some reference books I keep around my desk.
They are not necessarily the best possible tools, but they’re gateways to who I am at my best in different disciplines.
I think a big part of the issue is a certain German presumptuousness.
There’s a general sense that Germany is a prosperous, influential country. The reason for that must be that things are done correctly in Germany.
I think this is an inherited attitude that doesn’t really correspond to reality anymore as systems are crumbling and a trip to many other European countries (including those Germans grew up to view as a barbaric hinterland or as holiday destinations) shows them that even small towns can have fast mobile internet, that you can pay by card at market vendors, and that the government can use computers.
One is that he worked as a middling chef for decades and only had his breakout success in his 40s. I think this made him much different than your generic essayist/journalist, and also let him speak to different kinds of people. He became a symbol that international travel, trying new things, and living life to its fullest doesn’t have to be an early 20s thing followed by decades of monotony.
He was also uniquely principled (he’d finance some episodes himself) and speak to people on all sides politically, without pandering to either.
You may admire the reckless behavior and carefree attitude of the rockstar in part because you’d like to worry a bit less about what other people thought or whether that party will compromise your sleep score.
Not because you’d like to do the exact same thing they do.
You can admire their dedication to the mission, their wealth, ingenuity, etc.
But it’s easy to forget that for many of them, it comes at the expense of seeing their kids, spending time with their spouse, and their own health.
I’d much rather sell a company for $20m and have a happy marriage than be a billionaire stuck in a cycle of divorces and estranged children.
I'm not arguing for zero zoning laws, but I'm saying that much of modernist urban planning primarily benefits the planner.
Is this person actively solving the problem, tried a bunch of approaches and ran into a road block where they're asking for help?
Or are they wishing the problem didn't exist, so they do as little as possible, hoping somebody else will solve the problem for them?
There's a big difference between:
"Hey, I saw this job at [company you work at], could you refer me please? I'm [lists skills and experience]"
and
"Hey I'm thinking of applying to [company you work at] for the product designer position and I want to make an impression, so I'm putting together a demo Figma with a couple of things I'd fix and how. I spotted those when I did the onboarding for your free trial. I'm curious if you could tell me whether [design flaw] is intentional to deter abuse or if that's something I could fix? Totally get if that's confidential"
The part where you're solving the problem instead of hoping someone else will solve it for you, that's much more important then how you word it.
1. Societies are heterogenous, and some groups are always benefitting from the status quo, which will then rebel against any potential changes.
For instance, in Germany (and multiple other European countries), forming a GmbH (LLC) requires in-person sessions with expensive notaries, who will read the entire paperwork out loud.
This made sense when it was invented in 1892 because company formation was extremely rare, always required large facilities and upfront purchases, etc.
Today, this requirement only benefits notaries, as I reckon extremely few people would voluntarily hire a notary to form their company.
But of course notaries would be against making their services legally required! And of course, if anyone suggests loosening requirements, they will spell out all of the terrible fraud that would happen if company formation didn't require a thorough legal review.
That's an example for something relatively obvious. It gets much harder with genuinely complex topics (pensions? tax reform? healthcare?).
2. Many initiatives to improve society fail because they're viewed through a prism of administration/policy, not the actual people impacted.
Take modern urbanism, which cleanly separates where people live (residential areas), where they work (commercial districts), where they shop (malls), and where they spend their free time (recreational centers, etc.)
This beautifully serves people who plan and administer because it makes their work massively easier vs. a tangle of apartment buildings with a coffee shop and clothing boutique on the ground floor and a playground for the kids in front.
It checks the boxes of daily life, but the people living through it feel alienated from their community with these single-purpose urban areas.
3. We have a bias towards actionism.
We tend to think that big, complicated problems need big, complicated solutions, and politicians (who are evaluated on perception, not results) need to be seen designing and implementing those big, complicated solutions.
Massive amounts of infant mortality were prevented by simply making doctors wash their hands, and Semmelweis (who came up with the hypothesis) was so ridiculed for that idea that he died in an insane asylum.
It's sadly often unacceptable to say "we're going to make a little tweak that'll have a ton of downstream effects and solve this massive problem in 5 years".
Right now it's silly to default to frontier models, but it won't bankrupt your company. I believe in the short-medium term future, we'll need to be more deliberate about model choices.
In the long-term, of course, tech costs tend to plummet. Is there a future where in 15 years, my Apple Watch locally runs an Opus 4.8-class model? Maybe. And that would obviate this whole discussion.
1. We're still in the "$5 airport Uber" era of LLMs. They're heavily subsidized, and everyone still complains about costs.
2. There hasn't been a real incentive to work on cost optimization for data centers and the hardware they contain. When/if price hikes happen and send people scrambling to use other models or drastically reduce AI usage, this will suddenly need to happen.
3. We're massively overusing SOTA models. As long as you're on a subsidized subscription, you can use Claude Opus 4.8 high to write blog article meta descriptions. If you paid by token, you wouldn't do that.
4. Open models are a wildcard that could completely change the calculus.
We’re still in the “5$ airport Ubers thanks to VC money” era of AI
-the US is the only place where you can raise the kinds of money you need to run a lab like this.
-a government that won’t let you sell products to customers abroad will probably object even more to you moving abroad.
Even if you made the move abroad, that government might no longer let you access US data centers.
-This basically affects OpenAI and Anthropic, which make the only LLMs most people consider frontier nowadays. Since most open weights models rely on distillation of frontier models, it may genuinely entrench those companies more.
It may be playing into the hands of open source OAI/Anthropic dependencies start to look more dangerous, but it also makes building better OSS models harder.
The advantages the AI labs rely on might be less durable than a proprietary process in industrial manufacturing, but it’s still meaningful.
I think the bigger reckoning will come from a different angle: tokens will eventually need to cover cost.
That will likely mean multiplying prices compared to today. And companies already complain now!
Model orchestration and smaller models that can run locally or cheaply will become more important in my opinion.
Right now, you can still default to GPT/Claude and it’s kind of fine, but that will have to change.
The point of a ratatouille is not that it has precisely bell pepper, eggplant, tomato, and zucchini. It’s that it’s a stew of summer vegetables, and the point should be to figure out what the summer vegetables are for you.
We live in a world where we can buy summer vegetables in January (imported from across the world), so we don’t have to deal with those limitations.
I think it’s a fool’s errand to determine the precisely correct formula for the correct (tm) way of making x, y, or z.
This might be different if you work in a Michelin-starred restaurant or get a kick out of the scientific method.
But if you cook at home, the point is to make delicious things for the people you love.
It's also telling that the most popular videos are about building the most visually striking Cyberdecks and not about building what a Cyberdeck is actually useful for—that's what gets engagement on short-form video platforms.
But I think it's a massively positive thing overall:
-Women, LGBTQ folks, and other underrepresented groups are finding their way to these nerdier hobbies.
-People are getting tired of technology taking over their lives, specifically attention economy and surveillance tech.
-People are learning about electronics and understanding that there are other ways of doing things.
I fail to see the negative in this. Even if none of these cyberdecks are used for practical purposes, someone learned something new. And, even if their cyberdeck gathers dust, being conscious of their tech usage might change how they use their MacBook or the internet more generally.
I think what you're saying is a bit like criticizing someone for not being a self-sustaining farmer because they only grew their own vegetables one summer and then quit.
They may not only eat their own vegetables, but that experience may lead to them buying from farmer's markets vs. Big Food. And that's a net positive.
So I assume this is the non-Nvidia Groq.
Adding a new filter option for in-product analytics (assuming that data is being captured) is something AI can do reasonably well.
But things like “how should our orchestration layer be architected” isn’t a question of execution, so AI won’t be much use.