Escapado

For the past two years I have worked on a SaaS platform for the education space that we are about to launch (small 2 person start up) - it involves a lot of highly dynamic forms and fairly complex client side state and tons of rather complex components and workflows.

The stack for non-specifics we landed on for now coincides with a lot of what is mentioned in the blog post with a few extras:

Nextjs, oRPC, Next-Intl, Zod, Tanstack Form, Tanstack Query, Drizzle, SQLite, Tiptap, Better-Auth, Jest, Tailwind, a heavily customized Shadcn.

However there are a few more things I would like to mention.

Formatting: Oxfmt

Linting: Oxlint

E-Mails: react-email

Icon Library: Lucide

User HTML in React: Interweave

Date Handling: date-fns

Drag and Drop Handling: DnD Kit

Payment (MoR): Creem (we are EU based, so a EU MoR matters to us)

The blog post creates a good overview of technologies that exist out and some pros and cons but it's not so clear to me why the things that are recommended are recommended except for: Companies use them.

We went through multiple rewrites of large parts of the codebase, originally we started out using supabase for the backend, then moving to drizzle with self hosted postgres, then switching to sqlite. We started with react-hook-form and switched to Tanstack Form. We had our own UI library, then switched to Shadcn and customized that heavily. We used next-auth (now auth.js) and switched to better-auth, used server actions, then switched to oRPC.

Most of the decisions we made were either grounded in a very specific technical aspect related to our product or just out of curiousity of whether the grass would be greener on the other side (it was sometimes, not always).

In my normal day job I have also mostly worked on Nextjs apps for the past few years, eCommerce, internal dashboards and tools etc and I guess my takeaway is that most of these tools can get you where you want but depending on the shape of your problem(s), some lend themselves better than others and you will probably only find out if you try it out. If you are building a highly complex app, you will inevitably run into situations where all those libraries and frameworks will not provide you with the things you want or will expose primitives that make it cumbersome to work with. I will give one recent example:

In our application we worry a lot about potential loss of form data of dynamically created forms, so we need some sync solution for said form data and have some form of persistence. The all-encompassing solution for us would involve: Local sync to indexed db, including tab sync, websocket sync of change increments (not the entire form, they are quite large), retry on error / connection loss, ideally some state merging logic for the form state in case of divergence with last writer wins fallback, granular form validation with different representations of data for draft and submit states, atomic re-renders (e.g. save button gets disabled when sync is in flight but no full re-renders when the form object changes) and different representations of the form data (image URLs from the server snapshot look different than base64 representations in the form for example). This is not trivial. We looked at different LoFi solutions (Tinybase, different CRDT approaches, Zero, SignalDB) with SignalDB being the closest one to what we would like. Eventually we had to ditch that too and implement a lesser version of all of this for our v1 that syncs the entire exam state periodically over oRPC (+ debounced local storage) and a pretty massive hook that abstracts the sync logic over various forms and form builders we have to handle all sorts of edge cases. Tanstack form is not meant for this (dirty tracking does not line up if the "default state" changes), there is no concept of drafts and final forms, there are no storage adapters that do conflict resolution the way you might want because that is so specific to your app.

Another point I would like to make is that, while modern react stacks do afford the developer a pretty nice UX 90% of the time, sometimes that comes with some pretty annoying tradeoffs. Nextjs is quite heavy and slow. No cron jobs when self-hosting. Having your db schema defined in TS (using drizzle), tacking some extra Zod on it at the client/server boundary and carry it through into the frontend and eventually in Tanstack form is very nice on paper but for dynamic forms with arrays and more complex data types the TypeScript LSP will absolutely shit the bed, even TSGO can't keep up or outright crashes regularly for me. Tanstack form type inference for form fields is also lacking for complex array types for us. Using SQLite is great because no separate process is needed to host the DB, but you don't get CDC or realtime updates easily (maybe turso will help here in the future). Using shadcn is really cool to get up and running and build things that look generally professional and nice, but once you modify the files (which you will need to in order to keep up with the demands of the app) the upgrade path becomes painful and the shadcn dependency list and amount of DOM it sometimes produces of it is also ridiculous.

All that is to say, I think the difficult part of engineering even in 2026 happens at the outer edges of what you build and no framework or stack is gonna save you from all of that (and probably shouldn't!).

I agree with the sentiment but I wonder if a sufficiently large amount of sufficiently sophisticated benchmarks existed then I would be surprised if a model would only memorize those benchmarks while showing terrible real world performance. We are not there yet but maybe one day we will be.

"entirely possible" is a bit of a stretch. You can install some hacky WM and sketchybar but system settings, workspace three finger scroll speed, the finder app, window chrome, login screen etc are not something that can easily be changed. And the default apps are really not that great for power users. Calendar, Mail and Finder are all slow, dumbed down and only very superficially customizable. I daily drive an M2 Pro MBP and I was running a Linux desktop up until 2 years ago and I felt like there were barely any limits to customizing the latter while I have to fight macOS at every step if I want to do something that apple does not want me to do.

Is there a good reason browsers could and or should not support ts out of the box? I think even node does to some extend.

Anyway another thing that kept bugging me when reading this is that web apps and development and performance constraints can be so vastly different that I am not sure this pattern always holds up. What if you have an app that depends on a ton of global state, what if you need web sockets, what if you need a highly dynamic form with partial validation derived from a schema that your drizzle definition spat out. And 2 junior devs that benefit from end 2 end type safety and working in Typescript alone to be productive. Add syncing across tabs, offline first requirements, multiplayer rich text editing, sophisticated auth, caching, accessibility constraints, end 2 end monitoring and logging with open telemetry, animated page transitions scroll timeline animations, pdf generation of certain views and a design team breathing down your neck that has a thousand different extra bells and whistles it wants that are impossible to represent and maintain with tailwind alone. And then the marketing department comes around and wants to track everything, inject their A/B testing framework, wants dynamically updated translations without forgoing SSR and SEO benefits and all of that has to run on a 8€/month Hetzner VPS.

I understand that if you work on software that only needs a handful of those things you can get away with a better stack than what the author seems to loathe. But it you work on a large enough application, then you can’t escape complexity, even if you find some of it remedied by a well suited tech stack.

I would say you are both right in that if you have two competing variables (on-time for the defence vs calorie consumption), when the main causes of death before procreating were infectious disease and malnutrition before modern times, I would expect some equilibrium to be reached and we have not had that much time to evolve since caloric scarcity in the western world was a solved problem for large swaths of the population.

If in the future we could trade a few hundred extra calories per day for a great immune system (without auto-immune side effects) we would have found a nice cheat code!

This was interesting. Had a 5 minute chat with the outsider from the dishonored series. Just a one sentence prompt and its phrasing was at least 60% there, but less cold and nicer in a sense than the video game counterpart. Still an interesting experiment. But I also know that maybe 12-24 months down the line, once this is available in real time on device there will be an ungodly amount of smut coming from this.

My clients are. I do web dev for a living and I use these frameworks day in and day out. It’s not even that I dislike the dev ex on most of them and I’ve seen a lot of good code and bad code and I don’t even wanna blame anyone in particular for the situation we are in. I think my comment was more of a dig at the world we live in than anything else.

I hate the state of affairs. That said my guess is what we „gained“ is tons of telemetry, tracking and the likes, engineers not needing to think about performance to get a feature out, which absolutely lowers the bar to entry, high level abstractions and ux and visual bells and whistles of varying importance and quality (infinite scrolling, streaming updates, image blend modes, blur effects, scroll timeline animations etc). People creating Pokémon had to think about every bit in their texture atlas and carefully manage the hardware memory manually. Web devs now try not to forget to clean up event listeners in a useEffect that triggers on mouse move to generate data for an interaction heatmap for the marketing department while 25mb of 3rd party scripts make sure every data broker and their mother is well informed about your digital whereabouts.

So it is just the right wing neoliberal playbook then. Protect the rich and put everyone else against each other so we don’t focus on them. I want this to stop.

First off, the following is not meant to combative but I think this confuses me. If there was no cliff there was nothing to do administratively on that front, no new checking at all. We already have progressive tax systems in many countries. Adding 1% at the top end can’t be that difficult. My health insurance (in Germany) raises prices every year and most people don’t have a choice there either. Property taxes increase all the time. If we had a wealth tax and a higher (at the top) progressive capital gains tax it seems to me that the pitch would be politically even simpler: there are 5% that will pay a little more percentage wise from now on while retaining vast amounts of wealth and 95% that will pay less or much less. Genuinely I have heard that we can’t raise taxes on rich people because they will evade them but it also sounds like a lie repeated so often that we just take it as the truth. Didn’t a lot of countries or US states have higher marginal tax rates without seeing mass exodus of millionaires? Can’t lawmakers focus on plugging the loopholes rich people use? I mean our government is currently trying to go out of its way to make sure that unemployment benefits are only paid out to people who „really deserve them“ by tightening the rules around that and the political debates I see put incredible amounts of emphasis on „fairness“ and that we „have to do something about those who just profit off the system without contributing“ when it’s about that topic. The fervour is clearly not applied symmetrically.

To me it appears as though the success of the right wing politics everywhere is that they made socioeconomically disadvantaged people identify other socioeconomically disadvantaged people and the middle-class as the cause of their suffering while somehow becoming sympathetic to the uber rich in hopes to one day belong. And to me it’s clear that if we taxed wealth and high incomes fairly and removed the loopholes to level the playing field we would not even need these discussions to begin with because we simply had a well financed social society and the rich would still be rich, but maybe not so obscenely so.

I am also confused by cliffs. Maybe someone more knowledgeable than me could explain why you would ever want them for something like this instead of just having higher progressive tax rates for well off people to make up for it. Naively I would think that that’s significantly easier from and administrative point of view too.

Genuine question: How does it happen that a heart surgery costs 100k? 2 surgeons (200$/h) + 6 nurses(100$/h) for 10 hours would be 10k. Where do the other 100k come from? Is it the equipment cost? Consumables? After care? Or are the margins just ridiculous?

Interesting story. I want to agree with the general advice not to use it for that - especially if that is how you use it. And I want to preface this with: Don’t take this as advice, I just want to share my experience here. I tend to do it anyway and had fairly large success so far but I use the LLM differently if I have a health issue that bothers me. First I open Gemini, Claude and ChatGPT in their latest, highest thinking budget installment. Then I tell them about my symptoms I give a fairly detailed description of my person and my medical history. I prompt them specifically to ask detailed questions like a physician would and ask them to ask me to perform tests to rule out or zoom in on different hypothesis about what is might have. After going back and forth, if they all agree on a similar thing or a set of similar things I usually take this as a good sign I might be on the right track and check if I should talk to a professional or not (edging on the side of caution). If they can’t agree I would usually try to get an appointment to see a professional and try to get sooner rather than later if anything potentially dangerous popped up during the back and forth or if I feel sufficiently bad.

Now, I live in Germany where in the last 20 years our healthcare system has fallen victim to neoliberal capitalism and since I am publicly insured by choice I often have to wait for weeks to see a specialist so more often than not LLMs have helped me stay calm and help myself as best as I can. However I still view the output less as a the output or a medical professional and try to stay skeptic along the way. I feel like the augment my guesswork and judgement, but not replace it.

I keep hearing this but sometimes I am not 100% sure if they are _much_ better so asking honestly: Is there any reputable quantitative analysis of this in the context of language learning?

For example: I have spent the last two years in japan (I am in my 30s) and just got back to my home country. Went to a language school in the mornings there, immersed myself in the language a little but did not go all out on studying at home except for some Anki and the homework we got. I would spend 1 or 2 evenings per week talking to japanese people in my apartment building for practice. I just took the N2 exam before I left and just failed by 1 point, without any extra studying specifically for it. I could have conversations with people in my apartment complex, make phone calls to get stuff done and get the gist of most news I heard if they were not hyper-specific and I can read easy novels. If I open the NHK news website I am still lost on a bunch of stuff and have to look up a lot. But again, that was 2 years and I was neither particularly good nor bad compared to the other fellow students and I did not go all out full immersion - lots of my interactions were still with foreigners in the afternoon. Anyway, I for sure know more kanji than a 2nd grade elementary school student. I also can say more than a two year old kid. I know of course children learn to navigate a language without explicit study in their first years of life but the point still stands. If time spent studying was equal, how much of a difference remains?

Naive question: Are the current (from what I have heard not very effective) export restrictions of HPC GPUs to china truly productive in the long run if the goal is to retain an edge? As in, to me it seems that it just fuels an expansion of domestic capabilities and in the car and solar sector my impression is that china had already proven that it can absolutely perform on par or even better in many different metrics compared to western countries, given time and pressure. So while these chips are not on par with current or even last gen GPUs, I would not be surprised if china would catch up and even have a much higher incentive to do so, now that other countries try to control their access to key technologies.

I am not saying whether retaining an edge is good or bad or that I have a different answer if one thought it was good. Just curious what you guys think.

I have to second this. I study Japanese myself and the entire way the Japanese communicate is reflected so deeply in the language. There is so so much nuance to pretty much every sentence they speak and there are certain grammar points that carry more meaning in three syllables than what can be expressed in English or German in a full sentence. And ok turn this way of communicating shapes their culture too I believe. If I were to translate a German conversation into Japanese, even if I did so idiomatically it would most likely come off as a rude exchange, because of all the unapologetic directness in the source language.

Now compared to Berlin I think Hamburg is still pretty conservative and I am not in a position to make apt comparisons to cities in the US but I do have to disagree with the statement that Hamburg is more conservative.

I was born and raised in Hamburg and lived there or in adjacent parts most of my life. I also visited Munich quite a few times due to a long distance relationship and I would disagree that Hamburg is more conservative. tThe people in Munich vote for conservative parties at a greater rate than the people in Hamburg and Munich never felt even remotely as multi-cultural as Hamburg. I distinctly remember walking around München for the first time and being surprised by people’s reactions to seeing a black guy walking down the street. Some people would literally stop walking and stare. Almost no Middle Eastern people either in comparison. There is also a pretty strong divide between the north being much less religious. And one might argue that the people who are Christians are more often Protestant in the north which is arguably more progressive than the catholics in the south. If you look at Hamburg during may 1st, consider the Rote Flora building and the Schanzenviertel I think it’s quite clear that Hamburg has a pretty firmly established left-wing community. Granted if you go to Blankenese or the Neue Hafencity (areas for and of the wealthy) and talk to the people living there you might get a different picture. Anyways talking in averages I am not convinced your statement holds true today.

I think there is sort of a cultural rivalry where people from the north don’t want to get confused with the people from the south of Germany and vice versa. We make fun of their way they butcher the language and their festivities and traditional attire, and how they talk too much, and they make fun of us for being tight lipped humorless pricks.

N=1 datapoint here. I studied physics in university and before I started I was not aware that physics is basically just math where the results sometimes relate to reality. The pure math courses I took were the most difficult and in the beginning I loathed them, because it felt so unattainable to get any intuition, let alone real proper comprehension for all the concepts they threw at us. For a long time I felt like I was just hanging on by threads and especially if I compared myself to those who had some innate interest in math or generally some really good intuition on the abstract concepts (or even prior knowledge) it was really demotivating. But I also felt like I had no choice but to continue and as time went on the I grew fond of it. And the feeling of being overwhelmed changed - that is to say I still was completely lost every time a new topic was breached and I could not understand even half of the proofs in class - but I did not feel so defeated about it. And I grew to like the feeling of actually completing the work sheets they gave us every week. The process of solving them was often excruciating but if you did the sense of accomplishment is real. I think for most people higher math is really difficult and that is part of why it is interesting. Another aspect I had to accept over time is that even though you can state a mathematical fact or conjecture in just a hand full of symbols or a plain sentence it does not mean that truly understand it, its implications or how you got there can be understood the same way that other prose can be. Sometimes you have to stare at, contemplate and scribble around one equation for days until you understand whats up.

If there was any advice I would give, then it's probably similar advice on how to stop procrastinating on anything that is difficult. Establish a routine first - find a spot that you will only use for studying this (like a spot in a library), start small, divide and conquer, accept that you will not understand most things easily, reward yourself for the small wins along the way, find an accountability partner or someone to study with if that's your thing, make a regular schedule with regular times where this is what you do - consistency is key, even if its just for 5 minutes, stack it onto other habits, see yourself as a scholar of math - it is what you do, lean into the discomfort, as enduring that is a valuable skill in itself.

Yup, maintained an e-commerce site where the products were coming from a third party api and the products often had 200+ properties and we often needed certain combinations of them to be present to display them. We created schemas for all of them and also had to transform the data quite a bit and used union types extensively, so when displaying a product list with hundreds of these products, Zod would take some time(400+ ms) for parsing through that. Valibot took about 50ms. And the editor performance was also noticeably worse with Zod, taking up to three seconds for code completion suggestions to pop up or type inference to complete - but truth be told valibot was not significantly better here at the time.

I agree though, that filling your website with tracking crap is a stupid idea as well.

I am currently living in Tokyo and for the most part smoking is prohibited except for designated smoking areas. And while compared to Germany way less people seem to smoke and mostly follow the rules I see about one or two people during my commute every day who will vape or smoke either while walking or standing on the corner of a non-busy street. So not everyone cares but luckily most people do and I wish it was like this everywhere!

Thanks for pointing me to that. One thing that stands out about that argument though is that voting is already discriminatory, right? Permanent residents and minors are not allowed to vote (the latter because we take age as a proxy of competency, no?), despite facing the consequences of elections just as anyone else does. I do understand that a risk for misuse absolutely exists, but at the same time it looks like populism, social media abuse, smear campaigns, science denial and plain old corruption in sheep's clothing are rampant enough that we can agree that many many votes are cast by misled people, who would have made another choice if they really understood what they voted for. I guess it would boil down to the difficult question of which harm is greater.

That's an interesting perspective, but I wonder if we can't have both.

I see. In a sense we are already doing that. Minors can not vote (and if I am correct the reasoning is that they don't have the competency to cast a proper vote) and even foreign permanent residents can't either, even though the outcome of the elections totally influences their lives. In a sense these not-allowed groups are already totally dependent on the people that are allowed to vote.

I guess my argument boils down to: We already discriminate. My thoughts are that the way we do it is not optimal.

Honest question/thought experiment: if we only elected people who are qualified for their job (assume we can measure competence at least in some dimensions like we do for a myriad of other professions before we allow people to work in them) and if the election process was set up in a way where when casting your ballot you have to take a multiple choice quiz which tests for basic knowledge on what you will vote for and the country you’re in (as in “what is the household budget roughly, is this candidate in favour or against x, did the crime rate increase or decrease nominally” take these as rough examples of what I mean), to ensure that the people who vote for something have some clue what they are voting for and the broader context it’s embedded in (we require a license to drive a car, this would be akin to have a having a license to vote) would that remedy the situation a little? The idea would be that informed people would vote for informed people. Could you imagine this being a net benefit or not? I would assume it would make democracies significantly better than they are now. Imagine going to a doctors office to find out your doctor is a Plummer and he was voted into this job and that the people working for him and handling your prescription is a random assortment of people he seems to like.

Agreed and I would add to the list the ungodly amount of hours they work. I know per hour they might not be as productive (or simply held back by antiquated processes within their companies) but they seem to make up for it in part by working so much overtime.

I have not investigated this as much as I should have but if I remember correctly there were cases where the NN approach was yielding smaller solutions. Would be a great follow up to the thesis.

The way this was turned into an optimization problem was to assume a NN to input an identity matrix and then have a custom layer in there to generate S(2) unitaries (exponential form) for which the phase parameters are then the learned parameters in the NN. Similarly for global XX gates. Then from this the final unitary can be computed and compared against the desired unitary and a loss function can be derived. I remember it was a little fiddly to implement these custom layers in tensorflow since many of the functions didn’t work for imaginary numbers. But yeah in short a circuit structure was assumed (alternating between a global XX and single qubit operations) for which the phases of their generating matrices were the learned parameters of the network. Then multiple topologies (how many steps in the QC) could be validated. I think the coolest result was that NNs consistently outperformed other optimisation strategies to learn those parameters.

I didn’t include this in my thesis but from what I remember looking at hundreds if not thousands of matrices and their QC solutions some symmetries would immediately make it so that way less gates would be needed but it could also be deceptive and completely depended on which gate set you have available. If you take the matrix for the quantum furier transform for example with a gate set of phase gates and a hadamard gate then for the 100% solution you n hadamard gates and n! phase gates for an n qubit circuit even though the matrix is highly symmetrical. If your gate set was Clifford + toffli you would be able to do it with n*log(n). And then depending on how close you wanna approximate it you could get away with even less. But I have not gone into further analysis on which symmetries would have which effect on which gate set and whether it would influence predictivity. But it would be fun to investigate for sure!

University of Hamburg in Germany.

My master thesis was on using machine learning techniques to synthesise quantum circuits. Since any operation on a QC can be represented as a unitary matrix my research topic was that, using ML, given a set of gates, how many and in what arrangement of them you could generate or at least approximate this matrix. Another aspect was, given a unitary matrix, could a neural network predict a number of gates needed to simulate that matrix as a QC and thereby give us a measure of complexity. It was a lot of fun to test different algorithms from genetic algorithms to neural network architectures. Back then NNs were a lot smaller and I trained them mostly on one GPU and since the matrices get exponentially bigger with the amount of qbits in the circuit it was only possible for me to investigate small circuits with less than a dozen qubits but it was still nice to see that in principle this worked quite well.