> and hate how much extra bullshit I didn't ask for they always add to the output.
I can recommend for that problem to make the "jumps" smaller, e.g. "Add a react component for the profile section, just put a placeholder for now" instead of "add a user profile".
With coding LLMs there's a bit of a hidden "zoom" functionality by doing that, which can help calibrating the speed/involvment/thinking you and the LLM does.
Three things I can suggest to try, having struggled with something similiar:
1. Look at it as a completely different discipline, dont consider it leverage for coding - it's it's own thing.
2. Try using it on something you just want to exist, not something you want to build or are interested in understanding.
3. Make the "jumps" smaller. Don't oneshot the project. Do the thinking yourself, and treat it as a junior programmer: "Let's now add react components for the profile section and mount them. Dont wire them up yet" instead of "Build the profile section". This also helps finding the right speed so that you can keep up with what's happening in the codebase
> Try using it on something you just want to exist, not something you want to build or are interested in understanding.
I don't get any enjoyment from "building something without understanding" — what would I learn from such a thing? How could I trust it to be secure or to not fall over when i enter a weird character? How can I trust something I do not understand or have not read the foundations of? Furthermore, why would I consider myself to have built it?
When I enter a building, I know that an engineer with a degree, or even a team of them, have meticulously built this building taking into account the material stresses of the ground, the fault lines, the stresses of the materials of construction, the wear amounts, etc.
When I make a program, I do the same thing. Either I make something for understanding, OR I make something robust to be used. I want to trust the software I'm using to not contain weird bugs that are difficult to find, as best as I can ensure that. I want to ensure that the code is clean, because code is communication, and communication is an art form — so my code should be clean, readable, and communicative about the concepts that I use to build the thing. LLMs do not assure me of any of this, and the actively hamstring the communication aspect.
Finally, as someone surrounded by artists, who has made art herself, the "doing of it" has been drilled into me as the "making". I don't get the enjoyment of making something, because I wouldn't have made it! You can commission a painting from an artist, but it is hubris to point at a painting you bought or commissioned and go "I made that". But somehow it is acceptable to do this for LLMs. That is a baffling mindset to me!
>I don't get any enjoyment from "building something without understanding" — what would I learn from such a thing? How could I trust it to be secure or to not fall over when i enter a weird character? How can I trust something I do not understand or have not read the foundations of? Furthermore, why would I consider myself to have built it?
All of these questions are irrelevant if the objective is 'get this thing working'.
You seem to read a lot into what I wrote, so let me phrase it differently.
These are ways I'd suggest to approach working with LLMs if you enjoy building software, and are trying to find out how it can fit into your workflow.
If this isnt you, these suggestions probably wont work.
> I don't get any enjoyment from "building something without understanding".
That's not what I said. It's about your primary goal. Are you trying to learn technology xyz, and found a project so you can apply it vs you want a solution to your problem, and nothing exists, so you're building it.
What's really important is that wether you understand in the end what the LLM has written or not is 100% your decision.
You can be fully hands off, or you can be involved in every step.
> You can commission a painting from an artist, but it is hubris to point at a painting you bought or commissioned and go "I made that". But somehow it is acceptable to do this for LLMs. That is a baffling mindset to me!
The majority of the work on a lot of famous masterpieces of art was done by apprentices. Under the instruction of a master, but still. No different than someone coming up with a composition, and having AI do a first pass, then going in with photoshop and manually painting over the inadequate parts. Yet people will knob gobble renaissance artists and talk about lynching AI artists.
I've heard this analogy regurgitated multiple times now, and I wish people would not.
It's true that many master artists had workshops with apprenticeships. Because they were a trade.
By the time you were helping to paint portraits, you'd spent maybe a decade learning techniques and skill and doing the unimportant parts and working your way up from there.
It wasn't a half-assed, slop some paint around and let the master come fix it later. The people doing things like portrait work or copies of works were highly skilled and experienced.
Typing "an army of Garfields storming the beach at Normandy" into a website is not the same.
Anti-AI art folks don't care if you photobashed bits of AI composition and then totally painted over it in your own hand, the fact that AI was involved makes it dirty, evil, nasty, sinful and bad. Full stop. Anti-AI writing agents don't care if every word in a manuscript was human written, if you asked AI a question while writing it suddenly you're darth fucking vader.
The correct comparison for some jackass who just prompts something, then runs around calling it art is to a pre-schooler that scribbles blobs of indistinct color on a page, then calls it art. Compare apples to apples.
That's not what a strawman is lol. Me saying the analogy sucks is just criticism.
If you feel judged about using AI, then your choices are (1) don't use it or (2) don't tell people you use it or (3) stop caring what other people think.
Have the courage of your own convictions and own your own actions.
Lately I've been interesting in biosignals, biofeedback and biosynchronization.
I've been really frustrated with the state of Heart Rate Variability (HRV) research and HRV apps, particularly those that claim to be "biofeedback" but are really just guided breathing exercises by people who seem to have the lights on and nobody home. [1]
I could have spent a lot of time reading the docs to understand the Web Bluetooth API and facing up to the stress that getting anything with Bluetooth working with a PC is super hit and miss so estimating the time I'd expect a high risk of spending hours rebooting my computer and otherwise futzing around to debug connection problems.
Although it's supposedly really easy to do this with the Web Bluetooth API I amazingly couldn't find any examples which made all the more apprehensive that there was some reason it doesn't work. [2]
As it was Junie coded me a simple webapp that pulled R-R intervals from my Polar H10 heart rate monitor in 20 minutes and it worked the first time. And in a few days, I've already got an HRV demo app that is superior to the commercial ones in numerous ways... And I understand how it works 100%.
I wouldn't call it vibe coding because I had my feet on the ground the whole time.
[1] for instance I am used to doing meditation practices with my eyes closed and not holding a 'freakin phone in my hand. why they expect me to look at a phone to pace my breathing when it could talk to be or beep at me is beyond me. for that matter why they try to estimate respiration by looking at my face when they could get if off the accelerometer if i put in on my chest when i am lying down is also beyond me.
[2] let's see, people don't think anything is meaningful if it doesn't involve an app, nobody's gotten a grant to do biofeedback research since 1979 so the last grad student to take a class on the subject is retiring right about now...
>When I enter a building, I know that an engineer with a degree, or even a team of them, have meticulously built this building taking into account the material stresses of the ground, the fault lines, the stresses of the materials of construction, the wear amounts, etc.
You can bet that "AI" is coming for this too. The lawsuits that will result when buildings crumble and kill people because an LLM "hallucinated" will be tragic, but maybe we'll learn from it. But we probably won't.
Have you heard of the Horizon IT Post Office Scandal[0]?
> Between 1999 and 2015, more than 900 subpostmasters were wrongfully convicted of theft, fraud and false accounting based on faulty Horizon data, with about 700 of these prosecutions carried out by the Post Office. Other subpostmasters were prosecuted but not convicted, forced to cover illusory shortfalls caused by Horizon with their own money, or had their contracts terminated.
>
> Although many subpostmasters had reported problems with the new software, and Fujitsu was aware that Horizon contained software bugs as early as 1999, the Post Office insisted that Horizon was robust and failed to disclose knowledge of the faults in the system during criminal and civil cases.
(content warning for the article about that for suicide)
Now think of places where LLMs are being deployed:
- accountancy[1][2]
- management systems similar to Horizon IT
- medical workers using it to pass their coursework (A friend of mine is doing a nursing degree in the USA and they are encouraged to use Gemini, and she's already seen someone on the same course use it to complete their medical ethics homework...)
- Ordinary people checking drug interactions[3], learning about pickling (and almost getting botulism), talking to LLMs and getting poisoned by bromide[4]
I build a lot of custom tools, things with like a couple of users. I get a lot of personal satisfaction writing that code.
I think comments on YouTube like "anyone still here in $CURRENT_YEAR" are low effort noise, I don't care about learning how to write a web extension (web work is my day job) so I got Claude to write one for me. I don't care who wrote it, I just wanted it to exist.
I’ve wanted a good markdown editor with automatic synchronization. I used to used inkdrop. Which I stopped using when the developer/owner raised the price to $120/year.
In a couple hours with Claude code, I built a replacement that does everything I want, exactly the way I want. Plus, it integrates native AI chat to create/manage/refine notes and ideas, and it plugs into a knowledge RAG system that I also built using Claude code.
What more could I ask for? This is a tool I wanted for a long time but never wanted to spend the dozens of hours dealing with the various pieces of tech I simply don’t care about long-term.
Wireguard is _really_ simple in that sense though. If you're not doing anything complicated it's very easy to set up & maintain, and basically just works.
You can also buy quite a few routers now that have it built in, so you literally just tick a checkbox, then scan a QR code/copy a file to each client device, done.
My ISP-provided router (Free, in France) has WG built-in. But other than performance being abysmal, its main pain point is not supporting subnet routing.
So if all you want is to connect your phone / laptop while away to the local home network, it's fine. If you want to run a tunnel between two locations with multiple IPs on the remote side, you're SoL.
Strong disagree. For most parts travelling is a non-event these days.
A train that splits, on the way to the airport where there will be a lot of non-german speaking people, and for some reason only shows it on the platform is insane.
Having a train that splits on that route is already bad enough, but you HAVE to emphasize it on the train.
I know that I need to pay attention to this, because I've grown up with DB pulling all sorts of fucked up shit, but we should not accept that this is reasonable.
You get repeated hints like 20 times outside and inside the train. Announcements (also in english, recorded, by an english native speaker, repeatedly), the train displays explain it and when you get to the station where the train splits, every display in that train shows you whether you're in the right carriage and you get an extra announcement exclusive to the carriages that go somewhere else that you should change now if you want to go to the airport.
From the top of my head I know three cities which have peculiarities when it comes to public transportation to the Airport. In two cases, it's obvious they do this to push the private train to foreigners, at 5x the ticket rate.
This sort of used to be the case with Heathrow Express in London. There was a lot of signage that suggested to the unwary that Heathrow Express was the "right" way to get into London. Now, especially with the Elizabeth Line, while you can save a few minutes with Heathrow Express, that's really not a cost-effective alternative for a lot of people. (And Piccadilly may be a better option depending on your luggage and where you are staying.)
Stockholm (SL Bus/Train via Märsta 47kr vs. Arlanda Express 340 SEK)
Vienna (S-Bahn S7 4.40 EUR vs. City Airport Train 24.90 EUR)
While with the Stockholm one, the public transport option is cheap but a little bit more complicated (there are convenient medium priced options too), the Vienna one is really just branding and a non-obvious exit to the train station.
Ah, Stockholm. Been there as a student, went to Uppsala the day I was flying from Arlanda, and the only trains to the airport were the insanely expensive Arlanda Expresses. Had to take the normal train to Märsta, then walk the five kilometers to the airport. Fun times.
London for sure. All three of the Stanstead, Gatwick and Heathrow Express services are an absolute rip off compared with the alternatives that don’t take much longer.
Many people use them out of ignorance, expense accounts or they have the disposable income not to care.
Seems like there's a few abstractions mixed up, the problems have nothing to do with SSE.
You can store the state in the SSE connection and have the problems described, and if you don't like those, you can move thr state to something distributed/persisted.
Pubsub is just a layer on top of SSE or websockets, cause guess how it'd end up sending things to the browser
Yeah I didn't really get that... PubSub is more of a design pattern... you still have to get the data transported to the browser (via WebSockets, SSE, etc.)
They are definitely a PE firm. They buy up struggling companies with the aim to revitalise them, or otherwise recoup the cost of investment+ profit. They have switched to mainly relying on traditional debt rather than outside investor money recently but that doesn't make them not PE.
In fact this is much like the older form of PE, where efficiency gains were the main objective.
Bigger PE firms now usually focus on roll-up strategies (buy loads of similar companies and merge, say car washes is big right now for example, as well as dental, vet and family doctor/GP practices) as well as utilising bucket loads of leverage to amplify gains. This does not however make what bending spoons is doing not PE.
But PE firms don't have their own workforce, Bending Spoon does, which is why their model differs from, say, Apollo.
The fact they use some of the same tools doesn't mean they are doing the same thing. The majority of Blending Spoon's employees are devs, not finance people.
"They buy up struggling companies with the aim to revitalise them, or otherwise recoup the cost of investment+ profit."
1) Nope, they are focused on taking advantage of customer lock-in to raise prices, while reducing operating expenses to increase cash flows. There may be some initial reinvestment to increase surplus of its users, before raising prices substantially.
2) "recoup the cost of investment+ profit"? Yeah lets see if that pans out. The acquisition price is assumed to be under a going-concern basis in perpetuity, if they muck things up with the choices they make the acquisitions have a limited life to increase and capture those cash flows to deliver a positive NPV investment. The demand for the firms products are not perfectly inelastic w.r.t to price.
I can recommend for that problem to make the "jumps" smaller, e.g. "Add a react component for the profile section, just put a placeholder for now" instead of "add a user profile".
With coding LLMs there's a bit of a hidden "zoom" functionality by doing that, which can help calibrating the speed/involvment/thinking you and the LLM does.
reply