Gaslight-driven development (tonsky.me)

>Well, now there is a new way to serve our silicon overlords. LLMs started to have opinions on how your API should look

we have code review by LLM. There is no point or a way to argue. Just submit to the wishes of the overlord, resistance is futile.

I like seeing what users are currently viewing the same page, but man the constant jostling of users coming and going made it hard to read the post.

Related: https://news.ycombinator.com/item?id=44491071

Yes, but please separate plannning from reviewing, let alone real coding.

> We see the same at Instant: for example, we used tx.update for both inserting and updating entities, but LLMs kept writing tx.create instead. Guess what: we now have tx.create, too.

Good. Think of all the dev hours that must’ve been wasted by humans who were confused by this too.

If it were somehow a human that was consistently and confidently handing out made up programming advice about one's products, would companies still respond by just adding whatever imagined feature and writing a vaguely bemused blog post about it?

> for example, we used tx.update for both inserting and updating entities, but LLMs kept writing tx.create instead. Guess what: we now have tx.create, too.

If a function can both insert and update, it should be called "put". Using "update" is misleading.

I aint reading all this with the animated live user bar at the bottom.

I don't agree with the thesis of this post. It is begging the question of if we have to do what computers want.

> Millions of people create accounts, confirm emails, ... not because they particularly want to or even need to.

These were design choices made by humans, not computers.

To those that still believe that a bunch of data loaded into memory, where data can be anything from a scientific article to a message between two lovers, getting triggered for an output with input and a basic for loop can represent anything intelligence, i have some bad news for you like damn ya'll don't you know git(hub) & huggingface? Ofcourse the drawback of that is that you are not contributing to AGI KEK!

This feels like the beginning of a wonderful friendship between me and the LLMS. I work as a fractional CTO. One of the things that frustrate me is when my clients have various idiosyncratic naming conventions on things, eg there’s a ”dev” and a ”prod” environment on AWS, but then there’s a ”test” and ”production” environment in Expo. It just needlessly consumes brain cycles, especially when you’re working with multiple clients. I guess it’s the same for the LLMs, just on a massive scale.

In general I think it’s great whenever some weight / synapse strength bits can be reallocated from idiosyncratic API naming / behavior towards real semantics.

Recently i had an interesting chat with my team around coding principles of the future.

I think the way people will write code will not be around following solid principles or making sure your cyclometric complexity is high or low, nor it would be about is your code readable or not.

I think future coding principles would be around whether your agentic ide can index it well to become context aware, does it fix into the context window or not. It will be around the model you use and thr code it can generate. We will index on maintainability of the code, as code will become disposable as rate of change will increase dramatically. It will be around whether your vibed prompts matches the code thats already generated to reach some accuracy or generate enough serendipity.

I had to use the reader mode to be able to read this article

Have fun! I'm going to keep trying to build new things that are more excellent and original than the median training data.

Is there a general name and framing we could apply to these “AI” that is equally as accurate but sheds all of the human biases associated with the terms?

Like… it’s just a really, really, really good autocomplete and sometimes I find thinking of it that way cleans up my whole mental model for its use.

Maybe it's spite-driven development, but I'd love to hear about someone who, upon learning that LLMs are suggesting endpoints in their API that don't exist, implements them specifically to respond with a status code[0] of "421: Misdirected Request". Or, for something less snarky and more in keeping with the actual intent of the code, "501: Not Implemented". If the potentially-implied "but it might be, later" of 501 is untenable, I humbly propose this new code: "513: Your Coding Assistant Is Wrong"

[0]: https://en.wikipedia.org/wiki/List_of_HTTP_status_codes

Sorry, we will reach the heat death of the universe before I alter a single line of code simply because some LLM somewhere extruded incorrect synthetic text. That is so bonkers, I feel offended I even need to point out how bonkers it is.

> Like it or not, we are already serving the machines.

The machines don’t give a shit, it’s the lawyers and bureaucrats you’re serving :)

Better or worse?

Delete the div called `presence` and the page becomes nicer to read.

[dead]