undefeated

> A 2012 paper on dishonesty — published in the Proceedings of the National Academy of Sciences and coauthored by Harvard researcher Francesca Gino — was also retracted over concerns data in the study were fabricated.

Mfw when the paper on dishonesty in fact dishonest

A genuinely impressive and novel use of machine learning? In this economy?

Looks really impressive, although one issue I do see is consistency. For example, it seems to highlight `null` as a value, unless it is on the left side of an `===` expression, since it probably hasn't seen that in training.

Was it ever actually tested on languages not in the training set? Would be interesting to see if a model this small can actually generalize...

I can kind of see it, but hetzner costs, what, 5€/m for your own server with 10TB traffic included? How is a 4€/m markup remotely justified for what amounts to setting up an apache server and pointing a domain at it? And cutting traffic to a twentieth?

> always-on, local access to your machine's files, apps, and sessions

Uh... how about..., no...? What?

> Controllable from any device, anywhere

...

> There is a kill switch

Oh great, sounds like you're confident this is safe then!

Someone wake me up, please

Unfortunately, it doesn't look like the EU will actually be able to put any meaningful guardrails on AI, as evidenced by the AI act. I honestly don't know whether it's an off-the-scales level of incompentence or just blatant corruption (I vote both), but in any event, I think AI regulation has already failed, somewhat irreversably.

Uhm... duh?

> or cannot control any of the inputs to the llm

Seeing as LLMs are non-deterministic, I think even this is not enough of a restriction.

Unlikely, because the free version of ChatGPT isn't really making them any money, so less tokens is actually better — which I assume is why anthropic pushes Haiku models on free users which are not just more quantized but also less wordy.

I think part of the problem is that you need a model to classify the data, which needs to be trained on data that wasn't classified (or a dramatically smaller set of human-classified data), so it's effectively impossible to escape this sort of input bias.

Tangentially, I'd be far from the first to point out that these LLMs are now polluting their own training data, which makes filtering simulatenously all the more important and impossible.

I never had a conversation like that — probably because I personally rarely use LLMs to actually generate code for me — but I've somehow subconciously learned to do this myself, especially with clarifying questions.

If I find myself needing to ask a clarifying question, I always edit the previous message to ask the next question because the models seem to always force what they said in their clarification into further responses.

It's... odd... to find myself conditioned, by the LLM, to the proper manners of conditioning the LLM.

Complaining about JSON and then proceeding to write your API in Dart is... interesting.

Surely you've never read the job description for a programming position

It benefits no one to acknowledge that months were spent and the only outcome is millions of dollars wated. Sure, they could try to hold the contractor accountable, but that would put them at risk of being held accountable themselves by their own superiors for hiring that contractor in the first place. So they pretend everything works great, and pay accordingly

Then again, we'd probably end up with all the same bullshit on Threads

At least we got some of my all-time favorite GitHub issues tho

News stories are generally considered off-topic https://news.ycombinator.com/newsguidelines.html

Looks cool! It would probably be useful if there were a way to include dependencies that aren't shipp packages though. Otherwise, since most projects one might depend on right now don't have shipp support, there's a risk of ending up having to use other dependency management solutions alongside shipp, making everything even _more_ complicated rather than simpler, right?

Anything but the metric system

Agreed. I recently had to write a package with some quite complex type inference requirements, so I had to resort to classes.

I've since had to issue 17 patches (it's been 3 weeks) even though the package has 100% test coverage.

Was it only a BNF parser generator or did it also help with implementing the language's actual functionality?

I've never heard of such a thing, but it sounds amazing!

Out of curiosity: How exactly did it work (for users)? I assume you'd write a grammar on the one pane, but how did you teach the tool to convert it into the output language? Was that embedded into the grammar?

I think it's more like "Company X is big and we have to be like company X if we want to be big too"

Not that I know of, but definitely something I'm gonna put on my 'future project ideas' list.