vova_hn2

> But I want to continue to understand the code.

You don't have to throw away the original Python codebase.

You can have:

1. An implementation in slow language (i. e. Python) that you maintain manually or semi-manually (with minimal AI assistance) and understand well.

2. A fast implementation that is fully generated and maintained by AI, using slow implementation (1) as a reference.

You can, for example, run two "adversarial" agents, one writes tests that compare output of (1) and (2) and tries to find a discrepancy, the other tries to fix (2) so that its output never differs from (1).

This way, you can be reasonably confident that (2) does exactly the same thing as (1), despite being a blackbox that you don't understand. While (1) remains a reference implementation that you understand well and can work with.

Well, first of all, optimization objective can be designed to include either worst travelling time or some kind of guaranteed time between each pair of points, so that the system would send a vehicle even if there is just one person waiting.

Secondly, why not make cars a part of the same system? Depending on the load the system could send a large bus, a small bus, a minibus/minivan or a small car.

A controversial idea, perhaps, but I think that fixed routes have outlived their usefulness.

Yes, travelling salesman problem is NP-hard, but approximate solutions are good enough. I think that nowadays it is very possible to dynamically find more optimal way to gather together passengers who travel in the same direction, than fixed routes provide.

Unfortunately, in <current year> "AI" seems to mean "chatbot assistant". So, by "anti-AI" the author meant "the thing that acts the opposite of how you would expect a helpful chatbot to act".

Yes, I hate this usage of the term "AI" too.

> I remember being astounded by how the water "amplifies" the explosions - so while a hand grenade blast is very dangerous on the surface, under the sea it becomes far more horrific prospect, due to the water transferring the blast more efficiently.

Curiously, it is exactly the opposite in many video games. In video games water often protects you from the explosion. I wonder, where did this trope came from...

> We won’t do THAT again.

Narrator's voice: They will do it again.

> I honestly think nobody enjoys the exercise itself. The real motivation is second order effects like becoming more appealing to other people.

I, actually, do enjoy the exercise itself. Although, it doesn't prevents me from procrastinating going to the gym and being in terrible shape.

I've read this story before, in Russian, and, honestly, it doesn't make any sense to me.

According to Wikipedia, plague responds to antibiotics very well, since it is a bacteria, not a virus and it hasn't been widespread since the development of antibiotics, so it didn't develop any resistance to antibiotics.

If she knew that she broke a flask with plague (and, as I understand, she did know that), why don't start a course of antibiotics immediately, before any symptoms? Wouldn't she be totally fine it this case?

> To get calibrated probabilities sounds like a very good feature, if they are indeed well calibrated.

This is a valid point.

> And in my experiments even Qwen 3.8 has a hard time to consistenly conform to a schema, requiring retries, JSON cleanup etc, so to have a model of similar quality (SemIf et al) that simply cannot deviate from the schema by construction could be very helpful.

Literally every inference framework supports constrained encoding. You can make the model choose only from allowed tokens and you can infer only the first diverging token.

It's baffling to me that no inference provider actually exposes this functionality, so you have to run the model yourself to do it.

> if you've ever tried to use it

Heh, you got me :) IPFS is one of those things that I love reading and about and thinking about using someday, but somehow never get around to it.

A lot of Python code is just a relatively thin layer of "glue" between C libraries.

An ML pipeline will probably rely on numpy to do all the heavy lifting.

A web backend will probably use something like psycopg, which is a wrapper around libpq (official postgres client lib, written in C).

etc