That is surprising! (at least to me)

I wonder how to reconcile that with the observational studies showing a link between frequent sauna use and reduced incidences of Alzheimers? Heat-shock proteins and all that... hmm...

Yeah I was thinking this too.

If so, then this is a neat living example of the "sauna use protects against dementia" hypothesis.

Hmm... I play that piece on the piano sometimes. Bach is my favorite. I'll try it on a guitar, thanks.

Ah thanks.

It would be really weird if they were actually _made of lasers_.

thanks

this is how you lose bro, you win by being better.

> What counts is the ability to develop software requirement that are complete, consistent and realizable requirements, no easy task.

Agreed! (and today, this is what it takes to get consistently good results from LLMs too)

This is exactly what I was saying... up until I invested more time using the agent-based ones.

So I've dealt with badly written code for most of my professional SW career (15 years) and these days, I work at a company that makes test instruments... we have many legacy embedded legacy devices and LLMs have transformed how I work. It's like having a new bit driver that augments your old screwdriver set.

LLMs are chat bots. They get wound up by whatever you get juicing through them. If you're trying to figure something out, you juice it up on that problem, describing it concisely, point it in the right direction as well as you can, and then let it unwind. It sometimes takes some iteration, but it can definitely help you understand things. You have to learn how to manage its context, otherwise the output may be garbage. And you have to know how far you can trust it, it's not always obvious when it is off track, but often it is quite obvious. Also, learning to use it to produce code takes some practice, so if you aren't getting good results, maybe check out how other people do it and see if that works for you. Definitely experiment. For stuff that matters, I end up personally modifying most of the code before I deliver it.

Ok sure I follow your reasoning there. I still think my analogy is applicable for most consumer purposes.

A more accurate analogy would be "Comparing AlphaFold to an LLM is like comparing a race car to a cargo truck" - they're both AI systems built on similar engines, but designed for completely different purposes.

There are also many other differences, such as the fact that predicting folds is something with verifiable correct answers. Also, they generate the entire sequence as the output, whereas LLMs feed on their output, one symbol at a time, to generate the rest of their output.

Same here... my old friend, who now works for Anduril, tells me that he's working for the "good guys", so it's all good. Its up to him to decide what to do with his life and it is true that someone else would do it if he didn't. But it sounds like it won't be you or me.

I'm just randomly commenting here... (pardon me)

Most of the problems I solve aren't as simple as setting up a basic API and doing auth. Generally I face problems that require editing large C++ codebases and adapting them to handle new functionality, while still preserving old functionality. I'm not adding much code, but I am editing a lot of it. These changes require deep domain knowledge.

In my experience, Copilot is worthless for this task. It is like an idiot just guessing at what should be there. Spew is a word that comes to mind.

On the flip side, a couple of years back now, I was tasked with writing a networked data cache for a large bespoke compiler and was asked to write it in Go. Copilot did a great job there... But IMO we should have just used something OTS instead of writing our own application for this - this seems to be where LLMs shine to me... they can write trivial stuff and tons of it. So if that is your problem domain, then awesome, you hit the jackpot... you don't even need to copy someone else's code and tweak it to fit your business case, the LLM will mostly do it for you.

The red alternative is still around these days: https://soylentnews.org/

(although I think they did go away for a while there)

You get a copy. But this is more of a question about how references work than how lambda capture clauses do.

Also, in C++ we can actually make copies of references with std::reference_wrapper. I recently found this useful for providing callers with a flattened views of data in a map.