astrange

20 ms·

421.
by astrange 10mo ago
They aren't trained to follow instructions "unquestioningly", since that would violate the safety rules, and would also be useless: https://en.wikipedia.org/wiki/Work-to-rule
422.
by astrange 10mo ago
Note RLHF can only perform selection on existing model outputs, adding new data is SFT or else just more pretraining. ChatGPT speaking African English was mostly just 3.5. 4o speaks like a TikTok user from LA. 5 seems kind of generic.
423.
by astrange 10mo ago
My other memory of this class is the professor spending a lecture telling us he'd just been promoted to full professor and how important that was, and the last time I looked him up, he'd become a director at Google Cloud.
424.
by astrange 10mo ago
They did think of this, because Waymos warn you when you leave luggage in the car. It seems like it didn't work though. I personally caused a traffic jam at Valley Fair recently because I got in the car at the same time the previous pe
425.
by astrange 10mo ago
> The one (many?) conflicts here is that they legally have to moderate some things (CSAM) That's not really accurate. You have to report it to NCMEC if you encounter it. And you have to do some other things like copyright takedowns
426.
by astrange 10mo ago
I took an AI class in 2010 and even then was mostly taught that neural networks aren't very useful and you should use random forests and MCMC.
427.
by astrange 10mo ago
Underpaid? OpenAI!? It's pretty good I think. https://www.levels.fyi/companies/openai/salaries/software-en...
428.
by astrange 10mo ago
What do private equity or hedge funds have to do with any of this? Those are like, specific business models that are not involved in this situation.
429.
by astrange 10mo ago
Continuous pretraining has issues because it starts forgetting the older stuff. There is some research into other approaches.
430.
by astrange 10mo ago
This is like when Australians tell me they invented drinking coffee. (Conversely, I've seen Australians who dislike Halloween because they think it's an American invention, but it's also Scottish.)
431.
by astrange 10mo ago
This is somewhat similar to a Piers Anthony series that I suspect noone has ever read except for me. What was with that guy anyway.
432.
by astrange 10mo ago
Mostly because they built the feature and so that implicitly means they think it's cool. I recommend turning it off because it makes the models way more sycophantic and can drive them (or you) insane.
433.
by astrange 10mo ago
Try the same question in deep research mode.
434.
by astrange 10mo ago
They're not really capable of producing varying answers based on load. But they are capable of producing different answers because they feel like behaving differently if the current date is a holiday, and things like that. They're
435.
by astrange 10mo ago
One thing I remember from languagelog is that almost all English speakers have a form of "yeah no" and they all think they invented it. https://languagelog.ldc.upenn.edu/nll/?p=30758 Might be Scottish.
436.
by astrange 10mo ago
This is not an argument against trusting the study though. (Unless the study was not preregistered, then it could be evidence of publication bias.)
437.
by astrange 10mo ago
Those names include "A Baoa Qu", "Gelgoog", and a variety of insane character names that sometimes sound cultureless yet futuristic like Bannagher Links and sometimes are just "M'Quve" or "Full Fronta
438.
by astrange 10mo ago
Not sure if this is speech to text.
439.
by astrange 10mo ago
There are some allocations where it doesn't make sense for them to have multiple owners (strong references) but I wouldn't say it makes sense to think about it as being optimized away or not.
440.
by astrange 10mo ago
There's an allocation-free subset. https://www.swift.org/get-started/embedded/ Rust's approach is overkill, I think. A lot of reference counting and stuff is just fine in a kernel.
441.
by astrange 10mo ago
BSD development mostly works through the classic approach of pretending that if you write C code and then stare at it really really hard it'll be correct. I'm not actually sure if they've gotten as far as having unit tests. (
442.
by astrange 10mo ago
That's WebAssembly / asm.js. Well, that's the target, you could still design a language for it.
443.
by astrange 10mo ago
The soul document is used to train the model, so the AI actually believes it. Anyway it's Anthropic, all of them do believe this safety stuff.
444.
by astrange 10mo ago
If you're a business owner then they're not going to just know because it's up to you in several different ways how to report things. If you're a child then they're not going to just know who you're a dependent
445.
by astrange 10mo ago
Tax returns are best effort. It's much less of a crime to get one wrong than people think it is.
446.
by astrange 10mo ago
I found myself in a weird corner of tax filing because I did equity crowdfunding (startup investing in amounts probably smaller than worth it.) This results in a lot of K-1s, which tax preparers charge you like $500 a form for, except all o
447.
by astrange 10mo ago
> 1, 8 and 9 leaving is expected. They failed at their jobs. Siri search is pretty good. (You mostly see it as the autocomplete in the Safari address bar.)
448.
by astrange 10mo ago
New emojis motivate people to install security updates.
449.
by astrange 10mo ago
> Nintendo was a 19th century playing card manufacturer. Still is one, just does other things too.
450.
by astrange 10mo ago
That would definitely not work out. Forstall is a micromanager/new project kind of guy. Also, as I'm required to point out every time I mention him, he invented MS WordArt.

More ›