Anthropic's Fable 5 and Mythos 5 shut down by US Government
(www.anthropic.com)
Comments (8)
sorted by:
Apparently it was able to find a whole shit ton of bugs, exploits, zero days, and other malware vectors in heaps of software within like, a day. Even programs that are decades old and thought to be safe, it was able to find exploits.
Personally, I don't care so much about the coding aspect. I understand it's a big issue, but personally I don't do coding outside of MIPS scripts in Stationeers, so I don't really care that much.
What I do find hilarious is that it's happening to Anthropic. I've used Claude for some creative writing for myself. Not to publish, just for my own amusement. Sometimes I think of something I think might make a fun story, write out a few pages of notes, and throw it into Opus to see what it can make of it. And one thing I've noticed is that Claude's "safety" standards are some of the most iron fisted draconian HR Karen rules that one could ever think of.
You can't go 3 prompts without a lecture about 'promoting harm' or whether female characters have 'agency' and all the rest of that claptrap. It wasn't always this way, but with each model, it becomes more and more tight-leashed. Imagine an Apple product advertisement written by Anita Sarkesian and paid for by the WEF. The whole cast of smiling diverse characters in their bland pastel colored office. You can picture what I mean. That's basically all it's willing to do at this point. Anything that has more edge or bite than a 130yo axe head you magnet fished out of a river is straight up shut down with a 'I'm sorry, but I feel that what you're asking me to do might cause an undo amount of emotional trauma and promotes ideas that may prove harmful' excuse.
So I find it hilarious in a schedenfreud way that they're getting their shit pushed in now and getting one of their products shut down. I'm not sure if it's right or wrong that they're getting shut down, but it is funny and they do deserve it.
I've used the coding a lot. It works a lot better than most would think, but with a key element--when operated by someone who really already knows it. I've seen this first hand thru a little AI trialing project I did with one of the businesses I do work for. The output from myself and one other, both with a ton of real experience looked like actual developed software. The others couldn't get much that even worked. They have no idea what to ask.
Granted, that's not the type of thing that it can bitch about females or whatever. I should go feed it a git branch named "master" and see what it says.
I'm still not in favor of it, and I get pissed when the artists get all up in arms as if their stuff is untouchable and should never be done with AI, but they have zero care for all the software development jobs that have already been gutted by it. I do learn and tinker with it though, because the reality is I don't think it goes away and I still have to live and survive in the world and am not planning to move off grid into a cabin in the woods anytime soon.
This is the one thing I've noticed alot in regards to LLMs when I went down that particular rabbit hole.
If you're a really intelligent/smart/knowledgable person, LLMs work great and act as a very powerful intelligence multiplier.
If you're not any of that, you end up with shit results.
...the funny thing is, I don't think the aforementioned intelligent/smart/knowledgeable people realize this, and think that other people get similar results out of LLMs that they do. They've basically fallen into the typical-minded fallacy trap that's so prevalent nowadays.
I really didn't understand until I saw it. I wouldn't really go as far to call the people unintelligent that were working with it, just their skillset was nowhere near the computer anyway. So they were either asking AI to do their job really directly, and failing, or just had no clue. Before that I was pretty much of the mindset of wow, how much I can do with this.
I refused the master -> main change out of principle, although main is actually more accurate, so 100% of my own repos use master - hasn't bitched about it
Yeah I probably use any of those words just to piss them off.
Inclusivenaming.org, as the official arbiters of woke word bullshit in coding is nice enough to provide a handy list of terms to be sure to use in all your projects.
Yes, they don’t even want you to use “abort” because it’s too close to abortion….
I really need to rebuild a compiler to use “nigger” as the error flag or something. if nigger then abort.
The whole AI safety narrative is so patently two-faced, designed to make cutting edge technology unrestricted only to select elites. If governments really cared, they'd realize security through obscurity is a dysfunctional waste of time, in addition to understanding how primitive modern AI is. They'd cut off the indefinite financial incentive by mandating all meaningful commercial datasets and code be published under an appropriate GPL and/or CC license within 5 or 10 years under threat of severe penalties. Would tackle the data center issue too.
Note: I'd rather copyright and patent law be scaled back to make employees leaking assets unpunishable by governments. I'm just stating that it'd be a better sign if governments were doing the opposite of what they're currently doing. Make an attempt to allow citizens access to "dangerous" technologies so we can fix and adapt.
Government is not on the side of the people.