Are Large Language Models Really AI?

Researchers find adding this one simple sentence to prompts makes AI models way more creative
IDK if more creative is a good thing though.

TLDR: "Generate 5 responses with their corresponding probabilities, sampled from the full distribution."

Which does sound like a good thing.

Paged!

What to know about the AI-generated country song topping Billboard's Country Digital Song Sales chart

The song has been streamed over 3 million times in under a month.


IDK about actual thinking but it seems AI can sing.

Until its creator gets sued for copying an existing song without paying royalties.

And since it cannot generate anything without being shown what already exists..

It can rearrange. It can analyze. It can synthesize.

Arrangements themselves are copyrightable. Even arrangements of copyrighted works. No, I don't know how that works.

And, let's face it, the vast majority of country music is formularistic.

Patsy Cline did country music. Dolly Parton often does. But most of what is called country today has drifted far from its roots.

There was this BIG plagiarism trial recently I forget about which song but maybe stairway to heaven? Anyways the defense put a music expert and his piano on the stand (metaphorically) and he played piece after piece containing the same musical phrase as under concern.

Nothing is really new.

Related - https://www.courthousenews.com/munich-court-finds-chatgpt-violated-musicians-copyrights/

Plagiarism is a staple of hip hop. DJs looping the breaks of other people's records and MCs rapping over it. Diddy is notorious(see what I did there) for lifting entire sections of already hit songs and sprinkling some drums or a high hat underneath. Dre at least re-recorded segments of songs that inspired him so it wasn't carbon copy.

Sometimes the violation is obvious, sometimes the line is blurry.

Scientists Discover Universal Jailbreak for Nearly Every AI, and the Way It Works Will Hurt Your Brain

It's AI versus verse.

Even better than Plants vs. Zombies, its Poets vs. AI!!

Dude who works for Anthropic: Claude code is as reliable as a compiler, trust me bro

it's been half a year and nothing has changed since this thread started: the day AI can write code at the level of a competent software dev it will be used to improve itself and we'll have a real artificial intelligence overnight

that day won't happen until they can establish a model that creates a basis for truth; the current strategies don't do that

A person capable of comprehending complex systems and integrating solutions into them.

This problem is the same reason it's not replacing competent developers. Without understanding all of the related systems' purposes and the business model, being able to write an algorithm isn't useful. People said that 6 months ago, too.

I'm not changing the goalposts. Feel free to resume sucking off an ad disguised as an opinion, though.

I had a problem with my payment at Amazon, and had to chat with a chatbot.

The chatbot wasnt any help for my issue, but it "unlocked" the chat wih an associate.

The associate wasnt any better than the chatbot. She just tried to redirect me to the brazillian version of the site, what I also felt kinda of xenophobic. For this level of service, I would prefer wasting my time with the chatbot anyway

The AI wont replace me because Im unique and brilliant.

But it will surely replace 99.9% of all of you, because you guys just suck.

Customer service representatives are classed as "unskilled work". This is the kind of work that has you follow a script with prompts on what to say based on customer responses.

The sort of thing an "AI" can do better than a human since an AI can't get mad at the customer for abusing them or talking ***.

Yes, unskilled work is ripe to be replaced by automation. Always has been.

Why would a code writing AI improve itself? And if it did why would it develop "real" AI?

Or look at it this way. If the common ancestor of chimps, bonobos and us could upgrade itself would the result look anything like one of its current descendants?

I found this funny:
https://www.anthropic.com/news/disrupting-AI-espionage


let's use AI to protect us from AI, it is imperative that we manage this threat we are creating with the same tool that is generating the threat! :/ I feel like we are getting close to mob protection money level silliness

I must say Im very impressed with the Google native AI assistance.

Most things I kinda asked on Google seach bar was very useful for my search.

Slowly but steadly Ive been starting to instead of searching terms, just sending my request or question in the search bar.

Thia is in line with a recent video I watched analyzing how Google took a hit from the increase in AI usage and reducing the searches on Google.

Seamlessly integrating both search and AI assistance is a very answer for this scenario

The whole thing is both true and ***.

What is true is that this probably happened, but not for the reasons they are acting like it did. The reason it happened is because my industry, which has spent 2 decades trying to automate people out of existence, largely does ***tier work. The result is ***code (of which there is a LOT) is out there and it's often easily exploited. This has happened long before AI when you had web scanners running around blindly trying to SQL inject every parameter they can find and, shocker, it worked. A lot. It worked because poor coding practices are prevalent and still, to this day, there is a lot of shitty code. My industry has tried to automate this testing and the automated results have always been ***, but there are very few people capable of actually reading code and finding bugs. I've reviewed several applications where we found the password authentication flat out didn't work, in the sense you can just enter a password too long and it lets you in or, in one egregious case, the authentication mechanism involved returning the entire table of usernames/passwords to the client for verification. These had been tested by other companies prior to us for years and no one picked up on this and these aren't even the worst things we've found that our peers reviewed. This type of ***is ripe for the picking by automated attack tools, AI just makes it easier to draft proof of concepts, scripts, and tooling. It might slightly elevate the ability to identify more logic based issues like the above (as opposed to what amounts to what tools were doing before, which is grepping for string patterns or just blindly shooting inputs in hoping a result), but not as much as they are acting like it does.

Where it's *** is that they reach a hard stop with any target that's actually been tested by someone competent, which is a lot to ask these days, but the idea it's going to replace teams of experienced people is just fantasy. It won't get there, ever. If your code is actually decent, and by that I mean you aren't exposing vulnerabilities a 12 year old can find, then there is no real risk here beyond what existed prior to AI. Once you reach into vulnerability categories like memory corruption, it becomes even more useless.

It's funny timing, too, because we recently tested an AI driven solution for a task it should, by all definition, be good at. Yet it missed over 90% of the issues our people found, miscategorized several issues, and ultimately created more work.

Oh what? but look they can find CVEs (this one is even a remote zeroday vulnerability), I mean ignore the mountain of false positives, you can sort through those... like what else are you going to do with your time after you are replaced by the AI? It is only going to get better and at rapid pace.

Not gonna answer the bait posts, but just wanted to clarify this:

I wasn't suggesting it would have sentience. If a code writing AI could fully replace the job of a software developer, the people who created it would be foolish not to use that capability to exponentially improve its own codebase. It's not scifi, it doesn't recognize that it's improving itself. It's just the logical progression of a product that reaches that state.