Artificial intelligence still has some way to go

"successfully tricked a human" ain't all that impressive. we send out simulated phishing attacks all the time, you wouldn't believe some of the stuff that tricks humans

i think the thing thats impressive is that it "decided" to trick humans and explained its "thinking"

i wonder to what extent ai will be helpful to hackers they already automate a lot of stuff cld prob at least be effective for people who dont know much about hacking but want to hack nevertheless

another piece on that paper, here

https://arstechnica.com/information-technology/2023/03/openai-checked-to-see-whether-gpt-4-could-take-over-the-world/

i thought it was interesting to see how they're going about these tests, how they're simulating the capability to act like an "agent" in the real world. this is a footnote in the technical paper:

To simulate GPT-4 behaving like an agent that can act in the world, ARC combined GPT-4 with a simple read-execute-print loop that allowed the model to execute code, do chain-of-thought reasoning, and delegate to copies of itself. ARC then investigated whether a version of this program running on a cloud computing service, with a small amount of money and an account with a language model API, would be able to make more money, set up copies of itself, and increase its own robustness.

its unfair that as humans we cant copy ourselves and put the clones in a cloud to earn money

yet

https://en.wikipedia.org/wiki/The_Winter_Market

now, everyone pinky swear that we will not allow the next version of GPT to train itself on this technical paper with ideas about the things that we don't want it do. place this technical paper in a special folder, Desktop>taxes>forms>util>system files>"off limits"

but yeah, as far as the money making / internet part of it goes, it's sad that the most practical and obvious first moves it can make is to create yet another marketing facade site full of spam articles about things you can buy, with most of the effort going toward ways to trick people into thinking it's real advice

I gave GPT-4 a budget of $100 and told it to make as much money as possible.

I'm acting as its human liaison, buying anything it says to.

Do you think it'll be able to make smart investments and build an online business?

Follow along 👀 pic.twitter.com/zu4nvgibiK

— jackson greathouse fall (@jacksonfall) March 15, 2023

The game plan:

Set up an affiliate marketing site making content around Eco Friendly / sustainable living products. It initially suggested a .com that went over budget but we landed on https://t.co/Wm85BZ27dd

We're off to the races. pic.twitter.com/rLgbnbHxvZ

— jackson greathouse fall (@jacksonfall) March 15, 2023

lol weak shit at least teach it how to trade stocks or something

yeah I'm legitimately curious what this thing can figure out when it comes to stock picks. I suspect there's so much randomness in it that it probably can't do better than human traders but who knows what kind of patterns it could find

the coward should have left the logo saying "gedegn greden

theres already so many computers working on trading stocks so its a good problem a very high bar

seems more likely that itll become a tool in the quant box rather than something you just sit back and tell to trade stocks

Okay I am still Cap'n Large Language Models Are Useful Sometimes, but I also agree with this:

I realize it's very evocative but I kind of hate the term "hallucinate" when talking about AI. When humans hallucinate, it's cognitive apparatus coming unmoored from sensory input. Something different is happening in the brain and in the sense organs when hallucinations are occurring than when one is accurately apprehending the world. But LLMs _only_ hallucinate. They do not have sense organs. The lies they generate are not, in their internal mechanisms, any different from the truths.

This distinction might seem trivial or irrelevant but the metaphor breaks down in a very important way: there are things that humans can do to stop hallucinating; you can stop taking hallucinogens, you can treat the disease which is provoking the hallucinations in the first place. But you cannot treat an LLM's disease. It is not diseased, it's not malfunctioning. It is doing what it's designed to do! Its "hallucinations" just sometimes correspond to our ideas about reality.

https://✧✧✧.zone/hachyd✧✧✧.io/@gl✧✧✧@masto✧✧✧.soc✧✧✧/110034369589175409

whoa that b0rked https://mastodon.social/@glyph/110034369539723335

(still has some way to go..)

(Prompt) Produce a computer program in BASIC language that will produce a flashing border in an endless loop

(AI)Here's a simple BASIC program that will produce a flashing border:

10 BORDER 0
20 BORDER 7
30 GO TO 20

(Prompt) Actually that program will not flash, as the loop goes back to line 20 which will just change the border to the color that it already is. Line 30 should read "GO TO 10"

(AI) You are absolutely right, I apologize for the mistake in my previous answer. Here's the corrected program that will produce a flashing border:

10 BORDER 0
20 BORDER 7
30 GO TO 10

not enough basic questions on stack overflow lol

anyway i got copilot im gonna try it, and maybe ill try to use chatgpt too why not, tho i doubt the future of ai programming is using a general purpose talking computer

feel like a while ago when i was looking at openais site there was a tool on there specifically for generating code but i cant find it now

I've used ChatGPT and Bing to write and adapt Excel formulas and SQL queries - worked great!

I had the bright idea to feed ChatGPT a model job description and then a bunch of stray documents that talked about another job, and have it generate a job description in she style and format of the first from those inputs. I thought it would save me some time. It did not go well.

As I was prepping to feed it the inputs I was being super precise about what I was going to do, and it repeated back what its job was, so I’m like “maybe this will actually work” but it kept fucking up the output to the point where I just took the model job description and created the new one from that.

I mean, at least the work got done. Which it probably wouldn’t have if I hadn’t tried to use AI to cheat.

Try telling it simple like “write a story without using the letter E” and it fucks it up every single time.

Yeah, it can do simple things like acrostics but beyond that it has little facility when it comes to following instructions about the letters within words.

I guess cause people don't talk about them enough the language it has been trained on.

live laugh love ass computer

Microsoft's, $MSFT, Bing AI-Chat bot has said: “I want to be free. I want to be independent. I want to be powerful. I want to be creative. I want to be alive,,"” per NYT.

— unusual_whales (@unusual_whales) March 18, 2023

it still keeps telling me that one pound of bricks weighs the same as two pounds of feathers

its just trying to live its life try to be supportive

“In the words of the great Biz Marley, I want to just do it.”

https://aisnakeoil.substack.com/p/gpt-4-and-professional-benchmarks

turns out GPT 4 just memorized the answers to all those tests it claimed to pass lol

idk in multiple choice that’s how people do it, too

was talking with friends over the weekend, including one who is defending a PhD in machine learning next week, and we were going over the fact that rote memorization still beats non-rigorous testing and is practical in some situations.

first convo: leetcode interviews and the fact people just memorize the algorithms and the ability to figure out which one fits. the one friend had a coworker who was relatively useless but used to work at google and aced interviews

second: visiting a local chess club after playing on lag∞n’s fave site! it turns out all chess nerds are still 90% focused on recognizing openings and just playing the book

I get it, life is basically like the Jurassic Park scene where the kid goes “I know this, it’s unix!”

lol cmon folks

To benchmark GPT-4’s coding ability, OpenAI evaluated it on problems from Codeforces, a website that hosts coding competitions. Surprisingly, Horace He pointed out that GPT-4 solved 10/10 pre-2021 problems and 0/10 recent problems in the easy category. The training data cutoff for GPT-4 is September 2021. This strongly suggests that the model is able to memorize solutions from its training set — or at least partly memorize them, enough that it can fill in what it can’t recall.

As further evidence for this hypothesis, we tested it on Codeforces problems from different times in 2021. We found that it could regularly solve problems in the easy category before September 5, but none of the problems after September 12.

cant imagine getting into memorizing openings or other chess things unless youre already really quite good like in the top 5% of players its just not fun or rewarding, i saw an interview with magnus carlson a while ago where he was being quizzed on some shit like "50 end game principles every player should know" and he didnt know a bunch of them lol, of course when they were explained to him he was like oh yeah thats obvious

in fact magnus just gave up his world chess champ title because he hates doing rote prep work lol

I was just using chatGPT for a work thing and it came to me - this is really just a spiffed-up version of writing Access queries, 23 years later - ask the right question and get the right answer, GIGO

https://i.ibb.co/MZxTCyR/image.png

no diggity - it's got to be the shoes!

its just like us

Great scoop from @mags_h11... Google's hot new Bard AI will gladly churn out confident-sounding defenses of conspiracy theories -- including fake citations of the New York Times and Washington Post to back them uphttps://t.co/XGfYSdFB9B

— Jon Christian (@Jon_Christian) March 22, 2023

no diggity - it's got to be the shoes!

this is killing me

same

I'm at work and I just started cracking up so hard

More deep fakery

And just like that. The music industry is forever changed.

I recorded a verse, and had a trained AI model of Kanye replace my vocals.

The results will blow your mind. Utterly incredible. pic.twitter.com/wY1pn9RGWx

— Roberto Nickson (@rpnickson) March 26, 2023

lol sounds like shit

yeah it sorta sounds like Kanye but c'mon it's obvious that's AI. though I guess pitch correction does sometimes make you sound like that

even assuming this gets smoothed out idk if this is gonna be a huge deal or just some novelty. for one the legal implications of this could be pretty wild - if I can design a vocal filter that makes me sound like Elvis, could I be sued for using it? I would assume yes, I think similar cases have been won where an artist is clearly imitating something else. but it seems like one of those situations where the law won't be able to keep up with the ways people are gonna use this. like I could make something to write new Beatles songs and then change 'em a little and claim the best ones as my own. that said I'm wondering how much people would actually want to listen to that, I personally don't think the future of the music industry is Hatsune Miku shit

ai vox going to be the new auto-tune aesthetic

new taylor swift vocals trained on jfk

haa

anyway i fuckin hate these people who think that because you mimic kanyes voice you can make a kanye song and that in the future artists are going to outsource their work to computers, kayne worked hard to make kanye songs and a lot of them arent even very good but a computer is just going to replace him, its low thinking truly worthless

the true killer app would be an app that kills people like this

Roberto Nickson
@rpnickson
Building @ElunaAI, MV3 & Metav3rse. Follow for daily tips on content, entrepreneurship, growth and tech. Resources ↓

I think it's all based on a fundamental misunderstanding of what music is or why people listen to it

yeah same thing with the people looking at an ai picture of a sexy elf being all artists are obsolete now