Artificial intelligence still has some way to go

poor Taosif

i think it's going to be interesting to see if the horizon thing results wider awareness of the likely harms of algorithmic decision making relative to e.g. the debate in the US.

if so it comes at the worst possible time from the point of view of silicon valley vcs trying to sell ai dogshit to the uk government while governments are writing "ai law". IIUC the palantir / nhs thing is a done deal, but if the horizon scandal was in the news a year ago that's the kind of deal that could have been held up or abandoned in this climate.

The dogshit is really coming through, even on low levels. Last week a little provider was pitching an AI guidance solution to the gov programme I'm in at the moment - took 20 seconds to answer first question and immediately got 'no that's wrong' from an expert. Refused to answer second question. This was with questions written by the provider.
It was clearly just a GPT trained on a bit of specific guidance, so no solutions for hallucination or consistency when asked. Didn't even get into 'what are you charging and what value are you adding?'.
Did a frankly amazing job of making guidance writers (me) look good and necessary. Agreed to try and help with improvements, so that's useful for the CV.
No sense in getting complacent, but a useful bit of calibration for me.

theres just no way for llms to know what theyre talking about, its guessing all the way down

People under Dunning Kruger effect using tools that have no self awareness, but are almost definitionally a sort of DK effect, and also a data aggregation training model that has no ability to actually distinguish true from false. It’s like what can go wrong

pretty interesting how slow the “progress” on llms has gotten eh friends

using the latest ai tech to generate graphical resources

https://i.imgur.com/o3jkafn.png

(illustrator does have ai tools now, have not tried yet)

true ai success story from last night:

https://i.imgur.com/5v0GVje.png

i have found chatgpt useful for loads of things like that, but maybe i'm just more forgetful than yall

pretty interesting how slow the “progress” on llms has gotten eh friends

― Humanitarian Pause (Tracer Hand), Tuesday, January 16, 2024 12:20 PM (one hour ago)

gpt4 (the paid version) is definitely better than the publicly available gpt3.5 version (based on a few examples that friends with paid subscripts have shown me)

xp damn it gave you the first google result for “documentary punk bands los angeles 1980s” we’ve got one instance where it measures up to

i like to give my pal chatgpt a chance to warm up on a couple of easy ones before i get to the inevitable recommendation for today's lucky numbers

its fine for things that it doesnt matter if its wrong, which is not most things but is if youre trying to remember the name of something because youll know if the answer is right when you see it, the repo man answer is funny cause the movie is def not about la punk bands but it does have a sick soundtrack that meets the definition

lets say youd never heard about those movies someone had just told you that there was such a thing as cool 80s LA punk bands and you asked chatgpt to recommend you a movie about them, which would be kind of a weird thing to do but still, and it said repo man and youre watching it waiting for it to be about cool 80s LA punk bands but instead its about aliens guys who repossess cars and plate of shrimp youd be like wtf, tho you wouldve seen a great film so maybe its ok lol

here are some other cool things we could check out

Music Albums:

"London Calling" by The Clash
"Unknown Pleasures" by Joy Division
"Rumours" by Fleetwood Mac
"Purple Rain" by Prince and The Revolution

TV Shows:

"Twin Peaks" (original series)
"The Twilight Zone" (original series)
"Black Mirror"
"Stranger Things"

Books:

"1984" by George Orwell
"Brave New World" by Aldous Huxley
"Neuromancer" by William Gibson
"Siddhartha" by Hermann Hesse

Video Games:

"Fallout: New Vegas"
"The Legend of Zelda: Ocarina of Time"
"BioShock"
"Mass Effect 2"

Artists/Artwork:

Salvador Dalí's surrealistic paintings
Banksy's street art
Keith Haring's pop art

Movies (Apart from "Repo Man"):

"Blade Runner"
"A Clockwork Orange"
"Donnie Darko"
"Akira" (anime)

Cult Classics:

"The Big Lebowski"
"Fight Club"
"Fear and Loathing in Las Vegas"

Podcasts:

"Serial"
"Welcome to Night Vale"
"Radiolab"
"How I Built This"

damn, really hitting the deep cuts

chatgpt is basic and thats ok

I'd propose along the lines of the Turing test, a kind of metric based on whether you would trust the answer more than from, say Rudy Giuliani.

Like, is Repo Man worse than the answer Giuliani would give? Would you trust Giuliani's medical advice more?

Rudy Giuliani's favorite movie is North

With a little nudging, ChatGPT is really good at generating absurdist comedic sketches (e.g. the transcript I generated about a newly-ripped George Burns going on the Tonight Show ostensibly to promote Oh God! but instead smashing Johnny's desk in a fit of roid rage). This is literally the only shit that this shit should do or be able to do, and the quicker the world realizes this the better.

like good meaning bad

I got ChatGPT to generate an article about Justin Trudeau eating a chipmunk on live TV, I think we should encourage that type of usage

"A few months ago I told the American people Justin Trudeau did not eat a chipmunk on live TV. My heart and my best intentions tell me that's true, but the facts and evidence tell me it is not."

Borowitz AI

ai is pretty amazing at coming up with the worst possible shit i dont think a human could do this on their own without killing themselves first

https://www.youtube.com/watch?v=QZVXmtzTbmE

You know that saying about how capitalism privatizes the gains and socializes the losses?

In my experience AI tends to automate the fun parts of creative work and leave the boring parts to humans.

has this story been linked yet?
https://www.404media.co/google-search-really-has-gotten-worse-researchers-find/

"The researchers, from Leipzig University, Bauhaus-University Weimar, and the Center for Scalable Data Analytics and Artificial Intelligence, set out to answer the question "Is Google Getting Worse?" by studying search results for 7,392 product-review terms across Google, Bing, and DuckDuckGo over the course of a year."

my experience as a professional fact-checker is that google has definitely silting up with unreliable answers over the last year or so (peppa pig not being seven foot tall etc) -- it's not impossible to cast questions in a more detailed way to get to more authoritative sources but i've had a LOT of practice

Charlie Brown did not, in any way, shape or form, have hoes

are the comments on that Tom Brady standup video AI generated as well? I'm not hearing any actual jokes!! just clicking around and hearing 2 minutes about how "Beast mode" is the opposite of "Goblin mode"

more lies!

A-LIE moar laik

(peppa pig not being seven foot tall etc)

schrodinger's peppa pig's height measurement

https://i.imgur.com/ZezfhaG.png

190 million results (all with different answers)

Jennie Johnson is the AI representation of elite recruiting. With real-time global job insights, extensive industry knowledge, and 24/7 availability, she tirelessly connects candidates to their perfect roles. Deep expertise, always on hand to help. Experience the future of recruitment with Jennie.

love 2 tirelessly connect

Recruiters are already barely human a lot of the the time so this might work fine

Experience the future of recruitment with Jennie.

Experience the future of recruitment with Jennie.

i just had an apple
i just had an apple
i just had an apple but thanks

lol

i got the same jennie linkedin email too lol

they called her Jennie LinkedIn

i'm inspired to work at this cutting edge AI company

Get ready to unleash your inner artist. Starting this week, AI Art is available in public preview for U.S. customers on Fire TV.

This #generativeAI experience combines the power of Alexa with a fine-tuned Amazon Titan image generation model that allows customers to create and… pic.twitter.com/frt8WJ2GKM

— Adam Selipsky (@aselipsky) January 18, 2024

looks like absolute dogshit!

this is a hilarious idea

I am tempted to wrangle a way to scrape the images on this thread and set as a random login screen:
Post your AI art itt

A system developed by Google’s DeepMind has set a new record for AI performance on geometry problems. DeepMind’s AlphaGeometry managed to solve 25 of the 30 geometry problems drawn from the International Mathematical Olympiad between 2000 and 2022.

That puts the software ahead of the vast majority of young mathematicians and just shy of IMO gold medalists. DeepMind estimates that the average gold medalist would have solved 26 out of 30 problems. Many view the IMO as the world’s most prestigious math competition for high school students.

“Because language models excel at identifying general patterns and relationships in data, they can quickly predict potentially useful constructs, but often lack the ability to reason rigorously or explain their decisions,” DeepMind writes. To overcome this difficulty, DeepMind paired a language model with a more traditional symbolic deduction engine that performs algebraic and geometric reasoning.

they invented a computer that does math as well as a 17 year old nerd

math nerds tend to peak early. historically math geniuses flame out by 35 or sooner.

i'm actually impressed by a computer being as good as the best 17-year-olds in the world at solving really complicated geometry questions on the fly! being as good as the best 20-year-old in the world at geometry doesn't seem far off. i don't know the world/age rankings of geometry pros but i assume that if you're one of the best 20 year olds you kinda made it as a pro and get endorsements