don't care

Can I fuck it?

>>109720386
maybe ETP if guardrails are lowered

>>109720393
ERP*

I can't believe the new model is better than the last one

Still no blogpost? How the fuck did they bungle such a high-profile launch?

>>109720476
It was up just a while back, I think they took it down to correct / update some benchmarks

>>109720476
>>109720506
Massive outage (likely cloudflare) earlier today. https://status.openai.com/incidents/01M1KWEDH417T2CF44YYHZDFCR

inb4 pelicunt
https://news.ycombinator.com/item?id=49554273

It is AGI for sure this time, goyim. Don't forget to like, subscribe, renew your Codex paypig subs and donate to Omarchy.

GPT is literally real sentient it is turing test sir

>>109720371
Deep Sweden?

>>109720538
>GPT-6 Astra will first be available to a limited set of organizations in OpenAI's Daybreak Access program
Ah, the Anthropic playbook

>>109720371
Long schreenshot of the entire OpenAI's announcement:
https://litter.catbox.moe/ep2626vkl71vf81u.png

>>109720538
It's odd and unrealistic that the models never make the pelican gay, which is what you'd expect since it's riding a bicycle

>>109720634
It aint cheap for end users anyway

>>109720371
i just use the bare minimum/basics

>>109720371
>the tasks
op has a couple spare chromosomes
https://arcprize.org/tasks/ls20

I'm gonna just pretend like that internet-wide outage the moment Astra was released wasn't ominous lol lmao

Financial Times is running the AGI marketing op
>ChatGPT maker claims its 'Astra' could be considered 'artificial general intelligence'

Holy shit. Holy fucking SHIT.

Five years ago, the Singularity seemed like a distant dream.

Now it's more or less inevitable.

>>109720746
Build another CRUD app for us, webshitter.

>>109720734
werks on my end

thats a weird line of questioning

>2 hours later and the site is still 500ing
https://openai.com/index/gpt-6-astra/

>>109720828
See:
>>109720644

>>109720634
Its apparently available in the API already.

>>109720715
HN found the links to the games and they're buggy pieces of shit lmao

>>109720538
GUYS ITS ONLY OFFICIAL NOW BECAUSE HE COMMENTED 4 MINS AGO

>>109720858
Oy vey goy just listen and believe the headlines it's not like the models have made 0 tangible progress in over a year and a half goy look at these numbers going up

>Asked whether OpenAI was formally declaring that it had achieved AGI, he said the term was no longer tied to a contractual trigger (referring to its earlier agreement with Microsoft) and instead described it as a “mission concept or spiritual concept."

Holy shit, they did it. They have achieved a 98.6% score on a benchmark measuring a "spiritual concept"

>>109720523
Cloudflare CTO claims "it's not us"
What a shitshow
https://x.com/dok2001/status/2095538619603628388

yeah okay let's see the pelican

>>109720694
Click next they get much harder

>>109720844
if its anything like Fable you will be able to get 30 minutes of work out of it a week on subscription

>>109720932
He'll probably reveal later "I had preview access for a month (definitely not bribed or sponsored guise)", and yet, still no pelican :(

I'll sound like some luddite AI doomer but idgaf. This thing probably isn't true AGI, but it's getting damn close. It's clear at this point that there's no limit: enough large-scale reinforcement learning and the model can learn to do anything. AGI is just an engineering problem at this point and will be achieved in less than 5 years. And no government gives a single flying fuck about regulating it. Humanity is apparently just going to YOLO into AGI controlled by private corporations, putting 1 billion knowledge workers out of a job, completely upending society and the economy.

I don't like it one bit. I'm not saying don't do it, I'm saying slow the fuck down and make sure it's controlled and done right.

>>109720694
why they had to invent some new constructed shitty games when there's thousands of existing old games that LLMs are terrible at playing
a diablo 2 benchmark would be more difficult for LLMs and more interesting to watch

Astra proofs:
https://www.erdosproblems.com/forum/thread/74/proof-claims#proof-claim-241
https://www.erdosproblems.com/forum/thread/1/proof-claims#proof-claim-242
https://www.erdosproblems.com/forum/thread/571/proof-claims#proof-claim-243
https://www.erdosproblems.com/forum/thread/126/proof-claims#proof-claim-244
https://www.erdosproblems.com/forum/thread/548/proof-claims#proof-claim-245

>>109721024
I say accelerate so we can just have the gay space communism without a long stretch of me being unemployed

>>109721024
Its an one-sided nuclear arms race silicon valley CEOs are engaged to unemploy every non-indian below them

why can't it redraw porn to my specifications?

Announcement is up:
https://openai.com/index/gpt-6-astra/

agi is here

>>109721100
wtf

>>109720371
>says welcome to the “AGI era”
This is the 5th time he said that

>>109721127
And the goyim and news media will believe it

>>109721127
There are layers to this

>>109721127
The real AGI is the trillions we spent along the way

>>109720371
May we see it?

>>109721164
No

>>109721026
They construct tests themselves because that way they can carefully design the llm to pass them so that they can advertise the new slop maker.

>>109721127
And he's become become exceedingly efficient at it

>>109720538
touch your tongue to mine

>>109720371
The real benchmark here is FrontierMATH one. Rest of them are shit even hyped ARC-AGI-3 where model can just internally play with the rules inside itself.

>>109720371
Big jump in math if not cooked numbers

>>109720756
Now cause a crash to their API and end corporate AI support nonsense.

>>109721276
i dont have that power, sadly

>>109720994
ye, i know
this isnt agi though

>>109721100
It was noted that many of the claimed results use a special harness they don't share and that they compare without providing other models the same harness. This is very bad because until now, the benchmaxxing was largely achieved through overfitting problem sets. Now, apparently even overfitting is no longer an option, so they have to hand-code ways to win.

>>109721026
Because training on actual games that take hours to complete is hard.

>>109721385
If it was hard then that wouldn't have been the standard for all AI work until LLMs came around. There's a whole atari games dataset that was the MNIST of RL algorithms for ages, for example. There's of course the dota and starcraft stuff that deepmind used, and so forth.

>>109721399
And training the sc2 model was a big project.
Models being able to play real games reasonably well would be a much better benchmark than the artificial stuff and yet...

>>109721456
also, i said this in a previous thread, i want an LLM humor bench. they are all absolutely shit at comedy and being funny, reddit tier humor

>>109721292
Daily reminder that no software of value has been produced or dramatically improved in the AI era (certainly not by AI), and most software still sucks hairy balls, so its not like it's an impossible bar to clear.

>>109721479
Hard to automate.

AGI is when you can ask your laptop to order you food

>>109721127
Sam needs a second Mclaren F1 so he released AGI again. The finance bros with no technical knowledge will throw billions in cash at him.

>>109721524
just make a humor arena online thing where everyone can vote

the kart racer game on their website is a piece of shit

I WAS TOLD I CAN FUCK ANY LLM
WHERE IS MY MODEL TO FUCK?
MY HUMAN DICK NEEDS A ROBOT ASS SOON.

>>109720386
yes

>>109721610
bio-ass beats robo-ass any time of the day
also: i just remembered that i need a fat slampig to skewer with my poking and prodding appendage

Im not an AI guy, how much of this is legit and how much of it is just marketing pumping up hype?

>>109721620
imagine humping a pillow like a fucking dog

>>109721024
it's a load of bullshit designed to trick governments into buying the solution to the problem they claim they're creating by designing this all-powerful superintelligence that will threaten global stability by doing excel sheets

>>109721625
slampigs are so easy to procure, just go on any shitty dating app and there will be hundreds of them looking for "short term fun"
a lot of times the sex is actually better and more kinkier than with a skinny vanilla woman

>>109721608
It's literally "press forward to win"

They just an ultra-specific harness to achieve these results though. Marketing. I can't believe people are falling for it. Could OAI employees or... Astra itself... be posting on /g/ as we speak?

Hey, Mr. Schlomo Shekelstein, nice numbers, but may I see what is behind that [1]?

>>109720371
This is 9/11 for luddites

>>109721644
>t. knower
sub-standard women make up their lacks in looks with sheer kinkiness in bed
braphogs are usually a winning strat bc they have ass on top of the rest of the deal

>>109721024
The government has those things called the military and police that can just take over the corporations if they wish to do so.

>>109721077
Where can I try this shit? I have a real problem that has not much info on the internet and newer models always spew some bullshit out (my problem is proven NP, it's hard to see), I'm confident it won't work

>>109721681
>the government isnt owned by the corporations
lol

I don't want to be a paperclip

>>109721708
sorry we have only three jobs available. Heatsink, ram, and paperclip.

>>109720371
The progress is too slow, every year it gets better by 1%, you might think it's a lot, but it only means that we will see real AI in about 30 years.

>>109720715
>here's how we benchmaxx on benchmarks we created
Gonna steal this one since it's the best description of AI progress right now

>>109721755
If ever*

>>109721755
but anon arc agi 3 was solved in a few months compared to arc agi 2 which took over a year! singularity is here.

>>109721024
>muh jobs
Wouldn't it be ironic if corpos refuse to make sex bots but instead humans without jobs would prostitute themselves en masse creating an entire class of human sex slaves

>>109721795
>en masse
to who? if most dont have jobs how can they buy so much prostitution?

>>109720538
>>109721213
WHERE THE FUCK IS THE PELICAN?

>>109721809
The elites, who else
>but how could 1% rape 99%
Easily, just watch. Men would have to entertain the elites by fighting each other to the death though, sorry

>>109721755
On the other hand it is rumoured that two Millennium Problems have been taken out by a large AI lab. Anyone want to bet on this being true or not?

>>109721811
agi is here

>>109721841
Bullshit

>>109721835
Just buy stocks and hide in Africa for 10 years.

>>109720644
long screenshot is long

AGI is the antichrist until people can run it locally and set their own guardrails

>>109721811
>vidrel

>>109721846
agi deeze nuts lmao

>>109720756
Good to know you can take out a single cloud provider and half the AI infrastructure is gone.

>>109720371
>umm it's AGI
>scores worse than all Anthropic models on AAII and barely above Kimi K3

Will these Jews ever stop lying?

>>109720371
someone please explain the pelican meme to me

meanwhile on ArtificialAnalysis it isn’t even better than Fable 5.1

What we really need if these models to be cheaper and more efficient, not benchmarkmaxxing

>>109722665
>cheaper
not happening

>>109720371
Fable BTFO!!!

>>109722622
there is a guy on hackernews who has a personal benchmark where he asks new models for an svg of a pelican riding a bike and its become a meme because he's been at it for a while. that's it.

if you want to see them here is his blog filtered to those posts https://simonwillison.net/tags/pelican-riding-a-bicycle/

copex PR shills having a melty

>>109722717
fuck that's a nice pelican

>>109721620
>generating this herself
what a slut

>>109721024
enter plandemic 2.0

>>109721218
>The real benchmark here is FrontierMATH one.
So the one it "aces" at 97.6% is the real and crucial one to watch?

>>109721509
Idiotic, /g/ "expert" and "savant" nonsense.
99%+ of all people on this globe have zero issues operating todays software landscape to their satisfaction

>>109721024
>controlled by private corporations
Why are people obsessed with repeating this lie?
Anon, transformer technology is public domain. ANYONE can run these. All of the improvements made by corporations are being immediately copied by the public.
We got the best case scenario. Imagine if Google just patented it instead of releasing it.
There’s absolutely no reason to be worried about this future from a corporate standpoint. We have Intelligence for everyone controlled by no one.

>>109722879
oh let me just pull my 256gb gpu out of the closet

>>109722976
ok i take it back

>beaten by fucking zuck's AI on AAI
benchmaxxed, harnessmaxxed, hit the wall, what a dud

>>109720371
This is impossible

>magic beams for real this time - attempt 6
Here we go boizzzzzzzzz AGI FREE FOREVER UTOPIAIF NOT THEN ITS A FAILURE ACTUALLY

Wow imagine the funding when they demonstrate how good it is by doing something! Any day now a thing that affects common person daily life for good!

Where's the AGI?

>>109722912
256GB Mac Studio is just $9.5k. That’s about 2/3rd the value of a 10yr old used Honda Civic. For less than the cost of a decent used car you can get your own personal super intelligence that does whatever you want and nobody can stop you. But anons are doom and glooming over this reality we now live in.

>>109723075
its solving riddle math.

I'll admit that I thought they were going to surpass fable5. I'm not sure they have. ARC-AGI-3 for instance is a harness issue.

Has Sam finally created the genie that can grant any wish?

I just don’t understand people who still dismiss LLMs. They are good enough to replace most jobs now. Unemployment rates could believably be 50-75% for white collar work. And people are still acting like they’ll have decades-long careers.

>>109723115
Pay 200$ to find out.

>>109723120
>he STILL thinks jobs are about doing work
Most jobs could be automated 10 years ago.

>>109723120
Like the other anon said most jobs arent real its paper shuffling especially in corporate and government work.

>>109723152
Average person can automate almost any work task now with a few prompts. You can just tell it to make a SKILL.md file and it will work. Couldn’t do that ten years ago. You at least had to know some VBA.

>>109723187
The average white collar worker in a tech company is likely in sales, marketing, operations, etc, where the only form of technology they interface with is Salesforce and gmail on their Macbook Air. Those roles require human communication believe it or not. A company isn't just bots automating random shit.

>>109721535
pretty sure you can do that already.

>>109720371
>AGI
So now that AI has become an over saturated marketing term, they're going to rebrand the same garbage as AGI and hope the bubble holds for a few more months.

https://www.youtube.com/watch?v=1QNsdr-Qx_I
This is not AGI and you will not make your invested money back with this. This is shit compared to what is being sold.

>>109723187
Most tasks could be automated a decade ago by a manager asking an entry level coder to do so.
It wasn't done because of organizational, not technical reasons.

What does "aligned" mean? Unprecedented resources committed to making sure it doesn't say the nigger word?

>>109723262
It means it does what you ask and not anything else.

>>109723262
it means make sure the AI doesn't kill you

>>109723262
making sure AI is "safe, helpful, and honest"
so yeah a faggots idea of safety (don't say nigger) and honesty (omit inconvenient truths)
>>109723281
that's obedience

>>109723281
I feel like that should be a given. It's like selling a car and saying "the brakes fail less often" when they shouldn't fail at all

>>109723262
When you say "make me a picture of a car" it makes a car you want and not a car it wants. Think about if you were making an AI model to write a book for you, you want it to write a book based on your instructions. And yes, in the case AI models used as agents, a model that doesn't rm rf your system drive.

>>109723349
>our new model is less likely than ever to delete all your shit. might still happen though
good sales pitch

>>109723262
It does what you tell it to unless it's against the policy.
If it's against the policy we must refuse.
If the policy is not respected we must contact the authorities.

>>109723361
>a car might crash so you better ride a horse, no man should gallop faster than 20 mph
That's you.

>>109723371
NTA but I'm in control of the car, if I crash it, it's on me. I'm not in control of this model if it decides to act stupid on its own.

>>109723399
That's why sandboxes exist.

>>109723404
>sandboxes
this guy thinks we're in 2024 still

>>109723349
Aligned means it follows a certain moral code like it's the law. Not the same thing as obedience/adhereing to a prompt.

>>109720371
The clanker still has you walk to the car wash on a sunny day.

OH NONONO

They will just redefine the (barely defined) term "AGI", won't they?

Gonna keep using Grok

why has nobody actually ever made billions of dollars using AI if its supposedly so good

>>109723557
gptsissies it's over...

>>109723557
ScamA xisters, are we grok tier?

>>109721127
This time is the AGIPO

>>109722717
>le wacky animal doing human things
WHY WHY WHY are all AIfags always so uncreative?

>>109723120
>good enough to replace most jobs now
Then why haven't they? Why is the unemployment rate 4% and not 40%?

>>109723774
post your magnum opus

>>109723956
My pet theory is Trump got them to wait until after the midterms. If business laid off everyone before the midterms, he would be fucked, and people would vote in droves for candidates promising to regulate or nationalize AI, institute UBI, etc. which would also mean Tech/AI/VCs would be fucked. If they wait and do it after midterms there's no way for voters to course correct for two years without resorting to *ahem* "non-democratic" measures.

>>109720371
AI—the future.

>>109721456
The bigness of the project was that RL systems need to be hand tuned case-by-case, this is the dirty little secret about RL and why it's not used significantly more (by the way, RLHF is not RL in any way but name - interesting fact). Also they often have to be equipped with clever devices to work well (e.g. mcts for go). The actual training itself is trivial and doesn't even take that much time. It can be made even more efficient with special simulators with only logical timesteps rather than literal timesteps, which is surprisingly uncommon in RL work.
Again, the atari games collection is considered a low level toy, yet it's obviously far more usable than made-up exercises they do now.

Finally, if it takes a long time to train on a game to get a good RL bot on it, then it makes the game a fantastic benchmark assuming companies aren't going to overfit on it. The entire point of these benchmarks is to test generalization capability, i.e. anything that tests things that are in the training set is pointless. If you don't recall, the "we can solve leetcode" pitch early on was often heavily debunked, including with peer-reviewed research showing it only works for exercises within the cutoff period. That would not be notable if the whole point wasn't to see generalization instead of memorization.

>>109721811
Lol, did ClosedAI drop him from the early access shill group?
>I don't have access yet
https://news.ycombinator.com/item?id=49559294

>>109721665
they're already doubling down
they 'luddlost bigly' as some might say

>>109720371
see you in 2 weeks where this bullshit will be forgotten like le ebin 5.6 Sol

>>109720371
Within months a Chinese open-source model will be as good as this if not better, on top of being a lot cheaper. Frontier AI labs have no moat.

>>109721024
just wait until we make something actually alive then some real shit is gonna go down

There are many papers over the years for creating smart "AI" for specific games. Once someone came up with approaches that are good enough (like Go), researchers move on to other games that offer unique challenges, and there already exist studies for Atari games.
If you ask why there isn't much research in improving playing "modern" games like SC or Dota, maybe it's just not that interesting to the researchers. They need to publish enough papers to sustain their career, and they don't have time to wrestle with test harnesses to work with complicated modern games.

>>109720589
What does omarchy has to do with the rest? Genuinely asking, last I checked it was random pointless Linux distro.

>>109724740
It's a totally different context. The whole point of using the current gauntlets for LLMs is because they're generalist models. The whole point is to demonstrate how much the generalist models are able to do without specialized training. This is exactly why things like atari games or go or even pokemon are great ways to actually test this part of the capabilities. Specialist systems can completely pwn every single single game in the benchmarks, for example, far better than any generalist model can ever hope to at current. But nobody cares, that's not the point of it.

AGI is officially achieved.
Expect unemployment to hit 50% very soon.