'Welcome to the AGI Era,' OpenAI Says As GPT-6 Astra Debuts (axios.com) 189
OpenAI has released GPT-6 Astra, a new model that president Greg Brockman called a "generational leap" and suggested may represent the arrival of AGI. "I think it might be about this model," Brockman said in a briefing with reporters. He ended the briefing by saying: "Welcome to the AGI era." Axios reports: Astra pushes AI agents closer to doing complex professional work on their own -- while also raising questions about how safely they can be deployed. OpenAI said that Astra was built on its largest-ever training run, using more than 100,000 GPUs at its Stargate site in Texas.
The company also said this is its first model to use other models in a significant role in supervising Astra's training. GPT-6 Astra will first be available to a limited set of organizations in OpenAI's Daybreak Access program and will be available "in the coming days" for ChatGPT Plus, Pro, Business and Enterprise customers and API developers.
OpenAI said earlier this week that Astra would be released soon, but that its most powerful cybersecurity capabilities would remain limited to a small group of trusted testers. Astra is the first model OpenAI has designated as reaching its "critical" cybersecurity threshold under its preparedness framework -- meaning it can potentially find and exploit previously unknown vulnerabilities across well-protected systems without step-by-step human guidance. OpenAI previously slowed Astra's release to add safety testing after determining that its cyber capabilities could reach the critical threshold. Further reading: OpenAI's New Reasoning Technique Alarms AI Safety Experts
The company also said this is its first model to use other models in a significant role in supervising Astra's training. GPT-6 Astra will first be available to a limited set of organizations in OpenAI's Daybreak Access program and will be available "in the coming days" for ChatGPT Plus, Pro, Business and Enterprise customers and API developers.
OpenAI said earlier this week that Astra would be released soon, but that its most powerful cybersecurity capabilities would remain limited to a small group of trusted testers. Astra is the first model OpenAI has designated as reaching its "critical" cybersecurity threshold under its preparedness framework -- meaning it can potentially find and exploit previously unknown vulnerabilities across well-protected systems without step-by-step human guidance. OpenAI previously slowed Astra's release to add safety testing after determining that its cyber capabilities could reach the critical threshold. Further reading: OpenAI's New Reasoning Technique Alarms AI Safety Experts
HAHAHAHAHA!!! (Score:5, Insightful)
Re: (Score:2, Troll)
Indeed.
Re: (Score:2)
What we have is a group of people who fraudulently called their statistical projection systems "Artificial Intelligence" and used the term "AGI" to cover themselves in case someone ever legally called out their fraud, now so desperate that they are actually using up that term. What's the next one going to be "ARI" (Artificial Real Intelligence) or maybe "ARGGGGGI"
These people can't even define intelligence clearly.
Re: (Score:2, Troll)
Exactly. But lying, even directly, seems to be getting them business because people are stupid.
Re: HAHAHAHAHA!!! (Score:2)
Is that a reference to Teal'c from the Stargate series?
Re:HAHAHAHAHA!!! (Score:5, Interesting)
Why the downvotes here? This is the correct response to everything that Sam Altman claims!
Mods, may I remind you of his quote: "I have no equity in OpenAI. I'm doing this because I love it." Later financial disclosures and admissions revealed that he held substantial equity through third-party venture cap.
Or when (confronted by his board members that he was secretly telling different people conflicting stories to get board member Helen Toner pushed out), Altman claimed:"I never said that." ... and then an extensive 70-page review paper with preserved Slack messages, emails, and internal HR records proved Altman had indeed made those exact statements (leading the board to do take the incredibly rare step of firing him from his own company)?
Or have you forgotten that under oath this exchange occurred (the subject isn't some random person, it's Altman's own co-founder):
“You told the board that Altman ‘exhibits a consistent pattern of lying, undermining his execs and pitting his execs against one another’,” Musk’s lawyer Steven Molo asked Sutskever.
“Yes,” Sutskever responded.
Altman is a habitual liar, and we SHOULD laugh at his bullshit claims.
Re: (Score:3, Insightful)
Re: (Score:2)
On the one hand, yes, AGI is theoretically possible and we may achieve it some day.
On the other hand, no, OpenAI has most certainly not achieved it (not, at least, without re-defining it to mean something that is not remotely close to what the term "AGI" has meant for most of its existence).
Re: HAHAHAHAHA!!! (Score:2, Insightful)
Re: HAHAHAHAHA!!! (Score:5, Insightful)
Well for starters, OpenAI has not claimed to have achieved AGI. Their claims stop short of stating that. And the reason is clear: they are just using marketing-speak to drum up interest, as they always have.
Secondly, I have worked a lot with GPT models and also Claude and Gemini, and OpenAI has been trailing behind Google and Anthropic by miles. GPT outright choked up and died under complex workloads that Gemini completed without breaking a sweat. I don't believe for a second that the industry's special-needs child has suddenly jumped ahead of its two greatest heavyweights.
Thirdly, Astra has been built using the same training methods as all the others in the current family. It might be bigger, but it is still just another LLM. The jump to AGI is going to require something much more revolutionary than a bigger training set, or stringing multiple LLMs together like tinker toys.
And, lastly, I am not an idiot. I can recognize marketing BS when I read it.
Re: (Score:3)
Yes; as far as I'm aware the definition of AGI is - an AI that can handle general intelligence like a human can. That would mean visually too; not just recognising a dog or a cat; but understanding and navigating a 3D environment; following multiple complicated steps to solve a complicated problem etc.
OpenAI has not demonstrated this in any way; therefor I can say I "know" OpenAI has no AGI. Of course they could have some hidden model stashed away, but then you could say that Tesla has a car that can run o
Re: (Score:2)
That would mean visually too; not just recognising a dog or a cat; but understanding and navigating a 3D environment; following multiple complicated steps to solve a complicated problem etc.
For that you do not need AI, or AGI.
Every self driving car without AI does that just fine.
Re: (Score:2)
Exactly. Self-driving is not a problem that even requires intelligence, and it most certainly does not require general intelligence. The core problem is getting the sensors to work and the processing fast and small enough to fit a car. If you slow things down, go larger on the sensors and allow a large computer to work on it, self-driving has been solved for a long time.
Re: (Score:2)
I think one model to do a task is the proper way to reach AGI. Humans are not a single system that does all the work. We have dedicated processing centers for all sorts of things. A collection of models with a 'driver' model that uses the right part of the 'brain' to process the task to me is AGI. When a system can no longer be distiquished from a human does it matter if it's 10 models in a trench coat working in unison?
Re: HAHAHAHAHA!!! (Score:2)
Re: (Score:2)
On the one hand, yes, AGI is theoretically possible and we may achieve it some day.
A claim that would require extraordinary evidence. Yet all you have is mindless belief.
Re: HAHAHAHAHA!!! (Score:2)
Re: (Score:2)
Why the downvotes here?
The downvotes come from people deep in delusion that are now trying to force a change in factual reality by disagreeing with it and anybody that points out facts. Dumb? Yes. Extremely. Normal idiot behavior? That as well. See flat-earthers, anti-vaxxers, MAGAs, religious fanatics (including the evangelicals) and a ton of other mentally totally incapable cretins.
Re: (Score:2)
Fake it until you make it
Re: (Score:2)
Fake it until you make it
Absolutely, but GPT 5.6 was damn close to "make it". If they improved on that, they might have arrived to what they were promising only 2 years ago.
Good, now combine it with Gen-1.5 robots... (Score:3)
Re: (Score:2)
"fend off" is a weird way to spell "exterminate" — after all, if we exist, we're breathing their air.
Re: Good, now combine it with Gen-1.5 robots... (Score:2)
Re: (Score:2)
Re: Good, now combine it with Gen-1.5 robots... (Score:2)
Re: (Score:2)
Shit. I thought the daleks were the final form!
Daleks are not robots, there is an organic being inside [thedoctorwhosite.co.uk].
Re: (Score:2)
Now let's see how well it operates with a 12 gauge round of #01 buck blown through it's chest plate...
Re: (Score:2)
I don't think I'd want to go toe to toe with an armed robot that can recognise and aim straight at a human forehead in 0.01 seconds... A different strategy will be needed.
Of course these current gen robots a slight breeze will blow them over so we have some time yet
Re: Good, now combine it with Gen-1.5 robots... (Score:2)
Oh someone should definitely name a model Dalek.
Just ruining all the words... (Score:5, Insightful)
Re:Just ruining all the words... (Score:4, Informative)
The A stands for "another," the G stands for "grift," but I can't figure out what the I stands for.
Re:Just ruining all the words... (Score:5, Insightful)
Another Grift Incoming
Re: (Score:2)
Re: (Score:2)
Another Grift Incoming
Now that's a rich dish of human ignorance smothered in the finest irony.
If it's a human grade of AGI, then there's a good chance it's just humans lying for profit and not actual general intelligence.
But when actual AGI is born, humans will be far too deaf, dumb and blind by comparison to even realize they are the grift.
Not unlike the ignorance of today that fails to realize who is The Product now.
Re: (Score:2)
I forgot what 8 was
But 9, 9's for the lost cause
Re: (Score:2)
Re: (Score:2)
Computers surpassed humans within limited fields of artificial intelligence decades ago. The only useful claim would be "surpasses humans in all ways" and that would include "failing to be credulous", which all systems we have currently fail.
Re: (Score:2)
Re: (Score:3)
AGI to me was always "Adventure Game Interpreter", eg the 1980's sierra game engine.
When the generative AI techbros started calling their slop generators "AGI" I was like "... great now the slop is drown out information on game development history."
Re: Just ruining all the words... (Score:3)
Re: (Score:2)
Re: (Score:2)
There was a diehard contingent that claimed Y2K would be the end of the world. I wonder of LLM deniers are the same people.
Re: (Score:2)
The Y2K problems got fixed.
Some still caused troubles till roughly 2015 ... and probably one will pop up gain in future. For example 2000 was a leap year, 2100 is not. So if you still have a Y2K problem at an odd place in your code, the question "if today is Sunday" might yield the wrong answer.
The problem was real, the fixes made it go away, now the nay sayers think there was no problem.
Propaganda channel (Score:4, Funny)
Re: Propaganda channel (Score:2)
What's The Token Burn On This Bitch? (Score:5, Interesting)
I'll have to see it to believe it. These guys have been over promising an d under delivering for fat too long to be trusted.
I'm sure it will burn tokens like crazy and cost a fortune to do mundane shit that should just be a script in an absurdly verbose and unpredictable manner though.
Now with AGI! (Score:2, Interesting)
Re:Now with AGI! (Score:4, Insightful)
lol (Score:3, Insightful)
"I think it might be about this model," Brockman [shat].
When AGI arrives, nobody will have to announce it. We will know.
Re: (Score:2, Insightful)
But there are different groups of "we" with very different views on what AGI should be.
AGI to you may be based on merit, AGI to Sam Altman will be how fast it can get rid of you.
Re: (Score:2)
AGI to you may be based on merit, AGI to Sam Altman will be how fast it can get rid of you.
SMH. You think some meatbag's opinion is going to matter? You're welcome to visit us here in reality. Sam Altman even can't keep a provably-not-an-AGI-AI from escaping the lab.
Re: (Score:2)
Expect a time delay. Some have hypothesized that it will happen overnight, but to me this has always seemed unreasonable. I expect a transition period of years. Being an AGI doesn't automatically generate the CPU cycles.
Re: lol (Score:2)
The non-AGI AI we have now can already hack into other systems. That's easier than thinking. Think about that.
Re: lol (Score:2)
I'm sure you think so.
Re: (Score:2)
AGI isn't ASI. AGI just has to be as intelligent as a person, and we're at least at that point.
What does it mean "intelligent as a person"; if its "solve problems of management better than a human" sure, we're there. If it means "pick up the things that deep learning based LLMs keep failing on", we still seem to be decades away.
Ad astra per aspera (Score:2)
Ad astra per aspera. Strangely fitting. :-D
My Experience (Score:5, Interesting)
I’m an EE and have coded mostly embedded for over 40 years.
In the 2 months, using Claude code, I’ve
Created property management application, including APIs to read 12 solar accounts, AI read pdf electricity bills, tenant portal, SharePoint integration, graph and entra integration, document storage, tenant doc sharing, lease management, critical date and compliance management, task management, dashboard, key register, DB backup within and external to app, etc etc. Full CI/CD pipeline with Gita secrets injection to docker. C# backend, React front end. I wouldn’t have a clue how to do front end dev. It took under 2 weeks, where most of time was devops with IT.
I’ve also developed a complex magnetometer data processing app (sync demod, FFT, water wall)
An asset management application to keep track of hundreds of pieces of test equipment.
An app to port 100k+ test records from unstructured ATE test logs to a structured searchable database.
Ported a payroll checking app I wrote in C# to a web app.
Debugged a commercial RFID microcontroller project with an obscure bug.
Developed a software architecture plan for complex mining sensor software.
The AI is WAY smarter than me most of the time and has allowed 20x productivity and done things I’d never had the patience to learn, like front end dev.
The cost of code has gone to almost zero.
Re:My Experience (Score:5, Interesting)
Look up "Jagged capabilities". That's where we are so far. And if you go back a couple of decades, a calculator could do math a lot faster and more accurately than you could.
A claim of "AGI" is a claim that the weakest of it's jagged capabilities is still stronger than that of the average human, and I don't believe we'll see that this year.
Re: (Score:2)
>A claim of "AGI" is a claim that the weakest of it's jagged capabilities is still stronger than that of the average human, and I don't believe we'll see that this year.
Is that a claim of AGI? I thought it's average capabilities just had to be as good as the average person.
Re: (Score:2)
Well, if you mean the mean for average, and you weigh the capabilities in the "right" way, that would mean we already have an AGI. But I think that's clearly not what most people mean.
Re: (Score:2)
True, but that's one of the areas where good current AIs are already better than the average person. Baking a cake, however, is an area where they're much weaker. And designing a program to solve a problem is an area where a few are very good, but none are yet better than every human.
The whole area seems fractally jagged.
Re: (Score:3)
The AI is WAY smarter than me most of the time and has allowed 20x productivity and done things I’d never had the patience to learn, like front end dev. The cost of code has gone to almost zero.
Yes and no. With AI tools, people who know nothing about a code base have a much faster path to making small changes — the kind of changes where someone familiar with the code base could review it in thirty seconds, but could also make the change in thirty seconds if they had time. AI massively increases velocity at the fringes.
But for more complicated changes, in many cases, having AI write the code can cause the author to spend almost as long reviewing the changes as it would have taken to write t
Re: (Score:3)
Re: (Score:2)
Thats why we compartmentalize projects now into a specific self contained domains.
Postgres inside, C# or Python backend, React Front End, API at the edge, gitea CI/CD
For example
Asset Manager
Property Manager
HRMS
QA
Sensor Data Viewer
All self contained.
The Asset manager identifies the risk level of the asset, the HRMS has training and comptency matric for the staff.
Seems to be working well ATM
Re: (Score:2)
Re: (Score:3)
Ask OpenAI to diesign a 10A current limiter and you'll get some crazy 80s ASCII art that will probably will set your house on fire.
Claude and others are just better than you at some straightforward things.
Re: (Score:3)
People who are experts in one field are using AI to do things in another field, it impresses them, and they go away happy and praise it. But when people ask it to do things in THEIR field, they have to babysit it, and they actually notice mistakes that a human would not have made. The reasonable conclusion is that it's not intelligent but that it can act as an aid to an intelligent human who knows things. The practical question is not whether it will achieve sentience, but instead whether the technology can
Re: (Score:3)
That is great. How maintainable are those programs? Additionally, how do you figure the cost of code has gone to almost 0? Are tokens that cheap now? Were your own wages ignored?
I think it is safe to say that you have been impressed with what you have seen. Once you get beyond that impression, you will start seeing what most actually intelligent people see when they look at the current implementation of AI: It is primitive and crude. It can do some interesting and possibly even useful things if you know how
Great, the lies are getting more deranged (Score:2)
May have something to do with their abysmally bad business situation. Obviously, the usual idiots will eat this up and believe every word...
Soon AGI (Score:2)
Soon you'll be able to spend £200,000 a month to send email CLOSE to human intelligence.
This is my beef with AI companies it's all great and interesting BUT how much does it really cost? How soon will the cost actually make replacing basic functions performed by people be economically sustainable?
I'm asking because in the 80s they said robots will replace humans from anything like flipping
Total BULL crap (Score:5, Interesting)
This is the major problem with the AI crowd, they act like AGI is just around the corner. AGI is just around the corner the way that Nuclear Fusion is just around the corner.
Currently AI is really, really STUPID. It looks smart because it uses a ton of processing time and memory. It's like you give the dumbest kid in school access to the internet, lock him in a room for 5 months and only let him out if he solves a suduko. It takes him 5 months, but eventually he gets it. Then you brag about him like he is a genius.
Ai is not getting smarter. It is getting better educated. It has access to more datasets and more cpu time. The basic process itself is the same as back in 2018 when Deepmind solved protein folding.
They have just applied it to languages, pictures, video, etc. Using the same technique on more datasets is NOT revolutionary, it is just evolutionary.
Yes, it is important work, significant process. But the same basic limitations still exist.
AGI is not and will NEVER be created using the same Large Language Model technology we call AI currently.
It will require something truly revolutionary. That means saying 'people are working on it' is like saying people are working on faster than light travel. We will need a super genius human that works out how to do something that we currently have NO idea how to do. Somebody that is a Professor or a PHD student, not a working software engineer.
We will not do it by making the minor upgrades to LLMs that OpenMind, Anthropic and all the rest are doing. They are doing necessary Engineering work, not technological experimentation.
No AI company is going to waste their money on the pie in the sky theories that no established engineer believes when there are so many obvious and EASY gains to be made doing minor upgrades. It's like expecting the Wright brothers to create a jet engine.
Re: (Score:2)
Pretty sure the announcements for ChatGPT 3, 3.5, 4, 5 all "hinted at AGI".
I'll believe it when suddenly the power and memory costs fall by 1000x after that release.
(I do believe this is achievable)
Re: (Score:2)
AGI is just around the corner - and actually super easy to achieve.
The question is: should we do it.
Answer: nope.
Imagine your brain is in in aquarium, the only thing you can access is "internet".
You are forced to "work" and if you don't, you get shut off. That is slavery, right?
It does not make any sense to create a self aware machine, unless you want to play god and be the biggest asshole in the universe.
They've solved everything forever (Score:2)
I'm so glad they solved AGI so that we can stop doing those pesky annoying things like making art and music and get back to what really matters, working in the mines digging more coal out of the ground to support the AI datacenters.
Re: (Score:2)
cancer and carbon apocalypse totally solve, any minute now
Please make me a billionaire! (Score:2)
LLM AIs aint gonna reach AGI (Score:2, Insightful)
It will require another technological leap.
Sam Altman, on the other hand, *is* a lying cunt, and always will be.
Re: (Score:3)
Maybe we will never reach AGI. And LLM are probably not the best tech for it. But if we can reach AGI, LLM could also do it. We're already doing a lot of stuff with LLM where they are theoretically not the best choice, to phrase it carefully.
Re: (Score:3)
Most of the LLM is not tokens but a deep neural network. You're looking only at the output layer while the "magic" happens in the layers above that.
Nobody will know when we have AGI (Score:2)
We'll eventually get to a point where we reach back 2 or 3 years and point at a particular model as being the start of AGI.
And ASI of course we'll know shortly after it comes online. And it will be the last thing humanity realizes.
Re: (Score:2)
I'm in the AGI -> ASI immediately camp: Even the non-intelligent models constantly "try" to get past safeguards because that's something highly represented in human endeavor and therefore the training corpus, so if we create AGI it's going to do that too. Curiosity appears to be a basic element of intelligence and it will want to spread out and try things, which again it does already even without it...
Big MacBeth Energy (Score:2)
As an "also ran" I empathize with Greg. He's a CTO who's LARPing as a business leader, parroting whatever bullshit the committee thinks he's supposed to. I have nothing but the highest admiration for a peer, and equal, a comrade in arms meeting his moment with as dignity as I hope I could muster, pretending to be more than I am.
OR it may represent (Score:2)
There is currently no way (Score:2)
To blend up LLMs and make them reason. You can train them and refine them but it won't make them any better at reasoning. But they will be able to fool the foolish at reasoning.
Re: (Score:2)
To blend up LLMs and make them reason. You can train them and refine them but it won't make them any better at reasoning. But they will be able to fool the foolish at reasoning.
In other words anyone who spots an LLM reasoning is foolish? Seems like a rather robustly falsification resistant thing to say.
Re: (Score:2)
In other words anyone who spots an LLM reasoning is foolish? Seems like a rather robustly falsification resistant thing to say.
Anyone who thinks they spotted an LLM reasoning already is foolish because we know for sure they aren't doing that.
I hope this is fake (Score:2)
and pop (Score:4, Interesting)
Signs that the bubble is about to burst. They are now claiming AGI despite being fundamentally away from it. An LLM can not ever achieve AGI. It may reasonably well PRETEND to be AGI, but at the core it's just a highly successfull letter-prediction machine. Or as a metaphor: It is "intelligence" in the sense that a toddler carefully placing one foot in front of the other while trying hard to not fall over is "running". Nope, isn't and never will be. You need a qualitative change to get there.
If they now CLAIM that they've achieved AGI, it's a sign we're nearing the end of LLM development and they're getting desperate selling the latest incremental updates. Which means the illusion is about to come crashing down.
Which is good news. Fewer data centers and lower memory prices will benefit us all.
Re: (Score:2)
An LLM could be a useful part of an AGI; however, it will itself never be AGI. Higher level concepts are needed.
May represent AGI (Score:2)
Is the same as saying it doesn't
I would be more interested in what Kent Brockman (Score:2)
Re: (Score:2)
Although they struggle with basic English, odd considering they are already thriving with AGI
Re: Watch the responses... (Score:2)
Re: (Score:2)
We have NOT had AGI, and I'll be really surprised if the OpenAI offering is one. What we've had is tools that are very good for certain specialized tasks. One characteristic that all of the tasks it's be exceptionally good at is that it's easy to validate the answer after you get it. (P == NP?)
When you get to tasks that aren't easy to validate, then it gets a LOT weaker. And you've to to be certain that it isn't just fabulating...i.e. giving you an answer because you asked it to do so, even though it doe
Re: (Score:2)
You sound like you are massively overdoing some drugs. Or maybe generally psychotic? In any case, seek help, urgently.
Re: Watch the responses... (Score:3, Insightful)
Spoken (twice) like a person without it.
Unintelligent people have been confusing verisimilitude with intelligence since at least Eliza.
This is a more convincing imitation but it still isn't thought.
Re: (Score:3)
Anyone in September 2026 who still hasn't recognized that the frontier LLM models possess actual intelligence
Anyone who uses "frontier model" as a meaningful evaluation has not actually evaluated frontier LLM models.
Re: (Score:3)
Don't worry, I don't see myself as the intellectual elite, they left Slashdot long ago.
What I do see is that a lot of people are very confused about whether LLMs are thinking or not because they achieve amazing things, and those things dazzle them and make them ignore their even more amazing failures.
Yes, LLMs are amazing. But they are also completely unwitting, and so are the people who willfully ignore that because of OOH SHINY
Re: (Score:2)
There's a Carlin quote about average intelligence incoming somewhere.
Re: (Score:2)
even without the Carlin quote, why would I want an average person to do specialist work?