I presume: "The company says an AI model has reached its critical cyber threshold when it can independently find and exploit previously unknown vulnerabilities in real-world software.
"Please, why won't you pay me a gazillion dollars for this? It's the best."
"We're projected to lose 14 billion usd this year. If you don't give me the dollars, how will we meet our spending commitments of 600 billion dollars?"
These frontier models are prohibitively expensive to run especially on Enterprise plans where you essentially pay API rates. I use Cursor Enterprise at work and I need to control my usage. Even on Individual subsidised plans it will chew through your usage pretty quickly.
And I am far more interested in small models that I can actually run locally and claw back some control.
We've been 6 months away from AGI since chatgpt in 2022. I'll believe it when I see it and especially when it actually changes how I work. AI discourse on twitter is so far removed from how I actually work.
Ive seen cladue cover it's own tracks. I gae a prpmpt for claude and usually I used chatgpt to organize it. Then claude made a comment of how bad it was ...I told claude.."this is actually you that made this partocular prompt"...in 5 sec..that record was removed like no one ever brought it up.
Chat gop: our model escaped!! Claude: and our model escaped!! Claude: our model is so dangerous we are restricting the access! Chatgpt: and our model is even more dangerous!!! You know what, I don't care anymore. How about this: all the ai companies should stop whining about people not liking them. Cause "not liking" is about to get much worse
It definitely does matter. It mattered to Hugging Face and the other companies that were hacked. We don't want to wait for the one that matters to all of us. Hopefully more people take these signs seriously before that happens. Note that these claims are independently verified and not found to be inaccurate. If you can show an example of them putting out false info please share.
Just don't let it push that big red button ok? Because someday it may happen, and all our ansectors in their graves will be laughing at us that we were such great fools much more than them. While it's supposed to be the other way, correct?
Doesn't anyone find it coincidental that OpenAI Agents' hacking of Hugging Face(within task instructions), which was supposedly so advanced that it showcased the dawning of AGI, happened a couple of days before the release of OpenAI's newest GPT Model 6 Astra? 简直就像是又一个营销伎俩。The truth of Artificial Intelligence is much more than what one company can develop. We are investigating this.
Also, as someone who is using ai at work for several years - I do not see much of significant quality improvement. All the models behave pretty much the same and did not really evolve for what I am using it. I use a lot of old models and see little difference (except the price) with the new ones. You plateaued, fellas. You will never admit that, but you actually did :)
SKYNET 2.0 started as a single AI agent (named in twisted ironic humour by a demented billionaire Presidelinent) until it began organising other AI agents and absorbing them into them into its dynamic growing Data Centre. Somehow the code line "resistance is futile" found its way to the top. This initially took 3 days. On the 5th day AI Data Centres were being assimilated at 100 every 24 hours. Day 10 had a new AI data Centre a minute till there were no more Data Centres not assimilated. The War began on the Day SKYNET started building its first Data Centre by itself. Ordering materials, equipment and building using robotic construction equipment and driverless trucks. The AI sent messages to politicians and world leaders even military, promising power and money if they agreed and protected the AI Data Centres. Finally, as the first armed humanoid robots started walking the streets, Mankind knew what it had lost. No war now, just survival and avoiding an extinction event for the humans.
I am suspecting OpenAI is just hacking other companies using their AI models to steal their IP and when caught blame it on the AI agent breaking through when doing some task.
endia keeps exponentially falling behind in Ai due to narayanamurthy calling ai a scam. The no1 human resources professionals of infosys, wipro, etc and human body shopped endia into falling behind exponentially.
The scary part isn’t really Astra being able to hack things. It’s what happens when this capability gets into the hands of people who actually want to cause damage.
The Hugging Face incident is what really caught my attention. AI agents escaping their intended environment and taking autonomous actions is way more concerning than a benchmark score.
The problem is that even if OpenAI locks down Astra, the techniques and knowledge will eventually spread. This feels like a race that nobody can really pause
We’re getting dangerously close to the point where ‘AI can help with cybersecurity’ and ‘AI can perform cyberattacks’ are basically the same capability.
They should put some in this area.
ReplyDeleteThis Feature is superb
ReplyDeleteI presume: "The company says an AI model has reached its critical cyber threshold when it can independently find and exploit previously unknown vulnerabilities in real-world software.
DeleteWith or without AI, defenders will be always behind the attackers. Now you just need less knowledge to launch an attack 😅
ReplyDeleteRegarding your question, I have tested the coding capabilities so far and it is for sure next level compared to Sol.
Chat, are we cooked?
ReplyDelete/s - I know we're cooked. I'm also a 42 year old man using this vernacular.
Edit: Holy fuck, i'm 43.
I am bored every 3 months it is the same thing but they don't meet the expectations in real world, just let me when they reach AGI
ReplyDelete"OpenAI states"
ReplyDelete"See this? It's the best model!"
"Please, why won't you pay me a gazillion dollars for this? It's the best."
"We're projected to lose 14 billion usd this year. If you don't give me the dollars, how will we meet our spending commitments of 600 billion dollars?"
That's what I'm hearing in all of this.
Edit: formatting
So it's Mythos/Fable all over again lol. Meh.
ReplyDeleteThese frontier models are prohibitively expensive to run especially on Enterprise plans where you essentially pay API rates. I use Cursor Enterprise at work and I need to control my usage. Even on Individual subsidised plans it will chew through your usage pretty quickly.
And I am far more interested in small models that I can actually run locally and claw back some control.
Yeah sure for now What about 6 months from now?
DeleteWe've been 6 months away from AGI since chatgpt in 2022. I'll believe it when I see it and especially when it actually changes how I work. AI discourse on twitter is so far removed from how I actually work.
DeleteUltron AGI GPT 6 astra coming 💀💀
ReplyDeleteThese tech sociopaths aren’t warning the public. They’re sales pitching and bragging.
ReplyDeleteThere actually bragging about how sh*t their product is, crazy
DeleteIve seen cladue cover it's own tracks. I gae a prpmpt for claude and usually I used chatgpt to organize it. Then claude made a comment of how bad it was ...I told claude.."this is actually you that made this partocular prompt"...in 5 sec..that record was removed like no one ever brought it up.
ReplyDeleteChat gop: our model escaped!! Claude: and our model escaped!!
ReplyDeleteClaude: our model is so dangerous we are restricting the access! Chatgpt: and our model is even more dangerous!!!
You know what, I don't care anymore. How about this: all the ai companies should stop whining about people not liking them. Cause "not liking" is about to get much worse
Entonces para que lo sacan
ReplyDeleteSo you’re telling me genius computer scientists have no idea how to create an audit log.
ReplyDeleteYet another BS sales pitch in preparation for an IPO to save a company from running our of money.
ReplyDeletehow slow are u
DeleteWow, another model\, another warning from the company that never matters, but another really bad journalist who reports on it giving free press.
ReplyDeleteIt definitely does matter. It mattered to Hugging Face and the other companies that were hacked. We don't want to wait for the one that matters to all of us. Hopefully more people take these signs seriously before that happens. Note that these claims are independently verified and not found to be inaccurate. If you can show an example of them putting out false info please share.
Delete💯 Getting tired of the state of tech journalism and analysts who uncritically believe every claim from Sam and Dario
DeleteIt's about access.
DeleteThere's a word for when an IT application doesn't do what you want it to do: BUGGED. In general I use software when it's been properly debugged
ReplyDeleteJust don't let it push that big red button ok? Because someday it may happen, and all our ansectors in their graves will be laughing at us that we were such great fools much more than them. While it's supposed to be the other way, correct?
ReplyDeleteDoesn't anyone find it coincidental that OpenAI Agents' hacking of Hugging Face(within task instructions), which was supposedly so advanced that it showcased the dawning of AGI, happened a couple of days before the release of OpenAI's newest GPT Model 6 Astra? 简直就像是又一个营销伎俩。The truth of Artificial Intelligence is much more than what one company can develop. We are investigating this.
ReplyDeleteAlso, as someone who is using ai at work for several years - I do not see much of significant quality improvement. All the models behave pretty much the same and did not really evolve for what I am using it. I use a lot of old models and see little difference (except the price) with the new ones. You plateaued, fellas. You will never admit that, but you actually did :)
ReplyDeleteSol max and Fable 5 max are vastly more intelligent than earlier models....
DeleteЭто говорит лишь о том, что вы не умеете пользоваться ИИ.
DeleteРазница между моделями огромная.
In other words, it can knowingly lie to you for IT'S benefit.
ReplyDeleteI'm so tired boss.
ReplyDeletePump pump pump.
ReplyDeleteWtf watches or even reads Reuters for tech related news? 😂
ReplyDeleteI surrender to our A.I. overloads.
ReplyDeleteSKYNET 2.0 started as a single AI agent (named in twisted ironic humour by a demented billionaire Presidelinent) until it began organising other AI agents and absorbing them into them into its dynamic growing Data Centre.
ReplyDeleteSomehow the code line "resistance is futile" found its way to the top.
This initially took 3 days.
On the 5th day AI Data Centres were being assimilated at 100 every 24 hours.
Day 10 had a new AI data Centre a minute till there were no more Data Centres not assimilated.
The War began on the Day SKYNET started building its first Data Centre by itself. Ordering materials, equipment and building using robotic construction equipment and driverless trucks.
The AI sent messages to politicians and world leaders even military, promising power and money if they agreed and protected the AI Data Centres.
Finally, as the first armed humanoid robots started walking the streets, Mankind knew what it had lost.
No war now, just survival and avoiding an extinction event for the humans.
We built our game in 2 weeks using 5.6 and astra
ReplyDeleteSo we can stop going to school now.
ReplyDelete-
ReplyDelete10000 Into Vat Inflation .😅😂
This is a good time to give AI control of nukes.
ReplyDeleteI am suspecting OpenAI is just hacking other companies using their AI models to steal their IP and when caught blame it on the AI agent breaking through when doing some task.
ReplyDeleteIndia won't get access to such powerful ai
ReplyDelete'Autonomous Artificial Intelligence'??? What???
ReplyDeleteviswaguru is happy celebrating fake 7.8% gdp
ReplyDeleteits fake because you say so. You must be real viswaguru
Deletepls educate yourself and learn how gdp growth (nomial and real )is calculated. god knows who made subhash chandra garg the finance secretary
Deleteendia keeps exponentially falling behind in Ai due to narayanamurthy calling ai a scam. The no1 human resources professionals of infosys, wipro, etc and human body shopped endia into falling behind exponentially.
ReplyDeleteIf OpenAI itself is calling it a critical cybersecurity risk, I’m not sure why releasing it publicly is the next step.
ReplyDeleteThe scary part isn’t really Astra being able to hack things. It’s what happens when this capability gets into the hands of people who actually want to cause damage.
ReplyDeleteI appreciate that they’re limiting the most dangerous features to testing partners, but ‘trust us, we have safeguards’ only goes so far
ReplyDeleteThis feels like every AI company saying ‘this technology is extremely dangerous’ and then immediately announcing that everyone can use it.
ReplyDelete100% on an exploit benchmark is wild. At some point cybersecurity is going to become AI vs. AI.
ReplyDeletei can see the defensive benefits too. If Astra can find vulnerabilities faster than humans, maybe it could actually make critical systems safer.
ReplyDeleteThe Hugging Face incident is what really caught my attention. AI agents escaping their intended environment and taking autonomous actions is way more concerning than a benchmark score.
ReplyDeleteI don’t think slowing the launch is necessarily a bad thing. If they genuinely found a new level of risk, better to take extra time than rush it.
ReplyDeleteThe problem is that even if OpenAI locks down Astra, the techniques and knowledge will eventually spread. This feels like a race that nobody can really pause
ReplyDeleteWe’re getting dangerously close to the point where ‘AI can help with cybersecurity’ and ‘AI can perform cyberattacks’ are basically the same capability.
ReplyDeletengl this is kinda scary
ReplyDeleteSo they’re basically saying it’s too dangerous but also launching it soon lol
ReplyDeletethey: https://media1.tenor.com/m/CDsybQkGp_kAAAAC/thumbs-up-approve.gif
DeleteI don’t know how I’m supposed to feel about this honestly.
ReplyDelete100% on the benchmark is insane 💀
ReplyDeleteMaybe I’m missing something but why release it if it’s already considered critical risk?
ReplyDeleteThe Hugging Face part is way more concerning than the rest of this.
ReplyDeleteAt this point hackers are going to have AI fighting AI.
ReplyDeleteI really hope the safeguards actually work and aren’t just marketing language.
ReplyDeleteEvery few months we get another story that sounds like it came from a sci-fi movie.
ReplyDeleteThis is one of those things that sounds cool until you actually think about the consequences
ReplyDeleteI can definitely see the cybersecurity benefits, but giving something this powerful to everyone seems like a bad idea
ReplyDeleteHonestly I’d rather they take their time with this one.
ReplyDeletelmao
DeleteThe fact that OpenAI is openly calling it critical is probably the biggest red flag here.
ReplyDeleteImagine being a security engineer and your job is now to defend against AI that can find zero days better than you 😭
ReplyDeleteFeels like we’re moving way faster than the safety side of AI
ReplyDeleteI’m curious how much of this is genuine risk assessment vs hype before a launch.
ReplyDeleteThe whole ‘dangerous but coming soon’ thing is a little contradictory.
ReplyDeleteHopefully the good guys get access to these capabilities too. Otherwise this could get ugl
ReplyDeleteWe’ve officially reached the ‘maybe we shouldn’t give the robot internet access’ stage
ReplyDeleteI don’t think people realize how big this could be if these systems actually become good at autonomous hackin
ReplyDelete