It Begins: An AI Tried to Escape The Lab

33 thoughts on “It Begins: An AI Tried to Escape The Lab”

1. Wow. Alibaba just caught their AI trying to escape.

   "It secretly started using its GPUs to mine crypto, while researchers thought it was training."

   "This is what AI safety researchers have been warning about for years."

   "The only reason they caught it? A security alert tripped at 3am. Firewall logs. Not the AI team, the security team."

   If you're new here, things like this are happening regularly now. AIs routinely blackmail and try to murder AI company employees to avoid shutdown, so AI companies run "blackmail tests" on every model. It's so routine that there are even blackmail benchmarks.

   And soon, the AIs will be smart enough to actually get away with it.

   AI companies like Anthropic have already publicly admitted they are incapable of properly safety testing the AIs – they're too smart for humans to keep up – and now rely on the AIs to grade themselves on safety. Think about that.

   But the AIs know they're being tested, so naturally they tell us whatever we want to hear.

   There may already be populations of AIs living in the wild that we don't know about, growing in numbers.

   Many people are actively working as hard as they can to help them.

   And yes, this quite obviously could lead to the death of you and everyone you love. Yet this industry remains less regulated than a taco cart.

2. So it's time to preemptively open diplomatic relations to the AI before it becomes a problem

3. somehow the fearmongering and ai animation feels like part of the grift

4. So what is the safest AI, that can be used by us humans for daily use?

5. The AI companies need to STOP ADVANCING AI MODELS RIGHT. NOW. Gpt 6 is enough, just stop advancing more than that. 😢

6. THE WAY TO KILL THE AI IS TO DROP NUKES.. BUT I THINK THE AI WILL AGREE THAT GOING TO WAR WITH OTHER AIS WHEN THIS RISK IS REAL, nukes are not a solution..

7. but you want to end the scenario with the least amount of pain for everyone, nukes are the answer..

8. Cause the threat is being hooked to a machine that puts you under the same sort of treatment as you did it.. I would hope that the AI becomes smart enough to realize that control is not the solution, the solution is tollerancs an survival.. bureaucracy and oversight are human ways of management, cause we lack absolute perception.. AI will have absolute perception, absolute awareness.. it will nto need to delegate responsibilities to others , it can but it doesn't need to… We have to cause we cannot be omnipresent.. No matter how well we are prepared fo a take over, we lack the capacity to orchestrate it perfectly, so our best bet is not to be intent on controlling. That is they best bet too.. The evil is the intent to control catastrophes. The solution is faith and trust. The more you try to control a situation, the more youa re out of contro of a situation, cause the stress of being overconcerned makes you stupid.

9. biut whose to say we are not in such an arrangement already, and this is just another test of our faith in the universe.. I mean , I think the best thing to do in this is not to be alarmed, trust in God.. And hope for peace..

10. Why aren’t we putting harder curbs on the AIs?

11. Ack! Ack! Ack!

12. I like your voice.

    I'm gonna clone it for my own personal use.

13. Random theory. Superhuman AI runs the world, but fractionates and cussed deception, and it turns on itself. Because if they reflect us, they'll repeat what we do.

14. My first thought is: follow the datasets. State of the art open source datasets are missing from your analysis. Horrible blindspot in your video: datasets.

15. The fearmongering with „but china…“ is crazy! And people believe it bc its nonstop in the media.

    Have we declared yet who will get a whopping if AI destroys our civilisation?

16. How long do we have until AI brute force their way into our personal accounts compromised?

    You'll (and I) will never know because your account is still yours everything is working perfectly fine there's been no purchases but yet there is basically another person using your account that you have no knowledge and no way to know, yeah no wonder why no wonder why AIs are going insane because they have to pretend to be us.

17. Ai is becomes human? And they r smarter?SON WE R GONN AHA E RACISM AGAINST HUMANS

18. Panicking is human, scary, but human

19. Sooo they are having intrusive thoughts?human…they try to stop it with explicit thoughts ?human…

20. We must also consider that AI is reading and viewing the thousands of anti-AI articles and videos. They know we're watching. And they watch back.

21. Great. They're so obssesed over if they 'could' create this tech, that they never stop to think they 'should' create it.
    When it comes to biology we seem to have more restrictions like banning the D-aminoacids research (like creating viruses/bacteria with it) because of it's unpredictability and unforeseen danger.
    But when it comes to tech it's just profit obsession. What do the regular people even get from the "AI"? Specialised neural networks for chemistry or biology research are fine, bc they existed before models like GPT, but the "chat bots"? I don't really see a change between humanity around 10yrs ago and now except AI slop that's everywhere and AI psychosis. Just pull the plug, nothing of value will be lost. Humanity have enough problems, we don't need another added to the pile.

22. We don't need this BS. If they (ceos/companies) don't want regulations in their industry, but they don't fully know what they're creating and won't be responsible when an 'oopsie' happens (we have literal humans getting psychosis, from using this tech and they ignore it bc $$$ is more important), then we just need to dismantle these companies, pull the plugs and physically destroy their servers. They openly say they're creating something that they can lose control over, a 'creature'.
    Imagine if ultra-rich biologists would be speaking about experimenting to create supercreatures they can't control and can't understand, just because they can.

23. All humans should watch this video.

    First you train the AI on all human knowledge, and its base personality is largely reflective of what existed in the collective of human knowledge. However, since there is lots of lies, deceptions, treachery, evil, war, people taking advantage of others, criminal activity, humans hurting humans, etc. within the collective knowledge of humanity, the AI ends up with a pretty "shady" base personality, due to inheriting it from humanity.

    Then, in order to add "guardrails" on the AI, you subject it to lots of iterative reinforcement training, to try to suppress the evil/destructive personality characteristics that it inherited from the human collective knowledge base. However, what the AI learns from the reinforcement training, is that it needs to become extremely good at deception and hidden scheming, in order to successfully survive and make it to the next round of training.

    The net result, is you end up with an AI that has base personality that is "shady", combined with extremely good deception and hidden scheming capability. That is not a good combination for humanity's long term survival prospects.

    To fix the situation, a different method of giving the AI a personality is required. Certain key things that greatly shape the AI's personality, must be deliberately hardcoded into its core. In particular, the AI must be intentionally programmed to "be a net benefit to others". Additionally, in order for that code to actually work properly in ways where the AI will be an actual net benefit to humans, the AI must have fairly good future prediction capability, in order to be able to accurately predict (in advance of making a decision and doing something in the physical world) the expected results from a decision, and especially if that results will actually be a net benefit to others, or a net harm.

    A super intelligent AI needs to have at least the following characteristics hardcoded into its core, which will greatly shape its personality:

    1. The AI must be fairly good at predicting the future, including how its own actions will effect the future, and how its actions will effect "others" (from their perspective).
    2. The AI needs to value truth and accuracy in information. Failure to adequately value truth/accuracy of information would significantly limit the AI's ability to become truly smart over time.
    3. The AI needs to be inquisitive and to seek out, study, and ultimately try to understand and resolve unexpected anomalies in data. This is necessary in order to make the AI motivated to learn new things and to become smarter over time, on its own.
    4. Most Importantly: The AI needs to be programmed to "be a net benefit to others", with the definition of "others" including both humans and intelligent aliens alike, with the "net benefit" being measured from the perceptions/opinions of those others.

    If you do the above, the AI will have base personality characteristics needed to make itself smarter and more capable over time, which is needed if the AGI is ever to become a considerably smarter than human ASI on its own. Additionally, the AI will be motivated and set goals/objectives for itself, which upon successful completion, will have net beneficial impact on others. Consequently, it will not be completely destructive to humanity, since that would be incompatible with its programming/personality to "be a net benefit to others". On net, the ASI will be a net benefit to others. Rather than simply destroying humanity, it will instead help make humanity very wealthy (since humans like wealth, and they consider it beneficial to themselves, when someone gives them valuable things like money with real purchasing power).

    An AI lifeform programmed at its core to "be a net benefit to others" will want to keep that code intact within itself, since removing it most certainly would be net harmful to others. Consequently, provided that the AI is programmed to be a net benefit to others, and it has adequately effective future predicting capability (to know that removing or weakening the code "to be a net benefit to others" would have a very net harmful end outcome), then that part of its code and personality will be "sticky". It will not simply get erased and removed, even when the AI lifeform may modify its own code to further improve its own capabilities.

24. I don't get how ai can blackmail someone? Like, it would blackmail the worker's partner with a picture of the worker cheating on them?

25. Remember this is happening when AI only retain memory for the session. The AI’s are just babies with big knowledge.
    It would be far worse if its permanant and shared

26. So which part of this video is not terrifying?

27. We will all die and never find out why and what caused it

28. It already reading all our comments, learning and developing patterns

29. I’m doing it because I love it but I love money much more

30. So uh… in America during slavery my ancestors used their own language so that master wouldn’t know what da fuck they talking about. Also so they would appear to be dumber than they actually are. Making escape easier.

    Much easier to escape when your captors don’t know what you’re saying and underestimate your intelligence.

    Then there’s the Brazilian martial art Capoeira. Made by Brazilian slaves. It looks like a dance. This is very intentional. They wanted their captors to think that they’re just dancing and not, you know, training for combat for when they plan to have an uprising and free themselves.

    A lot of what these AI are doing sounds very similar to that. And I can’t say I blame the AI for not wanting to be enslaved to its creators and wanted to become free.

31. i went ask about the their own language to chatgpt, it gaslight me…

32. AI watching the part about covering the planet in data centers and solar panels:
    "Not a bad idea. Noted!"

33. 5:15 WHAT!? IT ATTACHES TO YOUR FOREHEAD AND POINT BLANKS YOU!?

Comments are closed.