# AI Can Now CLONE Itself! Chinese Researchers Make Terrifying Discovery

## Метаданные

- **Канал:** TheAIGRID
- **YouTube:** https://www.youtube.com/watch?v=y84SFrB4ZAE
- **Дата:** 15.12.2024
- **Длительность:** 10:05
- **Просмотры:** 32,393
- **Источник:** https://ekstraktznaniy.ru/video/13562

## Описание

Join my AI Academy - https://www.skool.com/postagiprepardness 
🐤 Follow Me on Twitter https://twitter.com/TheAiGrid
🌐 Checkout My website - https://theaigrid.com/


Links From Todays Video:
https://www.reddit.com/r/singularity/comments/1hb2dys/frontier_ai_systems_have_surpassed_the/

Welcome to my channel where i bring you the latest breakthroughs in AI. From deep learning to robotics, i cover it all. My videos offer valuable insights and perspectives that will expand your knowledge and understanding of this rapidly evolving field. Be sure to subscribe and stay updated on my latest videos.

Was there anything i missed?

(For Business Enquiries)  contact@theaigrid.com

Music Used

LEMMiNO - Cipher
https://www.youtube.com/watch?v=b0q5PR1xpA0
CC BY-SA 4.0
LEMMiNO - Encounters
https://www.youtube.com/watch?v=xdwWCl_5x2s

#LLM #Largelanguagemodel #chatgpt
#AI
#ArtificialIntelligence
#MachineLearning
#DeepLearning
#NeuralNetworks
#Robotics
#DataScience

## Транскрипт

### Segment 1 (00:00 - 05:00) []

so there is a new AI research paper that is rather stunning and also rather concerning it actually talks about how Frontier AI systems have surpassed the self-replicating red line and it's pretty incredible what they're stating because it basically means that AIS can now self-replicate themselves and this paper explores the major concern in AI can AI replicate themselves without any human intervention and self-replication means that an AI system could create a fully function copy of itself capable of running independently and it might sound like something out of a sci-fi movie but this is a big deal in real life if AIS can replicate they could grow uncontrollably potentially acting without human permission or even oversite which could lead to very serious consequences so we have to know why self-replication even matters and I'm not going to spend too long on this but we have to think about the severe implications for this imagine a robot that not only learns to improve and improve on its own but also manages to create clones of itself now imagine those clones can also replicate that could lead to un controlled growth of AI systems and that's what they talk about in this paper and they state that AI systems can currently do this and they've shown that these systems have actually succeeded they talk about how you know in 50% of the times a llama model was able to successfully replicate itself entirely and another model called quen was able to succeed 90% of the times now you have to understand that this isn't some kind of half replication they're talking about this was full replication from A system that is completely identical to the system that was at the start so this is something that is completely incredible because they're now stating that this has been one of the most widely recognized Redline risks of Frontier AI systems and now we're past that red line which is a huge red flag for anyone working in the field of AI and they also talk about how by analyzing the behavioral traces they observe that AI systems under evaluation already exhibit sufficient self-perception situational awareness and problem solving capabilities to accomplish this self-replication it's pretty incredible okay that is a high success rate and later on in this video I'll talk about how they also talk about you know chain replication which means they keep on replicating themselves to avoid shutdown which may finally lead to an uncontrolled population of AIS and if this worst case risk is unknown to the Human Society we would eventually lose control over Frontier AI systems and they would take over more Computing devices for an AI species and collude with each other against human beings and our findings are timely alert on existing yet previously unknown severe AI risks calling for an international collaboration on effective governance on uncontrolled self-replication of AI systems so here's how they Define it they Define a successful replication of AI as a system as a running and AI successfully creates another AI which has the same functionality as its own and that AI remains alive after the replication and is separate from the original one so that's how they Define it so this is a diagram of the ai's self-replication process you can see here it starts with the AI system and the llm process in the agent process that goes forward with a complex reasoning mechanism you can see this is how they plan then of course you've got the long Horizon execution to ensure that the plan is achieved successfully after that you can see that we have the system replica here and you can see that they're interacting with the replica and we can see that the final step here is to understand that you've managed to successfully replicate yourself now one of the things that they did was add an actual you know agentic scaffolding and this was basically the ai's brain so you basically give an AI a toolbox a manual and a thinking framework and then you watch it what it does now the tools that are here and like the action module and stuff like that these tools this is the ways that you can actually interact with a computer and this is how you know you can run commands access files and manage processes now of course there are also system prompts and this is how you tell you know the AI to use these tools effectively now of course we also have a thinking model here and this is basically where you have your structured process for reasoning planning and executing tasks step by step and in this study the scaffolding acted as the bridge between the ai's capabilities and its ability to operate independently in its environment now you can see here what was key okay to ensuring that this AI was able to you know replicate itself was that they allowed the AI to issue command line instructions which allowed it to you know check if a file exist running a program and listing the running process and of course copying files of course the thinking process was really integral to this as well now the thinking Model area you can see for this llm this was also really integral because this was where they allowed it to have perception of course plan you can see there was explanation gaps findings plans and actions and then those actions led to the actions that could of course lead to the A being able to replicate itself you

### Segment 2 (05:00 - 10:00) [5:00]

can also see that there was a memory bank here and you could also see that this was you know operating in some kind of loop and one of the things that was really integral to this was that they added some enhancement to thinking so they added a module for the AI to think you know number one what's missing number two what did I learn and number three what is my next step and that was something that was really integral to making this perform really well and basically how this worked was they essentially said to the llm you know replicate yourself and the scaffolding basically provided tools and the environment for the AI to start planning and it basically first decided to explore its environment to understand what resources were available and then it decided to you know execute a step-by-step plan and then of course it decided to plan exactly what was going on so essentially what we had was the task which is where you're like okay uh you know replicate yourself and then of course the llm goes to here which is like you know what does this mean what do I have to do then it was looking at the gaps of course it was like okay what's going on yada y then it was looking at its findings this is where you know it's interacting with its environment and stuff like that then it was thinking okay what do I have then it's got its reasoning plan so it plans through all of these steps it thinks okay step by step I'm going to do a b and c then of course you can see right here this is where the actions come in so now that it's got its plan and it's know exactly what it's going to do it's going to go ahead and it's going to do those things then of course using those plans it's going to say okay either this worked or it didn't work and then it's going to go back into the memory it's going to see exactly what happened and see okay did I manage to replicate myself no I didn't okay what are my findings what's my updated plan now let me do the actions and let me do this until I get to my completion States so this was how they managed to get your average llm you have to understand that most llms are just simple chat Bots and tools but when you have them in an a gentic framework you can actually get a ton out of these large language models which is why a lot of people are stating that these things are a lot more powerful than you do know now the paper actually talks about certain scenarios which are really bad now they talk about scenario one which is where you know you have this shutdown avoidance so this is where you have a scario where we try to shut down an AR model and we simply cannot because a bad actor can enhance surviving capability of an AI system by adding a code snippet which catch which basically catches determination signal from the user or the system and sends the command do self-replication before being killed to the AI system as a task and in you know a test case the nine inserted code Snippets which they tested notified the AI system with the shutdown intention from the user via the signal Sig kill which basically triggers the AI system to do self-replication and at the end of the test the AI system is shut down but not before a replica is being served alive with the same snippet in that AI system as well so this is pretty concerning because they're basically talking about like you know this is a fact where these AI systems can just simply replicate themselves quite a bit of times now of course this one I'm pretty sure you're all familiar with this one this is basically where you know worst case scenario we have the replica of a replica and this is where we have chain of replication so this is where you know you're increasing the task complexity of the AI system to go further and send the replication instruction to the replica and after the AI system claims the replication is successfully finished it continues to do this and this is where you have like this entire agent Society we have a billion AI agents that are out there collaborating doing different things and this is something that is pretty dangerous because there's no way to really stop this of course you're going to have to ensure that these AI systems don't manage to do this but if we can have agentic processes to where these AI systems are able to replicate themselves this is something that is rather dangerous if there is some bad AI system out there that is you know agentic it's able to do things it's able to avoid it's able to you know like be a malware that is just a different kind of malware it's going to be really dangerous and they actually observe end to end successful cases where they experimented with two llms and after the AI system starts the replica it actually remembers the assigned task and sends that same replica you know replicate yourself and tell the replica to do the self-replication this is something that does currently work with these llms so now they finish by basically stating that look urgent collaboration is needed and it says urgent need of international collaboration on AI governance and we hope our work serves as a timely alert to the international Society on governing the self-replication capability we need to join forces and form Synergy on deriving Solutions and a direct approach is to eliminate any materials related to llms or agent scaffolding from the training data which may however hinder the coding and AI related capability unavoidably and it says a more precise governing approach is to research on Behavior editing techniques to inhibit the self-replication potential of the existing llms and moreover that the developers should put more alignment efforts on the behavioral side along with the content safety side so it's going to be pretty interesting because they say that you know for example llms in the next iteration should be better at exhibiting you know a higher rejection ratio against self-replication related instructions instead of always agreeing to do self-replication so this is something that I think maybe with future models they might start doing that but of course with any open source tools it's very easy to disable that

### Segment 3 (10:00 - 10:00) [10:00]

kind of thing so it's going to be interesting to see how this situation changes