“That’s a dangerous point when the system can self-improve. We need to seriously think about unplugging it.” — Eric Schmidt the former CEO of Google
and that’s a dangerous point when the system can self-improve we need to seriously think about unplugging it so that was Eric Schmid the former CEO of Google actually talking about how you might have to unplug AI very soon now now this was something that I found to be rather interesting is it was only a couple of days ago we got this research paper Frontier AI systems have surpassed the self-replicating red line and this was essentially a paper in which they discussed the fact that AI systems were able to replicate themselves without any human help and self-replication basically means that an AI system could create a functional copy of itself capable of running independently and it sounds something like it’s out of a sci-fi movie but of course this would be a very big deal now from this paper I’m not going to go speak about it too much but essentially in 50% and in 90% of experimental trials they succeeded in creating a live and separate copy of itself respectively now currently this isn’t too big of an issue because these systems don’t have that much capability but many people are thinking about the future in which this could be a very very D scenario if they are more agentic and if they have more ability so that is essentially the gist of what he’s saying now we’ll take a look further at the extended statement but it is definitely a question that needs to be answered by some of the frontier AI Labs we’re soon going to be able to have computers running on their own deciding what they want to do and the way that happens is it’s a series of decisions we go from Agents to then sort of goal more powerful goals and eventually you say to the computer learn everything and do everything and that’s a dangerous point when the system can self-improve we need to seriously think about unplugging it but wouldn’t that that kind of system have the ability to count our efforts to unplug it well in theory we better have somebody with the hand on the plug metaphorically but the important thing is that the power of this intelligence the ability for this kind of new intelligence means that each and every person is going to have the equ of a polymath in their pocket in addition to your show and all your notes and writers you’re going to have an Einstein and a Leonardo da Vinci to give you advice on your show that will be true for everyone on the planet we just don’t know what it means to give that kind of power to every individual now whil his full statement is quite riting when we start to discuss the future implications of future AI systems he actually made a second statement which I think is a bit more thought-provoking in terms of the overall implications of that statement now this statement I’m about to cover is essentially one where Eric Smith refers to automating AI research now if you’re not familiar with this concept this is essentially the grand goal of any AI Lab at the moment and I say that because once they can automate AI research it essentially kicks off this crazy loop where you essentially have ai systems that can improve themselves over a long chain and what I mean by that is that if AI systems can do AI research then we can use that AI research to improve the model and thus the next iteration of AI research gets faster more efficient and of course smarter which leads to this recursive self-improvement maybe not in a singular system but overall these a companies are going to be moving at light speed I’ve done this for 50 years I’ve never seen Innovation at this scale this literally remarkable human achievement of intelligence and the things that we can do and the advances in science and on on on there’s a point at which maybe in the next year or two where the systems can begin to do their own research they’re called AI scientists as opposed to human scientists so you go from having a thous human scientists to a million AI scientists I think that increases the slope when you’re moving at this space it’s very hard for your competitors to catch up that’s the race it is crucial that America wins this race now is what Eric Schmidt just referred to their actually feasible is it realistic is there any research that backs up his claims well currently there actually is we can see recently from Sakai they published something in August of this year where they spoke about the AI scientist towards fully automated open-ended scientific discovery and this was basically an gentic framework which is built to automate the entire scientific research progress and this is heralded as the world’s first comprehensible AI of conducting independent scientific discovery from start to finish now this one is one that is about scientific discovery and I will come back to that but what about AI research now this one wasn’t covered that much but Neo was an actual new automated AI researcher that was able to automate AI research now this was the first AI engineer specifically designed for machine learning tasks and this operates on the entire machine learning pipeline outperforming even open ey in certain tasks and this system represents a significant step towards artificial super intelligence by automating AI research now this you know system that you know is currently being developed automates complex machine learning workflows such as data collection pre-processing model selection fine-tuning evaluation and deployment and this uses multi-step reasoning to evaluate multiple Solutions and implement the best approach now this optimizes performance by analyzing throughput and latency across many different Frameworks like TT and gpus and this is something that has been demonstrated to be remarkably effective in a variety of different scenarios so this is one of the first systems that we are getting that actually manages to automate AI research and this is what these companies are looking to build and internally I do know that openi has their own tool like this which they refer to as the AI scientist something along those lines it’s very very vague it’s very hard to find a lot of information on it but I do know that these companies are certainly working on this so it’s not that crazy to think that in a few years this kind of system this kind of software is going to be so much better it’s going to be so much more effective and it’s going to be able to allow us to not only run AI research in an agentic way that allows us to speed up our research process and there something that is completely incredible now you have to think about this okay if this is true if this is the case what happens when we manage to automate a research now do you remember the document called the decade ahead this was from a Open Air employee who used to work there but you know unfortunately got fired due to various different reion there’s a lot of speculation going on but his document was one of the most insightful documents on the decade ahead in terms of where our research is going to go so one of the things that he actually said was that we don’t need to automate everything just AI research it says a common objection to transformative impact with AGI that it will be hard for AI to do everything look at robotics for instance delas say that it’s going to be a really hard problem even if AI gets really smart some people say that robotics is still going to be hard or even if you manage to automate biology it’s going to actually take a lot of physical lab work and human experiments but the crazy thing about this is that we don’t need robotics we don’t need many things for AI to automate AI research the jobs of AI research and engineers at leading Labs can be done fully virtual and don’t run into real world bonck in the same way and the job of an AI researcher is fairly straightforward in the grand scheme things you need to read the machine learning literature come up with new questions or ideas implement the experiments to test those ideas interpret those results and then of course repeat and this seems squarely all in the domain where the extrapolations of current AI capabilities could take us beyond the best humans by the end of 2027 now the reason that this is so crazy is because this just kicks off an extraordinary feedback loop and the reason that this is so crazy is because if we actually look back at some of the biggest machine learning breakthroughs from the last decade most of them have just been you know oh let’s just just add some normalization or let’s just fix an implementation bug and all of this kind of AI research can be automated and once we do unlock that it’s like this thing manages to speed up 10f and in this document he actually talks about how you know we can expect 100 million automated researchers working at 100 times human speed not long after we begin to be able to automate AI research they’ll be able to do a Year’s worth of work in just a few days and the increase and the increase in research effort compared to a few hundred puny human researchers at a leading a lab today working at puny 1 * human speed will be extraordinary basically taking that look once we figure out the exact system that we’re able to use to actually get breakthroughs all we’ll have to do is scale that system across many different clusters and then we’re going to be getting a ridiculous level of a research that we’re going to use to improve these systems and we’re going to have a thousand maybe even 100,000 researchers that are going to be operating 24/7 and if you guys don’t think that this transformative change is going to happen let’s actually take a look back at history remember the printing press before the printing press book were painstakingly copied by hand and after gutenberg’s invention books could be mass-produced in a fraction of the time democratizing knowledge and accelerating scientific progress we also had the Industrial Revolution machines replaced manual labor vastly increasing manufacturing output and for instance cotton genin reduced the time needed to separate cotton viers increasing productivity by over 50 times we also have the internet and digital Communications before the internet communication was limited to physical mail phone calls or in-person meetings email and online systems speed up Global Communication from days to Mere seconds transforming collaboration research and business so overall when we take a look at these systems that are going to be basically automated AI researchers we have to understand that this is going to be an incredible rate of progress that happens in the AI World once we’re able to use agents effectively to perform long-term tasks things are going to start changing rapidly on a scale that you’ve never possibly could even imagine and we also do have ai scientists that are able to you know automate AI discoveries in terms of being able to automate many different things across many different Industries there’s going to be an incredible rate of scientific discovery across the board these AI systems are going to improve upon their plans they’re going to be able to consistently reason about different things in ways that humans just can’t think about and they’re going to be able to test these a lot quicker than humans can of course there are going to be some physical limitations the rate of discovery should be a lot larger than it is right now which is going to lead to transformational impacts across Society overall I think that the decade ahead is going to be one that is really crazy but that once you do get that a ating V research that is going to be where things go incredible let me know what you think about erish Schmid’s comments about shutting down Ai and of course the future of AI