“There is so much of AI we don’t understand… there are emerging properties that we don’t understand. We don’t understand how those machines develop those properties… we have no clue how it arrived at it… we have no clue most of the time how the machines do what they do. We don’t. “
“There were three barriers that, we all, all computer scientists that worked on AI: we all agreed there were three barriers that we should never cross, and the first was don’t put them on the open internet until you are absolutely certain they are safe okay? and… we said we should never cross is don’t teach them to write code and don’t have agents prompting them.”
“It really is about a point of no return where if we cross that point of no return we have very very little chance to bring the genie back into the bottle what is the point of no return the most important of which of course is the point of Singularity and Singularity is a moment where you have an AGI that is much smarter than humans.”
I think the reality of the matter is uh it is so much power so much power that if it falls in the wrong hands and it is bound to fall in the wrong hands unless we start paying enough attention right and that’s my My Cry Out To The World Is let’s pay enough attention so that it doesn’t fall in the wrong hands it would lead to a very bad place the third you know and and the biggest reason in my view uh of um of us needing to worry hopefully hopefully we will all be wrong and be surprised is that there were three barriers that we all compute all computer scientists or that worked on AI we all agreed there were three barriers that we should never cross and and the first was don’t put them on the open internet until you are absolutely certain they are safe okay and you know it’s like fdaa will tell you don’t swallow a drug until we’ve tested it right uh you know and and I and I really respect Sam Altman’s view of you know uh developing it in you know in public in front of everyone and to discover things now that could uh you know that we could fix when the challenge is small in isolation of the other Tool uh this is a very good idea but the other two barriers we said we should never cross is don’t teach them to write code and don’t have agents prompted them right so what you have today is you have a very intelligent machine that is capable of writing code so it can develop its own siblings if you want okay that is known frequently to uh to to outperform human developers so I think 75 of the code uh was no sorry 25 of the code uh given to chat GPT to be reviewed uh was improved to around two and a half times faster okay so so they can develop better code than us okay and and basically now what we’re doing is we’re not only limiting their learning the learning of those machines to humans so they’re not learning from us anymore they’re learning from other AIs and there are staggering statistics around the size of data that is developed by other AIs to train AIs in the data set.
I mean now now we could if we if we decide now we could simply switch off all of that Madness switch off your Instagram recommendation engine your Tick Tock recommendation engine your ad engine on uh Google your data distribution engine on Google you can also switch off chat GPT and you know a million other AIS and then we can all go and sit out in nature and really enjoy our time honestly we won’t miss any of it at all I’ll tell you that very openly I mean the reality of the matter is that Humanity keeps developing more and more and more because we get bored with what we have okay and we think that we can do better with an automated call center agent when in reality it’s not about better it’s just about more profitable okay and and the reality here is that we could but will we no we want why because of the first inevitable before because of the trust issue between all of us and because we need the AI policemen just as much as we need the you know as as we fear the AI Criminal.
nce your GPT again or bar your response to you it’s not referring to the entire data set from which it learned to give you the answer it’s referring to the abstracted knowledge that it created based on massive amounts of data that it had to consume okay and when and and when you see it that way you you understand that just like we needed the Mainframe at the early years of the computers and now you can do amazing things on your smartphone the direction will be that we will more and more have uh smaller systems that can do AI which basically means two developers in a garage in Singapore can develop something and release it on the open internet uh you know again you and I I don’t know if you coded uh any any uh Transformers or uh or or you know or a deep deep neural networks and so on uh but they’re not that complicated I think the code of chat of of gpt4 in in general is around 4 000 lines the core code right it’s it’s not a big deal when when I when I coded banking systems in my early years on kobel on you know uh on MDS machines or as 400 machines it was hundreds of thousands of lines of code okay uh so so there the the possibility for us why why has it become so much less is so much better because it’s all algorithms it’s not it’s all mathematics we I think this is a very important thing to differentiate for people when I coded computers in my early years those machines were dumb and stupid like an idiot they had an IQ of one literally no IQ at all okay developers transformed human intelligence to the machine we solved the problem and then we instructed the machine exactly what to do to solve it itself right so you know when when we understood how a general ledger works we understood it as humans and then we told the machine add this subtract that reconcile this way and then the machine could do it very very very fast which appeared very intelligent but it was totally a Mechanical Turk it was just repeating the same task over and over and over in you know in very fast speed we don’t do that anymore we don’t tell the machine what to do we tell the machine how to find out what it needs to do so we give it algorithms and the algorithms are very straightforward when you you know let’s let’s take the the the simplest way of deep learning when we started deep learning what we did is we had basically three Bots if you want one is uh what we call the maker uh the other is the student the the final AI that we want to to build and a one that’s called the teacher okay and we would say um you know tell them to look for a bird in a picture okay and they would identify a few parameters you know um edges and how the how do they see the edge and the difference in color between two pixels and so on and so forth and then they would detect the shape of a bird and basically we would build a code and and call it a student we would build multiple instances of it and then show it a million photos and say is it a bird is it not a bird is it a bird is it not a bird and the machines would randomly answer at the beginning it’s literally like the throw of a dice okay and you know some of them will get it wrong every time some of them will get it writes 51 of the time and one of them will get it right sixty percent of the time probably by pure luck okay the teacher is performing those tests and then the maker would discard all of the stupid ones and take the one code that got it right and continue to improve it okay so the code was simply a punishment and reward code it was saying guess what this is and if you guess it right we will reward you okay and and basically the machine the algorithm would then continue to improve and improve and improve uh until until it became very good at detecting birds and cats and pictures and so on and so forth when when we came to Transformers and why GPT and Bard and so on are so amazing is because we used something that was called reinforcement learning with human feedback so basically we allowed instead of discarding the bad ones okay we found a way which Jeffrey Hinton the you know who recently left Google was very prominent at you know uh promoting early on we found a way just like with humans to give the machine feedback you know show it a picture and then it would say this is a cat and we would say no it’s not it’s actually a bird what do you need to change in your algorithm of them okay so that it would the answer would have become a bird okay and so the machine would go backwards with that feedback and and and you know and change its own thinking so that the answer is correct and then we would show it another picture another picture and we keep doing this so quickly on billions or millions or tens of thousands of machines of you know millions of instances until eventually it becomes amazing just like a child just like you give a child a simple puzzle okay nobody ever told the child no no no no darling look at the cylinder turn it to its side look at the cross section it will look like a circle look at the board and find a matching shape that is a circle if you put the cylinder through the circle it will go through that’s old programming okay new programming which every child achieve intelligence achieves intelligence with is you give them a cylinder and a puzzle board and they will try they’ll try to fit it in the start it won’t they’ll try again it won’t they’ll throw it away and get angry then they catch it again and try in the Square it won’t and then when it goes through the cylinder something in the child’s brain sorry through the circle there’s something in this child’s brain says this is this works okay the only difference is a child will try five times a minute or five times you know 50 times a minute a computer system will try 50 000 times a second okay and so very very quickly they achieve those intelligences and as they do we we we don’t really need to code a lot because the heart of the code is an algorithm it’s an equation okay and and Mathematics is much more efficient than instructions so if if I tell you Tom uh when you leave home make sure that your um you know distance is no more than the day of the months multiplied by two away from your home and make sure that you don’t consume any more fuel then your height divided by four okay or then then your body temperature divided by seven whatever that is okay with those two equations I don’t need to give you any instructions anymore you can always look at your fuel consumption and your distance and say oh I’m falling out of the algorithm with very very few lines of code I just gave you two lines of code
In the existential risk scenarios uh one of our better scenarios believe it or not is that AI ignores us all together Believe It or Not uh it’s a much better scenario than AI being annoyed by us or AI killing us by mistake.
The alignment alignment problem I just address it perhaps with my other side not the engineer and the uh algorithmic thinking that I did address the problem with my whole life right the the the challenge uh has been that those who have developed AI who believed in what is known as the solution to the control problem okay and the control problem is in Humanity’s arrogance we still believe today that we will find a way to either augment AI with our biology so that they become our slaves or to box them or tripwire them or whatever so that they never cross the limits that we give them and and we can discuss this in detail if you want but in my personal view you can never control something that’s a billion times smarter than you.
“Every task we’ve ever assigned to AI it became better than us so when with that in mind when we have a superpower coming to the planet I’d like to have the superpower have our best interest in mind I’ll I’d like to have the superpower itself work for Humanity.”
“What we know so far is that our Behavior affects its decisions okay and what we know so far fact is that data affects it more than code so what creates the intelligence of Bard is the large data set that is trained on it’s not just the code that is that that develops its intelligence the larger the data set this is why when you ask open Ai and others where is most of the investment in gpt5 going it’s going to be new formats and bigger data sets but learning the the data is really where most of the of the intelligence comes from so if we can influence the data that it’s fed we will influence Its Behavior”
“we need to take control we absolutely need to take control we’re not and taking control is not just about the code and the the the control code it’s also about the data it’s also about the date okay and the data is not just books the data includes human behavior every time you swipe on Instagram you’re telling AI something.”