AI agents: The scientist's new superpower | Stefan Harrer | TEDxSydney Salon

[Music] [Applause] Marie K started as lab assistant of aie beel and became more famous than him for her pioneering work on radioactivity she won two Nobel prizes Charles Darin was assistant to geologist Adam swick and surpassed his Fame for his theory of evolution Francis Crick worked with Linus Pauling before co-discovery the double helic structure of DNA the code of Life what if we could give every scientist on Earth access to lab assistants of such caliber assistants that work with them and learn from them that know all about their field and ways of working that automatically

stay up to date on everything their field produces data methods tools and know how to access and use them assistance that can create critique and validate new research ideas and directions and assistance that speak every language on Earth and can be spoken to like humans thanks to generative artificial intelligence this is not a dream a Cambrian EXP explosion of AI assistance has started to fundamentally change how scientists work and this is changing the face of science the scientific method is one of Humanity's greatest accomplishments we observe nature we formulate a hypothesis we run experiments to

test it and then we look at the outcomes and decide whether our hypothesis holds up or whether we need to change it and then we repeat the scientific method has given Humanity the tools to fight disease to understand economies and to run Labs on Mars the scientific method will not change but it has been hitting a glass ceiling the complexity and scale of many scientific problems is too overwhelming for science to get a grasp on them this is particularly true in biology the science it studies living systems biological data is messy and it blows out

of all proportions for even the smallest biological systems experimental work to study them is very tedious and timec consuming let me paint you a picture to explain what I mean proteins are the smallest building blocks and workhorses of Life some of them speed up chemical reactions others give cells their shape and yet others transport substances through organisms there are over 826 million unique types of proteins known across species one single human cell contains about 20 million protein molecules the number can change depending on the type of cell there are 36 trillion cells in a human

body that means that at any given point in time there is about one sextilion that's a one with 21 zeros protein circulating around the human body going about their business if you had as many grains of rice you could cover the entire surface area of the Earth to a depth of 1 and a half met now imagine each one of those grains moving around with purpose appearing disappearing reappearing now try to understand what the movements of a few of those grains in Fiji might have to do with the movements of a few others up at

the North Pole using this analogy one single human cell would cover only the surface area of two basketball courts but even that system is so large and complex that we cannot currently model it using our scientific toolbox we do not understand life it's not the scientific method that fails us here it's a sheer size and diversity of the system we try to study that stifles us whether it is trying to understand life on our planet or the enormity of of the universe that we live in conventional scientific methods can only get us so far we

need to give scientists superpowers for years now so-called narrow AI has lent scientists Helping Hands whether it was discovering warning signals for incoming seizures in the brain activity data of epilepsy patients or whether it is scanning through hundreds of millions of radio signal Snippets looking for techno signatures of extraterrestrial life or whether it is predicting the 3D structure of every protein known to mankind there are many examples where AI models have been trained to perform tasks that human scientists either couldn't do at all or not at the scale of the systems they intend to study

one of those models Google deep Minds Alpha fold has already already had such profound scientific impact that its developers have just been awarded the Nobel Prize in chemistry last month however as bespoke as these AI models are they are narrow that means they cannot perform any tasks they have not specifically been trained for and furthermore scientists need to know that these models exist how they work what they can do how to access them how to run them and then how to interpret the results that they produce all of these are highly complex tasks that require

specially trained data scientists to work alongside domain scientists for long periods of time this assistance is a luxury that most scientists simply do not have only 16% of scientists worldwide and only 18% of biologists currently use AI as part of their scientific work building and deploying Cutting Edge AI models has become an elite expensive and timec consuming science in itself the same is true for many more methods and techniques which scientists have traditionally used to get to the top of their fields reading for decades scientific papers until you have your very own Ura moment and

come up with that new research idea waiting for month sometimes forever for that opportunity to brainstorm with the extremely smart but also extremely busy colleague training for years one researcher career at a time how to perform complex experiments these limitations have set the boundaries of what scientists could accomplish in their lifetimes and of what Humanity could accomplish using science until now a once in a millennium window of opportunity has just opened up to give every scientist a completely new type of tool with which they can overcome these limitations Archimedes famously said give me a firm

place to stand on and a lever and I shall move the earth I believe that generative AI has put science in that place and finally allows us to forge that lever here's why at the beginning of last year generative EI had gotten close to ingesting pretty much all publicly available digital information there was it started to excel at performing a seemingly unrelated endless variety of tasks that it had never been trained for the era of bespoke chatbots had begun about a year ago generative started to learn how to use tools to solve problems these tools

could be other AI or non-ai models data or actual physical tools and robots generative AI assistants or AI agents as they're also called were born then a couple months ago these AI agents learned how to plan and reason to solve problems in ways that no other AI was able to until then AI agents had become strategic and just a couple weeks ago it was shown for the very first time empirically that scientists who had used AI agents as aspiring Partners to brainstorm new research ideas could come up with more novel ideas than scientists who had

not used ai ai agents had become creative from hypothesis generation to experimental design to analyzing outputs with AI agents as helpers scientists for the first time can app apply all scientific methods and Engineering techniques ever developed to all scientific data ever produced this of course needs to be done ethically safely responsibly and it doesn't all happen at the same time some scientists and some scientific Fields charge ahead others follow But ultimately AI agents will bring them all together across their scientific Specialties blending their skills and amplifying them meet the astrobio chemist the paleoecological geneticist the

quantum bioinformatician generative AI has the power to bring about a new generation of polymath this will transform every scientific field but probably none sooner and stronger than the health and life sciences with AI agents in their hands biologists for the first time can scale the scientific method to a point where they can truly begin to unravel the mysteries of life and we have begun doing just that here at csro Australia's national science agency we are building AI agents that help our scientists to read the code of life to design new proteins to predict what they

can do and to plan and run lab experiments we believe that using AI agents will make our scientists fast gamechanging fast we expect that using AI agents we can bring down analytical time frames from 3 to 12 months to 4 to 5 days scientists over at startup in Silicon medicine are giving us an idea of the mindboggling impact that such a speed up can have on biology they have used AI to facilitate the entire endtoend drug Discovery and development Pipeline and a world first just brought a new drug candidate from its initial conception to clinical

trial stage in only about half the time that it would have taken to get there using conventional methods this will surely save hundreds of millions of dollars and years of work but impact goes beyond productivity gains here the drug candidate in question had been conceived by scientists who had used generative AI in the first place they used AI to come up with the idea Ai and science had teamed up and together accomplished something that neither one of them could have done by them themselves this is the beginning of a movement biontech the developers of the

very first covid-19 vaccine just recently announced that they're working on a specialized AI powered lab assistant that can automate scientific experimental workflows and the makers of alpha fold also recently shared that they as well are working on AI lab assistance for researchers and particularly biologists that can help them plan and run experiments and predict their outcomes at scale it's that scale that allows scientists to look at every single one of that one seilan proteins and to explore what each one of them does and why and where and when for the very first time in life

scientists can study life on its entire Spectrum from single molecules to entire organisms it's accelerating Google Deep Mind founder and CEO Demis sabis pictures all information that describes how the universe works as a tree of all knowledge he believes that AI can help us solve some of these great scientific root problems giving us access to new parts of the canopy and opening up entirely new areas of research I believe we can build on that Vision I also believe we can be even Bolder we can give every scientist access to trustworthy powerful lab assistance with which

they can explore the entire tree and make sense of it all generative AI agents are those lab assistants let's build them let's share them let's use them to stand on their shoulders and climb that tree AI agents will not replace human scientists history is full of moments when human creativity and curiosity have triggered mindboggling Innovation and breakthrough discoveries AI will not take away that Ingenuity that human brings that humans bring to the game but AI can enhance this Ingenuity and accelerate our journey towards solutions that would otherwise have been decades away or entirely Out Of

Reach the desire of biologist to understand life and the promise of generative AI to perhaps getting them there have brought Ai and science together in a watershed moment they're made from and for each other their success foreshadows a new era of Enlightenment AI will change the world of science and in doing so science will change the world [Applause] [Music]