New AI Breakthroughs Explained. It's ALL Accelerating!

just some days ago Google finally released so  much anticipated Gemini model and it beats GPT4 on   nearly each Benchmark which makes it the best LLM  to date I've worked on AI my whole life because   I've always felt would be the most beneficial  and consequential Technology For Humanity human   beings in our society we have five sensors and  the world we built and the media we consume is   in those different modalities so I'm super proud  and excited to announce the launch of the Gemini   era a first step towards a truly Universal AI  model if you watch all the demos that I watched   and read the paper it's clearly the next level  in LLMs what's interesting Google says that in   future it will be getting even more General even  more General than video and audio over time Gemini   will get more senses like actions and and touch  and this is definitely exciting and we we'll talk   about that but at first have you noticed that  the pace of technological progress is speeding   up several days ago deep Minds AI made 800 years  worth of Discovery their AI tool called GHoME   has discovered 2.2 million new crystals these  new materials can power future Technologies   from starting from our favorite Computer Chips to  solar panels and batteries now if you look at the   history of Technology if you consider anything  from mobile phones to computers there is always   the same Trend at first we built for a specific  solution like the first version of chat GPT and   then with time a single more powerful solution  takes over for example in the past we used to   have a separate device for everything for example  a voice recorder right digital camera a GPS but of   course now all of those things can be done by  an average smartphone the same trend is going   to happen in artificial intelligence as well  in the future it will be easier to create one   large single model than to create multiple small  models just because the biggest cost it what it   costs to pay human workers to create to develop  these models and I talked about this with the   CEO of Archetype AI exciting startup working on  expanding multimodality of large language models   now the question is do we think there's going  to be one one human like Planet size AI which   does everything I think people talk about AGI I  I doubtful that is going to happen but I think   they're going to be a very large model which do  a one large group of functionalities and large   language models is one of them maybe they going to  be large Vision models but we are building what we   called Large Behavior Models we in the company  build this new class of models which we called   LBM large behavior model which looks for patterns  of behavior in the real world captured by variety   of sensors what's recently became possible that  was not possible before is that we can create   generic foundational models and we can teach  these models to understand multimodel input such   as audio video um information about vibration  and also information coming from the radar for   example and in this data they are looking for  patterns or behavior over time and eventually   they will be able to answer any possible question  for example they would be able to identify why an   accident has happened in general there are two  things that can be scaled in AI it's search and   learning and many research papers suggest that  the creativity will actually arise from search   from AI exploring a huge space of possibilities  and actually there is one interesting example how   AI discovered unconventional creative solution  through research it's the famous move 37 made   by AlphaGo back in 2016 and it actually became  symbolic of ai's ability to innovate but was it   really creative let me know what you think  if you have a limited space to choose from   what creativity means a game of chess a game  of Go designed by humans right so it's fully   understandable right uh the universe which we live  in the way we map it out into the knowledge is by   definition is limiting the universe is much more  complex and much more difficult and there's a lot   of things we don't know if we Define creativity as  discover patterns which we never knew about well   I don't know the answer to these questions but  what I 100% certain is that it will accelerate uh   our capabilities and discovering new things and  seeing new things 100x 100 thousand x million x   I think it gives a superpower fundamentally this  models and now deep Minds AI discovered 2 million   new materials to be more specific 2 million new  crystals to do this they've created an AI tool   called graph networks for material exploration a  tool that can discover new materials very quickly   and again they did it using Graph Search and then  they synthesized some of these new materials in   Berkeley lab where AI was guiding robots through  this process how amazing and beautiful it is I'm a   huge fan of deep M's work in this field because  they're constantly pushing the envelope at   the frontier of science and if you agree please  share this video so more people can learn about it   we already know that by this point in time we  can use AI tools to create something cool for   ourselves for example build a website and for this  task Hostinger AI website builder is a fantastic   option creating a website with a Hostinger AI tool  is very easy you can actually build your website   with just four or five clicks you can either  choose from 15 pre-made designer templates or   generate a website from scratch using Hostinger  a a i website builder their AI tool constructs   a website for you based on one single prompt all  you need to do is to type in your brand name and   then provide a brief description of your brand  or business based on this information it then   chooses the most appropriate website interface  and what text and images to place on this website   once more Hostinger offers additional AI power  tools to further enhance your website use the AI   logo maker to create a logo or you can leverage  the AI heat map to predict areas of your site   where visitors will pay the most attention the  process is very user friendly and you don't need   any technical knowledge or background to manage it  if you've been thinking about creating a website   for a long time now it's the time to secure an  extra 10% off on all annual plans use the link   in the description below and use the coupon  Anastasia 10 as a checkout now Google finally   released Gemini and it's essentially a combination  of their famous AlphaGo and llm and the basic idea   behind AlphaGo is that it combines deep learning  and Mont Carlo Tree search to play the game of   Go and Gemini's multimodal capabilities is just  another level traditionally multimodal models are   created by stitching together text only vision  only and audio only models in a suboptimal way   at a secondary stage Gemini is multimodal from the  ground up so it can seamlessly have a conversation   across modalities and give you the best possible  response Gemini is our largest and most capable   model it means that Gemini can understand the  world around us in the way that we do uh and   absorb any type of input and output so not just  text like most models but also code audio image   image and video both objects are round and  flat there you go both are food the Orange   is a healthier choice than the cookie citrus can  be calming and so can the spin of the fidget toy   they are both examples of popular toys that have  been enjoyed by children and adults alike for many   years Gemini is based on this paper a family of  Highly capable multimodal models so they really   Gemini in three versions Ultra Pro and Nano  Ultra is the most capable one Nano is a one   which you can run locally at age for example on  your phone and pro is their equivalent of GPT 3.5   this version is already available in Bart and soon  it will be available in search ads and their other   products honestly the benchmarks looks rather  surprising here as a bench benchmarks comparing   Gemini Ultra to GPT 4 in general capabilities it  received 90% and GPT4 received 86.4% if we look   at math and code benchmarks it scored higher in  basic and challenging math problems and it's much   better at code generation in total it outperformed  GPT4 in seven out of these eight benchmarks and on   all multimodel benchmarks it surpassed GPT4 which  makes Gemini definitely the best llm to date just   have a look at this demo so we were looking at  this study from 2022 the authors had created a   data set by reviewing tens of thousands of  scientific papers in genetics they found a   few hundred papers that contained the relevant  information extracted it by hand and collected   it in a table studies like this can take a lot  of time we needed to update this data set with   what's new over the last couple of years but  that's over 200,000 new Open Access papers   added to this domain since 20121 we couldn't do  this manually so we asked Gemini to help us out   and over a lunch break Gemini read 200,000 papers  for us filtered it down to 250 and extracted their   data so now we have a refreshed version of this  data set with our refresh data set we can now ask   Gemini to update a graph from the original study  we first gave Gemini a screenshot of this figure   then we asked it to generate the code required to  plot it and by feeding this code our new data set   we get our updated figure you can see that this  figure now includes data up until 2023 oh wow I'm   deeply impressed by this and the demos showing  Gemini's video understanding capabilities are   also pretty cool the duck is in the middle of  the ocean there is no land nearby for the the   duck to rest or find food it is unlikely to find  Ducks here naturally both objects are round and   flat there you go both are food the Orange is  a healthier choice than the cookie citrus can   be calming and so can the spin of the fidget toy  they are both examples of popular toys that have   been enjoyed by children and adults alike for  many years what's interesting Google is already   looking at combining Gemini with robotics and  then adding different senses to it and as you   know many great minds are working on bringing AI  with this multimodal capabilities into physical   world which I'm truly excited about and I think  we will see this development happening very fast   because the pace of AI progress is accelerating  in all of this the most interesting question   to me is Will AI models be able to discover  completely new Solutions not the combination   of uh already existing things which generative AI  is already doing pretty well now but completely   new Solutions because that's what we need in  order to push the science hopefully in the   right direction I see this model as a tool and  less of the um kind of some sort of AGI Style whatever intelligence know uh it's going to be  our cognitive enhancement and cognitive amplifier   which makes humans way so we will be able to  solve problems and discover things we've never   seen before never heard about as you know me I  don't like to speculate still many of you were   messaging me and asking about the Q star and what  this can potentially mean as you know about uh   one or two weeks ago writers were writing about  the internal open AI Q* project there's another   twist in the open AI Saga involving a possible  breakthrough in artificial intelligence according   to one of the sources the project involved  is called Q* and some internally believed   progress made on Q* could be a breakthrough in  the startup search for super intelligence also   known as artificial general intelligence what's  interesting in an interview with wire some days   ago Sam Altman called it an unfortunate leak now  what is the A* model many suspect that it has to   do with the Q learning but you know Q-learning  for deep learning is nothing new it was already   described by Deep Mind in this paper from 2013  which is honestly an eternity for this field   in two words it's a goal oriented re enforcement  learning algorithm and I talked about it before   and the star from Q* seems to come from A* search  Graph Search algorithm which is so commonly used   in computer science there is a lot of excitement  around this topic and many people suspect that   AGI Artificial General intelligence is right  around the corner while I understand all of   this excitement and where it's coming from I find  it really hard to believe uh that that's the case   so what's next it's quite obvious that it has  been a huge acceleration of change recently and   a huge acceleration of technological progress  that I do feel you know it has been one just   one year since ChatGPT was released and by this  point it's already old news and now we already   talking of general intelligence I think that  from this this moment on the changes will be   happening only faster what now takes years will  take merely month in the future and the rate of   change we'll be experiencing in 20 years from now  will be at least four times faster unbelievable   like Ray Kurzweil said the whole 20th century is  equivalent to 20 years of progress at today's rate   of progress and because of this exponential  growth the 21st century will be equivalent   to 20,000 years of progress at today's rate  of progress which is so interesting to think   about and this also makes me so happy to live  at this moment so guys get ready for a lot of   exciting Innovations and also a lot of cognitive  dissonance because honestly I expect unexpected to   happen if you enjoyed this video please consider  giving it a like subscribing to the channel and   the most important please share it with your  friends and on social media I always see your   reposts and this makes me very happy thank you  so much and I will see you in the next video ciao

Channel: Anastasi In Tech

Views: 133,267

Keywords: AI explained, Gemini Explained, AI breakthroughs, New AI, Deepmind Gnome, Breakthrough in Artificial Intelligence

Id: n29WWr4g6sc

Length: 17min 20sec (1040 seconds)

Published: Mon Dec 11 2023