Generative AI Full Course – Gemini Pro, OpenAI, Llama, Langchain, Pinecone, Vector Databases & More - Ep214

right now it has comecombined everything di it has combineduh for the data analysis part also ithas combined that code interpretorfunctionalities and top ai certificationsall right andrecently it was launched and over thereyou can probably do tasks that arerelated to images that are related totext okay here in gini when we see rightyou are able to combine everything textimages videos audios and code today Iwill show you a lot of example withrespect to text and images okay we'llsee some amazing use cases you can alsodo it with the the help of PDF you cando with multiple things as such okay andall the task that are related to NLPlike chat with your PDF and all you canalso do with thisnow the most important thing whygini is the most capable AI model thatis because of this result okay now seehere you can see something called ashuman expert MML now what is this mmuokay if you probably search for whatexactly is MML okay mlu if you see thatit is nothing but massive massivemultitask languageunderstanding so with respect to humansright it is basically able to say thatover here it has achieved see Gemini isthe first model to outperform humanexperts onmlu massive multitask languageunderstanding one of the most popularmethod to test the knowledge andproblems solving abilities of AI andhere it is also said that it has crossedthe crossed the Benchmark of GPT 4 alsoright so if you probably see this seeunderstand Guys these all are veryimportant things to understand becausebenchmarking is done on which thing thatyou really need to get an idea about andbecause of this benchmarking you willget a clear idea and you can assume howgood this specific model is and anymodel so tomorrow if you probablytalking about Lama 2 if you're talkingabout Mistral it will be benchmarkingbased on this capabilities so if youprobably want to work in the field ofgenerative AI I think this benchmarkingis super important and you should learnabout this okay or get an idea about itbecause tomorrow if you're reading aresearch paper how can you say that thismodel is better than the other modelokay so here you can can probably see inmlu it is nothing but representation ofquestions in 57 subjects right it hasbeen able to get this accuracy 90% gp4it was somewhere around86.4 then in case of reasoning you cansee some results where it was greaterthan GPT in two different things like inbig bench hard drop in h Swag like incommon sense reasoning for everydayeveryday task it did not achieve thatmuch accuracy when compared to gp4 Okayso the reason see I'm telling you whyall these things you should know becauseyou should get an idea about it okay theother thing is that with respect to mamaths right basic arithmeticmanipulation final grade school mathproblem it was able to uh get a goodaccuracy of 90 4.4 but when you hadchallenging math problem it is not ableto get that much accuracy rightsomewhere around 53.2 but it is far morebetter than gp4 gp4 was able to getsomewhere around 52.9 okay similarlywith respect to the code evaluation hereyou can also see that it has crossedthis gp4 like python code generationpython code generation new heldout dataset human eval like not leaked on theweb right so completely a new generatedcode right so in short it says that andthis is completely from the researchright gini surises the state of artperformance on the range of multimodelbenchmarks so you are getting thisspecific information and there are otherinformation see this is somethingrelated to text okay this is somethingrelated to text now here if you go downit is something related to multimodelnow whenever I say multimodel what doesthat mean it is basically talking withrespect to images with respect to videowith respect to audio and here also youcan see that it has proven well withrespect to all the Benchmark whencompared to GPT 4V right so here you cansee 59.4 4 77.8 82.3 90.9 80.353.0 and here you can see all the otherreadings right if you want to read thetechnical report here you can probablygo ahead and read it entirely okay itwill be talking about how it hasbasically done the fine tuning and alland all and all it's just like aresearch paper like why it is basicallysaid as a so here you can probably seehere is a solution of a physics Problemby a student uh the information is givenover here that it was not able to getthe answer whether the answer is corrector not all the information clearly youable to see right again it is based onTransformer only encoder decoder so thatis the reason right why I say in theroad map of generative AI if you want tolearn something you really need to havea good base about Transformer and BTright if you understand what is encoderdecoder how it works right thendefinitely all the further things you'llalso be able to understand it right soguys till here everybody's clear rightnow we'll talk about what all differentsizes are there and which model isavailable for us so if everything isclear please do hit like if you're ableto hear me out andall so uh shini says sir what is thedifference between B and gin see B likehow we have chat GPT applicationsimilarly we have Google B in Google Bin the back end we used to use pal toright now they can change Palm to Geminiright where it will also support imagesand text it's all together anapplication so can I get a quick yes ifyou able tounderstandyes something quick yes come on guys Ishould be able to hear I think there's aless energy over there come on let'slet's make this session amazing see mymain aim is to make this session a goodone for you right you should be able tounderstand things your Basicsfundamental should be strong tomorrowwhen whatever model you should see youshould be able to understand so hit likeokay hit like hit something give asmiley I'll feel happy more energy willcome when I'm explaining because we alsoneed to do end to end project right nowokay nowum so one question is that how togenerate image data using Gemini stillthose features have not

(Editor: swift)