Could Chat GPT Talk to Whales?

this is the sound of the biggest toothedpredator on the planeta song being broadcast from the deepestreaches of the sea a call that soundsless like a harmonious Melody like wehear from other whales but more like adigital data transferit's the sound that haunted Sailors forcenturies a sound which they thought tobe the cries of the ghosts of drownedSailors calling out to them amazingly itwasn't until 1957 that scientists evenrealized that these sounds were in factcoming from whales and it wasn't untilthe 1970s that they realized that theseundulating clicking bellowing noiseswere a form of communication and nowresearchers think sperm whales could beour best chance at breaking a barrierwe've never broken that they might holdthe key to unlocking the first ever caseof inter-species communication becauseif you were to bet on one animal thathad something that was even close tohuman language it would be the spermwhale over the years scientists havelearned that their brains are extremelydeveloped that they have Rich social andfamily lives culture and an intricatecommunication system to support it alltheir lives in many ways are similar toours and their language could havedeveloped for similar reasons plusthey've had their massive brains fortens of millions of years longer thanwe've had our current sized brains andwith a vast repertoire of differentsounds with intricate patterns it seemsas though sperm whale language could beas complex as our ownspeaking with other animals has longbeen a part of the collective humanimagination part of folklore fables andfantasy and many different cultures butnow for the first time we havetechnology that might be able toactually break the coderecently we've all seen the huge stridesin the field of natural languageprocessing with chat gpt3 giving a clearidea of what is possible when it comesto computers in human language astechnology like this has advanceddramatically over the last 10 yearsscientists now think they can applythese techniques to the field ofinter-species communication specificallyto sperm whales because sperm whales andother species aren't just singing to oneanother they seem to be communicatingspecific messages maybe even usinglanguage in a way that we couldunderstandfor thousands of years humans haveconsidered language to be one of thedefining characteristics of our speciesand something that other animalscouldn't possibly approximate so wehardly bothered to investigate all theways that animals actually arecommunicatingtoday with better technology we'restarting to realize that many specieshave complex forms of conveyinginformationprairie dogs have distinctive alarmcalls for different predators and haveeven used different calls to distinguishthe size shape color and speed of thosePredators male and female putty-nosedmonkeys each produce their own alarmcalls when leopards approach the femaleis calling to get males to go on thedefense and the males calling to Signalthey're ready for a fightbat squeaks aren't just echolocation tohelp them hunt they sometimes containinformation about the speaker and whichother bat is being addressed even tinyjumping spiders have a form ofcommunication through vibrationsthe animal kingdom is clearly full ofspecies communicating to one another butwhether or not they're using language ismuch harder to recognizewhen Linguistics developed as an area ofstudies specifically interested inanalyzing the differences between humanlanguage and the mechanics of howlanguage works and is acquired it seemedclear that nothing else on Earth wascommunicating the way we arehuman language has grammar theoverarching structure of each languageand syntax which is the order in whichwords are spoken in order to conveymeaning we have the ability to createnew words and use them to communicateinformation about things that happenedelsewhere or in the past or future humanlanguages are complex and nuanced whatwe don't know is if the same is true ofany other animal but there's one speciesthat might be able to help us solve thatproblem as we've learned more and moreabout sperm whales they seem to be theperfect candidate to put our non-humantranslation abilities to the test spermwhales make the loudest noises of anyliving creature up to 230 decibels forreference when sounds are made above thewater human eardrums rupture when noisesare louder than 150 decibelssperm whale communities also usediscrete sequences of clicks that getrepeated in specific patterns calledcodas researchers believe this is thebasic communication unit for theirlanguage and codas differ between whalesin different regions for the whales offthe coast of Dominica a typical Coda hasfive clicks and lasts for around 4seconds they have recognizable patternsmeasured by inter-click intervals codasare mostly produced during periods ofsocialization not when the whales arehunting or engaged in other activitiesand calves can take up to two years toproduce recognizable codas and beforethat they Babble just like human babiesall of these elements seem like theingredients to a language that we mightbe able to make sense of thanks tomachine learning and artificialintelligence but how exactly cancomputers help us with this never beforesolved puzzle thanks to seti we may beabout to find outthe world's largest ever interspeciescommunication effort was launched inMarch 2020 with a project known as thecitation translation initiative or setiif that sounds familiar it might bebecause there's another moonshot projectknown as seti the search forextraterrestrial intelligence in thecase of this new citation projectscientists will be collecting data froma group of sperm whales around theCaribbean island of Dominica the whalesof that area have already been studiedfor more than 15 years thanks to theDominica sperm whale project but beforewe can understand how computers mighttranslate whale language let's firstdive into how they've done so for humanlanguagessince 2014 Machine translation haslargely relied on something called anencoder decoder deep learning modelwhich is a system that uses two separateneural networks the encoder is the firstneural network and is where you input asentence in say English the encodertakes each word in the sentence andturns these into a sequence of numberswhich are multi-dimensional vectorrepresentations for each word which arecalled word embeddings the second neuralnetwork decoder will take this sequenceof numbers and output a sentence in adifferent language but for this to workit needs human supervision it relies onknown pairs of inputs and outputs asystem like this wouldn't work totranslate any language that's new orunknown and for several years this wasthe best machine learning could do thatis until 2018 a group at Facebookresearch found it's possible to docompletely unsupervised translationyou do this by taking all the words fromone language and calculating theirstatistical properties more or less howoften that word is used and with whatother words it's often correlated youcan then assign all the words to a pointCloud kind of like a 3D Galaxy based onthese statistical properties then if youdo the same for a totally differentlanguage and compare the point cloudsthey have nearly identical structures aword located in a particular position inthe galaxy in one language would be inthe same location in any other languageyou can take languages as dissimilar asEnglish and Chinese and the structure inthe point clouds will still line up andthis means you can translate languageswithout needing a Rosetta StoneTom muscle a documentary filmmaker andthe author of how to speak whale hasfollowed the work of researchers at setifor years he is particularly thrilledabout how this development could applyto sperm whales that's what all thescientists excited because they thoughtif we can then get a big enough coreperson a big enough amount of uhrecordings and that gives you a weigh-inwith no bilingual dictionaries which iswhat the situation we have with otherspecies and it doesn't mean that we'regoing to make a cloud in sperm whale andthat it's going to match the human onebut maybe there'll be some shapes thatdo maybe some parts of the universe thatthey represent in their communicationdoes match ours but those differencesand similarities are going to give us anidea of what the different languagegalaxies that exist in in life are theAI will be trained not just to predictwhat a single whale might say next butwe'll also learn to predict what asecond whale might say like in aconversation the result will be like awhale chat bot but there's one caveat toall of this for it to work you need aton of data like so much data forreference chat gpt3 needed to be trainedwith over 500 billion words to be ableto do what it does this is the largestneural network that's ever been trainedmaybe we don't need chat GPT 3 levels ofdata but can we even get close to enoughdata from sperm whales for translationto be possiblein order to estimate how much data spermwhale communication the seti projectmight be able to gather we can do asimple back of the envelope calculationscientists are studying a 20 squarekilometer area off the coast of Dominicawhere 50 to 400 whales are observeddepending on the seasonsperm whales vocalize almost constantlyhowever 75 of the clicks are used forecholocation the remaining 25 percent ofthe clicks are used for communicationthe coda clicksa typical Coda from the whales in thisregion is about one click per second sodividing the number of seconds in a yearby four times the number of whales givesus 400 million to 4 billion clicks ayear approximately so not nearly closeto chat gpt3 levels but still enough towork with at the high end it's betweenGoogle's Birch algorithm and chat gbt2levels but how exactly will researcherscollect this data it's a task that willrequire a host of autonomous andsemi-autonomous devices thatcontinuously record the whales fromabove below and all around themone source of data will be from tetheredbuoy arrays these devices will collectmassive amounts of background whalebioacoustic data these are Big devicesand work in a similar way to militaryequipment that detects submarines overlarge areas and distances each arraywill consist of audio recording devicesthat will be positioned at differentdepths in 200 to 300 meter increments upto 1200 meters which is the depth thatsperm whales are known to hunt andattaching certain codas to certaindepths could be a particularly goodplace to start looking for patterns intheir language a study done in 2016reported that sperm whales seem to usecertain codas at certain depths somewere used on the ascent from a dive someon The Descent and some at the surfaceimmediately after returning with manyhydrophone arrays in the waterscientists will also be able to trackthe movements of specific whales overspace and time in addition to the statichydrophones scientists are attachingrecording devices to the whalesthemselves using suction cups these tagsprovide the most detailed recordings ofthe sounds that they make and alsorecord signals from inertial andpressure sensors which will allowscientists to reconstruct the whale'smotion into diving patternsthe tags are especially critical anduniquely identifying whales andconversations and associating thebehavioral patterns with the recordingsand at last are the autonomous soundrecording Robots free swimming andpassively floating aquatic drones willrecord audio and video from multipleanimals simultaneously to observebehaviors and conversations within agroup of whales this will hopefullycover any blind spots from the first twomethods all of this data will becombined into a complex array of datawhich can then be parsed out bydifferent algorithms so what they'rereally doing is they're working with awhole different Suite of machinelearning tools some of them will belooking through all the audio and theirjob will be to figure out when a spermwhale is making a sound or not is thatthe sound of a landslide is that thesound of a sperm whale then they'll haveto filter out is the sperm whaleecholocating or is it a sperm whilemaking a communication vocalization likea coder and then they'll have toseparate them out then they'll have tofigure out who's talking and where werethey talking and what else was going onbecause that sort of metadata you knowwas it the one that then fought with theother one was it the one that washunting was it the one that was nursingit's young and what relationship does ithave to the other sperm whales who wereinvolved in that vocalization exchangewho could have heard it what was the seadoing then they'll try and figure outokay well how different it is to thekind of things that have been saidbefore and how many different kinds ofsounds do they make and that's when youstart to sort of have hypotheses aboutmeaning what do you think they might becommunicating aboutresearchers have already collectedaround a hundred thousand sperm whalecodas and tied each call to specificwhales an algorithm was created toanalyze these codas and categorize themby which whale made which call and itwas accurate more than 94 of the timebut this experiment will not be withoutits challenges there are so manyvariables in any given Coda that it willbe extremely difficult to parse what isMeaningful and what is not for examplemany of the clicks have almost the samestructure but with slight variationslike codas with a different amplitude ora different frequency are thesevariations akin to differences inpronunciation or are these fundamentalvariations that have different meaningsand does the unsupervised machinelearning translation model rely onsomething innately human for it to workwe might use the words in a sentence todescribe how we're feeling for example Ifeel happy today but to translate thisinto whale assumes they have the samesense of self that we do that they havesimilar emotions as we do that theyperceive time in the same way we dothere are some words that are morelikely to be common between us wordslike mother baby Friend or Foe but justas a whale might not understand theconcept behind a word like today whatmight a whale have a word for that wecan't even conceptualizeso even if the seti researchers are ableto collect all of this information aboutsperm whales it's impossible to saywhether they'll be able to translate thecodas in a way that would make sense tous as language but even if we can'tdirectly translate everything thatthey're saying researchers hope thatthere will still be some patterns we canrecognize some fragments of languagethat can tell us more about their livestheir families or even their intentionsor problems because the attempt tounderstand whales whether successful ornot might also be a way for us to carefor and help thema thing that's really struck me throughmy whole career was the more that youcan relate to other species the more youcare about them all ocean ecosystems areunder threat from climate change andcetaceans are among those groups theMarine Heat Wave known as The Blob thatformed over the Pacific Ocean from 2014to 2016 resulted in crashing fish andwhale populations even without extremeheat waves the oceans are still gettingwarmer which means cold water specieslike the white beaked dolphin are havingto find new habitats the availability offood is changing and runoff from theshoreline could eventually lead tocetaceans being poisoned by differentchemicals and on top of this theshipping industry has regular impacts oncetacean populations shipping lanes thatoverlap with important whale habitatscan lead to whales being struck and thenoise of vessels can disrupt hunting andvocalizations whales and dolphins arefacing a difficult future in large partdue to human activity language is afterall the tool that propelled humans toinnovate and solve so many of theproblems that once Afflicted us it's anevolutionary advantage that likelyallowed us to dominate The Globe and itsorigins in our species is one of thegreatest mysteries of modern sciencedid language suddenly develop in humansor did we earn our voices through theslow incremental process of naturalselection that marks all other aspectsof evolution did language only everdevelop in our species Homo Sapien ordid the other early humans use languagetoo this is a hotly debated extremelyexciting corner of science Linguisticsand Archeology and to understand thegreat debate and hear the mostcompelling theories you should watch ourbrand new episode of becoming human onnebula how humans started speaking thisis the fourth episode in our Flagshipseries about human evolution which takesyou through the most important steps inour Evolution that made us theincredible strange Apes that we aretoday for this series we've created ourown world in 3D to show you theartifacts fossils and archaeologicaldigs all in one place like a magicalMuseum of all of the most incredibleaspects of early human in archeologythis is a series that wouldn't havereally made sense to put here on YouTubeit's a series we wanted to take our timewith and a series that's morearchaeological and more philosophicalthan what we normally do here but it's asubject so important to the innatedesire we all have of trying tounderstand our own humanity and that'swhy we created this series for nebulanebula is a streaming platform that wascreated by us by the educational YouTubecontent creators who wanted toexperiment play and just create moresometimes the content we want to make isexperimental too long too short orcontains themes that would get itdemonetized instantly on YouTubesometimes we want to make content thatdoesn't exactly fit on our regularYouTube channels or sometimes like inthe case of becoming human we wanted tomake a beautiful Standalone piecesigning up to nebula is also the singlebest thing you can do to support thischannel by signing up you're directlysupporting us getting to watch all ofyour favorite content ad free and alsogetting access to tons of exclusivecontent nebula is now also offering anebula first content where you can getsome of your favorite creators showsweeks or months before they're onYouTube like jet lag the game orepisodes from Johnny Harris legal eagleand many more and on top of the originalcontent and nebula First videos if yousign up with my link you get free accessto nebula classes where creators hostclasses on how to make the content youknow and love all of the classes aretaught by the creators on nebula so youcan learn how Alex the low spec gamermakes videos on a budget or how foreigntells stories or how Patrick Willemsmakes movies your favorite creatorsteaching you how they create if you'veever wanted to become a content creatorvideo maker or are just curious how itall gets done this is a great place tostartso if you sign up for nebula using thelink below you can get nebula for alittle over two dollars and fifty centsa month or 30 dollars for the entireyear a portion of the subscription feegoes directly to us and the rest goestowards building this platform to becomethe best streaming platform out therefor viewers and creators It's Quicklybecoming the hub for online creators andnow is the best time to join