Core include appears below:
You don understand artificial intelligence yet?
For di past six months, chatbots like ChatGPT and image generators like Midjourney don sharp-sharp become popular.
But artificial intelligence (AI) or "machine learning" models don dey ground for some time nau.
For dis beginner guide we go waka pass chatbots to look di different types of AI - and see how e don dey already work for our lives.
[Note to translators: Please keep your translation of "training" between the HTML codes . Same with other examples you will find further down]
Di instruction fit be "find all di pictures wey get face" or "arrange dis sounds"
Di program fo come search for patterns inside di data dem give am to achieve im goals.
Dem fit need correct am for di way like "dat no be face" and "those two sounds dey different", but wetin di program dey learn from di data and di clues e dey collect na wetin dey forn di AI model- and na di training materials na wetin go determine wetin di AI fit do.
One wey to understand how dis training process fit create different types of AI na to torchlight different animals.
Ova millions of years, di natural enviroment make animals develop specific abilities, for similar way, di millions of cycles wey AI dey make for di training data go arrange di way e dey develop and lead to specialist AI models.
So wetin be some examples of how we don train AI to get different skills
Think of chatbot like parrot. E dey mimic and fit repeat words wey e don hear get context of but e no fullt sabi dia meaning.
Chatbots dey do di same, but more sophisticated, and dey close to change how we dey take releate to di wrtitten word.
But how dis chatbots dem sabi how to write?
Dem be a type of AI wey deym dey call large language models (LLMs) and dem dey train dem with plenti plenti text.
LLM no just dey look di single word vut sentences too and compare di way dem dey use words and phrases for one passage to oda examples across all im training data.
With di use of dis billions of comparisons between words and phrases, e fit read question and generate ansa - like di predictive text for your phone but for ogbonge scale.
Di amaxzing tin about large language models ve sat dem fit learn di rules of grammar and work out wetin words mena demselves without say human help dem.
If you don use Alexa , Siri or any oda type of vice recognition system, then you doun dey use AI
Imagine rabbit with im big ears, wey dey created to catch di tiny difference for sound.
Di AI go record di sounds wen you tok, comot background noise, arrange your speech for phonetic units, wey be di individual sounds wey dey make up spoken word, e go come match dem for library of language sounds.
Dem go come turn wetin you tok to trdt wia any listening errors go dey fixed before dem give you ansa.
Dem dey call dis kain artificial intelligence say e be natural language processinf.
Na di technology wey dey behind evritin like wen you tok "yes" to confam phone banking transaction, or wen you ask your phon to tell you di weather ova di next few days for place you dey travel go.
Your phone don ever arrange your pictures for folders name am tins like "for di beach" or "nights out" for you before?
Then you don use AI without say you know. AI algorithm look di patterns for your pictures come group am for you.
Dem don train dis programs by say dem look through plenti images wey get small description.
If you show one image-recogntion AI enough pictures wey dem label bicycle, soon e go fit sabi wetin bicycle look like and how e take different from boat or motor.
Sometines Ai dey trained to find tiny differences beween pictures wey resemble.
Dis na how facial recognition dey work, e dey find small-small relationship for your face wey make am differnet and unique wen you compare am with any oda face for di panet.
Dis same kain algorithms don dey trained with medical scans so dem go fit find life-threatening tumours and e fit work on thousands of scans for di time wey e go take for consultant to make decision on just one.
Recently dem adapt image recognition put for AI models wey don learn di chamelein likr power to dey manipulate patterns and colours.
Dis image-generating AI dem dey turn complex visual patterns wey dem collect from millions od photographs and drawings and turn dem to brand new images.
You fit ask AI make e create photographic picture for sometin wey no ever happun - for instance, foo of pesin wey dey waka for di surface of Mars.
Or you fit creatively form di image style: "Make portrait of di English football manager, wey dem paint for di style of Picasso."
Di latest Ais go start di process of generatinf dis new image with collection of random coloured pixels.
E go come look dis random dots to see wetin resemble di pattern e bin learn for training - na di patterns wey dey build different objects.
Small-small di pattern go dey boosted with di addition of more layers of random dots, as e maintain di dots wey dey build di pattern but troway odas until finally e start to form sometin.
E go develop all di patterns like "Mars surface", "astronaut" and "walking" togeda and you don get new image be dat.
Because di new image dey built from layers of random pixels, di result na somtin wey never exisit before but still be pikin of billions of patterns e learn form di original training images.
Society now don dey torchlight wetin dis mean for tins like copyright and weda e dey ethical to create artworks trained on di hard work of real artists, designers and photo graphers.
Self-driving cars don dey part of di tok-tok on AI for decades bow and science fiction don make am popular for pipo mind.
Self-driving AI also dey known as autonomous driving and di cars get cameras, radar and range-sensing lasers
Tink of dragonfly wey get 360-degree vision and sensors for di wings wey dey help am moe and adjust constantly as dem dey fly.
Na like dat, di AI model dey carry data wey im get from im sensors ro take identify objects an find out weda dem dey move and if dem dey move, which kain moving object dem be - weda dem be anoda car, bicycle, pesin wey dey waka or sometin else.
Thousands and thousands of hours of training to take understand wetin beta driing look like don make di AI to fit make decisions and take action for real life on how to drive motor and avoid accident.
Predictive algorithms fit don find am hard to handle di unpredictable nature of human drivers, but driverless cars don gbab millions of miles of data from real riafs. For San Francisco, dem don dey carry paying passengers.
Autonomous drivinf na also anoda public example sat new technologies need face abd cibquer more than just technical wahala.
Goment legislation and safety regulations with deep fear ova wetin go happun if we give control to machines still be roadblocks for fully automated future for our roads.
Some AIs dey handle numbers, to collect and combine dem for volume to create plenti information, wey di product fit dey very valuable
Plenti tins fit dey about you for your financial and social actions especially online wey fit make predictions about your behaviour.
Your supermarket loyalty card dey tack your habits and tastes by wetin you dey buy weekly. Di credit agencies dey track how much you get for bank and owe for your credit cards.
Netflix and Amazon dey track how many hours of content wey you stream last night. You social media accounts know how many videos wey you comment on today.
And no be only you, dis numbers dem dey for everibodi and na why A models fit use am check for social trends.
Dis AI models dey already shape ypur life, from helping you to see weda you go fit collect dat loan to influencing wetin you go buy with di kain ads wey dem go allow you see online.
E go dey possible to combine some of dis skills to one AI hybrid model?
Dis na exactly wetin one of di most recent advances for AI dey do.
Dem call am multimodal AI and e dey allow a model to look different kain data like images, text, audio or video, and torchlight new patterns between dem.
Dis multimodal approach na one of di reasons why di skills from ChatGPT3 wey dem only train with text and ChatGPT4 wey dem add images to di training, jump.
Di idea say one single AI model fo fit process any kain data and like dat, e go fit do any kain work, weda na to translate between languages or to design new drugs na wetin dem dey call artificial general intelligence (AGI).
Somepipo see am as di ultimate aim of AI research, but odas see am as di start to those science fiction dystopias wey be say we go create intelligence wey pass our understanding sotay, we no fit control am.
Till recently, di key process for training most AIs dey called "supervised learning".
Huge sets of training data we humans give labels dey used so di AI go fit find di patterns for di data.
Dem come ask di AI make e apply di patterns to some new data and come give feedback for im accuracy.
For example, imagine say you give AIa dozen pictures - six dey labelled "car" and six dey labelled "van".
Next tell di AI to work out one visual pattern wey go sort out di cars and vans to two groups.
Now wetin you tink say go happun wen you ask am to categorize dis picture?
Unfortunately, e be like di AI tink say dis na van - no too intelligent.
Now show am dis
And e tell you say na car.
E clear wetin dey wrong.
From di small number of images wey dem use train am, di AI don decide say na colour be di strongest way to seperate cars and vans.
But wetin dey amazing be ay di AI program decide dis one all by itslf - and we fit adjust im decsion making.
We fit tell am say di way e indentify di two new objects no dey correct - and force am to find new pattern for di images
But more importantly, we fit correct di bias for our training data by giving am more differnet images.
Dis two simple tins wey dem dey do togeda and on grand scale na how most AI systems dey get training to fit make very difficult decisions.
Many of di most recent breathroughs for AI na sake of deep learning.
For simple terms, dis na wia complex algorithms and huge datasets mean say di AI go fit learn without human guidance.
ChatGPT na di most popular example.
Di amount of text for di internet and inside digitised ooks plenti sotay ova months ChatGPT bin dey learn how dem go take combine words for meaningful way by imself.
Imagine say you get plenti books for foreign language, and maybe some get images.
Las-las you fit find out say na di same word dey appear for page wia drawing or picture of tree dey, and anoda word for picture of house.
And you go see say word deynear those words wey fit mean "a" or maybe "di" and so on.
Dis na di deep learning model, wey dem dey also call unsupervised learning.
E dey use eplenti plenti computing power wey allow di AI to memorise plei words, alone, in groups, in sentences, across pages, - and den dem go read and compare how dem dey use dem ova and ova again in less than a second.
Di fast advances wey dey made by deep learning models for di last year don drive new ginger and fear ova wetin artificial intelligence fit do , and sign no dey say e go slow down.
Di promise and warning wey science fiction don dey shout don suddenly reach us and we dey find ourselves inside world wia AI don dey show dia strange inhuman abilities.
Written by Paul Sargeant
Design by Jenny Law, Debie Loizou, Patricia Ofuono and Oli Powell
Development by Assiz Pereira and Alli Shultes
Additional editing by Emma Atkinson and Bella Hurrell. Ogbonge thanks to Maryam Ahmed wey provide guidance on machine learning models.
Fotos: Getty Images
...there we go.
More warnings come from sabi pipo say AI fit finish human beings