Artificial intelligence fit show say e sabi tings with di way wey machine learning dey work.
Even as dis way of lanin' dey follow wetin pipo don set, dem never explain well-well why sabi-sabi dey come out from all dis steps and di way artificial intelligence dey built.
For dis article, as we go tink deep about wetin lanin' truly be, I go try find out why sabi-sabi dey show face.
As we go deep inside di idea of lanin', we go see say artificial intelligence and our brain-brain get one kind nature wey make dem fit lan how to lan.
Dis one dey show say one kind system wey dem fit call Natural Born Frameworker dey exist.
Laning with Body and Laning with Tọnge
We dey come sabi di world around us and make our power strong by lookin' things with our eye and movin' our body.
Dis one too be one kind of lanin', wey dem fit call laning with body.
For di other side, when we generally dey tink about lanin', we fit imagine say we dey make our knowledge plenty by readin' books or listenin' to teacher explain.
Apart from lanin' wey dey based on school work like dat, we also dey get different different knowledge from talkin' with friends, online news, and other places.
Dis kind of lanin' no be about seein' picture for eye and memorizin' am or lanin' by movin' body, but na lanin' with tọnge.
Metacognitive Laning and Metaphysical Laning
For laning wey dey use tọnge, sometimes na to repeat am plenty times before you fit remember one knowledge, and sometimes you fit lan am after just one or two tries.
Or, some knowledge, even if you no remember am completely, you fit still use am by findin' de details from book or internet when you need am.
For di sense of gettin' and usin' knowledge well-well when e needed, dem fit consider both of dis ways as laning.
Among dem, knowledge wey you no fit remember without plenty repetition, dem fit call am metacognitive knowledge. Di process of lanin' de concept itself na metacognitive laning.
Dis one be like physical laning, where repetition dey involved for seein' things with our eyes or movin' our bodies. Dem fit also put dis ones for metacognitive laning.
For di other side, gettin' knowledge wey you fit remember with few trials or use by lookin' am up sharp-sharp, dem fit call am metaphysical laning.
For dis case, concepts wey dem don lan before through metacognitive laning fit be used to lan new knowledge as types of those concepts or as combinations of concepts.
Since concepts wey dem don already sabi through metacognitive laning fit be used, metaphysical laning no need repetition.
Natural Tọnge Machine Laning
Make we take dis one apply am to machine laning for artificial intelligence.
Normally, neural networks wey dem dey use for machine laning dey do metacognitive laning, wey involve say dem dey lan concepts by repeatam.
For di other side, big language models wey fit do natural tọnge processing like humans fit do laning with tọnge.
During di pre-training and fine-tuning of big language models, metacognitive laning wey dey based on tọnge dey happen.
Then, one trained big language model fit answer by usin' knowledge wey dey inside di sentence wey dem put in, wey mean say e dey do immediate metaphysical laning.
Dis power for tọnge-based metaphysical laning dey allow big language models to use new knowledge without say dem go lan am over and over.
Dem fit compare dis one with old numerical machine laning, wey dey adjust model parameters by repeatam, and dem fit call am natural tọnge machine laning.
Natural Tọnge as Di Metaphysical Interface
Natural tọnge dey sit for di place wey dey separate metacognitive laning from metaphysical laning.
De tory wey dey sweet for natural tọnge na say you fit lan am through metacognitive laning, and on top am, metaphysical laning fit happen.
Other Metaphysical Interfaces Apart from Natural Tọnge
For real, metacognitive laning and metaphysical laning dey also happen for physical laning. For example, person wey sabi sport well fit quickly fit into a new sport wey e never play before.
Likewise, person wey sabi biology well fit quickly understand di characteristics of a new kind of animal or plant when e see am.
So, even for physical laning, one metaphysical interface dey exist wey get similar position to natural tọnge.
Framework
Wetin dey for dis interfaces na one framework wey different from basic concepts or knowledge; e dey define how dem relate and structure, and e dey make new structuring possible.
As different different metacognitive knowledge dey acquired through metacognitive laning, sometimes e fit dey possible to lan di framework for di metaphysical interface from di way dis metacognitive knowledge pieces dey connect.
One framework wey come from physical laning dey make am possible to immediately get new knowledge through metaphysical laning after person don sabi am well. But, di knowledge wey dem get through dis kind metaphysical laning no easy to tell other pipo.
For di other hand, di framework wey come from laning with tọnge na natural tọnge itself.
So, knowledge wey dem acquire through metaphysical laning by lanin' di natural tọnge framework fit directly enter another person's tọnge acquisition.
Dis one no just apply to knowledge wey mainly dey based on tọnge acquisition, like textbooks or online news.
One experienced soccer player wey dey try baseball for de first time fit sabi explain de metaphysical knowledge of baseball wey e acquire, and tell am to other experienced soccer players. Dis one mean say if pipo get di same metacognitive knowledge, dem fit use words to tell wetin dem dey call "tips" or "tricks".
Also, person fit use tọnge tell other biologists about one new animal or plant wey e see, and by so doin', share dat knowledge.
So, natural tọnge don show face say e be very strong framework wey dey for di metaphysical interface.
Virtual Framework
On top of natural tọnge, anoda framework fit dey acquired.
Dis one include frameworks wey specific to particular area or metaphysical frameworks.
For different academic fields, business areas, and everyday life, plenty different frameworks wey specific to particular area dey.
Scholars fit make new discoveries inside their specialized frameworks and easily tell other scholars wey get di same framework about these discoveries as knowledge.
De framework itself fit sometimes be expressed for natural tọnge, for dat case, individuals or big language models wey get a natural tọnge framework fit acquire and understand am.
Business models and cooking recipes na also examples of such frameworks wey specific to particular area wey fit be expressed for natural tọnge.
Again, mathematical formulas, programming languages, and business analysis frameworks na formal frameworks.
Dem fit also express or explain dis ones for natural tọnge.
Such frameworks wey specific to particular area and formal frameworks wey dem build on top of natural tọnge fit be called virtual frameworks.
Dis one go easy to understand if you imagine one virtual machine wey dey run anoda OS on top one physical computer. Anoda framework dey work on top of natural tọnge, wey dey serve as di main framework.
Native Framework
For de start, na through natural tọnge dem go understand dis virtual framework, but as you dey practice am, e go pass explain-explain and understand-understand through natural tọnge. E go just start to work directly as one metaphysical interface framework wey dem build on top of metacognitive knowledge.
Dem fit call dis one a native framework.
Natural tọnge na, in a way, a native framework, but na only for your mother tongue. Generally, languages wey no be your mother tongue, dem dey acquire dem as virtual frameworks. As you dey sabi am well-well, dem dey come close to bein' a native framework.
De same thing dey apply to domain-specific frameworks and formal frameworks. Mathematicians fit dey communicate natively with each other using mathematical formulas, and programmers fit understand wetin each other want just by lookin' source code without comments.
Dis one dey suggest say de move from a virtual framework to a native framework fit also apply to big language models.
De idea of findin' virtual frameworks wey dem dey use plenty, generatin' plenty example data using those virtual frameworks, and then fine-tuning dem to become native frameworks go be worth tryin' sharp-sharp.
Natural Born Frameworker
If we look at dis, we go realize say big language models fit dey learn all dis special and formal frameworks not just when dem dey fine-tuning, but also when dem dey do pre-training.
Again, for dat process, e fit be say dem no dey learn special or formal frameworks naturally from de start. Instead, dem first dey learn de natural tọnge framework, and then, while dem dey sabi am well or after dem sabi am, dem go learn special or formal frameworks and make dem become native frameworks.
If we go deep inside dis idea of framework learning wey dey happen small-small, e fit also be say natural tọnge learning itself na one parallel pipeline of framework learning wey dey very detailed and incremental.
Dat mean say, from de plenty text wey dem give as learning data during pre-training, big language models fit dey learn not just individual concepts, but also some very simple rules of natural tọnge as frameworks. Then, dem go use dis simple frameworks as foundation, and dem go fit dey learn rules wey dey small-small more complex over and over.
For dis way, from when dem just dey learn individual word concepts, dem go fit acquire compound words and basic grammar, then understand sentences, and eventually learn complex elements like literary techniques and ways of expressin' things.
Dem fit understand dis as a model of framework learning wey get layers and plenty parts, where one framework dey serve as de foundation for learning de next one.
Dis one dey show de picture of big language models as Natural Born Frameworkers, wey naturally get de system to learn frameworks from de start.
Atenshon Mekanism
Di technology wey dey make de Natural Born Frameworker happen na de atenshon mekanism.
De atenshon mekanism be like say e dey choose tokens wey dem suppose focus on inside one context. E dey make di relationships between tokens clear. Dis one na exactly wetin a framework be: e dey abstract by keepin' important concepts while e dey make de relationships between those concepts clear.
By changin' dis selection for each token, e dey possible to dynamically change frameworks too.
Dis one dey allow us to explain why de atenshon mekanism na a very important technology for de way big language models dey grow, by usin' de model of de Natural Born Frameworker.
Kọnklusion
If dis system dey truly happen for de pre-training process of big language models, then de way dis models dey work, wey bin dey hard to understand before, go come fit be explained.
Dis explanation cover di metacognitive and metaphysical laning wey we don discuss, de framework as a metaphysical interface, natural tọnge wey dey make tọnge acquisition and virtual frameworks possible, and de atenshon mekanism wey dey bring de Natural Born Frameworker to life.
Again, two other things go come out from dis one.
First, natural tọnge get one kyn structure wey good well-well for developin' complex frameworks small-small from simple ones into native frameworks.
If natural tọnge first come out for human societies for a simple way and gradually grow to get a more complex and rich structure, then dis one na normal result.
Also, a structure wey dey allow quick laning go be better. If we assume say many societies with different natural tọnges bin dey compete, de idea say de natural tọnge wey good pass for laning don survive till today na easy to prove.
Tinkin' about wetin natural tọnge be go lead us to de second thing: say we humans too be Natural Born Frameworkers.
Even if de particular foundations and systems dey different, our brains too must get one system, like de atenshon mekanism, wey dey learn and flexibly change frameworks small-small.