Meanwhile after ChatGPT -that has
initiated some amusing online reactions, GPT-4 -OpenAI’s powerful language generative
model, is to be released very soon. I read that it is going to make astonishing
leaps in performance of memory -retain and refer back, and summarizing -distill
essentialities in a text. The issue is Large Language Models (LLMs) like GPTs
are running out of training data. There are lots of data generated online but very less
is within acceptable quality threshold. They are moving towards transcription
of spoken content of formal meetings. I find GPT progress exceedingly interesting.
These started in 2017 with iterative machine learning transformers through neural
network architecture. I find them very real in the way they acquire language. I
got my frame on writing, and ofcourse keep updating, by iterative addition through
constant reading. Better writers and incisive writings mean better frame to
consolidate. Unlike GPT human brain takes years to work, particularly when
there was no frame to begin with. Though I began choosing randomly very soon
conscious part of the brain kicked in to evaluate and found the writings wanting
in quality, nuance and span of idea. Indian writing -with rare exceptions,
lacked the verve and vigor, mostly cosmetic and planted with word play. As also
English classics like Austen were too boring and out of my reality or any
reality I could fathom. Russian and Latin American writings were where things
were happening. I was looking for contemporary writings -and that is how I bumped
into Nobel prize writers which seems a definitive reference in a chaos filled
world of writing. Ofcourse now I have moved out of fiction -it has become a luxury, into nonfiction -there is so much to catch up in the fast changing world. Without awareness of grammar (not that I didn’t try, I really
couldn’t understand at all, as also it starts to intrude into writing and
thinking hence I let it go) iterative consolidation is the only way out. I identify
myself with GPT a lot. Like GPT I sometimes write fascinating stuff but also make
silly embarrassing mistakes since I get into my own loop and brain gets tripped.
But unlike GPT human mind has limitations but fascinating possibilities. You are
not working on words from models of contexts and references alone but also through experiences
and thinking. GPT at this stage is emulating third rate Indian writers who use superficial
play of words to impress the system to make easy money and influence. Intermittently
impressive but mostly mediocre. GPT can easily be better than most Indian writers
(who can be safely taken as benchmark of mediocrity!).
There is also research going on
to augment data through visual language or multimodal access afterall much of human
speech is complemented by visual cues and representations. When I see something,
I am acquiring data that is converting the experience through emotions which
are then imprinted into memory. Many a times I find it difficult to convert
into words, and even if I do, it doesn’t really carry the whole feeling of experience.
It seems to be lacking and futile. Even in best of writers it is very much hinged to partial
emotions it excites in readers. Ofcourse consensus is high on great writers (is there primal quantum entangling in human brains that aligns emotions? Crazy thought!),
but it is never complete, and many a times it could even be readers limitations.
It does create diversity of experience but doesn’t address the limitation of language.
So, is the language lacking since it is in a linear time frame whereas experience
encompass linearity and has input from past and future -oscillating in
time, to add to the richness of memory. AlphaZero has exhibited this ability to
look deep into time and backforth for possible outcomes. It has accessed
reality that human emotions are able to sense but brain is incapable to articulate.
GPT is open ended, subjective frame, while AlphaZero is close ended system,
clear objective, but the way AlphaZero experience time in limited frame holds
lots of potential.
