I vs Algorithms 3 – Age of GPTs
Let’s look at few scenarios.
2002
Pratik
and Vivek stood outside the Dean’s office anxiously.
They
were charged with copying a dissertation paper from each other, while a unique
submission was required.
The
office assistant ushered them in at the scheduled time.
The
Dean thundered, “When you know this is a course deliverable, why did you
indulge in this copying act?”
“Sir,
it was my dissertation paper. I had shared with Vivek as he wanted to review
certain concepts.” Pratik pleaded.
“No
Sir, this is my original document.” Vivek defended.
“Unless
whoever has copied owns up, I am going to suspend marks for both of you.”
“Sir,
I have back-up documents of tests that I had run and the results are captured
in my paper. If you may please allow me, I can share the documents and prove my
case.”, Pratik requested.
2019
“Your
patent application is rejected, as several sections were found to be extracts
from a journal published in 2008.”
The
letter from patent registration office landed on CEO's desk.
“Sir,
we had done fact-checks and plagiarism checks before submission.”, the Legal
head tried to reason.
The
CEO with eyes still glued to the letter, “Please issue a termination letter to
our Head - Research and the team that submitted this application.”
2025
“How
do you explain this trading strategy turning negative returns for third
consecutive month, when broad index is up 9.7%?”
The
CEO enquired the Fund Manager.
“Sir,
we are investigating the trading model. Even the CIO has engaged his team.”,
the Fund Manager looked across the table at the CIO.
The
CEO asked impatiently, “How is the technology team involved in your trading
strategy? There are no execution error or trade delivery issues.”
“Sir,
we are investigating the codes as one of the team members had used a coding AI
tool, without authorisation.”, the CIO spoke with regret.
The
learning in humans and machines
The
scenarios state an elementary human need – “to reduce effort”, for recurrent
actions or events, and for new discoveries.
Be
it the college dissertation paper, the patent application or, the trading
strategy, the desire to reduce effort as intent precedes the method deployed to
achieve the goal. There are inherent conflicts, temptations and incentives that
may impair judgment about sanctity of method.
Learning
on the contrary is a difficult path that requires rigor, research, application
and validation. Learning may reduce effort, but learning inherently is a
demanding multi-step process.
Learning
when applied to machines, start with development of mathematical models called
algorithms to a variety of inputs like data, text, images or, sound. Much like
human cognition, machines learn with algorithms that when fed with tons of data
from MB, GB to PB or Zetta bytes, can turn these into artificial intelligence,
or AI based on a multi-step testing and validation, mathematically speaking.
What
are the points of failure in this learning, reduce effort and AI mass
deployment?
Unsupervised
and unknown
In
recent memory, when Covid-19 started there was no reference to determining how
quickly would it spread or, how would lockdowns and vaccination flatten the
infection spread.
Similar
situations arise in fields of academic research, medicine, space, business et
al.
The
input data or signals in these situations may be a combination of qualitative
and quantitative. Implementing algorithms with these data inputs have to be
passed thru learning stages.
In
unknown situations, the algorithms shall require assisting with validation
through domain experts, or a certain understanding of what the results indicate.
Imagine
Christopher Columbus used an AI-navigator and set the destination as India. The
navigator is supposed to take coordinates and direct the ship to India.
Columbus and his team have never been on that route and hence dependent on the
navigator. Upon landing on the shore, the team celebrate sighting India, when
it ended up being America.
In
unknown and unsupervised situations like here, not all outcomes may end up so
favourable.
What
was used as a tool to guide, or reduce effort of arduous discovery and mapping
of the route to India, carried risk that may be disproportionately higher
depending on area of deployment.
Unbridled
AI
Implementing
AI so far has been in areas of research and by technology/IT-ITES companies to
consumer platforms all within realms of industrial applications, with their own
guardrails and spillovers. Consumers are feeders to the giant AI engines that
work behind social media, streaming content to search engines.
With
Generative AI, ChatGPT and upcoming GPT stores, large-language-models (LLMs)
have now made their way into code building, imaging, creatives, blogging,
3D-printing to host of consumer applications.
It’s
as easy as an app-store download and use.
What
does it change?
The
learning effect which is knowledge acquired from a set of actions repeated, or
for discovery has been the basis of continuity and self-correction in human
civilisation.
Large-scale
development of GPTs as a class of AI, and ease of consumer access has
implications on the foundation of learning and accountability around data,
security and authenticity.
A
college project around monetary policy implications on economic growth may turn
from a careful study of policy actions and consequent effect, into a GPT prompt
and report ready in few minutes. Even referencing the sources shall prove a
challenge.
Besides the potential impurities in input data being taken into models from public sources that may not pass verifiability, there are vulnerability points at storage, refresh and even the algorithms being used.
With consumer access there is a risk of mass adoption of verified to unverified AI models, even code stacks.
An
autonomous car given its physical size and impact on traffic safety may carry a
higher hazard value of AI deployment than a simple text editor. A trading
strategy gone wrong may cause a market anomaly or, undetected may cause a
crash, with significant financial loss. Errors or risks that seem trivial in
early stages, may develop into large impact outcomes.
There
is already growing need for greater monitoring, plagiarism tools, fact-checkers
to filters to detect AI-generated content.
Re-thinking
GPT
ChatGPT
may be the Cambrian moment in growth of common language as input to consumer
and industrial AI. Use of common language as input opens it up for large-scale
adoption.
In
the last 12 months, there has been curiosity, hype, new marketing language,
rapid adoption to growing concerns around its effect on jobs to veracity of its
usage.
Writers
Guild of America representing over 11,000 writers in Hollywood went on strike
for reasons including guarantees that AI would not impact their compensation
and protect jobs.
The
proposed European Union AI Act is a step towards bringing data, accountability
and security with framework around oversight, implementation and ethical
applications.
As
the world gets digitally inter-connected, AI becomes an inevitability.
However,
anything that is not supervisable, verifiable or, explainable is a random
event, cannot be AI.
This
understanding is vital; our history in future depends on it.
Previous
posts in this series:
I
versus Algorithms (bluepeepal.blogspot.com) – 2019
I
versus Algorithms 2 – Who moved my home page? (bluepeepal.blogspot.com) – 2020
Comments
Post a Comment