Industry executives and experts share their predictions for 2024. Read them in this 16th annual VMblog.com series exclusive.
Feeding the (AI) beast is all the RAG(e)
By Steve Wallo,
CTO, Vcinity
Data-intensive workloads in the real
life
Artificial intelligence (AI) and machine
learning (ML) have already captured the attention of the world. The thrilling
possibilities of technological revolution lurking beyond a close horizon have
proven to amass people from every industry and walk of life alike-those ready
to increase productivity and cultivate innovation, efficiency, and
competitiveness within almost any and every vertical. Tools from ChatGTP to
Amazon Bedrock are making it easier to experiment with analytics and Generative
AI, and we see industry leaders activating their analytics and AI journeys. For
instance, the Chicago Transit Authority (CTA) recently announced it was
partnering with Google on AI-powered ‘Chat
with CTA' bot to ease rider experiences and CocaCola used GenAI to
create a new advertisement for their trademark product in
record time. Seems the possibilities are endless. But... what can go wrong? Unfortunately, a lot.
Hello, hallucinations
The highly popularized adage "with
great power comes great responsibility" really rings true in the realm of
analytics, AI, and ML. As we see the power of AI being harnessed in countless
industries, we also see an uptick of faulty AI outputs... or as we like to call
them, hallucinations.
Without accurate, diverse, and up-to-date data being fed into the AI engine (as
we warmly refer to the as the AI beast), the output will be erroneous. And when
organizations act based on bad insights-they can take a major hit when
misguided decisions are made. Here is
where the more responsibility part comes into play-it is crucial to implement
safeguards within those models to prevent AI hallucinations. These can serve as
foundation guidelines to maintain the integrity of data inputs-obtaining the
right amount, diversity, and freshness thereof-as well as defining risk
tolerances, thereby ensuring the outputs stay within acceptable boundaries and
accuracy. That's
no small charge on its own and before you even consider how to get ahold of
that data, you need to understand what data you need to get ahold of.
So many data options...
That's
where retrieval-augmented generation (RAG) comes into play, providing
organizations a framework for how to optimize the output of a large language
model (LLM) with more contextualized and up-to-date data-as well as where to
get it. To guarantee the fastest possible data ingest into AI/ML models for the
most accurate outputs, organizations must first fix their access to widespread
(pun intended) data.
...In so many places... in so little
time...
The concept of democratized data is
nothing new but the actual execution of it isn't quite there. Yet, this will be a mission-critical
capability for those leaders seeking successful implementation of AI, ML, and
Gen AI. Data creation is booming from all over-outside the data center, at the
edge, in the cloud. Points of data consumption, whether a uniquely skilled data
scientist working from home, scattered branch offices, or a centralized yet
hybrid infrastructure, are equally as dispersed. Getting any dataset-regardless
of its level of perishability, diversity, sensitivity-from point of production
to point of consumption can be challenging, considering geographic distance,
software and hardware interoperability, networks. Throw in the scale and speed
of data access necessitated by today's
data-intensive workloads... and we got ourselves a roadblock. How do we get
around that?
...A perfect use for high-performance
remote data access technology
To effectively and correctly feed the AI
beast, we need both a guideline to ensure the AI model has the right
composition of data to improve large language model (LLM) outputs, as well as
an ability to enable timely, ubiquitous access to data when it's needed. The future of tech and the major focus for the
next few years, will be just that-solving for data curation to optimize AI/ML
model output.
This can be done today by technologies
that allow organizations to move data at outrageous speeds and seamlessly
integrate with organizations' existing infrastructure for ubiquitous data
management. AI and ML innovations need an accelerated data solution that fuels
these initiatives with the fastest data movement and access. This kind of
technologies will be at the forefront of accurate and near real-time data
curation.
The rapid expansion of disperse and
diverse data creation is a wealth of opportunity-prospectively through the
pairing with analytics, AI, and ML-to improve the way we live, work, and play.
Yet, there will be a turning point for organizations to address the holes (or,
in this case, hallucinations) of that expansion. Safeguarding the future of AI
means elevating the composition of your data inputs for accurate outputs. The overall success of an organization's AI journeys starts and ends with
the (right) data.
##
ABOUT THE AUTHOR
Steve Wallo currently serves as
Vcinity's
CTO, overseeing resources related to the insertion of advanced technologies and
strategies into customer architectures and future IT decision methodologies. He
is responsible for bridging future IT trends into the company's existing portfolio capabilities and future offerings.
Prior to Vcinity, Wallo was the CTO at Brocade Federal, responsible for
articulating Brocade's
innovations, strategies, and architectures in the rapidly evolving federal IT
space for mission success. Wallo has served the U.S Government as the chief
architect for the NAVAIR Air Combat Test and Evaluation Facility High
Performance Computing Center.