Unlocking AI's Potential: A Fun-Filled Journey

My journey with AI has been a fascinating one, marked by both evolution and innovation. In the early days, AI was less of a buzzword and more of a practical tool. I recall working on KNN Models within Excel sheets, where AI was simply a basic classification model. However, my first foray into real AI was in 2014 when I embarked on a hobby project for my son. The project involved using basic OPENCV libraries to track his eyes during a quiz, triggering a loud buzzer whenever he tried to cheat. Fast forward to 2023, I find myself immersed in the incredible world of the Falcon7b model – a game-changer in its own right. This model strikes a perfect balance, not overly cumbersome to strain infrastructure yet substantial enough to deliver exceptional results. Language Models (LLMs) represent a significant shift in paradigm to replace human intelligence with machine intelligence. Having conquered numerical and inferential intelligence with calculators and CNN/RNN, the frontier of language ability is the last major stronghold of human intelligence yet to be fully conquered by machines.

Currently, I'm engrossed in experimenting with various transformer-based models, particularly derivatives of BERT, notably RoBERTa. This endeavor led to the development of an enterprise chat engine. However, in the rapidly changing landscape of technology, a model studied just six months ago can seem obsolete.

Some important questions, which I want to ponder over through this brief note are as under

1.      Where are we right now?

As I see it, we are on the cusp of achieving remarkable advancements in AI, but we're not quite there yet. We are atleast 3-5 years away from the point where AI will become as ubiquitous as internet/email.

2.      Where is the gap?

The primary gap lies not in algorithms or methodologies but in infrastructure. Even moderately sized models like the 7b require a substantial 15-20 GB of GPU, rendering higher models prohibitively expensive. Google Colab provides free complimentary access up to a 15GB storage threshold, beyond which it transitions to a cost-intensive service tier. In terms of models related to AI field, one notable challenge and irritant is the prevalence of CUDA, an integral component in many models, primarily associated with NVIDIA GPUs. It is noteworthy that this robust integration of CUDA into numerous AI frameworks and models has engendered the whole ecosystem. There is an implied alignment with NVIDIA hardware, inadvertently steering users towards acquiring NVIDIA GPUs. Well that’s a story for some other day and I don’t want to spat on CUDA here.

3.      What are the security risk to our internal data?

Perhaps the most critical question organizations must address pertains to the security risks associated with their internal data. Many individuals inadvertently share sensitive data with entities like OpenAI through resources like ChatGPT, raising concerns that these data points could potentially find their way to competitors or malicious actors via OpenAI APIs. I've consistently held the view that Google posed a significant threat to privacy, but OpenAI has now taken its place.

4.      What should organisation target for in medium term (2-4 years)?

While there is no one-size-fits-all strategy, substantial gains in productivity can be realized by harnessing AI. For instance, mundane tasks such as data analysis or responding to generic customer queries can readily be handled by AI in its current form. A standard benchmark can be prepared for all the task. The benchmark can involve simple matric such any task which has not involved any innovation from last 3 months should be treated as a mundane task and should be handed over to AI. It can even be seemingly tech related tasks such as coding, data refining etc.

5.      What’s the way ahead for organizations in short term (03 months – 02 years)?

In 2023-24, not adopting AI is no longer a viable option. However, it's nearly suicidal to explore web-based open-source AI tools due to lack of transparency and data privacy related issues. Ideally, organizations should work towards developing in-house models, but this necessitates deep technical expertise in areas like encoder-decoders and vector dynamics. A middle-ground approach is often the most practical, where existing resources can be utilized without compromising data integrity. Numerous open-source models are available for fine-tuning to an organization's specific needs, and this process doesn't demand extensive technical expertise. Platforms like Hugging Face offer a wealth of models under one roof.

Throughout my journey, I have been primarily self-taught, placing immense faith in the potential of peer learning. Consequently, I extend my willingness to engage with not-for-profit organizations and individuals, offering guidance on harnessing the power of AI in both daily life and aspirational projects. The present era brims with exciting opportunities, and I wholeheartedly invite you to join me in exploring its boundless potential. It’s an exciting time to live and let’s explore it together.