Adversarial AI

 

With all the innovations and wonders that AI is going to unfurl on our lives, let’s not forget that AI can also be used adversarially to harm against us.  Bad actors might exploit the vulnerabilities in AI systems to disrupt and devastate the real intention of the systems.

Some examples of adversarial AI attacks include:

1)  Image recognition in automobiles might be corrupted to misinterpret a ‘Stop’ sign as a ‘Yield’ sign and might cause an accident.

2)  Algorithm of financial system might be manipulated to cause stock market crash and destabilize economy.

3)  Cybersecurity attack on corporate IT systems to disrupt the company’s daily operation.

The MITRE organization, a consortium by government, industry, and academia, has prepared ATLAS™ (Adversarial Threat Landscape for Artificial-Intelligence Systems) as a comprehensive knowledge base of adversary techniques on AI systems.  MITRE's objective is to increase awareness of the evolving vulnerabilities that might exist in AI systems.

It's concerning that there're so many ways bad actors can exploit AI systems for nefarious purposes.  As a result, AI developers need to be aware of evolving vulnerabilities and take important steps to ensure their AI models are built with strong safety protocol.

LLM Visualization

 

I came across Brendan Bycroft’s LLM visualization and was mesmerized by his stunning work.

Viewing Brendan's work is a treat intellectually and visually.  The description will explain the steps in detail and the visualization will aid in elucidating the entire process of large language model construct.  You will appreciate the dedication that Brendan had put into his magnificent work by making learning a complex subject like LLM algorithm this rewarding and enjoyable.  So go ahead explore and be wowed!

Dexa.ai

 

Podcasting has become a popular platform for people to get informed.  It’s a growing phenomenon in term of content as well as the number of listeners, with an estimated 100 million listeners in 2024 in the US alone.  But how do you search for content in your favorite podcast channel?

Introducing Dexa.ai, an AI-powered search tool designed to make it easier to find specific content within podcasts.  Dexa is building a new search experience where it analyzes and transcribes podcast conversations with precision.  If searching on Google will give you a list of web links, then searching on Dexa will provide a list of podcasts of who said what and when in any conversations.

For example, here’s the search for Joe Rogan’s podcast conversations on the subject of how to stay fit.  The result shows a summary of Joe Rogan's opinions on the subject and the links along with the precise time within the episodes.  I found the search result relevant and the experience refreshing with the user-friendly UX design.

Dexa.ai is a different search engine that can revolutionize the way people seek knowledge in an audio format.  Will Dexa eventually pose an existential threat to Google’s dominance?  Perhaps it’s too early to tell, but one thing for sure is that AI is going to upend many industries and companies in the years to come.

LSTM vs. Transformers

 

Now that LSTM and Transformer architectures have been discussed in my previous posts, let's ask GPT-4 to summarize and highlight the differences between these two architectures.  

Transformer has become the de-facto architecture for natural language processing (NLP) due to its parallel processing capability, parameter efficiency, and innovative attention mechanism.  That explains for the T in GPT, which stands for Generative Pre-Trained Transformer.  By the way, NLP is an important component of AI that focuses on the interaction between human language (spoken or written) and computer.