There is never a dull moment since Gen AI crashed into our lives. It is very difficult to keep up with the amount of innovation and new startups exploring this space. If, like me, you have not been reading Gen AI news 24/7 then the recommended readings in this article could give some well rounded overview of what’s been happening recently and what’s driving some of the new buzz words.
While the initial Generative AI feats were quite impressive a few years ago, it was limited to the extent and date of training data it was exposed to. If you needed some updated information the model was going to hallucinate. The introduction of frameworks like retrieval augmented generation RAG which allows the LLM model to work on data sources provided by the user (for example a document or database) and answer based on that) has allowed for much more focused and productive outputs from these models.
The Agentic AI concept which further abstracts nodes as individual agents orchestrated by an overall “graph”, allows to combined different “roles” of LLM agents and introduces the use of tools (ex: generate python code or send email etc.) to further create useful workflows. This allows LLM to move from being an interesting chat bot to an agent that can take actions and operate workflows.
It is a very quickly changing domain and to keep up with it requires a serious time investment but I believe we are at a stage where it is not a luxury, most of us need to be aware of these developments and start to think about how to make use of them in our personal and professional lives.
If we take an analogy from the age of the personal computer (PC) coming into most of our lives (for me that was the early nineties) and how computer literacy became a must in about 10 years, it was in everything related to education and work. How many engineers and corporate professionals can imagine their work without a spreadsheet for the past 20 + years? What was initially considered as a job challenging technology quickly became the norm and computer literacy today is no longer a big achievement.
A similar effect happened shortly after while the PC became a tool for the masses, which is the availability of internet access (mid to late nineties) and the buildup of global broadband connectivity in another 10 years or so. Today, we look at our multiple connected devices we need to function in the workplace and at home and in our vehicles and quickly realize this all happened so quickly. For good or for worse, connectivity has become a commodity like electricity, and my generation had some villages in the early eighties that did not have consistent electrical power.
The weaving of AI into our lives will probably be even more ubiquitous and it’s important that we build the proper learning and literacy programs to know how to be proficient users and also continue to have people interested in diving deeper into the technology and developing it further. The biggest nightmare is that we become simply too lazy and complacent to continue developing our skills and rely on all these AI helpers eager to answer our every question.
I am specially concerned for the younger generation, as our ways and sources of learning are very different from theirs. When I was a kid I had to rely on books or teachers/mentors to learn any new skill. For the past 15 years, learning was available and accessible freely online along with a stream of nonsense and garbage on video sharing and social media platforms. Now kids can do research with AI and skip all the hard work needed, is this better for them? Will they develop higher intellectual functions as a result or will they simply become dependent on this technology? It’s very difficult for me to know the answer to that, I only know that it’s better for them to understand how AI works, the best way to use it, and when not to use it as a basis.
How to start learning Gen AI?
The good news is that there is a lot of information available. Google has published a series of whitepapers and python notebooks on Kaggle to keep us busy for some time. I will refer some of them below.
Google Paper on prompt engineering best practices. Explains key approaches to promoting your LLM such as Zero-shot, few-shot, Chain of Thought and how to draft a prompt that helps you and the LLM get the job done.
https://www.kaggle.com/whitepaper-prompt-engineering
Google Paper on embeddings and vectors stores. Embeddings allow representing multi-model data (text, pictures, video, speech) in a unified vector representation which preserves the relationship between different objects. This is essential for Retrieval in RAG. A vector store is the library that stores all the embeddings .
https://www.kaggle.com/whitepaper-embeddings-and-vector-stores
Google Paper on agents. Think of your favorite LLM as representing one helpwe that can do specific tasks with the help of the right prompts. Now imagine you can have multiple helpers or agents, working together to fulfil a bigger plan, like researching a topic and writing a report. This is possible through agentic AI frameworks that allow us to define different nodes (instances of LLM that are primed for certain role or function), connections between nodes (edges) and overall workflow logic and tools that allow the agents to utilize certain functions like web search, generate python code. This is what a lot of hype is about and it certainly is an impressive step in Gen AI’s evolution, I am sure there will be a lot of innovation around this.
https://www.kaggle.com/whitepaper-agents
Google Paper on agents companion. A more advanced look at agents, highly recommended to read it.
https://www.kaggle.com/whitepaper-agent-companion
Google Paper on solving domain-specific problems using LLMs. showcases two exampes, the first in cybersecurity operations enabled through agentic AI, RAG, tool use. The second in medical Q&A system that can support diagnostics and care through instruction fine-tuning and prompting strategies.
https://www.kaggle.com/whitepaper-solving-domains-specific-problems-using-llms
A very interesting paper about an agent model STORM that spins out a research assistant able to take a specific topic and research it from the perspective of multiple analysts, each with different persona, objective, background that will ask questions to an “expert AI system” that relies on web search and wikipedia , possibly RAG to answer the analysts questions. Then other agents summarize and combine the output in a properly formatted research with references.
https://arxiv.org/abs/2402.14207
Tips
If you get stuck or don’t have time to read all this, Gen AI is at your service. You can use notebookLM from google to upload any PDF , video link, website and start chatting with it, asking questions, asking for explanations, summarizing etc. There is also a studio function where NotebookLM will generate a podcast style audio with 2 people talking about the content, summarizing and commenting…
https://notebooklm.google.com/
What next?
Have you started thinking about how you would use agents? I know I have, I am eager to start tinkering and I am constantly impressed by actual innovations appearing left and right in this domain.
Wasla 
Leave a comment