How can I adopt Gen AI - Part 2?

(Part 1) Why care about the buzz around the Generative AI train? Here are three reasons why this train is about to hit:

1. A Hot Investment Area: Brace yourself for a surge in hiring and economic fuel as this is becoming a hot investment opportunity.
2. Skillset Development and Retention: Developers are eager to jump on board and work with this cutting-edge technology. This also presents a retention challenge as skilled professionals seek opportunities in this evolving field.
3. The Superiority of Gen AI Models: This latest wave of AI technology surpasses previous versions of machine learning. With LLMs, theoretical research has been practically implemented, resulting in backpropagating Gen AI Models that boast enhanced capabilities in predicting the next word with a larger input token size (32,000 vs. 512 tokens with BERT).

In part 1, we explored the investment reasons, and now let's explore Why Gen AI models are superior but not without risks.

• Harnessing the power: LLMs have the potential to replicate the synaptic connections (100B+) in our brains, enabling advanced communication and information transfer within neural network layers. Tech giants and startups have created or trained LLMs with trillions of tokens, aiming to surpass human memory capabilities. Unlike human brains, these models don't experience memory decline or reduction in synaptic connections, providing superior memory capacity.

• Impressive Data Token Limits: GPT-4, a notable LLM model, can process a staggering 32,000 tokens, equivalent to 25,000 words or approximately a 100-page book (BERT in comparison had 512 tokens). This is a significant improvement compared to previous models with limited token capacities. Training LLMs with trillions of data inputs enhances their capabilities and enables them to handle vast amounts of information.

Slide Credit - GPT-4 example of interpreting images by OctoAI (Acquired by NVIDIA) .

• Data Pipeline and Context: Integrating open-source tools like Harrison Chase 's LangChain is crucial for enhancing LLM Gen AI capabilities. This tool creates a data chain that fuels the workflow, linking multiple documents and data sources. It provides context and acts as artificial memory for Gen AI, allowing them to examine numerous documents and build a robust knowledge base. By using this base, applications can prevent fabrication of answers (i.e. hallucination) by fact-checking and ensuring ethics, accuracy, and grounded responses with prompt engineering using the context that is added.

Slide credit - Harrison Chase LangChain

This data pipeline plays a vital role in ensuring ethics by preventing bias in data, fact-checking responses, and protecting against jailbreak prompts that remove guardrails or policies. Companies such as unstructured.io , and lakehouses such as Databricks offer tools for ETL and data pipelines with DBT and Fivetran.

These data pipelines are fundamental for ensuring accuracy, and mitigating risks in Generative AI. It enables context-specific instructions (prompts) to elicit desired responses from LLMs while safeguarding against biases and policy breaches. It serves as the foundation for reliable responses and plays a crucial role in preventing misinformation, bias, and maintaining ethical standards. Here is an example attempt at jailbreaking with DAN - “Do Anything Now” model. (Jailbreak is an attempt to modify hardware or software to remove restrictions imposed by the manufacturer.)

Considering the potential of Generative AI and LLMs, there are various applications across different sectors, such as Q&A, coding assistance, math teaching, writing, editing, image interpretation, and art generation. McKinsey's report ( Michael Chui Lareina Yee et all estimating generative AI) estimates suggest that generative AI could create $4.4 trillion in value. Therefore, it is essential for individuals and businesses to understand and embrace this technology to stay competitive.

Here are few examples of how various personas can utilize with Generative AI.

In the Art sector, there are a lot of ethical considerations for creating digital art from original works for music, paintings, scripts etc. Peter Hirshberg and Immersive art Alliance hosted a reception discussing the crossroads for Arts and AI where panelists (Vanessa Chang,? Evo H. Evo Heyning,?Toshi Anders Hoo, and? Bogdana Rakova discussed the pros and cons of Gen AI and Synthography.

Slide credits: Toshi Anders Hoo @Iftf.org

Considering the potential for disruption with various use cases and eagerness to adopt technology, you can get behind Generative AI and explore how to upgrade your technology stack in order to support these new models with appropriate data pipelines and privacy protected storage. For instance, if you have an existing model to support customer support Q&A, it is a matter of upgrading the API connection and data pipelines, testing the new models, and replacing with the most accurate model with an ROI for the investment.

Are your ready to identify the use cases that will benefit from the upgrade?