AI - a State of the Union

As the initial excitement and hype surrounding artificial intelligence (AI) begin to settle, it's time to take a step back and assess the current state of AI from an engineering perspective. This blog is the first in a multi-part series that will explore various aspects of AI, including chatbots, financial inclusivity, software lifecycle delivery, privacy and ethics. Our aim is to examine what is useful to our customers, how AI can be integrated into existing product lines, and key considerations for implementing AI in enterprise settings.

The foundation of many of today's AI models can be traced back to the 2017 paper "Attention is all you need", which outlined the neural network model/architecture that is underpinning many of today's AI models. The launch of ChatGPT 3 in November 2022 further popularized AI capabilities, making terms like "LLM" (Large Language Model) commonplace. Since then, AI has found applications across various sectors, from software development and education to legal research and beyond. Major tech companies have been quick to adapt to this new landscape. Microsoft, Meta, Google, Amazon, Apple and others have all adopted different strategies to get ahead in this lucrative space.

• Microsoft partnered very successfully with OpenAI - everything is "co-piloted"
• Google caught up and after a few hiccups Gemini offers a good alternative to GPT
• Meta has been an early adopter and continues to evolve its models, with Llama as its flagship, with the added benefit that they have kept their model open
• Amazon seems to have decided to be the workhorse AI rather than create their own models
• Apple with their recently launched "Apple Intelligence"remains product centric and doesn't seem to want to compete in the "model" market at this time

The AI startup ecosystem has exploded, with hundreds of new .ai domain names being registered daily. While well-funded entrants like Anthropic and mistral, offer LLM models to compete with the larger companies, many others offer AI-driven value-aded services, which we'll explore in the following parts of this blog series.

Many parts of the world have responded to this rapid growth with new regulations and guidelines, and they have implemented their own regulatory standards on this topic, such as the EU AI Act, and the "Shared Approach to the Responsible Use of Artificial Intelligence in Government" in Canada.

As the field matures, the focus is shifting from achieving higher accuracy or benchmark scores to addressing two key business sustainability factors: cost efficiency and customer demand.

AI Advancements means greater access and easier adoption, not just by bold startups but also regulated industries. However, challenges remain:

Unpredictability:

Although significant advances have been made to prevent AIs from "going off the rails," and guardrails are now either embedded in existing solutions or can be added via frameworks (e.g., NVIDIA MeMo) -- this still doesn't make the outputs of most LLMs fully predictable or explainable. This can affect not only the functional outcome of a solution but also have non-functional impacts on performance, reliability, and other aspects of the system.

Evolution:

We need to account for new or evolved disciplines in the build-out of a solution. Prompt design has emerged as an important new discipline. But it is likely that time will show an evolution is needed in user eXperience (UX) and design disciplines to cater to the specific interactions needed when a user "works" with an AI solution.

Quality Assurance:

This is another discipline that will also evolve. Validating a system where the outcome is not completely predictable brings new challenges. QA strategies will have to adapt to include this added complexity, and the QA approach may need to rely more on probabilities rather than fixed outcomes for validation. The goal may shift from meeting zero defects to achieving a success threshold expressed as a percentage of expected outcomes. Expected outcomes themselves may need to be defined differently, and validation is likely to involve other AI systems.

We will explore these topics and more in the rest of this series, so stay tuned.