geordee

Age of Agents

“What’s an agent?”, my friend asked. To me it’s quite obvious from the word itself. Agent is an entity which has “agencies” - power or capacity to fulfil an action or a service. And agents act on behalf of someone. Today, we want AI models to be the agents - agents to perform actions on our desktop or through the internet of things.

How does it work? The AI models, acting as agents, need knowledge and tools. The knowledge to answer questions, make decisions, or to learn how to use tools. Tools to do things, make changes, send emails, book...

Ready, Set, AI

These days it’s difficult to talk about technology without mentioning AI. I have been quiet about AI for a while, as the enthusiasm gets built up through newer models, larger capacities, and more attractive demonstrations and use cases.

Yet, I have been using AI tools regularly, getting excited about every new feature and slowly waning off the interest as I start using it for some serious work. It’s a pattern I have observed with code completions, vibe coding, writing assistance, audio summaries, image generation and countless other use cases on the world-wide web today. The problem I faced was often...

AI Front Desk

In 1995, I got my first email id. And we had access to the world wide web. We were browsing the internet on dumb terminals without a mouse, using a browser called Lynx. The Internet was a place where we go to read random stuff and kill time.

One of the first technology transformations I witnessed and experienced was Dotcom in the 2000s. Everyone built websites for just about everything - for business, shops, people, pets - even the toasters were tweeting. By the end of the decade, we had some best practices around building and hosting websites, user interface...

Textbooks for AI

By now all of us are familiar with Generative AI. Type a question into ChatGPT, and we wait for the magical response, often better that we can find or write or even think. Some of us have replaced Google with ChatGPT. If the topic is not current affairs, ChatGPT seems to provide better answers and that too without the ads and clickbait.

World Knowledge

ChatGPT, and other competing models, are pre-trained on vast content from the Internet, which it correlates and creates a semantic understanding of the world. I will refer to this as world knowledge. It is no...

Sorry! I cannot help you with that.

My last article was over a year ago. It was about the rise of Generative AI. A lot of things changed in one year. AI models have become much more capable, and fast. And, the hype took a feverish turn, and now it is subsiding for good.

AI is extremely successful in the consumer space. Friends of mine who are not in the technology industry are putting AI to good use - to write, create, record, summarise. They can live with the inaccuracies and ambiguities of the AI responses. They themselves are human in the loop.

During the past one...

The AI Spring

We are living in the AI Spring.

All the big players were holding back their research due to the experimental nature and rough edges of current implementations. OpenAI led one of the earliest moves, along with Stability.ai. Meta released LLaMA, a foundation model for natural language processing, under an open-source license. Stanford fine-tuned LLaMA and built Alpaca, which led to many other fine-tuning initiatives. IBM has released Watsonx. Baidu, Alibaba are testing their large language models. And probably Tencent too.

ChatGPT and Midjourney’s Discord interface made AI accessible to the public. The ecosystem is also getting a boost. Vector databases...

The Curse of Agile Product Management

I was involved in building a large enterprise solution - composed of many products and libraries providing features and capabilities such UI, API, authentication, ledger, databases, enterprise search, microservice runtimes and so on. Almost all the components are product-managed, built and released in an Agile fashion - with fast iterations and frequent releases. This in itself is noble and good (for marketing) but poses a few challenges to large-scale solutions. We need to find that narrow space where all the component products and versions work together. Occasionally the products have dependencies on platforms or libraries which are moving at a...

Prompting the AI

Recently I came across a problem, typically solved by machine learning. A standard problem, addressed through multiple research papers, of capturing the image of an object and figuring out the HS Code - an international classification maintained by the World Customs Organization for traded products.

A typical solution consists of a computer vision solution translating the image to a set of labels and list of objects identified. These object names and labels are then passed into a classification algorithm to arrive at from over five thousand product groups or classes. The challenge is to get a reliable match since the...

Contract First

Digital as a capability has climbed up and down the hype curve and is now seen as an established practice across enterprises. One of the common patterns is to build microservices to solve business problems. There are a few reasons why microservices might be a good choice. Consider these reasons as different facets or dimensions of an application. If an architectural choice is recommended from multiple perspectives or dimensions, it might be worth considering as the dominant alternative.

From the business perspective, a service-based architecture brings in modularity. Traditionally, ERPs tend to be comprehensive one-stop solutions for businesses. This helped...

Trustworthy AI

In 2014, I conceptualized a product to assist legal professionals in India. Central and state governments are responsible for legislations in India. But, when the laws are pronounced and declared by the judiciary, it becomes part of the wider legal framework that includes judgments, and this is known as judicial legislation. This improves the understanding and applicability of laws and aids to interpret in a variety of circumstances. However, it comes with its own complexity - such as hierarchy of courts, applicability in the region, similarity in the nature of cases and so on.

The product idea was not quite...