Going off-ramp for high-stakes logic in AI environment.
As enterprises adopt artificial intelligence (AI), one mistake that they do is, to apply large language models (LLM) to every single problem. As we have discussed in my articles before, since generative AI is probabilistic,
Sustainable architecture as a quality attribute requirement.
AI models of the day require heavy power usage to run; they are computationally expensive. Even for a tiny task, a big Large Language Model (LLM) would end up consuming huge amounts of energy. Just as we have
quality attributes (also referred to by some as non-functional requirements, though as an architect, I disagree with the usage of this term) such as efficiency (performance), security, reliability, that an Enterprise Architecture (EA) is supposed to support,
Architecture for Non-Deterministic Outcomes
As architects, we need to design feedback loops for monitoring to check if our AI models start moving away from their original business purpose. AI-driven IT systems do what they think you mean, unlike our traditional IT systems
that is, they provide the most likely response based on patterns they have learnt. Over time, as new data enters the system or the model interacts with more users, the way it responds can subtly change. Read more here … https://t.co/5i0OsykSsi #AI#NonDeterminism#Probability
Auditing the Shadow AI.
Enterprise Architects have dealt with Shadow IT before. Shadow IT used to result from teams or departments bringing their own software systems without the awareness of the IT team. In the rush to leverage today’s AI, also partly due to the AI hype
going on, we see employees already using unvetted AI tools to handle corporate data that is sensitive. This is Shadow AI that enterprises are facing today. The wall for data protection, is being breached. Increasingly, people in the enterprises use publicly available AI tools,
Enterprise Architecture (EA) has followed an API-First approach, when it comes to integrating the IT building blocks in the EA. And, over time, API based integration has been perfected so systems can talk with each other seamlessly. These APIs exposed by IT systems, allowed
IT system A calling IT system B's API in a pre-determined way, we need a software system (Agent) that understands the goal and finds the best way to achieve it across multiple IT systems...
Knowing When NOT to Use Today’s AI.. In the current rush to integrate today’s #AI into every business process, a valuable skill for you as an Enterprise Architect is knowing when to say ‘no’ to its usage. Not every business problem needs a predicting brain, some only need a
calculator. While today’s AI models excel at identifying patterns and generating content or predictions, they still operate on probabilities (stochastics) rather than certainty or precision. Read more here https://t.co/QKw6nKRCkI
own data over the years of its operations is used in the AI model, AI systems might end up offering less than relevant intelligence. Read more in my https://t.co/2u8VhUv7l4 article here https://t.co/ccnMpQoYoO
#RAG#RetrievalAugmentedGeneration
Corporate Memory is the AI Model in your Enterprise’s Data Architecture
In the early days of the recent #AI boom driven by progress in #NaturalLanguageProcessing ( #NLP ), the focus was on the massive, public AI models such as GPT, Gemini. But for a real enterprise, unless its