"The days are long, but the years are short."
If there were one saying that would capture the speed that AI has moved that would be it. The long days of total immersion into learning and feverishly trying to catch up with Information Retrieval, Machine Learning, Process Automation, Networking, and Neurology were strong enough to decaffeinate the strongest energy drinks available. The unrelenting persistence of new sub-Reddit's, AI this, and AI that, and opinions changed faster than many could keep up with.
What went from a casual question of "have you used ChatGPT?" turned into debates over what Llama model was better and PoCs that proved hundreds of thousands of dimensions didn't make your LLM better. Why is that? The simple answer is specialization. Think of dimensions as scenarios of possible outcomes without focus. If you don't know what you want then you must weed out more. If you have some kind of idea then you refine that until you have your perfect outcome.
A Large Language Model is the generalist. You can bring it almost anything and it will find its way to an answer. A Small Language Model is the specialist. Fewer dimensions, but the right ones, trained deep on a curated topic. The specialist doesn't know more than the generalist — it knows exactly what you need, better than anyone else in the room.
Here's all three in one example, framed as ordering dinner.
I don't want to cook dinner so I'm getting takeout.
Here are my large list of questions: The LLM
• What do I feel like having?
◦ A warm cooked meal?
◦ A sandwich?
◦ Some kind of soup?
◦ Drive through or takeaway?
◦ Sit down fast or sit down enjoyable?
◦ A lot for a low cost that means leftovers I may or may not eat?
◦ Do I like Indian, Chinese, Italian, American, Spanish, Thai, Pub food etc.?
◦ How far am I willing to travel?
◦ How much do I want to spend?
◦ Would I like to eat with friends, or just around a living crowd?
◦ Do I really want to be around people?
◦ What do I feel like having?
[Continue to iterate until you find what you want.]
Here's my short list of questions: The SLM
• I don't feel like cooking and I typically like
◦ Fresh food fast.
◦ Thai, Spanish, and Middle Eastern are my go-to's
◦ I'd like something that's large enough to have leftovers
◦ I don't feel like eating alone, but I don't want to interact with anyone
◦ What do I feel like having?
▪ Kabab
Then came a term nobody asked for an explanation of before they started using it. Agentic AI. The word "agentic" showed up the way most technical terms do. Someone said it in a meeting, it spread to LinkedIn, and suddenly everyone was nodding along. But tracing it back to what it actually means leads you somewhere familiar — bot networks and workflow automation. IT teams have been building triggered, sequential, automated processes for years. Task fires, condition is met, next task runs. That wasn't new. What changed was what was now sitting inside that workflow. Not a rule. Not a trigger condition. An AI that could read the situation, decide the next step, execute it, evaluate the result, and keep moving without being told what to do at every turn. You stopped being the operator. You became the supervisor. What does Agentic AI do? • I'm looking for something to eat and I've already had Thai takeaway this week. ◦ My navigation app opens with directions to the Kabab restaurant, and my order is placed while I'm en route.
Then came a term nobody asked for an explanation of before they started using it. Agentic AI. The word "agentic" showed up the way most technical terms do. Someone said it in a meeting, it spread to LinkedIn, and suddenly everyone was nodding along. But tracing it back to what it actually means leads you somewhere familiar — bot networks and workflow automation. IT teams have been building triggered, sequential, automated processes for years. Task fires, condition is met, next task runs. That wasn't new. What changed was what was now sitting inside that workflow. Not a rule. Not a trigger condition. An AI that could read the situation, decide the next step, execute it, evaluate the result, and keep moving without being told what to do at every turn. You stopped being the operator. You became the supervisor. What does Agentic AI do? • I'm looking for something to eat and I've already had Thai takeaway this week. ◦ My navigation app opens with directions to the Kabab restaurant, and my order is placed while I'm en route.
This is how in just four years the world went from skeptics who were still deciding if it was worth their time to Generative AI conversations that are indistinguishable from a person.
The popular belief was that we were decades away from AI being more than clunky machine learning behind a search bar. Those outside of the AI field were blown away with how much Natural Language Processing (NLP) changed the landscape. What this did opened the floodgates for anyone with patience or curiosity to start tinkering.
The skills, knowledge, and decades of experience turned into a conversation with an AI model to create something for you that requires skills you've never had. Recently an Anthropic Hack-a-Thon was won by a lawyer that beat out all the "real developers" proving that the most valuable skill is not what you know, but how you communicate. In one of the most ironic, twist of the AI revolution "The human element" is quite possibly the most valuable skill needed for AI.
As the days blur together the retrospection on years gone by will show what we are unable to forecast. What is clear is that unlike the light-bulb making lanterns obsolete, and cars doing what horses once did, the email replacing envelopes, and mobile phones replacing landlines, AI will not make humans obsolete. There may become a co-dependence, or maybe we will have Iron Man's Jarvis or Master Chief's Cortana. No matter what happens, today tells the story that human judgment is not a process that can be automated.