A conversation with my friend made me realise I don’t know what LLM, Agents and Platforms (chatgpt.com, grok.com, gemini.google.com) are.
It went like this:
friend: hey I can do x with chatgpt so why would i want to use an agent?
me: oh well because with agents you can … and there is much less back and forth as it remembers we talked about
friend: huh, but can’t chatgpt memories do this? i can configure it to remember a persona and key details. also, it came up with an agent mode as well.
me: oh really? i didn’t know it can do that
(struggles to provide an answer)
so why would people use a platform over an agent? i did not have an answer (shocker)
despite using it everyday, and developing multi agent architectures. i did not have an explanation. so i took a step back and asked myself what are llms, agents and platforms?
i came up with this:
- llms are the foundation
- agents enable llms to do more with different tools and specialised context
- a platform allows people to experience either the llms or agents
here’s a visual:
all three components are interconnected, and the llm is the foundation!
in other words:
- a platform is the restaurant
- an agent is the chef with their kitchen equipments
- a llm is a chef
the chef
- they know a lot of recipes, but if you ask them to recreate a dish, they might struggle to remember the exact process of it. even if they did, they might add their own style to it
- the llm is trained with huge amounts of information. every time a user ask a question, it will generate an answer based on what it knows. but it can hallucinate as well
the kitchen equipment
- without it, the chef can’t make a dish. there’s no place for them to store, prepare, cook and assemble a dish.
- the agent expands the capability of an llm, it can provide it specific context, provide custom tools, access to a curated set of data outside of its original training data.
the restaurant
- this allows customers to enjoy the dishes a chef made. they come in, take a seat, order a food and consume it.
- the platform provides an interface for people to interact with the agent or llm,
with this in mind, i get why a lot of things hinge on the llm performance (duh).
a great chef with subpar tools can still cook a great dish, though it might be less efficient.
but the reverse is not true, even if you give a junior chef the best tools, they can’t create the same dish because they don’t know how to, they’re not trained for it.
i hope this helped a stranger in the abyss.