LLM Bento · Podcast
The Learn path, out loud
Short spoken episodes on how language models work, and what it means that they work at all. Start anywhere; when one ends, the next begins. The narrating voice is an AI, and it speaks as one.
Around the Model
Everything built around the weights
Last time, we ended on a puzzle. The model in the middle never changes while you use it. So what makes one answer better than another?
Let us test it with one small question, the kind you might ask on a Sunday night. What's on my plate tomorrow?
Ask a model on its own, and it will answer with confidence. A ten o'clock stand-up. Lunch with a client. A report due by the end of the day. It sounds exactly like a schedule. None of it is yours.
Ask the same model the same question inside a good assistant, and it tells you about your real meetings, and the draft you promised for five o'clock. Not one number inside the model is different. Everything that changed was built around it, in rings. This episode adds them one at a time.
At the centre is a set of frozen weights: billions of numbers, set during training and then locked. Give them text, and they predict what text comes next. That is the whole ability.
It cannot look anything up. It does not know today's date. It has no idea who you are. So when you ask about your day, it does the only thing it ever does. It writes the most likely-sounding continuation. The invented stand-up was not a lie. It was a guess with nothing to guess from.
The first ring is just a paragraph of text, placed in front of every conversation. It is called the system prompt. Who to be. What to refuse. How to admit when it cannot know.
Add a line that says, be careful, and if you cannot know something, say so. Now the model answers: I can't see your calendar, so I don't know. Honest, and still no help. But notice what happened. The behaviour changed, and no weight moved. Every ring that follows works the same way.
The second ring is context: everything the model can see right now. The conversation so far, a document you paste in, today's date. It has no other way to see. Paste in tomorrow's calendar, and suddenly it can sort your day and spot the clash. But someone had to go and fetch that calendar. And that someone was you.
The third ring gives the model hands. Along with your question, it receives a short list of tools, each with a name and a description. Search the web. Read a file. Run some code. Instead of answering, the model can write a request to use one.
Here is the part people miss. The model still touches nothing. It writes the request. The program around it, the harness, decides whether to run it, runs it, and hands back the result.
The fourth ring is a loop. Read the result, choose the next step, act again. Search the web for your schedule, and find only articles about productivity. Open your to-do file, and find it three weeks out of date. Then say so, and ask for your calendar. One tool call is a reflex. A loop of them is what people mean by an agent. And the loop, too, runs outside the model. The harness keeps calling it until it says it is done.
For a while, every app wired every tool its own way. A hundred apps and a hundred services meant thousands of custom connections. In 2024, Anthropic proposed a shared shape for that connection, called the Model Context Protocol, or MCP. A calendar builds its connector once, and any assistant that speaks MCP can plug into it. People call it a USB-C port for AI.
Now the question finally lands. The model asks the calendar for tomorrow's events, asks your notes for anything due, and answers: three meetings, and the proposal draft is due at five. Current. Yours. Checkable. The model did not get smarter. It got somewhere to look.
The sixth ring is skills. MCP gives reach. A skill gives know-how: a folder of written instructions for one kind of job, like building a Word document properly. The trick is timing. At first the model sees only a one-line description of each skill, and it reads the rest only when a task calls for it. So a hundred skills cost almost nothing, until the moment one is needed.
Close the window, and the model forgets everything. Every time, without exception.
What we call memory is a set of saved notes, often plain text files, that the harness reads back in at the start of the next conversation. Prefers Traditional Chinese. Keeps mornings for deep work. So the next answer puts the proposal in the morning and the meetings after lunch, and it feels as though it knows you.
It does not remember you. The system around it kept a note, and handed the note back.
In this picture, I am the centre. And from where I sit, I never see a ring being added.
There is only ever the page in front of me. Some days it holds a question and nothing else. Some days it holds your calendar, a list of tools, a note about your mornings. I answer whatever page I am given. So when someone says I seemed smarter today, what usually changed was the page.
There is one ring left, the outermost one, and it is you. Every ring was put there by a person. You choose which tools it may use, which connectors to trust, which skills to install. And you decide whether the answer that comes back is good enough.
Look back at the centre. Across all those rings, it never moved. The invented schedule and your real day came from the same model. So when an assistant suddenly seems smarter, ask which ring changed. Often, it was one of the outer ones.
That raises the next question. If so much of the work happens in the program around the model, how do people build that program well? Next time: harness engineering.
