AI models change every month. This doesn't.

So what doesn’t change?

Every AI tool you’ll use as a PM is powered by a large language model (LLM). Claude Cowork, ChatGPT Codex, Lovable, etc. are all powered by LLMs and their core capabilities and limitations are the same.

Which means the highest-leverage thing you can understand isn't any individual tool. It’s understanding the LLM itself.

Here's the map of the core mechanics of LLMs you need to understand:

These items haven't changed since ChatGPT launched in 2022. Every new model release, every new tool, every technique sits on top of these same mechanics… even the buzzwords filling your feed right now like RAG, MCP, are just a layer on top of LLM’s core mechanics.

Ultimately what you should actually be trying to master is developing AI literacy. AI literacy means knowing what's available and possible with LLM-based AI tools, what their limitations are, and ways to work within (or around) those limitations. And developing AI literacy starts with really understanding LLMs.

Here’s an example:

Imagine leadership says "We need to synthesize thousands of pieces of customer feedback a month. What are possible AI solutions?”.

As a technically & AI literate PM, you’d be able to think through the problem:

1. What the available solutions are

With your knowledge of what’s available for building in AI, you might choose between:

• Setting up an agentic workflow using n8n or Zapier for tagging feedback and a weekly agent for summarizing

• Skip the pipeline entirely to export the feedback manually and work it into long-context chat LLMs (claude, NotebookLM) for clustering.

2. What the limits are

With an understanding of LLMs and their limitations, you know the agent will sometimes tag the same feedback differently on different runs, which means your week-over-week theme rends drift with the model’s variance, not your customer’s moods. LLMs are also prediction machines, so if leadership asks “what percent mention pricing?”, an LLM will give you a confident number that it generated, not actually computed.

Long-context LLM chats also have it’s own catch: each feedback that fits the window degrades in a long session. This has to do with how context windows work in LLMs. The model doesn't remember your feedback the way you'd remember a correction. Since LLMs do not have memory, it re-reads the entire conversation every single time you enter a prompt and weighs each part against everything else. The more there is to weigh, the less any one instruction counts.

3. What to do about them

You can design around LLM’s limitations by limiting how much it predicts. This means giving the agent a fixed tag list instead of letting it invent themes so you get consistency. Require verbatim quotes with source IDs for every theme (extraction instead of generation). And LLMs are notoriously bad at math so you design traditional software to do counting. The LLM tags → the spreadsheet computes.

Which option is right? That depends on your team, your risk tolerance, and the volume of feedback.

But nothing in that evaluation was a tooling question per se. The tools in both options will be different a year from now but the reasoning won't. That's AI literacy, and understanding LLMs is half of AI literacy. The other half is understanding the software systems LLMs live inside because every AI feature your team will ever ship is still built on APIs, data, and infrastructure. An agent is an LLM plus API calls plus data pipelines. A RAG system is an LLM plus a database plus retrieval logic.

AI is still software. Which means the technical foundations you build as a PM don't just survive the AI era but they're what make you fluent in it!