before you can work with AI fluently, you need to understand what you’re actually working with.
not at a PhD level. but enough to know why it behaves the way it does, where it’s reliable, and where it isn’t. that understanding changes how you work with it.
what generative AI actually is#
most AI you’ve encountered before generative AI was analytical — it classified, predicted, detected. generative AI does something different. it creates. text, images, code, audio — new content that didn’t exist before.
large language models like Claude are the most visible example. trained on vast amounts of text, they learn to predict and generate human language at a level that feels remarkably natural.
three things made this possible:
algorithmic breakthrough — the transformer architecture in 2017. enabled models to process entire sequences of text in parallel while understanding relationships between words across long passages. everything changed after this.
data — the internet. books. code. decades of human-generated text at massive scale. the raw material that made training possible.
compute — GPUs and TPUs powerful enough to train models with billions of parameters. scale unlocked capabilities nobody explicitly programmed.
how these models learn#
two stages:
pre-training — the model analyzes billions of examples of text, building a statistical map of language. it learns patterns, relationships, facts, reasoning styles — all by learning to predict what comes next.
fine-tuning — after pre-training, the model is further trained to follow instructions, be helpful, and avoid harmful outputs. this is where raw capability gets shaped into something useful and safe.
what they can and can’t do#
honest and balanced — because this matters for how you use them.
capabilities:
- versatile across language tasks — writing, summarising, translating, explaining
- maintains conversational context across long exchanges
- connects to external tools and data sources
- adapts to new tasks from examples without retraining
limitations:
- knowledge cutoff — no awareness of events after training date
- hallucinations — confidently states incorrect information. sounds plausible. isn’t always true.
- context window limits — maximum information it can hold at once. long conversations can cause earlier context to drop
- non-determinism — same prompt can produce different answers each time
- reasoning challenges — improving rapidly but still struggles with complex multi-step logic
- no access to private data — unless explicitly connected
for Security analysts — hallucinations and knowledge cutoffs are the most dangerous limitations. a confident-sounding wrong answer about a threat actor or CVE is worse than no answer.
the AI fluency framework — the 4Ds#
knowing what AI is gets you halfway. knowing how to work with it gets you the rest of the way.
the 4D framework is a set of four core competencies for working with AI effectively, efficiently, ethically, and safely.
Delegation — deciding what work to do yourself, what to collaborate on with AI, and what to let AI handle independently. requires understanding both your goals and AI’s capabilities.
Description — communicating clearly with AI. defining what you want, how you want it approached, and how you want AI to behave during the collaboration.
Discernment — evaluating what comes back. assessing the quality, accuracy, and appropriateness of AI outputs with a critical eye.
Diligence — taking responsibility for AI-assisted work. being transparent about AI’s role and accountable for the final output.
each post in this series goes deep on one of these. but they work together — fluency comes from applying all four, not just one.
three ways to work with AI#
the 4Ds apply across three different modes of AI interaction:
Automation — AI executes specific tasks based on your instructions. you define what needs doing, AI does it.
Augmentation — you and AI collaborate as thinking partners. iterative, back and forth, both contributing to the outcome.
Agency — you configure AI to work independently on your behalf. you shape its knowledge and behavior rather than specifying exact actions.
most Security analysts operate in automation mode — one-off prompts, specific outputs. fluency opens up augmentation and eventually agency. that’s where the real leverage is.
next up — delegation. the competency that starts before you ever open a prompt window.
which mode do you currently use most — automation, augmentation, or agency?
took ai help to clean up typos. my brain works faster than my fingers. xd
next up: AI Series #20 — “not everything should go to AI — the art of delegation” AI Fluency Cheat Sheet back to series index
