By the end of this article, you'll be able to:
Identify the major capabilities and limitations of current generative AI
Enjoying this series? Support the Newsletter - Buy Me A Coffee
So we now know what Generative AI is, how it differs from previous tech, and how we got to this point.
Up until now, weâve talked about mindset.
Weâve talked about collaboration.
Weâve talked about the skills that make you irreplaceable.
Now we need to get honest about the tool itself.
Because working well with AI is a lot like working with a new colleague.
You donât need them to be perfect. You need to know:
what theyâre great at,
where theyâre unreliable,
and when you need to step in.
Letâs start there.
What Generative AI Does Remarkably Well
Modern language models are better at writing than any Pulitzer Prize winner.
Not just spelling or grammar, but structure, tone and transformation.
They can:
Draft emails that sound like you,
Summarise long documents,
Translate between languages,
Explain complex topics across wildly different domains.
Whatâs especially striking is how flexible they are.
The same system that helps you:
brainstorm ideas for a birthday party
can also help you:
analyse business trends
understand scientific concepts
explore unfamiliar topics
All through conversation.
They donât need to be âretrainedâ for each task. They adapt based on context.
Many models can also:
remember details mentioned earlier in a conversation,
Refer back to them later,
build on what youâve already discussed
This fluid interaction feels closer to talking with a person than using software.
And increasingly, these systems can:
search the web
process files
interact with other tools
This expands what they can help with even further.
All of this makes AI incredibly useful.
But usefulness without understanding is where people get burned big time!
The Limits You Need to Respect
AI systems today have very real constraints.
Ignoring them doesnât make them go away; it shifts the risk onto you.
And risky investments are usually bad ones. (Not financial advice)
1. They are bound by their training
Every model has a knowledge cutoff.
Thatâs the dates after which it has no built-in awareness of events.
A useful metaphor:
Imagine someone who went on a long retreat with no internet access.
Theyâre intelligent. Theyâre articulate. But they donât know what happened after they left.
Unless you give the AI tools like web search, itâs operating with that same limitation.
2. They donât verify truth; they generate plausibility
AI models learn from large collections of human-created text.
That text contains:
insights
opinions
mistakes
inaccuracies
The model doesnât fact-check during training.
So sometimes it produces something that:
sounds confident
feels coherent
is completely wrong
This is what people call a hallucination.
Itâs not lying. Itâs pattern-matching too confidently.
Think of a friend who tells a story with total conviction - and gets the details wrong.
Thatâs AI on a bad day.
Which is why discernment matters.
3. They have limited working memory
AI doesnât see everything at once.
It operates within a context window (The "Chat" you open when using ChatGPT, for example)
This includes:
your prompt
its responses
anything else shared in that session
When the window fills up, older information can fall out.
This explains why:
AI sometimes forgets things
Long documents can be tricky
You need to manage context intentionally.
Itâs not careless. Itâs constrained.
4. They are non-deterministic by default
Ask the same question twice, and you might get slightly different answers.
Thatâs not a bug, itâs a feature.
AI generates text probabilistically, choosing what comes next based on likelihoods.
This variability is great for:
brainstorming
creativity
exploring options
But it requires caution when:
Accuracy matters
Consistency is critical
Decisions have real consequences
Some systems allow you to reduce randomness.
But awareness matters more than settings.
5. They still struggle with complex, multi-step reasoning
AI has improved a lot here, especially with newer âreasoningâ models.
But:
multi-step logic
detailed maths
deeply constrained problem-solving
Still requires oversight.
This is another place where you matter.
6. They canât access what they canât see
Even the smartest AI canât help if it doesnât have access to:
your internal data
proprietary tools
specific systems
Itâs like having a brilliant colleague who doesnât have the right permissions, access or autonomy.
Intelligence without access still has limits.
Why Does This Matter To You?
If you donât understand these boundaries, one of two things happens:
You either:
trust AI too much
or
dismiss it entirely
Both are mistakes.
Fluency lives in the middle.
Itâs knowing:
When AIâs strengths complement yours
When its weaknesses need intervention
When judgment must stay human
Humans bring:
critical thinking
context
ethics
responsibility
AI brings:
speed
scale
pattern recognition
tireless iteration
The most effective work happens when these strengths are combined deliberately.
What's the Point Of All This?
AI will keep changing.
Capabilities will expand. Some limits will shrink. Others will remain or shift.
Thatâs why fluency isnât about memorising what AI can do today.
Itâs about learning how to evaluate, test, and decide as things evolve.
Thatâs also why we keep coming back to the same four competencies:
Delegation. Description. Discernment. Diligence.
Theyâre how you stay grounded when the tools get powerful.
In the exercises ahead, youâll get direct experience with these strengths and limits.
Exercises
Reflection
Before moving on, take a moment to consider:
How does understanding the technical foundations of generative AI (like training data and pre-training/fine-tuning) change how you think about working with these systems?
What ethical considerations come to mind after learning about how these systems work and their current limitations?
Until next time,
I love you.