I’ve been using large language models intensively for about two years now. Long enough to develop opinions that have changed completely at least twice.
Here’s where I’ve landed: the framing of AI-as-automation is mostly wrong, and it’s causing people to either overestimate or underestimate what these tools actually do.
The Automation Frame Is a Trap
When people evaluate AI tools, they usually ask: “Can this replace X task?” That’s the automation frame. It produces a binary answer that misses the real value.
A better question: “Does this change what I can think about in an hour?”
LLMs, at their best, function as what I’d call cognitive prosthetics — they extend the native capacity of my thinking rather than replacing it. The analogy isn’t a robot that does the work; it’s a calculator that lets a mathematician think about problems too large for unaided arithmetic.
What This Looks Like in Practice
Let me be concrete. Here’s a workflow I’ve developed that’s changed how I approach complex problems:
1. Externalizing the thought space
When I’m working through a hard problem, I now start by describing it to a model in natural language — not to get an answer, but to force myself to articulate it precisely. The act of explaining is itself clarifying. The model’s response is secondary.
2. Stress-testing positions
Once I have a position I’m prepared to commit to, I ask the model to argue against it as strongly as possible. Not “what are some counterarguments” but “steel-man the opposing view as if you believed it.” This surfaces things I’d rationalized away.
3. Domain translation
Some of the best insights come from applying frameworks from one field to problems in another. I’ll describe a problem in software architecture and ask how a military logistics planner or an evolutionary biologist might approach it. Cross-domain translation is where LLMs are genuinely surprising.
4. First-draft scaffolding
I write everything myself, but I often ask for a structural scaffold first — “what are the 6-8 conceptual components of this topic?” — and then use that as a framework I’ll inevitably deviate from. Useful for organizing, not for content.
The Limits That Matter
LLMs have specific failure modes that are important to understand, because they’re non-obvious:
Confident hallucination: The model will state false things with the same tone as true things. This is most dangerous in technical domains where the output sounds right but contains subtle errors. Always verify claims in domains where precision matters.
Recency blindness: Training cutoffs mean models have no knowledge of recent events. More subtly, they’ve overfit to the distribution of text that existed at training time — which skews toward certain perspectives and assumes certain contexts.
Sycophancy: By default, models will agree with you and validate your reasoning even when you’re wrong. You have to explicitly prompt for disagreement, and even then it’s often gentler than a smart friend would be.
The Cognitive Partnership Model
The most productive frame I’ve found: treat the model as a knowledgeable collaborator who needs careful supervision, not a tool that outputs reliable facts.
This means:
- You bring the judgment; it brings the breadth
- You verify; it explores
- You decide; it generates options
The division of labor isn’t “I think, it works.” It’s “I direct, we explore, I synthesize.”
That’s a genuinely useful cognitive prosthetic. And it’s a fundamentally different technology than the one that most hype cycles are describing.