When the memory is the problem
Maybe AI knows too much to be helpful
The same capability that makes AI models useful is also what keeps them from doing certain jobs well.
When you load a model with everything you know about your business – every assumption, every framework, every belief you’ve built up over years of building – it becomes just as close to the problem as you are. Maybe closer.
It can surface all of that context instantly without the friction of being human and distracted by all the things. So if we look to Claude, for example, for objective or outside perspective, it struggles. We already put insider knowledge in there.
AI loses objectivity when you give it the same memories you have. And really, by dumping so much context into these models or harnesses, AI has more context than we can probably carry in one moment.
Business strategy and humans
I tried to work around this memory problem. I stripped back the context and curated what I provided in any given chat with Claude. I tried to strike the balance between enough context and too much context.
But without enough context, the answers are meaningless. And whatever context I gave it skewed the results somewhere. If there’s a clean Goldilocks just-enough-context option right down the middle, I have yet to find it.
Confrontation needed
What I’m looking for when doing strategic work, deep work is confrontation, not validation, and we all know how validating these models can be.
Ask ChatGPT and it’ll tell you I’m a freakin’ genius. Good thing I know better.
Here’s the exercise that I’ve been running:
We do sales calls, customer calls, customer research calls, and we’ve brought in a consultant to do some of that work for us.
To see what kind of insights and results we can get using AI, we put transcripts through multiple models. The analysis came back confirmational. Things I already believed or knew were neatly reorganized and reflected back at me. Turns out I’m still wicked smart and I have very good ideas.
Sigh.
What I couldn’t get was the thing a good outside perspective is supposed to give us: a willingness to push back on the things we hold sacred or we’re too close to see.
Giving any model all of the context that you can gets it as close or closer to the problem than you are. It makes it impossible for any of the models so far to have zero attachment to the ideas that came before.
Leading the witness
LLMs are guided by the questions we ask. Ask leading questions, get led answers. To get a model to think in wavy, crooked lines instead of straight ones, we have to point the way. Which means you already have to know which direction to go.
Which of course is counter to this entire point. If I knew all the channels and avenues that we needed to explore and consider, I wouldn’t be asking anyone or anything.
So we hired a consultant and things changed.
In which it gets awkward
The consultant ran interviews with people we probably wouldn’t have spoken to ourselves. She brought experience, no emotional stake in the outcome, and no memory of the framings we’d spent months building. She took the same volume of content we’d hand to an LLM and went down different pathways. She pushed back.
She didn’t tell us our baby is cute.
That honesty is honestly why we engaged her. I think something is valuable if you get nervous or slightly anxious before it happens. And calls with our consultant had me feeling a bit uneasy. I knew some hard choices were going to be presented to us, and while part of me wanted to go back to ChatGPT and find out again how smart I am and how clever I am, we all know that’s not a path to success.
What AI gets so right
AI excels at throughput. Damn it’s fast.
Going back through a full catalog of sales and research calls – further back or more thoroughly than we’d normally go – and finding through-lines across all of it. The speed makes something practical that would otherwise not happen if we were to do it by hand.
To compare what I was hearing with what Claude picked up, I took careful notes on several calls and the models came back with what I wrote down and then some.
Running this kind of analysis at speed with a pretty high degree of accuracy is an incredible superpower we all have now. But it’s also helped in a few other ways.
Competitive monitoring is useful. The market is moving so fast that it’s hard to keep up with who’s doing what and when and for whom. I stopped doing this as recurring tasks and I’ll run competitive monitoring ad hoc now. I find this to be incredibly helpful to keep track of who’s in market and what they’re doing.
These are the right jobs. Triage, volume, pattern recognition across data.
Strategy and testing the assumptions that actually drive direction requires objectivity. Objectivity with LLMs is a bit of a myth right now.
For day to day work, work that requires the context, LLMs have been a big productivity boost for me personally and for the company.
For strategic decision making, even strategic brainstorming, they fall flat. Maybe the context I’m giving it is just so full of my own beliefs and opinions that it can’t strip that out, but I don’t know how to get rid of that.
Objectively speaking
So my rule now is to look for humans to speak with when I need objective pushback. LLMs are great when they have the context, but that same context becomes a liability when you want to think different and be challenged.
This applies to big decisions and little decisions.
When objectivity is mission critical, you go with the human. Straight up.
TTFN,
Peter