“AI is like the world’s smartest person, but with amnesia all the time”
Been using this phrase with teams in the past few weeks and have found that it resonates. It helps explain the importance of Context in AI Product & Engineering.
Imagine waking up every morning with access to all the knowledge in the world, but zero memory of yesterday’s conversations, last week’s decisions, or that crucial meeting where everything went sideways. That’s AI, every single time you hit “new chat”.
Empathy for the machine is a concept I try to practice. What would it be like to have access to all the knowledge in the world and then get a query like, “please help write email to Bob about expense issue with vendor that Jane brought up yesterday”. And that’s it. Oh and you HAVE to respond. And try really hard to help the user and not ask them a series of endless questions. The AI is like a Michelin star chef, locked away in the world’s most magnificent pantry, with only a half-crumpled note and a single fortune cookie reading “some like it cold”. Impress Gordon Ramsey, ready set go.
Even when we think we’re providing AI with lots of context, alas there’s just so much stuff that lives in our head that is hard to put down on paper. Beyond the obvious situational details like what the issue actually is, who the vendor is etc., there are all the metaphysical, socio-cultural, and frankly vibes-based details that we don’t even think about. How to explain to the model the years long political drama that underpins the Bob & Jane relationship? That Bob always CCs the entire leadership team because of that one incident in 2019? That Jane has this habit of bringing things up “casually” that are actually five-alarm fires?
The machine has no way of knowing any of this. It’s trying to navigate a complex social and political minefield with three sentences of context and all the knowledge in the world which, it turns out, isn’t nearly enough. The context we don’t share is often more important than the context we do. It’s the organizational scar tissue, the unwritten rules, the “everyone knows that…” knowledge that everyone definitely does not know.
And so we all feel it; that feeling of the machine being so close to the answer, yet frustratingly far away. The response is almost right - professional format, mentions the vendor, appropriate length - but somehow completely misses the point. Like watching someone's key ring jangle with the right key while they alternate between futilely trying to pick the lock and furiously call every locksmith within a 50 mile radius.
We want to keep trying so we keep sending message after message, each one a desperate attempt to transmit years of institutional knowledge in a paragraph. We watch in horror as the machine becomes more and more incoherent, like watching a GPS confidently announce “You have arrived!” as it’s directions slam our car into a wall. The tragedy isn’t that AI lacks intelligence, it’s that we keep expecting it to read our minds while simultaneously forgetting it can’t even read our last email. Every interaction is a first date, but we’re asking it to plan our wedding.
The world’s smartest person with amnesia would at least know what they’ve forgotten. AI doesn’t even get that mercy. Our real job, it turns out, is learning to be the memory for a mind that cannot remember it needs one