Update Your Priors: The AI Capability Arbitrage
Re-test what AI can do today and demand more of it — old scar tissue is leaving alpha on the table
- Difficulty
- Easy
- Time to result
- ~ongoing to results
- Steps
- 4
- Confidence
- 90%
The hardest part of working reflexively with AI is that people's mental model of what the tools can do is stale — models that couldn't reason, generate clean images, or do data analysis a year ago now can. That scar tissue makes people set the bar too low. Deliberately discarding what you 'learned' the tool couldn't do, and demanding more of it today, is a real arbitrage.
Origin
Aparna Chennapragada; she cites Shopify's Tobi Lütke on 'reflexive AI usage' and uses a self-built Chrome extension as a forcing function.
Core principles
- 01Updating priors is genuinely hard because the models improved discontinuously
- 02Scar tissue from a bad experience months ago sets your expectations too low
- 03'The baby just grew up to be a 15-year-old in a month' — re-test capability constantly
- 04There's alpha in cutting against the grain and demanding more of today's AI
How to run it
- 1
Notice the stale prior
Recognize that your impression of a tool is from the last time you tried it — image generation full of misspellings, no real reasoning, no data analysis — and that impression is now likely wrong.
Watch out Being early is the same as being wrong — and being late by clinging to an old prior is its own cost.
- 2
Deliberately discard the scar tissue
Do the counterintuitive, against-the-grain thing: tell yourself to ignore what you learned the tool couldn't do a few months ago.
- 3
Install a forcing function
Build a cheap prompt to pause and redirect yourself — e.g. a Chrome extension that on every new tab asks 'How can you use AI to do what you're going to do right now?'
Pro tip Make the nudge trivially cheesy but ever-present; the friction it removes is the point.
- 4
Demand more of the AI and capture the alpha
Keep raising expectations of today's models on real tasks; the gap between the stale consensus and current capability is where the arbitrage lives.
In the wild
Chennapragada runs a self-built Chrome extension that surfaces 'How can you use AI to do what you're going to do right now?' on every new tab to force reflexive AI usage in work and personal life.
→ A concrete, low-cost habit that keeps re-testing where AI can now substitute for manual work.
Common mistakes
Freezing your prior after one bad try
Concluding a tool 'can't do X' from a months-old experience and never re-testing as the models leap forward.
Under-demanding of the model
Setting expectations low so you never discover the newly-unlocked capability that competitors are already exploiting.
Is it for you?
Best for
Operators, product builders, and knowledge workers trying to build a reflexive AI-usage habit
Not ideal for
Situations demanding verified reliability where uncritically trusting a model's newest claims would be risky
From the transcript
“the updating of the priors is really hard”
“the baby just grew up to be a 15year-old in a month”
“if you can kind of cut against the grain and say, "No, I won't have that scar tissue”
“There's a lot of alpha in in doing that”
“whenever I uh open a new tab it just says how can you use AI to do what you're going to do right now”
From the episode
Microsoft CPO: If you aren’t prototyping with AI, you’re doing it wrong
Aparna Chennapragada