Context Over Models: The Data-Management Bet
AI products win on getting good, timely, well-structured data to the model — not on the model itself
- Difficulty
- Advanced
- Time to result
- ~months to results
- Steps
- 3
- Confidence
- 92%
Clowes' central thesis: LLMs are simultaneously very smart and 'insanely dumb' because they only know what they were trained on or what you hand them in the moment, which they immediately forget. Since information has a decay rate, the durable advantage in any AI product or workflow is not the model or even the prompt — it is the context: getting good, timely, well-structured data to the model. That is a data-management problem, and it is where 90% of the effort goes.
Origin
Clowes' 'hot take'; he links it to Mikey Krieger's observation at the Lenny/friends summit that at Anthropic the product value came from model+data research, not UX tweaking.
Core principles
- 01Models know only what they were trained on or what you bring them in that millisecond, then forget it
- 02Information has a decay rate — new data loses decision value very quickly
- 03Models are 'limitless information eaters'; more relevant context generally makes them better
- 04Models are largely replaceable; prompts help but are secondary; context is the real answer
- 05Building AI is a data-management problem: good, timely, well-structured, accessible data
How to run it
- 1
Assume the model is a commodity
Accept that the model layer is mostly replaceable and the next one will be smarter. Do not build your differentiation on the model.
Pro tip Spend your scarce inspiration on data access, not on chasing the latest model.
Watch out 'If you think all I really need to do is get a model and wire it in... you're going to be very sad for a very long time.'
- 2
Map all the context the task actually needs
For your use case, enumerate every linked piece of data required for a smart answer. Follow the jumps of logic and how data links together, not just the obvious single source.
Pro tip Clowes' HCM-bot example: an AI needs employee data plus benefits, the legal situation in the person's country, and company policies and procedures — the linked context, not one table.
Watch out A naive 'suck some data in and whack it into the LLM' pipeline leaves you perpetually wrestling with data quality and timeliness.
- 3
Engineer the data pipeline as the product
Invest in getting good, high-quality, timely, well-structured data to the model reliably. Treat this as where 90% of the calories go.
Pro tip For internal PM use, even copy-pasting all relevant context beats relying on the model's memory.
Watch out Nobody wants to hear it — everyone fixates on the cool models — but the hard work is the data plumbing.
In the wild
Clowes cites Mikey Krieger describing two PM groups at Anthropic — one on UX/product, one on model research — who found 'all of the success came from the model research work' and the data provided to the model, so they put more of the team there rather than tweaking UX and buttons.
→ Product resources shifted toward model+data work because that is where the value originated.
Common mistakes
Betting on the model or the prompt
Models are mostly replaceable and prompts are only a secondary edge; teams that differentiate on them miss that 'the real answer is the context — all the data you're going to give it.'
Building a naive data pipeline
Assuming you can 'wire in' a model with a thin pipeline leads to constant wrestling with how to get good, timely, well-structured data — the actual hard problem, underestimated up front.
Is it for you?
Best for
Founders and product leaders building AI features who must decide where to invest scarce engineering effort
Not ideal for
Pure research settings where advancing the model itself is the actual product
From the transcript
“L&M can only be as good as the the the data they are given and how recent that data is”
“information has a Decay rate”
“it's obvious that the real answer is the context like all the context you're going to give it”
“that's where 90% of the calories go”
“these models are very very very smart but they're also like insanely dumb”
From the episode
Why great AI products are all about the data
Shaun Clowes (CPO Confluent, ex-Salesforce, Atlassian)