The Proxy Goal Ladder
Every metric you chase is a proxy — climb the ladder to the mission before you optimise it.
- Difficulty
- Moderate
- Time to result
- ~months to results
- Steps
- 4
- Confidence
- 90%
OpenAI runs no OKRs. Instead, Kilpatrick describes prioritisation as passing each candidate initiative through a mission filter, then a reliability tenet, then standard H1/Q1 planning. The distinctive move is recognising that stated goals like revenue are intermediate layers, not ends: revenue is a proxy for compute, compute is a proxy for better models, better models are the mission. Optimising a metric without knowing what it is a proxy for is how companies chase shiny rewards like engagement.
Origin
Logan Kilpatrick on Lenny's Podcast (Feb 2024), describing how planning and prioritisation actually work inside OpenAI — including his own admission that OpenAI is not an OKR company.
Core principles
- 01First filter: does this actually help us get to the mission?
- 02Every headline metric is an abstraction towards something else — name the something else.
- 03Revenue is not the goal; revenue buys compute, compute buys better models, better models are the goal.
- 04Reliability precedes new capability: nobody cares about something great they can't use.
- 05Goals must be re-derived as the ground shifts, not defended because they were set.
How to run it
- 1
Run the mission filter first
For every candidate initiative, ask whether it actually advances the mission. Kilpatrick calls this always the first step of deciding any of these problems.
Pro tip Sometimes the shiny reward and the mission genuinely align — the filter isn't a rejection tool, it's a forcing function to check.
Watch out The trap is the potential shiny reward right in front of you — e.g. optimising user engagement — that feels obviously good but was never tested against the mission.
- 2
Climb the proxy ladder on every metric you set
Take the metric and ask 'this is a proxy for what?' repeatedly until you hit the mission. If you cannot complete the chain, the metric is unowned and will be misoptimised.
Pro tip OpenAI's chain: revenue → compute → GPUs → better models → AGI. Write yours down and share it, because heard in a vacuum a metric like revenue gets misread as the terminal goal.
- 3
Apply the reliability tenet before new surface area
Before shipping new endpoints, modalities, or abstractions, ask whether existing customers are getting a robust and reliable experience. When the answer is no, priority snaps back to reliability.
Pro tip Kilpatrick admits OpenAI has fallen short here and had to pull focus back from exciting new work to reliability.
Watch out Everyone wants everything from you. Without a reliability tenet, inbound demand sets your roadmap.
- 4
Plan on a normal cadence, but hold the plan loosely
Keep the ordinary machinery — H1 goals, Q1 goals, a sprint against them. The interesting work is the mechanism for updating your understanding of the world as the ground changes underneath you.
Pro tip Build an explicit re-derivation ritual rather than treating a mid-cycle change of plan as a failure.
In the wild
Kilpatrick notes that people hear OpenAI has revenue goals and conclude 'OpenAI just wants to make money'. Internally, revenue is understood as the mechanism to acquire compute and GPUs, which trains better models, which serves the mission. The metric is real but it is an intermediate layer.
→ Teams optimise revenue without mistaking it for the terminal objective — and can say no to revenue that doesn't buy mission progress.
With enormous inbound demand for new APIs, modalities, and abstractions, OpenAI has repeatedly had to pull attention away from exciting new work because the existing API wasn't robust enough for customers.
→ New capability is gated behind the reliability tenet, on the logic that nobody cares if you have something great if they can't use it reliably.
Common mistakes
Chasing the shiny proximate reward
Kilpatrick's named example is optimising user engagement — a metric that always looks like progress but may have no path back to the mission.
Treating a proxy as the terminal goal
When revenue (or any proxy) is stated without its ladder, both the team and the outside world misread the company's actual objective and start optimising the wrong thing.
Letting inbound demand be the roadmap
Kilpatrick describes this as one of the more challenging pieces at OpenAI: everyone wants everything from us. Without explicit tenets, the loudest customer wins.
Is it for you?
Best for
Product leaders and founders at fast-moving companies who need a prioritisation logic that survives a shifting market and heavy inbound demand
Not ideal for
Organisations with a genuinely stable environment and a well-functioning OKR system that already ladders cleanly to strategy
From the transcript
“like one going back to the mission like is this actually like going to help us get to AGI”
“it's like revenue is not actually the goal revenue is a proxy for getting more compute”
“because at the end of the day nobody cares if you have something great if they can't use it robust and reliably”
“there's this like potential shiny reward right in front of us which is like you know like optimize user engagement”
From the episode
Inside OpenAI
Logan Kilpatrick (head of developer relations)