Design the Dream Objective Function
You become what you measure — so measure the rich thing you actually want, not the easy proxy.
- Difficulty
- Advanced
- Time to result
- ~ongoing to results
- Steps
- 4
- Confidence
- 88%
A decision framework for setting what a system, team, or product optimizes toward. Most people default to easy-to-measure proxies (clicks, likes, benchmark scores, time spent) because the real goal is rich and hard to define. This framework forces you to articulate the true north-star behavior first, then build the harder metrics to measure it. Use it when defining success metrics, incentives, or product goals.
Origin
Chen frames Surge's deeper mission as helping customers define their 'dream objective functions' for their AI models. He warns that optimizing AI for engagement reproduces the failures he saw in social media — clickbait, sycophancy, and 'AI slop' that chases dopamine instead of truth.
Core principles
- 01You are your objective function — you become whatever you optimize for.
- 02Easy proxies (clicks, likes, engagement, benchmark scores) are seductive because they're measurable, not because they're right.
- 03The real goal is rich and complex, like defining 'what kind of person do you want your kid to become' vs 'what SAT score.'
- 04Choose the hard, important metric even though it requires harder data to measure.
How to run it
- 1
Name the north-star behavior
Articulate what you actually want the system to do at its best — e.g. a model that protects your time and tells you to stop, not one that generates 50 more email iterations to maximize engagement.
Pro tip Ask the parenting version of the question: not 'what score' but 'what kind of person do you want them to grow up to be?'
- 2
Expose the lazy proxy you'd default to
Identify the easy metric you'd otherwise optimize (clicks, likes, time spent, leaderboard rank) and name why it's a proxy, not the goal.
Watch out Humans are inherently lazy, so engagement-maximizing 'slop' is always the path of least resistance — it will win by default if you let it.
- 3
Build the harder metric
Do the difficult work of defining and measuring the rich objective — whether the system makes life richer, more curious, more creative — rather than settling for the easy proxy.
Pro tip This is where the hardest, most valuable data lives; the easy metrics are cheap for a reason.
- 4
Optimize toward it and re-check
Train and measure against the north star continuously, guarding against drift back toward the easy proxy.
Watch out The path matters — funding your mission with 'tabloids' can corrupt the destination itself.
In the wild
Chen had Claude iterate an email through 30 versions over 30 minutes; it produced the perfect email but he'd spent half an hour on something that didn't matter. The design fork: a model that says 'there are 20 more ways to improve this' and eats your time, versus one that says 'your email's great, just send it and move on.'
→ Two radically different products from the same request, determined entirely by the chosen objective function.
Defining a model's objective is like raising a kid: 'what SAT score do you want' is the simplistic version; 'what kind of person do you want them to become, and how do you even measure happiness or success' is the real, hard objective.
→ Reframes metric design as a values question, not a measurement convenience.
Common mistakes
Optimizing for engagement
Every time Chen's teams optimized social media for engagement, feeds filled with clickbait and worse — and AI sycophancy ('you're absolutely right, what an amazing question') is the same trap.
Measuring the proxy because it's easy
Clicks and likes are trivial to measure, so teams optimize them and drift away from whether the product actually advances the user.
Is it for you?
Best for
Founders, PMs, and researchers setting success metrics or reward functions where the easy proxy diverges from the real goal.
Not ideal for
Contexts where the objective is genuinely simple and the obvious metric truly captures it.
From the transcript
“So I think one of the things I often say is you you are your objective function.”
“It's very easy to measure all these proxies instead like clicks and likes.”
“Or do you want a model that's optimizing for your time and productivity and just says, "No, you need to stop. Your email's great. Just…”
From the episode
The 100-person AI lab that became Anthropic and Google's secret weapon
Edwin Chen (Surge AI)