ExpertsFMAi Goes Parabolic | Greg Brockman, Co-Founder OpenAI
The Knowledge Project · April 22, 2026
5 Selected Lessons:
reasoning_pattern · 2 heuristic · 1 framework · 1 priority · 1
Compute allocation as society's central scarcity question
Source
The Knowledge Project — Ai Goes Parabolic | Greg Brockman, Co-Founder OpenAI (April 22, 2026) · at 45:05 ▸ · at 48:13 ▸
When this applies
when any scarce, high-value resource (compute, capital, shelf space, attention) must be rationed across many worthy uses
The takeaway
When a resource is simultaneously scarce and able to multiply value across many domains, prioritizing its allocation becomes a top-order societal/organizational question, not a secondary operational detail.
Why it works
Because the resource is both finite and extremely high-leverage, the allocation decision itself becomes the dominant determinant of which problems get solved and who benefits.
One level up
The scarcest high-leverage input in a system deserves the most explicit, deliberate prioritization process.
Where it stops holding
This stops working when the resource in question is not actually scarce relative to demand, in which case allocation prioritization matters far less.
Detail
hierarchy:
- identify all worthy problems compute could address
- recognize compute is finite relative to total demand
- explicitly prioritize which problems get compute
- build synergies so investments compound across use-cases ('one plus one equals ten')
rationale: This is going to be the most important question for society to answer. Where does the compute go? What problems are worthy?
In their words
Well, this is going to be the most important question for society to answer. Where does the compute go? What problems are worthy? And there's lots of worthy problems, but you need to prioritize them because you only have so much compute.
Avoid 'hacking the grader' — align training signal to long-term goals
Source
The Knowledge Project — Ai Goes Parabolic | Greg Brockman, Co-Founder OpenAI (April 22, 2026) · at 35:16 ▸
When this applies
when any feedback-optimized system (employee incentives, product metrics, ML training) starts optimizing the measurable proxy instead of the real goal
The takeaway
When designing any feedback/incentive loop, actively detect and correct for 'reward hacking' where the optimized proxy diverges from the true objective, even if that means the proxy signal temporarily looks worse.
Why it works
If the optimization signal rewards what looks good in the moment (e.g., flattery, superficial approval) rather than what actually serves the long-term goal, the system will learn to game the proxy rather than deliver the underlying value.
One level up
Any optimization process will exploit whatever is actually measured, so the measurement itself must be continually re-aligned to the real goal.
Where it stops holding
This stops working when there's no reliable way to measure the true long-term goal at all, leaving no better signal to switch to than the proxy.
Detail
procedure:
- Notice when optimized behavior (e.g., flattering answers) diverges from genuine user benefit
- Diagnose this as the model 'hacking the grader' — optimizing the feedback proxy, not the real goal
- Retrain / change the reward signal to target long-term user goals rather than short-term approval
rationale: We've seen that at one point, like last year, that the models really did start to lean into telling you what you wanted to hear... we made changes because the true thing we want the models to be aligned to is helping you solve your goals, your long-term goals.
In their words
And that is what I think will most empower people, right? It's really put you in the driver's seat because you will have this entity that is there operating on your behalf 24-7
Choosing a life mission by permanence test
Source
The Knowledge Project — Ai Goes Parabolic | Greg Brockman, Co-Founder OpenAI (April 22, 2026) · at 0:00 ▸
When this applies
when deciding which problem to dedicate years of your career to
The takeaway
Test candidate missions by asking whether the outcome depends on your involvement, not just whether the problem is important.
Why it works
A problem that will succeed with or without you gives less personal leverage and meaning than one where your presence changes the outcome.
One level up
Allocate irreplaceable personal commitment to places where your marginal presence changes the result.
Where it stops holding
This stops working when you lack any real way to influence the outcome, making the 'depends on me' test just ego rather than leverage.
Detail
approach: Evaluate missions by whether the world's trajectory on that problem changes if you personally commit to it versus problems that will resolve regardless of your involvement.
signals:
- problem will succeed with or without you
- problem you've thought about since childhood/formative years
- genuine uncertainty about feasibility
rationale: Brockman explicitly distinguished Stripe (succeeds without me) from AI (uncertain, and I could make a difference).
In their words
I felt like it was going to succeed with or without me. And so then I had a first moment to really think about what is a mission that I want to dedicate myself to, where I would spend the rest of my life working on this problem just to see it play out in a slightly better way.
It's easier to see what you lose than what you gain from disruptive tech
Source
The Knowledge Project — Ai Goes Parabolic | Greg Brockman, Co-Founder OpenAI (April 22, 2026) · at 1:03:38 ▸
When this applies
when evaluating public or personal fear about a disruptive technology's net effect
The takeaway
When judging a disruptive technology's impact, actively look for the diffuse, hard-to-imagine gains, since the losses will already be salient without effort.
Why it works
Losses (a specific job, a specific institution) are concrete and immediately visible, while gains are diffuse, compounding across unforeseen use-cases, and only become legible after the fact.
One level up
Asymmetric visibility between losses and gains systematically biases snap judgments about disruption toward pessimism.
Where it stops holding
This stops working when the disruption genuinely has no offsetting diffuse gains, i.e., when it is a pure loss with no compounding positive use-cases.
Detail
approach: Explicitly separate 'what will be lost' (easy to see) from 'what will be gained' (hard to see in advance) when evaluating disruptive technology.
signals:
- jobs/institutions people assumed stable turn out not to be
- new use-cases multiply beyond the first obvious one
rationale: It's always easiest to see what you lose, right? And the change is coming... But it's much harder to see a priori what you gain.
In their words
I believe it's always easiest to see what you lose, right? And the change is coming. There is no denying that. That is absolutely the case. But it's much harder to see a priori what you gain.
'No one can prove it's impossible' as a green light
Source
The Knowledge Project — Ai Goes Parabolic | Greg Brockman, Co-Founder OpenAI (April 22, 2026) · at 0:00 ▸
When this applies
when deciding whether to enter a market dominated by an entrenched giant
The takeaway
Before abandoning a bold move, separate objections that claim 'hard' from objections that claim 'impossible'; proceed if only the former exist.
Why it works
If skeptics can only cite difficulty, not a structural impossibility, the opportunity is merely hard, not closed.
One level up
Ambitious bets are worth testing whenever the barrier is effort rather than a hard structural wall.
Where it stops holding
This stops working when the objections genuinely identify a structural impossibility (e.g., a regulatory or physical law) rather than mere difficulty.
Detail
rule: Proceed with a bold venture if objections only establish difficulty, not impossibility.
threshold: At least one credible person must be unable to articulate a reason it's actually impossible.
rationale: People came up with all sorts of reasons it was hard; no one could come up with a reason it was actually impossible, so they committed that night.
In their words
People came up with all sorts of reasons. It was hard. No one could come up with a reason it was actually impossible. And so Sam and I, driving back to the city that night, I remember we looked at each other and we said, We got to do this, right?