the rsi pitch keeps forgetting to attach a unit economics slide

the slide that's always missing

jack clark put 60% odds on ai being recursively self-improving by the end of 2028. that quote is doing a lot of work in policy rooms right now. it's also missing the slide that would make it an investable forecast instead of a vibe: the cost-per-iteration curve, and the capability-gain-per-dollar curve, plotted on the same axes.

because the unsexy problem with rsi isn't alignment. it's that each iteration costs millions of dollars at frontier scale, and no lab has published evidence that self-bootstrapped capability gain is compounding faster than the compute bill. the pitch keeps getting made without the unit economics attached, and washington keeps treating the pitch as a forecast.

the frame: rsi is a capex argument before it's a safety argument

call this the rsi compute economics problem. for recursive self-improvement to matter as a near-term phenomenon, three curves have to cooperate: the cost of running a post-training iteration, the marginal capability gain per iteration, and the rate at which the model's own contributions reduce the cost or improve the gain of the next iteration. if any one of those curves goes the wrong way, you don't get a takeoff. you get a very expensive plateau.

the consensus read collapses all three into 'capability go up.' the sharper read is that rsi forecasts are a capex argument dressed as a safety argument. and once you see it that way, the policy implication flips.

receipts

start with what frontier rl post-training actually costs. credible estimates for a single frontier-scale rl run sit in the millions of dollars per iteration, and that's before you count the eval harness, the human raters where they're still in the loop, and the failed runs that don't make the blog post. the cost trend is up, not down, because reasoning-style rl burns vastly more tokens per training example than supervised fine-tuning ever did. token deflation on the inference side is real. token inflation on the training side is also real, and it's larger.

now look at what's actually been published on self-bootstrapped gain. the public artifacts are constitutional ai, rlaif, self-rewarding language models, and a steady drip of papers where a model grades or generates its own training data. each shows a real lift on a specific eval. none shows a compounding curve where iteration n+1 produces a larger jump than iteration n at constant or declining cost. the closest thing to a public compounding result is the o-series reasoning scaling story from openai, and even there the compounding is on inference compute, not on a self-improvement loop that reduces the cost of the next training run.

meanwhile the policy layer is moving faster than the empirical layer. the trump administration is reportedly weighing pre-release government review of new ai models. joe lonsdale named the dynamic on cnbc, pointing at an 'fda for ai' as exactly the kind of regulatory bureaucracy china would prefer the us adopt. the labs publicly forecasting rsi by 2028 are the same labs whose policy teams are sized to clear a federal pre-release review process. nobody else's are. this is the convergence trade: rsi timeline as permitting strategy. convince washington that recursive takeoff is imminent, and pre-release review becomes the only responsible response. pre-release review is also, conveniently, the most durable moat a frontier lab has ever been handed.

the steelman

the honest counter is that rsi doesn't need a published return curve to be real. internal results at frontier labs run 12 to 18 months ahead of anything that shows up in a paper, and the labs with the strongest rsi conviction are exactly the ones with the most internal signal. it's also true that the cost-per-iteration figure is the wrong denominator if a single successful iteration produces a model that meaningfully reduces the cost of the next one. that's the whole bet. and on a long enough timeline, the bet has to resolve one way or the other regardless of whether the slide deck shows up in 2026 or 2028.

fair. but 'trust the internal signal' is the same epistemic move that produced the gpt-4o sycophancy rollback, the agent demos that didn't survive contact with production workloads, and most of the autonomous-coding revenue projections quietly walked back over the last year. internal signal at frontier labs has a real track record. it is not a track record of conservative forecasting.

resolution

the useful posture is to separate two questions that keep getting fused. is recursive self-improvement physically possible on a timeline that matters? probably yes, eventually. is it economically demonstrated today in a way that justifies pre-release review architected by the labs forecasting it? not in any artifact you can cite. when you see an rsi timeline quoted in a policy context, ask for the cost curve. when a lab proposes a review regime calibrated to that timeline, ask whose competitors clear the bar.

the rsi pitch will eventually attach a unit economics slide. when it does, the curve will tell you whether you're looking at a takeoff or a permitting strategy. until then, treat the forecast as what it is: a capex argument that figured out how to bill itself as a safety argument.