“Foom” Is Not a Verb

Somewhere out there, a timeline thread is happening where "AI will improve itself" is treated like a natural law, like gravity, except with more spreadsheets and fewer experiments.

The reveal
The "intelligence explosion" idea goes back decades: a sufficiently capable system improves itself, that improvement yields more capability to improve itself, and the loop accelerates. People call the hypothetical takeoff "foom," because it sounds like something exploding, which is great branding for an argument that often forgets constraints exist.

The steelman (yes, but also no)
Yes: feedback loops are real. Better tools can build better tools. AI-assisted R&D could accelerate AI progress. No: "feedback loop exists" does not imply "runaway explosion." The loop has to be strong, fast, and hard to bottleneck—under real-world constraints like data, compute, energy, hardware supply chains, and organizational friction.

Prerequisites checklist (what RSI actually needs)

So what does recursive self-improvement actually need to work? First, you need end-to-end R&D autonomy—not just "writes code," but "runs the whole research cycle." The typical bottleneck here is reliability, integration, and long-horizon control. Second, you need fast iteration with tight loop times from hypothesis to training to eval to deployment, but training and inference costs plus experimentation time usually get in the way. Third, you need access to resources like compute, data, talent, capital, and hardware, which runs straight into geopolitics, export controls, and supply chains. Fourth, you need compounding gains where each loop materially improves the next loop, but diminishing returns, local optima, and eval ceilings tend to interfere. Finally, you need robust generalization where improvements transfer across domains and tasks, but overfitting to benchmarks and brittle systems are the usual suspects.

If your argument doesn't say which prerequisites are already met, it's not forecasting. It's spellcasting.

Receipts (the canonical "explosion" story and the modern constraints)

Let's trace the intellectual lineage here. I.J. Good's 1965 paper "Speculations Concerning the First Ultraintelligent Machine" gave us the original "ultraintelligent machine builds better machines → explosion" framing. Kaplan et al.'s 2020 "Scaling Laws for Neural Language Models" shows predictable capability gains with scale, but also implies the loop is bounded by compute and data allocation. Hoffmann et al.'s 2022 "Training Compute-Optimal Large Language Models" demonstrates that "just scale parameters" is often inefficient—data and compute tradeoffs matter. Narayanan & Kapoor's 2025 "AI as Normal Technology" provides a serious counterweight: transformative technologies still diffuse through institutions and constraints. The ESPAI 2023 survey of thousands of AI authors directly probes beliefs about acceleration from AI doing most R&D, and highlights uncertainty and framing effects.

Where the "foom" crowd tends to cheat
They treat the loop as frictionless. But compute doesn't appear via vibes; it appears via data centers, power, chips, and cooling. "Scaling laws" aren't prophecies; they're empirical curves with regime changes. The pun here is unavoidable: you can't just scale your way out of physics.

Disproof conditions (what would make "foom" feel less like a meme and more like a forecast)

What would actually make us update toward taking this seriously? If we observed a system autonomously executing full AI R&D cycles with short turnaround times, repeatedly, across labs, that would be the loop becoming real, not metaphorical. If we saw clear evidence of compounding improvements that reduce the marginal cost and time of the next improvement, that would be "explosion" structure, not just progress. If resource bottlenecks failed to bind—compute, data, hardware—despite large scale, the "constraints save us" story would weaken considerably.

Action line
When someone says "RSI means takeoff is inevitable," ask them for the loop diagram with timelines and bottlenecks labeled. If they can't draw it, they're not predicting the future; they're doing interpretability on their own feelings.

Anyway, back to touching grass—before the discourse recursively self-improves into another 200-comment thread.

Related reading

Why read about AGI timelines next? Because that's where "foom" goes to cosplay as certainty. Or dive into compute realities and cooling bills, since chips and watts are stubborn little alignment researchers. You might also want to check out the vibes versus evals piece, because benchmark wins are not the same as deployment reality.