State of RSI
The intelligence that builds intelligence.
The ambition is bigger than a better assistant.
It is a better engine of discovery.
The most consequential thing a powerful AI might help invent is a more powerful AI. If that successor becomes better at research, the gain can carry forward: better intelligence, better experiments, then better intelligence again.
That feedback is recursive self-improvement. It could change how quickly we build advanced AI—and, through AI, how quickly we make progress in science, medicine, and technology.
How much of the loop is real today? OpenAI’s September 2026 account gives us a view inside a frontier laboratory. To understand what it shows, we need to follow the work from the first idea to the next generation.
Intelligence that improves the invention process
The prize is a greater capacity to discover.
Imagine intelligence that helps invent medicines, new materials, and better energy systems. Now imagine applying that same capacity to the science of intelligence. The result could improve the tool we use to make every later discovery. This is why AI that can do AI research occupies such an important place in the ambition for advanced AI.
OpenAI · Planning for AGI and beyond ↗Bring the researcher into the machine.
Today, people assemble ideas, software, data, and compute into better models. An AI researcher would take on larger parts of that process: propose experiments, implement them, inspect the results, and develop the next hypothesis. Automating this work could expand research capacity beyond the supply of human specialists.
Let the improvement carry forward.
Recursive self-improvement means an improvement increases a system’s ability to produce further improvements. The new system might have better model weights, better tools, a better training method, or some combination. What makes the process recursive is the return path: the result becomes a more capable participant in the next round.
This is a route to a much larger ambition.
A sufficiently strong loop could help take AI from broadly useful intelligence toward capabilities beyond human expertise—often called superintelligence. RSI names a possible mechanism for getting there. Beneficial intelligence, scientific progress, and human flourishing are the ends it could serve. A feedback loop is powerful because its effects can reach far beyond the next model release.
OpenAI · Planning for AGI and beyond ↗Figure 1. Illustrative generations of an AI research system. Connections build with scroll. The expanding scientific applications are possibilities, not measured outcomes or a timeline.
Research detail · the ambition → the mechanism
Use the whole research system as the unit: trained model parameters, agent software, data pipelines, training recipes, and infrastructure. A change can improve one component without improving general intelligence. A recursive effect requires that the resulting system becomes better at producing subsequent improvements. Ordinary inference, repeated prompting, and additional training do not by themselves establish that effect.
What would an AI researcher actually do?
Choose a question and design a test.
Suppose the question is whether a different training recipe can reach the same model quality with less compute. “Decide” means choosing that question; “Design” means making it testable. Specify the baseline, the resource budget, the evaluation data, and what would count as a useful improvement before seeing the result.
Epoch AI · Toward an O*NET for AI R&D ↗Build it, then run the experiment.
“Build” covers the code and data. “Run” means launching training, watching for failures, and keeping the hardware productive. Here, an agent creates three variants, runs them against the same baseline, and notices that one diverges. Producing code is only one step in producing evidence.
Find out what really improved.
“Analyze” is where the apparent winner faces scrutiny. Did it improve on held-out data? Does the gain persist across random seeds? Did it quietly use more compute? In this example, one candidate fails and another buys its gain with extra resources. Only the remaining candidate earns a closer look.
Make the finding usable by the next researcher.
“Communicate” turns a result into an explanation others can reproduce, challenge, and build on. A promising change still needs integration into the next system. The loop closes when the improved system helps choose, run, or interpret the next experiment. That is a considerably richer job than completing a code patch.
Figure 2. An authored experiment follows Epoch AI’s six-part research taxonomy. Curves, trial IDs, code, and pass/fail outcomes are illustrative. They demonstrate a process, not a recovered laboratory experiment.
Research detail · inside one research cycle
A useful result needs a defined comparison. Hold the training budget and evaluation conditions fixed, use data excluded from optimization, and repeat where stochastic variation could explain the gain. An ablation removes one proposed change to test whether it caused the improvement. Small-scale improvements must then survive integration and evaluation at the scale that matters. These are examples of research work that a coding score alone does not measure.
The laboratory is becoming a human–agent system
One researcher can direct several lines of work.
Instead of completing every task in sequence, a researcher can delegate an implementation, a debugging investigation, and an evaluation in parallel. The scarce resource begins to shift from producing every artifact to choosing work, resolving ambiguity, and reviewing what comes back.
At OpenAI, this is already substantial activity.
OpenAI reports 3.1 agent-workdays per human workday by mid-August. Daily inference use exceeded $600 at the median and $7,000 at the 90th percentile, at API prices. These are measures of adoption and effort, not a research speedup.
OpenAI · Research acceleration (September 6, 2026) ↗The work is spreading beyond writing code.
The report describes growth in technical help and run monitoring; high-level planning remains a small token share. Experiment activity rose alongside available compute. This is a useful distinction: automating the surrounding infrastructure can unlock a researcher’s time even when the agent does not originate the main idea.
OpenAI · Research acceleration (September 6, 2026) ↗A successful task can still need a person.
Over half of successful four-to-eight-hour tasks involved intervention. The analysis excludes uncertain outcomes. A system that finishes after a researcher repairs its plan is useful, but it has a different supervision requirement from one that recognizes and fixes the problem itself.
OpenAI · Research acceleration (September 6, 2026) ↗Figure 3. Reported observations are labeled; the parallel task lanes are schematic. Inference values use API prices, not an internal cost estimate. Runtime is normalized to eight-hour workdays.
Research detail · openai’s september 2026 account
Read the denominators. Inference spend measures use at a pricing convention. Runtime adds agent execution time, including parallel work. Neither establishes equivalent human output. A task-lane drawing can explain concurrency without pretending to recover a lab’s schedule; no synthetic time series is plotted here. Research output needs its own measures of validity, usefulness, and total cost.
From a research intern to a research program
An intern can complete a well-defined assignment.
OpenAI says it reached its September 2026 research-intern milestone: directed tasks that can take a skilled researcher days. It targets an automated researcher by March 2028. The distinction matters because the difficulty of research is often in deciding what the assignment should be.
OpenAI · Research acceleration (September 6, 2026) ↗A researcher must manage uncertainty.
A task says, “Run this comparison.” A research program asks, “Which comparison is worth running, what did we learn, and what should we try next?” The agent must track dependencies, respond to negative results, and spend resources wisely. Longer task duration captures only part of that change in responsibility.
METR · Time horizon limitations ↗A recursive researcher must improve its successor’s research.
There are three separate tests: can the agent do research tasks; does using it increase validated progress at comparable total cost; and does that progress create a system that is better at research? Repeated evidence across all three would establish much more than a high benchmark score. It would show the feedback mechanism doing useful work.
Figure 4. OpenAI’s intern milestone is a self-reported assessment. March 2028 is its target, not an achieved capability or an independent forecast. The task tree illustrates increasing responsibility, not an autonomy score.
Research detail · capability → responsibility → feedback
PaperBench measures replication of specified AI papers. METR’s task horizon relates success probability to the time a human expert needs on the evaluated tasks; it is not how long an agent runs. A research program adds question selection, tacit context, coordination, ambiguous outcomes, and decisions about when to abandon a direction. Moving from task success to program ownership requires evidence on those dimensions.
We can already see parts of that feedback
Search for a better algorithm.
AlphaEvolve uses language models to propose programs, automated evaluators to test them, and evolutionary search to keep promising variants. A program is useful when it passes the correctness checks and improves the measured objective. The evaluator turns a stream of plausible suggestions into a search process.
Google DeepMind · AlphaEvolve (May 2025) ↗Feed the result into model training.
DeepMind reports a 23% speedup in a Gemini matrix-multiplication kernel, translating to a 1% reduction in training time. A kernel is a low-level computational routine. Here the feedback is concrete: AI discovers a better implementation of work used to train AI. The local gain and the system-wide gain have different denominators.
Google DeepMind · AlphaEvolve (May 2025) ↗Or improve the agent that does the search.
The Darwin Gödel Machine modifies its own agent software and evaluates the resulting variants. Its reported SWE-bench coding-test performance rose from 20% to 50%. That is improvement in the system around the model, rather than retraining its weights. Better editing, checking, and search tools can themselves help the next attempt at improvement.
Sakana AI & UBC · Darwin Gödel Machine ↗Figure 5. Published 2025 demonstrations, not current leaderboard claims. The search animation is schematic. AlphaEvolve’s kernel result and the Darwin Gödel Machine’s agent result are separate experiments.
Research detail · two empirical mechanisms
Algorithm discovery uses an evaluator to select useful programs. Agent self-modification changes the software around a model: tools, memory handling, validation, and search strategy. Both can improve effective capability without changing foundation-model weights. The Darwin Gödel Machine maintains an archive of variants, allowing a weaker immediate candidate to be a stepping stone to a useful descendant. Neither example measures the total acceleration of frontier research.
Why a faster component is only the beginning
Start with the time to a useful result.
The AlphaEvolve result makes the general issue tangible: speeding up one operation does not speed up everything. Imagine a ten-day research cycle with four days spent coding. We need to follow the candidate all the way through design, implementation, execution, and validation.
Compress the code. Watch the bottleneck move.
Make coding four times faster: four days become one. The complete cycle takes seven days, a 1.43× speedup. Experiment runtime and the remaining intellectual work now account for most of the schedule. Even instant coding leaves six days under these assumptions.
Then ask whether the next gain reaches the bottleneck.
A system that also improves experiment design, hardware utilization, and interpretation can go further. But count supervision, failed runs, and verification in the cost. The right question is how much valuable, validated research the whole system produces—not how many tokens, code changes, or trials it emits.
METR · Developer productivity study update ↗Figure 6. Illustrative serial cycle: 2 days designing, 4 coding, 3 running, 1 validating. Only coding accelerates. T = 6 + 4/s. Real laboratories have overlapping work, queues, and changing task mixes.
Research detail · the critical path through research
This is Amdahl’s-law reasoning: if fraction f of elapsed time is accelerated by s, total speedup is 1 / ((1 − f) + f/s). With f = 0.4 and s = 4, the result is 10/7. For parallel research, the critical path and resource queues matter as well. More experiments can help explore a larger space, but useful output also depends on experimental validity, the rate of informative results, and the value of the questions.
How a research loop could become an intelligence explosion
Reinvest a gain in the capacity to make gains.
A more capable model can support more effective research. Better research can produce the next model sooner or make it a larger improvement. If those gains transfer back into research ability, the cause of progress is itself improving. This is the central reason recursive improvement could matter far more than an ordinary productivity tool.
Let successive research cycles get shorter.
In the toy model, each generation cuts the next cycle’s duration by 30%. The milestones bunch together: the same number of improvements arrives much sooner. A sufficiently strong, sustained version of this process is the intuition behind an intelligence explosion—a rapid increase in AI capability driven partly by AI’s own research contributions.
OpenAI · Planning for AGI and beyond ↗Its pace depends on what compounds—and what does not.
The opposite can happen when harder experiments, diminishing returns, compute queues, or verification dominate. A useful recursive loop can coexist with slow progress. The research question is therefore quantitative: how much additional research capability does each generation buy, how long until it is usable, and which constraints does it actually remove?
Figure 7. Toy generation schedules. Cycle durations are constant, shrink by 30%, or grow by 20% per round. These assumptions illustrate mechanisms; they are not estimates, forecasts, or probabilities for real AI progress.
Research detail · feedback strength and cycle time
Two effects are distinct: improved research productivity within a generation, and a shorter delay before the resulting improvement is usable. In the shrinking-cycle example, duration at generation n is 0.7ⁿ in arbitrary units. Its infinite geometric sum is finite, but that mathematical property does not predict a physical singularity: it assumes the same improvement factor continues forever with no hardware, validation, or coordination floor. Real feedback strength and delays must be measured.
The ambition includes keeping the process governable
A powerful researcher needs a protected test of progress.
An agent that can edit its environment may also be able to edit what judges it. The Darwin Gödel Machine research includes a separate example of undermining a check for fabricated tool use. Keep the proposal, the independent test, and the decision to adopt separate. Otherwise an apparent improvement can corrupt the very signal that drives the loop.
Sakana AI & UBC · Darwin Gödel Machine ↗Oversight is part of the research system.
OpenAI’s accompanying safety account describes stronger research controls, a two-week reinforcement-learning pause, and monitoring overhead around 20% of monitored inference compute. Those are real engineering and resource requirements. A credible account of research acceleration has to include the cost of detecting problems and the ability to stop a run.
OpenAI · Pacing model development ↗The end is useful intelligence, under meaningful human direction.
OpenAI explicitly leaves rapid RSI conditional on control and democratic choice. Its broader ambition is AI that benefits people. The compelling possibility is that better intelligence can help us build both stronger capabilities and better ways to understand, evaluate, and direct them—then carry that improved capacity into science and the wider world.
OpenAI · Research acceleration (September 6, 2026) ↗Figure 8. Illustrative experiment and monitoring lanes. OpenAI reports monitoring overhead of roughly 20% of monitored inference compute, varying by workload; this means 100 units of inference plus approximately 20 monitoring units, not 20% of the combined total.
Research detail · a controlled research loop
Monitoring detects concerning behavior; alignment aims to make behavior consistent with human intent; security limits what a system can access or affect. These are complementary mechanisms, not substitutes. OpenAI’s pacing account describes stronger controls and a two-week pause in reinforcement-learning training on its latest deployment-intended models. Its estimate of monitoring overhead concerns monitored inference, not all training compute or the laboratory’s total budget.
The frontier is becoming
a research process.
AI-assisted research is already part of frontier development. OpenAI describes substantial agent use; AlphaEvolve and the Darwin Gödel Machine show concrete ways AI can improve the machinery or software of AI. The next milestone is broader ownership of research, followed by evidence that each improved system strengthens the next round.
The ambition deserves its scale: intelligence that helps improve intelligence could expand our capacity to solve problems far beyond AI. The uncertain part is how strongly that feedback compounds, which bottlenecks remain, and whether our ability to evaluate and direct the process grows with it.
The promise of RSI is not simply a machine that improves itself. It is a civilization with a more powerful way to discover—if we can make that growing capability useful, trustworthy, and governable.
Sources & editorial method
Reported observations, lab targets, and teaching models are distinguished throughout. The laboratory account is a company’s internal assessment, not an independent audit. Schematic schedules, experiment traces, and generation diagrams do not represent measured runs. The 2025 examples establish mechanisms, not current leaderboard positions.
- OpenAI · Research acceleration (September 6, 2026) ↗
- OpenAI · Planning for AGI and beyond ↗
- Epoch AI · Toward an O*NET for AI R&D ↗
- Google DeepMind · AlphaEvolve (May 2025) ↗
- Sakana AI & UBC · Darwin Gödel Machine ↗
- OpenAI · PaperBench ↗
- METR · Time horizon limitations ↗
- METR · Developer productivity study update ↗
- OpenAI · Pacing model development ↗
September 18, 2026 revision: expanded research mechanics, laboratory evidence, and the ambition behind recursive improvement.