Poetiq | A Poetiq Perspective on Recursive Self-Improvement

get in touch<br>hello@poetiq.ai

join our team<br>join@poetiq.ai

A Poetiq Perspective on Recursive Self-Improvement

RSI is the fastest path to superintelligence, because its gains compound at every step.

August 4, 2026

RSI is here, and it's fast

At Poetiq, we believe RSI is the single most important frontier in AI research today, because<br>it is the fastest path to superintelligence. The reason is simple: the gains from RSI<br>compound with every step. Humans don't become continuously smarter while improving models.<br>RSI systems can.

As Anthropic recently stated,1 AI that can improve<br>itself would be a “major development in the history of technology” — one with enormous<br>potential to do good in science, healthcare, and beyond. They also raise thoughtful questions<br>about how to keep these systems controllable as they accelerate, and we think they're right<br>to ask them.

RSI is already here; we've been watching it work for months at Poetiq. Most of the field is<br>still debating when it will arrive — we're past that question, focusing now on how to build<br>it well. Not all RSI is alike: the form it takes shapes both how good and how controllable<br>it can be, and we've deliberately chosen one of the more controllable forms. We've also<br>chosen a form that is fast; in less than a year we went from zero to automatically<br>achieving state-of-the-art results on major benchmarks.

In the following, we define RSI and reveal some details about our approach, self-optimizing<br>optimizers, for the first time. We compare our approach to other major current approaches to<br>RSI and discuss some of the evidence supporting the benefits of our self-optimizing<br>optimizers. We also share our perspective on safety and control, and give an overview of our<br>long-term vision.

Defining RSI

“RSI” has become one of the most overloaded terms in the field, so let's be precise. Not<br>every system that improves over time is doing RSI. Consider the three letters in turn.

(I)mprovement is the easy part: the system gets better at a task, generally<br>in an iterative loop. But this isn't a differentiator — plenty of classical and modern AI<br>systems clear this bar through better algorithms, more data, or longer training runs. What<br>actually distinguishes RSI lives in the R and the S .

(R)ecursive and (S)elf are the parts that matter.<br>Self means the system improves itself, not some external target. Recursive<br>means its newly improved capabilities become the tools it uses to drive the next round of<br>improvement. These come as a pair: a system that genuinely improves itself is, by the next<br>step, a better improver running the loop. Improve some other target and the gains never feed<br>back; improve yourself just once and there's no chain reaction. Put the two together and<br>every improvement raises the system's ability to make the next one.

Drop the R and the S and you fall back to standard iterative optimization. Keep both, and<br>the improvement compounds and the exponential gains are realized.

A concrete example of a method sometimes confused with RSI but which is<br>not RSI is autoresearch2 (and<br>the similar broader class of “Automated AI Scientists”) where the core AI system improves<br>some target LLM, but does not use that target LLM to generate subsequent improvements.<br>Closing the loop, so the model being improved is the one driving the improvement, would turn<br>it into genuine RSI.

The gains from RSI compound with every step. Humans don't become continuously smarter while<br>improving models. RSI systems can.

There's more than one way to build RSI

Even for genuine RSI, there are multiple options about what improves at each step. That<br>choice sets how fast and how efficiently — in terms of compute, cost, and data — the loop<br>can run.

Most approaches are LLM-centric : the language model itself is the unit of<br>improvement, and improvement flows through neural-network training. Anthropic's work is a<br>clear example: Claude Code helps build the next Claude, with a model-training step at each<br>turn. Similarly, OpenAI intends to have their own automated AI researcher doing a<br>“significant fraction” of their internal research by 20283, and Google is using AlphaEvolve to make training LLMs more efficient4. These approaches are powerful. But when the model is the thing you improve, every step<br>inherits the cost, latency, and energy of a training run. On an exponential curve, what<br>matters is how quickly you clear that slow early regime, and with months between steps, the<br>benefits of compounding take far longer to arrive.

Poetiq's approach: self-optimizing optimizers

Poetiq takes a whole-system view. We treat the LLM as a single component of a larger<br>reasoning system (code, prompts, exploration/exploitation strategies, and more) and what<br>improves in each iteration is the system, not necessarily the model's weights.

Concretely, the Poetiq Metasystem is a self-optimizing optimizer . Each<br>problem, task, and dataset it optimizes helps it to...