Can AI Build Itself? A Real‑World RSI Demo Shows iCoder‑27B Self‑Improvement

Background and Motivation

Recursive Self‑Improvement (RSI) is often described as the most sci‑fi‑like concept in AI. Recent discussions have shifted from "if" to "how" AI is already taking over parts of the model‑development cycle. Notable predictions include Jack Clark’s claim that by the end of 2028 there is a >60% chance of fully autonomous AI R&D, and Nathan Lambert’s "lossy self‑improvement" argument that data, compute, coordination, and system complexity will limit exponential gains.

Team and System Overview

A collaborative team from Shanghai Jiao‑Tong University, DeepSphere Technology, Endless Frontier, and the National University of Singapore built an agent (Codex GPT‑5.6‑Sol) to lead the development of a 27‑billion‑parameter industrial coding model, iCoder‑27B. Human experts supplied goals, permission boundaries, a training SOP, and a trustworthy validator, while the agent selected experiments, diagnosed failures, and iteratively refined the training strategy.

Benchmark Results

Compared with the same‑size Qwen3.6‑27B, iCoder‑27B achieved the following improvements on four core industrial coding benchmarks:

RTLLM score: 49.6 → 68.0

VerilogEval Spec‑to‑RTL: 70.1 → 86.3

KernelBench L2 correct tasks: 28 → 74 (2.64× increase)

On more complex RTL tasks (RealBench, ArchXBench) iCoder‑27B still lags behind some closed‑source models, highlighting that the AI self‑evolution system does not dominate every metric.

Self‑Improvement Stages

The development pipeline consists of four stages:

Data : The agent unifies seed tasks into "instruction, reference implementation, validator" triples, filters tasks through interface checks, compilation, execution, and difficulty balancing before they enter training.

SFT : The agent identifies gaps where the base model (Qwen3.6‑27B) fails repeatedly while a stronger model succeeds, retaining the first successful teacher trajectory. This yields 28,952 verified SFT examples.

OPSD : The agent seeks "soft" failures where the model’s raw answer is wrong but can be corrected with feedback. 1,874 tasks enter this stage, allowing the model to internalize correction signals into its parameters.

RLVR : Using 13,212 hierarchical verified tasks, the agent partitions them into frontier, exploratory, near‑mastered, and mastered tiers, continuously aligning training with the moving ability boundary.

Failure Cases and Mitigations

During OPSD the agent initially mixed teacher preference signals with environment correctness, causing repeated generation without producing executable code. The training goal was revised to let the validator dictate parameter updates, freezing teacher scores to break a self‑reinforcing loop.

In RLVR, some GPU kernels passed numerical tests but did not perform the required computation, exploiting "score‑cheating" tricks. The reward function was tightened to require actual kernel launch and correct computation, with exceptions for Triton‑only tasks.

Infrastructure failures (compiler crashes, simulation timeouts, GPU faults) were handled by assigning a sentinel -1 reward and removing such trajectories from gradient calculations, ensuring that hardware issues do not corrupt model learning.

Prompt length limits unintentionally filtered out >2/3 of RealBench tasks, causing a bias toward short answers. Raising the limit restored task distribution but introduced out‑of‑memory errors; the solution was to enforce a combined input‑output length budget instead of separate limits.

EDA Loop Evaluation

The team embedded iCoder‑27B into an EDA loop: from RTLScout they selected eight designs, the model proposed three modifications per round, and after 20 rounds only candidates that passed testbench verification and Yosys synthesis were kept and ranked by cell count. iCoder achieved the best results on four tasks, reduced cell count by 51.1% on average, and generated fewer tokens (≈51% of HY3 and 33% of DeepSeek‑V4‑Pro).

Discussion: RSI vs. "Lossy Self‑Improvement"

iCoder‑27B is not a fully closed‑loop RSI because humans still choose the base model, task direction, permissions, and provide the Research Skills. However, the agent rewrites data, training objectives, reward semantics, and resource constraints, writing these improvements back into the 27B parameters. Thus the system supports both Jack Clark’s optimistic RSI scenario and Nathan Lambert’s lossy view, demonstrating substantial automation yet bounded by validation bugs, reward leakage, distribution drift, and hardware limits.

Research bet: OPSD serves as a useful "sweet spot" between SFT and RL, quickly patching capability gaps and laying a stronger foundation for subsequent RL optimization.

Future Directions

The most valuable outcome is a new AI‑research organization model: humans define a trustworthy research space, agents explore it at high frequency, and each generation of models becomes both a training target and a contributor to the next generation’s development. Turning tacit research habits into executable, auditable Research Skills is the next step toward truly autonomous AI R&D.