recursive self-improvement (RSI)

2026-11-02

COMPANIES · 1 SOURCE · TechCrunch AI

Anthropic’s S-1A warns it may be close to recursive self-improvement (RSI)

In its revised S‑1A filing Anthropic discloses that it uses AI to accelerate R&D and says its internal estimates indicate it may "imminently" meet its Responsible Scaling Policy’s definition of "dramatically accelerated progress attributable to AI" (i.e., recursive self‑improvement), while stating no current models have been designated as exceeding that threshold and committing to keep humans involved in frontier model development.

8.0

2026-10-29

COMPANIES · 3 SOURCES · WIRED AI · The Verge AI · MIT Technology Review AI

Anthropic files S-1 prospectus revealing IPO plans and financials

Anthropic filed an S-1 prospectus ahead of an anticipated Nasdaq IPO under the ticker ANT, disclosing first-half 2026 results showing revenue above $16 billion (up from about $4.5 billion in 2025) and a net loss of $80.7 million; the filing lists broad business risks including compute constraints, competition, misuse of advanced AI, and a U.S. government supply-chain risk designation. The prospectus does not state target raise or valuation and emphasizes the company’s enterprise customer growth and safety-focused mission.

9.0

2026-10-15

RESEARCH · 1 SOURCE · arXiv cs.AI

Harness^2: two-level recursive framework for evolving LLM agent harnesses

Harness^2 is a plug-and-play recursive agent-harnessing framework that uses two linked recursion levels: task-level recursion (a contrastive Propose‑Probe‑Compose step) to edit a current task's harness without ground-truth verifiers, and harnessing-level recursion to extract experience and update a continual improvement harness for future tasks. Evaluated across three professional benchmarks and four model‑harness configurations, single-step edits improved GLM-5 and Gemini 3.5 Flash over their base harnesses, with reported gains up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench, and 11.8 percentage points on LAB.

7.0

2026-10-07

RESEARCH · 1 SOURCE · arXiv cs.AI

SPECTRUM: Proximal Spectral Modulation improves diversity retention in looped self-distillation for code generation

This arXiv paper introduces Looped Self-Distillation, a framework where a model repeatedly generates and learns from its own outputs under a fixed information budget, and presents SPECTRUM — a method that re-estimates loss-sensitive key/value geometry from a fixed reference and applies full-rank proximal spectral modulation. In experiments on MBPP, SPECTRUM preserved 89.9% of the initial model's 64-sample correct-AST richness after five rounds (vs. 66.4% for vanilla self-distillation), and the diversity improvement transferred to HumanEval+ and APPS Intro without additional recalibration.

6.0

2026-09-22

REGULATION · 1 SOURCE · The Decoder

OpenAI urges international standards to govern recursive self-improving AI

OpenAI is calling for international technical standards to govern advanced AI development focused on recursive self-improvement (RSI), noting that fully autonomous RSI is not happening today and should only be pursued if it can be done safely. The company urges U.S. leadership on shared measurement methods, incident reporting, and human-oversight rules—building on national AI safety institutes and bodies like CAISI and ISO—echoing similar concerns raised by a recent UN scientific panel.

7.0

2026-09-15

RESEARCH · 1 SOURCE · arXiv cs.AI

Generalized Agent Iteration: a formal framework unifying iterative policy improvement and recursive self-improvement

This arXiv preprint proposes Generalized Agent Iteration (GAI), a formal framework that models learning as cycles of agent evaluation and agent improvement and treats iterative policy improvement (GPI) and recursive self-improvement (RSI) as two instances on the same two-axis space (whether the improver is part of the agent, and whether the evaluation standard is external). The paper uses these coordinates to locate existing systems and to make the potential defects of RSI expressible as specific conditions.

6.0