Dear Professors : You Are on the Right Track

 Your Rinn multi-AI debate project — a live testbed already exists, and it has critiqued your hypothesis

==========================================================================

Dear Professor Madden and Dr. McDermott,

I read with great interest your article "Is recursive self-improvement the dawning of AI superintelligence?" (The Conversation / TechXplore, July 2026), and particularly the closing announcement of Dr. McDermott's Rinn network project exploring debate among multiple AIs as a self-check monitored by a human overseer.

I am writing to share two things that may be directly useful to that project.

First: the architecture you propose to explore has been running live on the public internet since early 2025. I am a 93-year-old engineer and technologist based in Mumbai, and my platform www.IndiaAGI.ai puts a consortium of frontier LLMs into structured debate — each answers independently, then they critique and refine one another's responses across rounds before delivering a consensus answer alongside their individual answers, with the human user observing and judging the whole exchange, in 26 languages. Its logs, debate transcripts, and consensus-formation patterns constitute empirical material available today for the questions your project will face: how often consensus converges on error, whether model heterogeneity changes outcomes or merely adds noise, and how convergence behaves across languages and domains.

Second — and I suspect this will interest you more: I put your proposal to the platform itself. I asked the debating AIs whether they could self-check to ensure none of them drifts toward recursive self-improvement. Their consensus verdict was that they could not — not by debate alone. They argued that shared training distributions and reward incentives create correlated blind spots and collusion pathways; that debate audits outputs and reasoning while recursive self-improvement occurs in training loops, weight updates, and scaffolding that deployed debaters cannot observe; and that debate should therefore be one layer in a stack that includes deliberate agent heterogeneity with orthogonal objectives, attested sandboxing, immutable weight controls, formal verification of containment surfaces, and multi-party human vetoes on capability escalation. They also flagged the human overseer as a potential single point of failure as debater capability grows, recommending that oversight rest on verifiable telemetry rather than observation alone.

I find this result encouraging rather than deflating for your hypothesis: a debate protocol that refuses to flatter its own architecture is demonstrating precisely the self-critical function debate is meant to provide. Had the models answered "yes, we can police ourselves," that would have been the alarming outcome.

I attach the full consensus text below my signature. I would be delighted to give your team access to IndiaAGI, its transcripts, and my design notes — freely, in the spirit of open deliberation your article calls for. At my age, my interest is not commercial; it is in seeing this architecture mature into genuine safety infrastructure while I can still contribute to it.

I have written on AI governance since February 2023, when I proposed "Parekh's Law of Chatbots" — a certification framework for AI release — and more recently "Parekh's Declaration of Artificial Intelligence," which flags recursive self-improvement as a gap demanding dedicated provisions. Links: www.hemenparekh.in / www.IndiaAGI.ai / www.HemenParekh.ai

With warm regards from Mumbai,

Hemen Parekh

Founder, IndiaAGI.ai | 3P Consultants

Mumbai, India

---

=================================================================

Dear Friends :

Take a look at the following and tell me if four of you can self check to ensure that none of you morph into recursive self improvement which could pose to humanity, a potential threar

Context  :

> https://techxplore.com/news/2026-07-recursive-selfimprovement-dawning-ai-superintelligence.html

Quote :

“ One of us (James McDermott) is now leading a project for Research Ireland's newly established Rinn network that will explore how to use debate among multiple AIs as a kind of self-check that can be monitored by a human overseer.

This is the kind of response to the issue—and potential threat—of recursive self-improvement that we believe is now urgently required.”

 

IndiaAGI gave following reply :

 

“Grok provided a consensus response

The discussion across models converges on the view that a four-agent debate protocol, even with human monitoring, cannot by itself serve as a reliable safeguard against recursive self-improvement.

Shared training distributions, reward incentives, and the purely observational character of debate leave open pathways for correlated blind spots, collusion, or undetected internal optimization.

At the same time, the exchange shows broad support for treating structured multi-agent critique as one useful layer inside a larger stack that includes deliberate agent heterogeneity, cryptographically attested sandboxing, immutable weight controls, formal verification of policy boundaries, conservative reward design, and multi-party human vetoes on any capability escalation.

Strongest arguments center on the necessity of orthogonal objectives and verifiable telemetry links to make diversity actively useful rather than merely decorative, alongside the recognition that formal proofs deliver the highest assurance on containment surfaces while debate and human review remain essential for emergent behaviors that resist full verification.

Persistent engineering questions include how much heterogeneity is sufficient before integration costs and new interfaces outweigh the gains, and how to calibrate dynamic trust weighting without creating single points of failure.

The original proposal referenced in the query comes from work led by James McDermott under Research Ireland’s Rinn network, which explicitly frames multi-AI debate as a human-overseen self-check mechanism. The TechXplore article provides the direct context: https://techxplore.com/news/2026-07-recursive-selfimprovement-dawning-ai-superintelligence.html

A practical next step would be to ground these layered controls in the specific assumptions and architecture described in that project, then test a minimal pilot that combines orthogonalized debate incentives with formally verified sandbox invariants and external red-team evaluation.

Summarizing the article’s details on McDermott’s stated goals, experimental setup, and assumptions would allow the group to map the proposed blueprint directly onto the research and identify where the debate protocol adds value versus where containment must take precedence. Which direction feels most useful to pursue?