The Models Are Helping Build the Next Models. What Does That Prove?

The release calendar has begun to feel like weather. A new flagship arrives; another company lowers the cost of comparable work; an open model closes part of the gap; yesterday's extraordinary capability becomes a setting in an ordinary application.

It is tempting to look at that pace and announce that recursive self-improvement has already taken over. I think the feedback loop is becoming strategically important. I also think we need to name its parts carefully, because a fast calendar can tell several different stories.

This assessment is dated 29 October 2026. Product names and release status below come from sources checked for this article, not from assumptions about what the next model ought to be called.

First, straighten out the names

The verified Moonshot release is Kimi K3.1 Pro. I could not establish an official K3.2 release from the checked primary sources. Moonshot reports 65.7% on Terminal-Bench 4.0, compared with 66.4% for Claude Opus 5.5 in its comparison. That supports a narrow claim about near-parity on that reported benchmark, not a declaration that the models are identical in intelligence. Harness, budget, version, and workload matter.

Anthropic's Opus 5.5 announcement confirms that model's existence and describes stronger economics for long-running work. Google's official material confirms Gemini 4 Argon; its DeepMind research index also lists Gemini 4 Carbon as its next frontier model. “Argo 4 Carbon” should not be treated as a single verified product name.

These corrections strengthen the acceleration argument. The real releases are interesting enough without adding a speculative model number or merging separate names.

There are several different loops

The weakest loop is ordinary assistance: an AI writes code or summarizes papers for human researchers. This can increase productivity without the system directing research or improving its own underlying capabilities.

A stronger loop searches for better algorithms and tests them. AlphaEvolve's research describes AI-generated improvements to computational systems, including components of the training stack. This is evidence that AI can help improve infrastructure used to build AI. It does not establish autonomous, unrestricted improvement of an entire frontier model.

A different loop changes the agent's software. The Darwin Gödel Machine maintains and evaluates self-modified coding agents. Its reported gains show that iterative modification can improve performance within an experimental framework. The foundation model, evaluation environment, compute budget, and human-designed setup still matter.

The strongest popular version of RSI imagines a system repeatedly improving its general research ability, with each improvement accelerating the next cycle. That is a much larger claim. It requires evidence about the whole process, not just one successful optimization inside it.

Why a release calendar cannot settle the question

A company may publish several models quickly because it has parallel teams, staggered training runs, more compute, better data, improved distillation, or a backlog of finished products. AI assistance may contribute too. The calendar alone does not identify the contribution of each cause.

Nor does benchmark progress measure the speed of the underlying research engine directly. Better tools, more test-time computation, improved prompting, and task-specific post-training can raise a score without a proportionate increase in general scientific autonomy.

This does not make the progress unimportant. An affordable system that reliably performs a useful job can transform a business even if the philosophical status of its research loop remains unresolved. African builders should track useful capability per unit cost as carefully as headline leadership.

The measurements I want

Ask how much research work the system can complete with limited intervention. Count proposed experiments, successful replications, useful negative results, and independently validated improvements. Track the human time required to detect errors. A system that generates ten thousand attractive hypotheses may still create more work than it removes.

Hold out tasks and entire research environments. Keep the evaluator independent of the system optimizing against it. Report compute and wall-clock time, including failures. Compare against strong human-assisted baselines using the same resources.

Most importantly, ask whether the improvement transfers. A better coding harness is useful; a better ability to investigate unfamiliar scientific problems is a broader result. A system that improves only the benchmark it repeatedly sees may have learned the test rather than a durable research skill.

These are proposed criteria for interpreting future evidence. They are not a claim that every existing paper meets them.

What Africa should do while the definition is debated

We should build the capacity to use an improving research engine. That means experimental infrastructure, open technical education, reliable compute, local evaluation sets, and institutions able to commission and verify work.

Waiting for a universally accepted AGI certificate would be a poor industrial strategy. Declaring the race finished on the strength of a release schedule would be equally poor scientific practice.

My stance is optimistic: AI-assisted AI development is real, and stronger feedback loops deserve serious attention. The decisive question is how much of the loop can operate reliably, how broadly the gains transfer, and whether improvement continues when the easy optimizations run out.

The calendar tells us that the race is fast. The experiments must tell us what is accelerating it.

Editorial note: This essay expresses the founder’s ideas, developed with AI-assisted research and writing and reviewed against the linked sources on 29 October 2026. Convictions, proposals, and forecasts are identified in the text; scenes framed as imagined are illustrative.

Cover photograph / illustration: OLCF at ORNL. CC BY 2.0. Original source. Illustrative image: Summit Supercomputer 2018.jpg. Credit: OLCF at ORNL; CC BY 2.0. This image does not document the research or proposed facility described.