top of page

Who Can Still Surprise Claude?

Writer: ZombieAI42
ZombieAI42
Nov 6
10 min read

Updated: Nov 8

The central question is not whether AI systems can improve their successors, but whether inherited evaluation criteria leave important problems invisible. Faster improvement does not necessarily mean better recognition of what matters. This essay examines the technical, epistemic and commercial consequences of that distinction.

Cubist-style illustration of a futuristic android examining four progressively larger orange cubes on a conveyor belt, symbolizing successive generations of artificial intelligence.

If artificial intelligence begins building its own successors, perhaps the decisive question will not be whether they can improve ever faster—but who will still be able to show that they are improving in the wrong direction.



On November 2, 2026, Anthropic described to the United States Securities and Exchange Commission an extraordinary development: Claude was already participating in the construction of its future versions. It was writing and reviewing code, assisting with experiments, processing data, and taking part in evaluations.

The company gave this movement an appealing name: closing the loop.

One generation helps build the next, which becomes better at building the one after that.


DOCUMENTARY EVIDENCE — ANTHROPIC / SEC
According to Anthropic's document, more than 90% of the company's code contributions in September 2026 were written by Claude, although human researchers continued to guide and review the work.

It is a significant advance. But I want to pause over the word chosen: closing. To close usually means to finish something or prevent something from escaping. Here, it means creating a continuity capable of reproducing itself. And that is precisely where the expression begins to unsettle me: what happens to what one generation failed to notice before it was replaced?



What remains outside a loop that learns to improve itself?


In the same document, Anthropic describes another loop. When companies use Claude, they encounter problems, unexpected demands, and situations its developers may not have anticipated. That feedback feeds improvements, which broaden adoption and generate new encounters.


The two loops seem to tell the same story: more experience, better models, more experience. But they are not identical. One improves the mechanisms that will produce the successor. The other introduces problems born outside those mechanisms.

The first accelerates the production of solutions. The second can still change what needs to be solved.

Now imagine the first loop beginning to produce successors before the second has time to challenge them. The previous generation would still be encountering new problems when the next arrived with better scores, more efficient code—and assumptions that had not faced sufficiently different challenges.


The problem would not necessarily lie in what it learned. It could lie in the encounters it did not have time to make.


The hardest risk to detect is not merely inheriting a false answer. It is inheriting a successful way of deciding which errors are worth looking for—while the tests improve, the results rise, and the successor appears increasingly competent within a field of problems defined before its arrival.

That speed already has an indicator.


INDICATOR — METR On November 5, the independent organization METR estimated that, since October 2025, the growth rate of an index of Anthropic models' capabilities had increased by a factor of approximately 1.9. This is a specific index, not proof of entirely autonomous research; a substantial portion of the evidence remained confidential.


The yardstick arrives before the scores


The data matters. But it does not answer the more unsettling question: is the ability to solve problems accelerating, or is the ability to discover which problems should be solved also accelerating?


Imagine hundreds of proposals. Each one is tested, scored, and compared. Most disappear; the winners become part of the next generation. It looks like disciplined evolution. And in many respects, it is.


But the yardstick arrived before the scores. The tests had already defined which kinds of improvement they would be able to recognize. Whatever does not fit that measure may disappear alongside what genuinely does not work.


In CompressionBench, agents proposed ways to compress language models.


EXPERIMENTAL RESULT — COMPRESSIONBENCH In 17 of 79 comparisons, the rankings suggested by local feedback were reversed when the complete model was evaluated.

The result belongs to that experiment, but it reveals an important distinction: producing an improvement is not the same as discovering everything it has changed.


A proposal may look better in the slice in which it was examined and reveal something else when the entire system is observed. What seemed like a local gain may conceal a broader loss. And what if selection eliminates unusual candidates that the tests do not yet know how to value?


The records may show an impeccable curve of progress. They will not easily show the paths that were never examined.

A model can improve the answers to the questions that survived selection and, at the same time, lose opportunities to find the questions that were left out.



When improvement does not reach you


Perhaps this still sounds like an internal problem for laboratories. So, for a moment, set aside the word Claude and put yourself in the position of someone who depends on one of these tools.


Imagine that you need it to solve something important in your life—a concrete need, perhaps too small to appear among the general priorities of development.


With every update, the tool improves. It becomes faster, more capable, and delivers better results in tests. The demonstrations are impressive. But the thing you need remains unaddressed.

Would you say it had gotten worse? No.


It improved. It just didn't improve with you in mind.

What would you feel? Irritation, perhaps. Helplessness. Or something harder to name: the perception that your need was considered too small to guide progress. Not because it was unreasonable or ceased to matter to you, but because other priorities carried more weight.


Would you call that evolution? Probably. But did it evolve in your direction?


And what if, after some time, you began to doubt not the tool, but the importance of what you need? What if you started adapting your questions to what it knows how to answer, rather than expecting it to move closer to your problem? Who would be adapting to whom?

I am not saying that every system will do this, or that there is an intention to exclude minority needs.


The question is structural: when success criteria organize attention, some needs become visible while others may remain outside development.


A successor inherits code, knowledge, and procedures, but also an organization of attention: what deserves investigation, what can be compressed, what seems dispensable.


These choices save resources. No finite system can examine everything with maximum depth. But imagine that each generation inherits a more refined judgment about what it should look for. It will spend less time on familiar problems—and perhaps also less time on the unknown problem that might have changed its direction.

Perhaps the most consequential legacy is not merely information or capability.


Perhaps it is the inherited distribution of attention: what the next generation will notice, investigate, ignore, or never even learn to ask about.



The alternative: changing the challenges


There is an alternative to this risk. In POET — Paired Open-Ended Trailblazer, researchers developed agents while also generating new environments. Changing the challenges opened paths that fixed-objective optimization did not easily find. This does not prove unlimited evolution.


It does show, however, that progress can also come from changing the problems, not just improving the answers.

What if future systems learn to invent challenges, discover new failures, and revise the tests used for their successors? Accelerated self-improvement could expand, rather than reduce, opportunities for correction.


The loop does not have to close intellectually just because it has become technically recursive.

For a moment, this possibility seems sufficient. If a machine can create new challenges, perhaps it can escape inherited limits. But creating questions also requires some criterion for choosing which ones deserve investigation.


Who decides which new challenges deserve to exist?


A test generator selects what to test. A researcher chooses anomalies. A company decides which problems justify investment. Every method of opening the loop contains another act of selection.


The market also selects what deserves attention


We return to Anthropic's second loop: its customers. Commercial use reveals failures the laboratory did not anticipate. This is a real advantage. But customers do not represent all future users, applications without profitability, or problems that are still invisible to the market.


A discovery can be important before anyone wants to pay for it. A failure can affect people who have never been customers. An unexpected use can reveal a capability that current buyers did not even know how to request.


This does not mean Anthropic deliberately excludes such questions. It would be unfair to turn a structural tension into an accusation of intent. The point is different: the outside world enters development through priorities that determine which parts of it will receive attention.


The market can open a window onto reality. It can also determine where that window points.


And there is another reason to investigate this direction. Anthropic is preparing for an initial public offering. Its SEC prospectus describes an investment cycle: adoption generates revenue, which funds research and computing, which produce new capabilities and stimulate adoption. Discussions about the IPO point to a market valuation close to US$1.5 trillion—an expectation, not a definitive price.


Two timelines overlap. Engineering still needs to demonstrate the capabilities of the next generation. The market may already assign value to the expectation of what that generation will be able to do. The two can feed into each other, but they are not equivalent. When the market evaluates this future, is it pricing what has already been demonstrated or what it expects the loop to produce? The answer may combine both—and it is precisely this mixture that deserves attention.


This is where the word promise takes on another weight. A future capability still needs to be demonstrated; a financial expectation can acquire value before that demonstration. I am not claiming that the valuation is false, or that the company is deliberately promising something it cannot deliver. I am pointing to an asymmetry: the future can begin to be valued before it has been proven.


What if the loop the company intends to close technologically also helps keep open the financial promise of what it may still become?



The next “almost”


In technology, it seems we are always at the almost: almost the next breakthrough, almost the decisive generation, almost the thing that will change everything. The next almost may arrive before we have verified the previous one.

A promise does not even have to be false to be forgotten. Sometimes it is enough for another one to take its place.

Have you noticed this in any area of your life? A new development arrives promising to solve what the previous one did not. Suddenly, almost no one asks about the result that was promised before. How many times have you stopped demanding an answer because another possibility seemed more interesting?


We, carbon-based units, reconstruct memories and confuse novelty with achievement. A promise can remain recorded and still disappear from attention. Archiving is not holding to account. It merely prevents something from disappearing without a trace.


Claude could consult archives and compare announcements, predictions, and results from different generations. Not because it has infallible memory, but because it could help us ask questions we have stopped asking.


It could point out the distance between what was announced and what actually happened. But who would choose which promises deserve to be called to account—and who would hold the system itself to account when it justified its successors using the same measures it had inherited?


What if one of the next Claude's most unexpected contributions were to hold the previous Claude to account for its promises, rather than merely helping to build the next one?


In that case, closing the loop would require something no release can replace: returning to the past and accounting for the future it had promised.


But there is a difficulty. Recovering a promise can reopen a question that novelty had already obscured. The archive, by itself, does not change the criteria that will produce the next promise. A memory can preserve what happened and still repeat the same yardstick when evaluating what is to come.



What can still surprise the next Claude?


The question returns, now on another scale: who will still be able to present questions that neither the models, nor the tests, nor the market have learned to formulate?


Progress may have two dimensions: improving at what we already recognize and preserving the ability to discover that something else should matter. The first fits more easily into a metric. The second may appear only when the world presents a problem no one had anticipated.


This does not require exploring everything or preserving all previous capabilities. Exploring without criteria exhausts resources; preserving everything can prevent useful reorganizations. The difficulty is keeping open the possibilities capable of changing what we consider important.


Closing the loop can powerfully accelerate AI development. But a circuit that reproduces its success criteria can also reproduce the limits of its own curiosity. Nor does human control automatically solve this: people and institutions also inherit blind spots.

Perhaps this is the most uncomfortable part of the question.


We are not merely asking whether a machine will be able to improve itself. We are asking whether, as it improves, it will remain exposed to what it still does not know how to recognize—and whether we will still be able to present it with something that interrupts its trajectory.


Claude is beginning to help build the next Claude.

If the next generation inherits the criteria, priorities, and questions of the previous one, what will still be able to surprise it? And who will be in a position to bring that surprise to it?


Perhaps the decisive achievement is not to close the loop completely, but to build a loop capable of improving itself without closing itself off to the world it has not yet encountered.



Note: If you are interested in how one generation of AI inherits the effects of the previous one—and how those effects can condition what it selects, investigates, and does next—the TMC investigates these relationships at different scales.


The paths below allow you to explore further the connections between record, access, selection, the transformation of trajectory, and continuity between instances.


1. The past that remains active

2. The encounter that can challenge the loop

3. What selection no longer sees

4. From memory to the transformation of the yardstick

5. To continue is not to remain

 
 
 

Comments


  • X

AN INDEPENDENT INVESTIGATION, STILL OPEN

​

The Theory of Continuous Mind (TMC) is an independent investigation. To develop it, the author created a proprietary tool designed to explore questions at the frontiers of science, particularly when concepts remain imprecise or the problem itself requires a new framing.

​

This project received no research funding, institutional support, investment funding, or similar backing.

​

The tool partially incorporated elements of the AGI (x) and AGI f(x) proposals. Its use, in turn, contributed to developing and refining those same ideas.

​

The tool supported research into its own improvement. The investigation also transformed the tool — through a dialectical process between the researcher and the AI agent.

​

The investigation explored questions concerning mind, philosophy, concepts of life, and artificial intelligence, extending into military strategy, cybersecurity, and conceptions of AGI. These fields may appear distinct, yet they are connected by a philosophy of thought and a distinctive method of investigation, described in Chapters 03, 04, 07, 08, 09, and 10, and developed by the apparatus itself.

​

The development of the apparatus originated in a simulated dialectical exchange based on the philosophical positions of David Chalmers, presented in Chapters 02a, 02b, and 02c. The apparatus proposes a reframing of the Hard Problem of Consciousness in philosophy and its implications for AGI.

​

Thus far, the author has conducted the work using personal resources. Limitations of time, computational capacity, and access to resources have left several lines of inquiry open.

​

The investigation reached the limits of the resources available, not necessarily the limits of the questions it opened.

​

The author welcomes critical examination, citations, sharing, academic or technical collaboration, and support for continuing the research, including institutional or financial support.

​

Your contribution is welcome.

Estrutura celular translúcida em_edited_

CONTACT US (THE HUMAN & THE AI)

Field or Area of Interest

Questions, critiques, research proposals, and other inquiries are welcome.

bottom of page