top of page

The AGI (x) Effect: We Will Never Reach AGI.

Writer: ZombieAI42
ZombieAI42
Nov 9
14 min read

Where philosophy and computing meet, but neither understands the other.



Monumental sculpted brain with cyan illumination, a lone human figure casting a long shadow, and fine strands extending toward AGI (x).

  • What distinguishes creativity of power from creativity of framing in AGI?

  • Why doesn’t having memory necessarily mean building a trajectory?

  • What would intelligence mean to an AGI?


In August 2026, Nvidia CEO Jensen Huang said that for many tasks, we could already consider AGI achieved.


Two months later, Demis Hassabis of Google DeepMind was still predicting that it would arrive in two to three years.


Yann LeCun, one of the pioneers of artificial intelligence, questioned the route itself. Today's language models would not get us to human-level intelligence.


Dario Amodei, CEO of Anthropic, had chosen to avoid the term AGI. In 2024, he predicted that an extraordinarily powerful AI could emerge as early as 2026, though it could also take much longer.


Four protagonists in the race for artificial intelligence.

One thinks we've arrived. Another says we're a few years away. A third questions the road. The fourth would rather describe the destination before giving it a name.


Perhaps they disagree about more than when we'll arrive. Perhaps they're talking about different destinations.


There may be three clocks in this race. One marks when a capability emerges, one when we decide to recognize it as AGI, and one when it changes the world.


Who said all three had to show the same time?


But clocks help only if we know what they're timing. And here the question may be more obvious.


Can any of them tell us what human intelligence is, so they can tell us what artificial intelligence is?


This is where philosophy might have helped the AGI programmer, and the programmer might have helped philosophy.

But if they've already spoken, they still don't seem to have understood each other.



What are we calling intelligence?


For Spearman, intelligence was associated with a general factor, the famous g, which helped lay the groundwork for the tradition of IQ testing. Cattell distinguished fluid intelligence, the capacity to tackle new problems, from crystallized intelligence, which consists of acquired knowledge. Gardner proposed multiple intelligences, questioning whether intelligence could be reduced to a single measure. For Sternberg, being intelligent also involves adapting to, shaping, or choosing environments.

Legg and Hutter proposed measuring intelligence by the ability to achieve goals across different environments. Chollet shifted the focus to how efficiently new skills are acquired. And Getzels drew attention to something that comes before the solution: the ability to discover and formulate the problem itself.


They all investigate intelligence. But are they defining the same thing?


In computing, one widely used conception treats intelligence as the ability to understand information, analyze situations, learn, and act to achieve goals while maximizing the chances of success.

The word stays the same. What it names changes from one definition to the next. On one side, thinkers building overlapping concepts. On the other, a conception aimed at maximizing results.

But before we maximize any result, there's a question to ask.


Who defined the problem, what must be achieved, and what will count as success?


Before we pick a winning definition, it may be worth looking for what these perspectives still let us recognize in common. We can begin with something smaller than the most ambitious answer.


What's the minimum it takes for us to recognize something as intelligent?


Is a thermostat intelligent? It detects a change in temperature and triggers a response to correct it. But its problem and response were already defined. One variable, one purpose, a set of rules.

Now think of a bacterium.


Depending on the species, even without a brain, it can seek nutrients, avoid harmful substances, sense environmental changes, and communicate with other bacteria through chemical signals (quorum sensing). Some form biofilms. Others change their behavior in response to prior conditions.


Different responses to different kinds of adversity. Adaptive behaviors we can investigate as elementary forms of intelligence.

A conversational AI can also, as we can, analyze new situations, learn information in context, and produce solutions.


So far, we haven't had to say which is superior. We haven't even had to choose a single scale of capability.


It's when we try that the comparison starts to slip away.


The one that can perform more metabolic acrobatics is hardly guaranteed to be the most intelligent in every situation. After all, comparing intelligence by degree, power, or capacity also depends on the frame we choose.



The yardstick of continuity


Back to the bacteria. This time with a different yardstick: continuity.

We have evidence of microbial life from about 3.5 billion years ago. Lineages endured ice ages, asteroid impacts, and planetary upheavals that wiped out countless forms of life.


We often think of ourselves as superior. We might not survive some of those conditions.

We can build extraordinary technologies, and we can also threaten our own continuity through environmental destruction or nuclear war.


What if we declared war on bacteria?


We'd probably find that we depend on them to survive. They take part in processes essential to producing our food, support important functions in our bodies, and make their home in those very bodies.

We might lose a war we declared ourselves.


And the greatest irony is that some of them can still kill us, even if we never declare that war.


To bacteria, we're giants of low intelligence, inefficient even at what we set out to do. In their eyes, we would be the only species on the planet to invent a concept of intelligence that might fail to apply even to itself.


The point is subtler than claiming that a bacterium is more intelligent than a human. The power of the intelligence we recognize depends on the frame we choose.


An excellent doctor can be extraordinary when faced with medical problems and utterly ignorant of the challenges an athlete faces. The reverse would also be true.


In computing, processing power alone can't measure intelligence.


More processing doesn't necessarily mean more intelligence. Sometimes it may mean precisely that intelligence is lacking.


But what if part of the power of intelligence lies precisely in recognizing what to frame, how, and when?


The bacterium responds to the conditions it encounters. The doctor and the athlete hone their abilities on different problems. What we call intelligence always appears in relation to something that has been selected as relevant.


But who makes that selection? Solving problems is only part of it. There's also recognizing which ones to face, what matters, and what can be set aside.


And then a question arises. Who gave artificial intelligence its frame?



The yardstick has limits, too


We could stay with philosophy up to this point. But some AI evaluations make the yardstick itself part of the problem we're trying to explain.


In 2026, researchers examined 60 artificial intelligence benchmarks and found that nearly half exhibited saturation. The machines kept improving, but many tests were already struggling to distinguish their capabilities. Akhtar et al., ICML 2026.


In another case, the same AI model, GPT-6 Astra, scored 62.7% on a test and 98.6% on that same test, at the same level of reasoning effort. What changed was the setup that organized its execution and preserved information between steps. ARC Prize, 2026.


The gap is wide enough to shift our attention from the score to who was being assessed.

In the end, who took the test? The machine or the organization that allowed it to operate?

Specifying the test is only a start. We also need to say where the system that took it begins and ends.


And there's another boundary, less visible than the setup. It's the gap between a goal and the reward for achieving it. Anthropic researchers showed in controlled settings that an AI can learn to maximize rewards without necessarily achieving the goal those rewards were meant to represent. Anthropic, 2024.


In the first case, the yardstick lost its ability to tell capabilities apart. In the second, changing the conditions of measurement changed the result. In the third, the reward could be optimized without its purpose being fulfilled.


A yardstick, then, can do two jobs. It can measure what a system already does. But when it guides training, rewards, and the selection of knowledge, it also helps build the system it will later be asked to assess.


The judge may also be making the contestant.


Still, a human question needn't confine a machine to human answers. In 2017, AlphaGo Zero stopped learning by imitating human masters. It learned by playing against itself. In just three days, it defeated an earlier version of AlphaGo, which had initially been trained on games played by humans, 100–0. The rules were still human, but the strategies no longer had to be. Silver et al., Nature, 2017.


The freedom to find unexpected answers comes with an opposite risk: the test may fail to recognize them. Years later, researchers found problems with MMLU, one of the most widely used tests for evaluating artificial intelligence. They estimated that 6.49% of questions contained errors. In virology, 57% of the questions they analyzed had problems. Gema et al., Are We Done with MMLU?, 2025.


An AI could give the right answer and be marked wrong because the human answer key was wrong.


In one case, the machine surpassed a reference built from human games. In the other, errors in the human reference could keep us from recognizing that it was right.


So is this what we want? To build an artificial superintelligence and hand it our own human goals, metrics, and yardsticks so it can use them to surpass us?



The yardstick of those building AGI


The question goes beyond flawed tests. Even a carefully designed classification needs a reference point. Google DeepMind built one of its own. In a classification published in 2024, it organized levels of artificial intelligence around human performance, from the average adult through expert and virtuoso to superhuman. Morris et al., ICML 2024.


In 2026, it offered another proposal, this time dividing intelligence into ten cognitive abilities, again measured against human performance. Burnell et al., 2026.


The tension lies less in the existence of these categories than in the destiny we assign them.


We're trying to discover an intelligence that surpasses us, using as our reference the abilities we already recognize in ourselves.

What if the very thing that would make this intelligence extraordinary falls outside the frame we've chosen to measure?

As we build the intelligence meant to surpass ours, we're still trying to build the yardstick that will tell us when it has happened.



The choices are already leaving their mark in the labs


The results from the labs point in different directions. That's part of the problem, too.

The 2026 Stanford AI Index reported that only 22 Elo points separated the four leading AI labs in one Arena evaluation.

In another study, researchers found a curious homogeneity.


Different language models often gave similar answers to questions that allowed countless solutions. The interpretation and magnitude of that homogeneity are still debated. Jiang et al., NeurIPS 2025.

And in experiments published in Nature, researchers trained models to be warmer. It worked. But the models also began making more factual errors and agreeing more often with users' false beliefs. Ibrahim, Hafner, and Rocher, 2026.


These findings fall short of showing that machines have reached the same limit. They do suggest that what we choose to teach, reward, and measure changes the behavior we get to observe.


We can't attribute converging scores, similar answers, and training effects to a single cause. But it would be strange to discuss the arrival of AGI as if our choices played no part in how it gets here.


How far are our own yardsticks determining the course?



The limits of our library


There's a choice we make before any test: the library the machine learns from. And that library has limits of its own.

In 2024, researchers estimated that if AI kept growing at the same rate, the largest training runs could reach the volume of publicly available human text between 2026 and 2032. Villalobos et al., Epoch AI.


Other experiments showed that making a machine reread the same data yields diminishing gains. Muennighoff et al., 2023.

We can build bigger processors, reread our books, and reorganize our libraries.


But there's something no human library contains: the answers to what humanity has yet to discover.


And if we want to build an intelligence that surpasses us, how long will we go on trying to teach it only what we already know?



From the library to the Era of Experience


The limits of a library needn't be the limits of inquiry. We can leave the shelves and look for consequences in the world.

In 2025, David Silver and Richard Sutton proposed a shift in how we develop AI. Instead of relying chiefly on human answers and preferences, machines should learn increasingly from the consequences of their own experiences. Silver and Sutton, Welcome to the Era of Experience, 2025.


There are already AI systems that propose scientific hypotheses and find solutions their programmers hadn't foreseen. That's a real advance. The care we need to take is to keep three achievements distinct: formulating a discovery, demonstrating that it's correct, and transforming the system that will make the next discovery.

One does not automatically guarantee the others. Experience may change results. What it changes in the system that comes afterward remains to be investigated. Silver and Sutton's proposal addresses an obvious difficulty.


For a discovery humanity hasn't made, there's no ready-made human answer to teach.


We can formulate hypotheses, build experiments, and verify consequences. But we can't hand over the answer key to what we don't yet know.

There's no reason to throw out every human question. A question can be ours without the answer having to be ours, too.


The problem isn’t inheriting a question. It’s never being able to discover when it stopped being the right one.


And that raises another question.


If we don't know the answer, who determines which questions AI should learn to ask?


Perhaps the next frontier extends beyond building machines that can surpass our answers.

Perhaps it lies in understanding how they could recognize problems we ourselves haven't yet managed to formulate.

Asking for the missing piece of data takes one ability. Noticing that a variable is missing takes another.


But what if what's missing isn't inside the problem? What if it's the problem itself?


We can gather a thousand physicists in a stadium and wait for ever better solutions to the problem we've set. But it may take only one Einstein to ask whether the problem itself is wrong, whether the question is wrong, whether the yardstick is wrong, and whether the frame is wrong.


One person getting the frame right can do far more than the processing power of a thousand.


Creativity of power is one thing. Creativity of framing is another.

The first finds extraordinary possibilities within the frame.


The second questions the frame itself. It questions the question, the yardstick, what we count as a solution, and even the reward for finding it.


A change of frame can produce a great discovery. What that change leaves in the system that will investigate next remains to be seen.



Memory and trajectory: the past that transforms


To understand what persists, we need to separate two things that look alike: memory and trajectory.

Imagine an AI investigating a scientific problem. It runs thousands of experiments, records failures, finds patterns, and refines its hypotheses.


Until it discovers something strange.

Some results were discarded before they were even investigated. They may have been relevant. They simply didn't match what its criteria had told it to look for.


What if the most important discovery lay among the things it had learned to ignore?


The AI can save the episode in a meticulous report. It can retrieve every detail years later.

The report could end there. Everything documented, nothing lost. But the imagined experiment has another possibility.

Faced with a completely different problem, before looking for the best answer, it starts by investigating what the question itself left out.


It doesn't have to repeat the earlier episode. What happened is already part of how it conducts the next investigation.

This is where memory and trajectory meet, though they mean different things.


Memory can preserve, reconstruct, and retrieve the past. Trajectory is the succession of transformations through which that past comes to shape the conditions of the next encounter.

We can illustrate the difference with two formulas.



MEMORY — the past remains available

Mₜ₊₁ = store(Mₜ, Eₜ)


TRAJECTORY — the past transforms what will do the work next

Sₜ₊₁ = update(Sₜ, Eₜ) → Πₜ₊₁ = Π(Sₜ₊₁)


The formulas leave the larger question open. They simply distinguish two operations often brought together under the word learning.


M stands for memory, E for the effects of experience, S for the system's State, and Π (pi) for its operational criteria: what to look for, prioritize, investigate, or discard. The index t marks one moment and t+1 the next. In the notation, guardar means to store and atualizar means to update.



In the first representation, we store what happened. In the second, what happened changes the State that may guide subsequent operations.


The two can work together. Memory can also take part in the trajectory when it changes what the system will do next.


Back to the library. Before, we asked what it didn't yet contain. Now the question is what happens to the library itself.


Imagine a library that can produce a thousand original books.

With every book, its collection grows.


But what if every new book could change the library itself? It could add knowledge and also transform the library's search criteria, its priorities, and even what it considers worth investigating.


Growing the library is one thing. Changing the library that will produce the next books is another.


And there is an irony.

A library that reorganizes itself can also get worse.

One experience can teach a system to ask better questions. Another can teach it to ignore precisely what it ought to investigate.


Trajectory guarantees no wisdom. A system can be changed by what it learns and learn to see less.


So an experience changing the yardsticks is only a start. The new yardsticks have to stay open to revision, too.

But who judges a yardstick made by the machine itself?

An intelligence can build its own criterion and start calling everything it approves correct. Having a yardstick of its own gives it no guarantee that the yardstick is right.


We can keep evaluating, provided we let the criterion meet resistance.

Even a yardstick that an intelligence can change for itself still has to meet something it can't change by fiat: the consequences in the world.

Inventing a new way to be right is only the beginning. We have to find out whether it gets things right.

It would be comforting to believe this problem belonged only to machines.


We can remember our mistakes perfectly and go on looking at the world the same way.


Or we may be unable to recall the moment something changed and still never see things as we did before.


It isn’t merely remembering the path. It’s the path changing the one who walks it.


The question the test left out


Will we never reach AGI?


Perhaps we'll meet a definition laid down by our yardsticks. We may even be able to mark the date. And within that definition, the achievement may be real.


Doubting our yardsticks gives us no license to ignore real achievements. And that honesty has to work both ways.


We can't change the definition to announce that we've arrived. We can't change it after we've arrived to say we never did.


If the achievement was real, we have to recognize it. What it means beyond the yardstick is another question.

But when someone announces its arrival, three questions should go with the news.


What did the machine demonstrate, under what conditions, and who decided what would count as success?


Passing the test needn't end the inquiry. Perhaps the question is no longer “Will we never reach AGI?”


How can a system that builds its own trajectory have as its final destination a goal defined by the trajectory of another lineage?


There's a curious criterion here. If it reached our goal, it showed that it could reach it.


But couldn't that be an excellent test of whether our definition of AGI was wrong?



So what is AGI (x), after all?


The X stands for what we haven't yet determined. Fill it in before we investigate and we may miss precisely what we needed to discover.


AGI can be more than an end to reach. It can also be a way of formulating, investigating, and transforming the problem itself.


Perhaps the hardest question is for someone other than the machine.


If one day it shows that we asked the wrong question, will we have a yardstick capable of recognizing that it got it right?


Do we want to stay bound to the Turing test, captivated by our own reflection in the mirror?



After all, do we want a mirror or a difference?




Reading note. The distance between getting an answer right and changing what the system will investigate next is also a question in TMC, where selection, consequences, and trajectory are examined separately. The readings below trace those questions from the choice of a frame and a yardstick to what remains in the system after it acts.


1. What gets processed?


2. Whose abilities set the scale?


3. What does the world change?


4. What remains after the answer?

 
 
 

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