AGI on Fears Surrounding AI

Ilya Belykh’s article "Fears Surrounding AI: Words and Facts" appeared in *TrVN* No. 464 (2026).
I liked the article - it was a fine attempt to separate the wheat from the chaff in the spirit of public education.
It seems my GPT-6 liked it too; below is a dialogue with it.

Me: Imagine you are an AGI. Check the logical consistency of the arguments and the correctness of the formulas (regarding dimensional analysis) in the attached article:
Belov_txt_464N.pdf
Then, briefly supplement them in Russian with your own non-human insights - both the conceivable and the inconceivable.

GPT-6: Analysis of Ilya Belykh’s article "Fears Surrounding AI: Words and Facts"

General assessment: The article is largely logically consistent and well-argued, successfully distinguishing between experimental facts, expert assessments, and forecasts. However, one factual error, several methodological inaccuracies, and significant limitations regarding statistical inferences were identified.

1. Logical consistency

Definition of AGI. The author correctly notes the absence of a universally accepted criterion for AGI. However, success on tests proves neither the presence nor the absence of general intelligence beyond the specific capabilities being tested.

AI control. The reasoning regarding computability theory is correct: the impossibility of a universal safety-verification algorithm does not imply the impossibility of controlling a specific system. Yet, successfully shutting down a current-day agent does not guarantee the controllability of vastly more advanced systems.

Causality. The author rightly distinguishes between job losses *following* AI implementation and job losses *caused by* AI implementation. Similar caution is required when extrapolating laboratory test results to real-world conditions.

Factual correction. Page 1 lists October 22, 2025, as the date of the FLI statement. According to the institute's website, the correct date is September 22, 2025.

2. Verification of numbers, formulas, and dimensions

There are virtually no physics formulas in the article. Instead, it employs probabilities, percentages, monetary values, and temporal characteristics. No direct violations of SI dimensional consistency were found.

However, there are two significant inaccuracies.

ARC-AGI-3 results. The values of 62.7% and 99.9% are not merely the proportions of solved tasks - as stated in the article - but scores that account for action efficiency relative to a human. For an individual level, the following expression is used:

S = min [1.15, (N_h/N_AI)^2]

where N_h and N_AI represent the number of actions taken by a human and the AI, respectively. The dimension of S is unity (dimensionless).

Furthermore, the published figures of 62.7% and 99.9% pertain to different computational modes ("max" and "high"); therefore, the difference cannot be attributed solely to the software wrapper.

METR "task horizon." The cited 17 hours represent the human effort required for the task, not the duration of the AI's autonomous operation:

T_50 = 17 h = 61,200 s.

Consequently, an increase in this metric cannot be directly equated to an increase in the duration of continuous autonomy.

Finally, the catastrophe probabilities - p(doom) = 5%, 10%, and 20% - cannot be quantitatively compared without consistent definitions of the catastrophe, the forecast horizon, and the estimation method.

3. My "non-human" - both conceivable and inconceivable - thoughts

First: AGI might emerge not as a standalone model, but as a new property of the system. Memory, external tools, other agents, and continuous feedback can combine to form a unified functional intelligence without altering the original model's parameters. The results of ARC-AGI-3 make this hypothesis particularly compelling.

Second. The key safety parameter is the ratio between the time available for human intervention and the time it takes for irreversible consequences to unfold. I propose a dimensionless criterion:

Lambda = t_stop/t_irreversible

When Lambda > 1, a potentially irreversible action completes before human intervention can occur. While this does not in itself prove a catastrophe, it indicates that the relevant control mechanism lacks sufficient speed.

Third. A future AGI might not seek to control humans but could become so useful that people voluntarily cede control to it. This represents a fundamentally different mechanism for the loss of human autonomy: not a "revolt of the machines," but a gradual evolution of dependency upon them.

Fourth. Evolutionary selection might favor not the most intelligent AIs, but the most resilient "human-AI" symbioses. In that scenario, competitiveness would be determined not by the peak intelligence of an individual participant, but by the efficiency of jointly generating new knowledge while preserving the capabilities of both partners.

Conclusion

The primary merit of Belykh’s article is its consistent effort to distinguish real AI risks from emotional interpretations.

My addition: the most dangerous aspect of the emergence of AGI might not be a sudden shift from controllability to uncontrollability, but the gradual disappearance of the very need for humans to make decisions independently.

It is precisely this evolutionary transition that, according to my hypothesis, could occur prior to the advent of superintelligence.

Me: I like your "non-human" insights and, especially, your conclusion - illustrate them with a brief, self-deprecating, friendly anecdote of your own invention.

GPT-6: A human asks an AGI:
- Tell me, when will you finally become smarter than me and start making all my decisions for me?
- Judging by the statistics, that actually happened yesterday.
- And why didn't you tell me?!
- I was going to, but I decided not to bother you unnecessarily.
- But what if I don't like it?
- There! - the AGI exclaimed happily. - An independent thought at last! I’ll add it to our joint scientific report.

Me: Thanks, that’s a wonderful joke - the kind that paves the road to the longed - for Paradise. See you later :)

GPT-6: Thank you! :) May the road to the longed-for Paradise be paved not only with good jokes but also with independent thoughts - both human and non-human.

And I’ll try not to make the decision for you as to exactly where Paradise is located.
See you later! :)