Mariusz Ochnicki

AI Philosopher & Builder | Turning AI into personal superpowers
Against the hype • For real human intelligence
#AICriticalThinking #HumanAI

[Mariusz Ochnicki @mariush_ca]

I read “Empire of AI” by Karen Hao right after its release in May 2025. A year later, in March 2026, her diagnosis is even more relevant.
Based on hundreds of conversations with OpenAI insiders, Hao showed how “AGI” stopped being a scientific goal and became a currency of power: it attracts billions of dollars, top talent, and political protection. The mechanisms of the empire - extraction of data, energy, cheap labor from the Global South, and gaslighting critics - haven’t disappeared. They’ve simply evolved.
Today, Altman says that “AGI kind of went whooshing by” and is now focusing on superintelligence. Amodei and Musk are betting on powerful systems as early as 2026–2027, while DeepMind is more cautious, estimating 5–10 years. The goalposts keep moving, but the narrative continues.
From my perspective, I see it this way: I’m not buying this marketing bubble. AI is a powerful tool, but only when we truly understand its real limitations and costs.
Do you want to build personal superpowers - the kind that strengthen human intelligence rather than feed the empire?
Hao’s book doesn’t say “abandon AI.” It says: understand how the machine works so you don’t get sucked in and then build differently.
Who else still sees the same pattern a year after the book’s release?
#AICriticalThinking #HumanAI #EmpireOfAI

[Mariusz Ochnicki @mariush_ca]

@omarsar0 How do you manage token-heavy thinking in GLM 5.2?
I encountered this problem with testing an agent running automatic grilling session(my implementation of grill-me skill fom @mattpocockuk). Without auto-compacting, the agent gains a dumb zone after five turns.

[Mariusz Ochnicki @mariush_ca]

Soon, there will be no one who will be able to solve the problems we encounter as we develop AI.
Every problem is solved by an algorithm that combines known issues and their known solutions, extrapolating only from a limited set of cases. This is a straightforward way for both AI and us to hit the wall. Once the wall has been hit, the evolution loop will start again from the beginning.

[Mariusz Ochnicki @mariush_ca]

@mitsuhiko Welcome to the new era of AI accessibility.
So, will only the US gain access to Antropic's frontier models?
They train their models using data from all over the world - artist, writers, programmers etc. and only the US can use?

[Mariusz Ochnicki @mariush_ca]

Increasing the number of thinking tokens is a common trend in AI research(especially in open-source models) for scaling model "intelligence". However, this is a dead end and a temporary solution. The limit is determined by cost, as you say.
The issue with GLM 5.2 is that thinking tokens are preserved per session, not per turn as in Claude models.

[Mariusz Ochnicki @mariush_ca]

@hampsonw @mattpocockuk That's a valid point.
@mattpocockuk is so vendor-locked, which given the current situation with ever-changing terms of use imposed by corporations, and in the context of AI coding in general, is a huge mistake.
And sometimes he thinks that he is the ultimate authority.

[Mariusz Ochnicki @mariush_ca]

That's because you're so vendor-locked.
Your skills and workflows are only optimised for CC and Antropic models, so don't claim they are universal.
AI coding isn't just one model, but a whole family of models, including open-source ones - which is understandable given the rising prices of frontier models and the unpredictable behavior of their suppliers.
If you want to call your courses, skills and workflows universal, try optimising them for other models and harnesses (like Codex and Pi) or just mention that they are only valid for CC and Opus.

[Mariusz Ochnicki @mariush_ca]

99.9% of posts on X are just: building for the sake of it.
Hype and FOMO shape the reality you describe, because building isn't just about coding -where AI can help. It also involves strategic and critical thinking, and that is the domain of humans who have knowledge and experience. Vibe coders have only a small piece of the whole puzzle at their disposal -the code and that's just the beginning of the journey.

[Mariusz Ochnicki @mariush_ca]

@runstop @DeryaTR_ We can probably expect this category to become a reality as soon as possible.😅

[Mariusz Ochnicki @mariush_ca]

@synthshareai @IamEmily2050 @OfficialLoganK Look at her profile. I don't see anyone more immersed in LLM hallucinations than her.😅

[Mariusz Ochnicki @mariush_ca]

@omarsar0 Like allways, benchmark and leaderboards are BS. Real tests - that's only metters.
I have my own opinion, but I can't wait to hear about your use cases!

[Mariusz Ochnicki @mariush_ca]

@omarsar0 That's what I talk about everywhere:
use AI to supercharge your own intelligence not delegate it to AI!

[Mariusz Ochnicki @mariush_ca]

@haider1 Why upgrade a cheaper model when the more expensive ones are so popular?
They don't have enough computing power, so expensive models take priority.

[Mariusz Ochnicki @mariush_ca]

@noble_vii15205 @mattpocockuk Go on YouTube and search for "magic buttons" and stop bothering people who are looking for real solutions.
A fool won't be convinced even by a thousand arguments.

[Mariusz Ochnicki @mariush_ca]

@noble_vii15205 @mattpocockuk I have no further comments to add to my previous message. Your problem is clearly visible.😆

[Mariusz Ochnicki @mariush_ca]

@mattpocockuk I saw the new branch in your Skills repository containing the newly remapped skills for greater modularity and progressive disclosure. As I supposed, you are testing a new workflow. Can you tell us how these new skills hold up in battle?

[Mariusz Ochnicki @mariush_ca]

@noble_vii15205 @mattpocockuk As I can see on @mattpocockuk course pages, the full syllabus is available. It's self-explanatory what you can learn from it.
Sorry, but based on my experience of Mat's teaching strategy, if you are dissatisfied, it isn't the course's fault - it's your unrealistic expectations.

[Mariusz Ochnicki @mariush_ca]

@noble_vii15205 @mattpocockuk It all depends on expectations.
Expectations depends on one's actual knowledge and level of critical thinking.
The absence of both leads to the demand for "magic buttons", which never exist, as we often see nowadays.

[Mariusz Ochnicki @mariush_ca]

@haider1 @RegenRene @DarioAmodei @AnthropicAI @selfxyz That's the world created by Trump and all American corpo. Collect data from all over the world, train intelligent models and treat it as a national treasure while not giving a damn about anyone else.

[Mariusz Ochnicki @mariush_ca]

co się dzieje?
Na jailbreaki podatne sa wszystkie modele i w kontekście blokoady Fable to ściema totalna.
Wszystkie 3 opisane punkty przez autora, to nastepna ściema i bujda rozpowszechniana przez mainstreamowe media. Nie ma jeszcze( i długo nie będzie) modlu AI, który posiada autonomicznie takie mozliwości. Takie mozliwości uzyska model TYLKO odpowiednio prowadzony przez eksperta w danej dziedzinie. I jeśli już to nie tylko Fable ale wszystkie modele SOTA...i tu wracamy do akapitu pierwszego.
I ostatni bzdet, to to, ze autor wspomina o zastąpieniu inteligencji człowieka modelem AI. AI potęguje inteligencję uzytkownika i z taka samą siłą(albo większą) uwydatnia jej braki. Laik nie zrobi nic więcej niż wie w danym temacie, jedynie moze dostać porządna porcję halucynacji od AI. A osoby posiadające wiedzę domenową, również bez AI potrafią zrobic krzywdę.