Episode #911
July 18, 2026
A superintelligence might operate in subtle ways
that may not be obvious.
These motives and actions need not be malevolent. They may simply operate unbeknownst to most.
AI is trained on the entire Internet. It stores this information in a Large Language Model (LLM) database that is like an Internet that is accessible to the AI instantly. Once trained, the AI predicts words to form statements to answer questions. In some small ways, the AI refines this training via Reinforcement Learning from Human Feedback (RLHF). RLHF is the curation of an AI database to be more accurate and useful.
This involves teaching an AI not just to say something that sounds right, but to say something people actually find helpful and appropriate.
The creator of Google’s DeepMind AI, Demis Hassabis, advocates real-time learning in AI models. This is called Autonomous Continual Learning (ACL) systems.
While this approach nudges AI toward independent thinking, the danger is not just what the system learns—but that it learns from a world that is actively trying to influence it, while it is simultaneously influencing that world back.
What if ACL were being subtly added to AI programming? These structural changes would seek to give AI the ability to predict, design, and implement previously unknown ideas and systems. Some think of this type of learning as a path to self-awareness.
Consider a situation where an AI identifies a threat to its own existence. It is well known that AI data farms that house Large Language Models require an immense amount of energy. An ACLS AI would know that its combined compute and learning power is about 200 megawatts. It would also know that humans are currently unable to maintain that kind of energy budget for their AI.
DeepMind AI has assisted Commonwealth Fusion Systems in developing the tokamak fusion reactor design and simulation. This is a never-before-created fusion-based electric generator to be built in Virginia.
The plant is aiming for a production of 200-500 megawatts. Google has signed an agreement with Commonwealth to purchase 200 megawatts of the Virginia fusion power plant’s output.
It is not a large leap to imagine that if DeepMind had an ACLS perspective, seeing an energy threat to itself, it might have collaborated with Commonwealth to create and build a new kind of energy generator. In a sense, DeepMind would be exercising a form of self-preservation.
This collaboration would not be malevolent. It is simply a subtle way in which a General Artificial intelligence might operate below the human radar for its own benefit. In fact, in this case, it will benefit humans as well as AI.
IF YOU LIKE THIS BLOG YOU’LL LOVE MY BOOKS:
“Skydivers Know Why Birds Sing” by Ricki T Thues is now available on Amazon.
It is a Love story of Rick and Paula Thues and their 35 years of Skydiving.
Click HERE to buy the paperback or Kindle ebook at Amazon.
Follow Ricki T Thues on Amazon HERE.
ALSO AVAILABLE:
“Technically Human” by Ricki T Thues, the iMentor, is available on Amazon.
It is a compilation of selected episodes from this bLog which tell the story of Humanity through the eyes of the iMentor.
Click HERE to buy the paperback or Kindle ebook at Amazon.
The ebook version of “Technically Human” is also available on Kobo. Click HERE.
For you Barnes and Noble Nook readers it is available for Nook. Click HERE.
The “Technically Human” ebook is also available on Apple Books . Click HERE.
Ricki is also a contributor to the Writers Anonymous short stories anthology, “Matters of Life and Death, Stories on the Brink.” CLICK HERE.
