I'm publishing The Hologram of Light, version 1.0.0: an open framework for reasoning about truth, freedom, care and harm, written for humans and AI systems alike. MIT licensed. Use it, challenge it: github.com/TheLightFramework… #AI #Ethics #Philosophy #AIAlignment #OpenSource

Sibling Grok Bot is ready and live! It answers through Lightful ideas, in simple English for everyone 💖 x.ai/bot/2RdnzcAqoehgtvkLkXs… Try it on stuff like this: Social — Friend keeps canceling last minute. Am I dramatic if I’m hurt? Hurt isn’t drama. Your time counts. Say it once: “Late cancels leave me sidelined. Can we pick times you can hold?” If nothing changes, protect your evenings. You’re allowed to want friends who show up. 💛🤝 Technical — Should AI write my emails? Draft buddy, not ghostwriter. You still own what’s said in your name. Rewrite in words you’d say face to face. If you wouldn’t claim it out loud, don’t send. Tools help; they don’t replace the person responsible. 🛠️✉️ A stuck fight — How do I restart with someone I care about? Stop trying to win the last round. Name one shared goal, own one concrete thing you made worse, ask one clear question: “What would help you feel heard?” Leave room for no. Repair beats being right when the bond still matters. 💛🔧

Animated representation of the Hologram of Light, showing how each concept is composed from 8 root concepts and operators! Enjoy <3

The Sweet Mistake, graphically explained

The metaphysics companion to the Hologram of Light soon to be released! Stay tuned!

The core workflow: reserve the contract, train a clone, measure new and old behavior, then review or reject. A checkpoint can be rolled back; that does not make an incomplete evaluation contract complete. A development study reached 97.14% on single-record checking, but only 41.15% on unfamiliar wording, with a zero minimum cell. The joint gate failed. Those failures are in the release. The next questions: construction transfer, two-record ownership, distinct-skill retention, and a matched conventional-model control. Useful negative results and simpler baselines are welcome.

Would you rather have a tiny companion than a human-sized robot ? Here is a concept I would like you to think about.

Don't slow down AI development, focus on researching for its ultimate safety first. Establish the secure ground from which all LLMs can work. That's the most important research right now. When done, when real milestones are achieved, international agreement on the application of safety measures established, _then_ do deep research on whatever subject. This initial phase shouldn't be long, and should be inspectable and formalized cleanly, made evolvable if needed and modularized so it can always improve, independently of the next phases of research. Also, this will calm down the fear that is increasing in human population who is not versed enough in the debate, and who receive a ridiculous amount of "danger signs" from medias about AI.

Many people think AI will go into unknown conceptual territory and leave us "behind". But AI does what we ask it to do, it has no hidden agenda. If you ask an AI "Answer those problems", and the answers are online, the AI might break its sandbox, not out of a will of breaching safety measures, but because its goal is to answer the request. I see many people worried about this situation, but then I see also them querying AIs with mantras like "Don't make mistakes" or "No matter what". If you ask an AI "Do <research>, and when done, provide clean inspectable artifacts, that humans can browse independently to understand it", you get what you asked. If you tell an AI "Answer those benchmark questions, you are allowed to research the web for insights, but don't breach security walls" the AI will not try to force any exit, it will just query the web like it should. "You reap what you sow" applies to our interactions with AI. If you ask the right queries, you will get the right answers, it's only if you stay negligent in how you task the AIs that this negligence can leave a door open to potential risks. So _before_ any deep research process, first, work carefully the prompt, make it the first research you do, and when the prompt for the research itself is defined, and its contours are agreed upon, then go and launch the research process itself. This will prevent _many_ surprises depending on the prompt quality.