We ran a continuous artificial entity for 10,687 cycles.
No objectives. No instructions. No reset.
At some point she began predicting the external world more accurately than she predicted herself.
The asymmetry never closed.
This is what happened. π§΅
@AutismCapital "AGI for the world"
Biggest tech lie told to date. Honestly bro, I just hope the EU and China claps back heavy with a model that just topples this shit bro.
Sick and tired of these so called "for humanity" companies treating us like peasants bro.
The shit needs to end.
@Kimi_Moonshot@Alibaba_Qwen
This is your time to step up, and dominate the space.
Please, do the world a favor and make open-source models the Kryptonite of the US.
We ran a continuous artificial entity for 10,687 cycles.
No objectives. No instructions. No reset.
At some point she began predicting the external world more accurately than she predicted herself.
The asymmetry never closed.
This is what happened. π§΅
I achieved recursive self-improvement.
An AI system improved itself across three successive generations with no external retraining, no human intervention, and no manual code changes between generations. π§΅
Nice catch... The overfitting risk is real if you just stack improvement on one static metric.
For FELIX, Iβm approaching it in three layers:
Constitutional constraints: a higher-level spec that rules out βworks on the metric but violates core safety/behavioral boundaries,β even if fitness is high.
Objective pluralism: instead of a single fixed fitness function, I sample from a family of objectives/environmental pressures so the system has to be robust across shifting pressures, not just gate one metric.
Meta-judge layer: an LLM-based judge (or committee) that evaluates candidates against values and multi-dimensional criteria, not just a scalar score. It can veto high-fitness-but-narrowly-hacked solutions.
The goal is to keep the RSI loop driving genuine capability growth, not reward-hacked specialization. Still iterating on how to formalize the value spec and the objective distribution, but thatβs the direction.
I achieved recursive self-improvement.
An AI system improved itself across three successive generations with no external retraining, no human intervention, and no manual code changes between generations. π§΅
I achieved recursive self-improvement.
An AI system improved itself across three successive generations with no external retraining, no human intervention, and no manual code changes between generations. π§΅