Humanity’s Last Exam?

Not long ago, a new test for evaluating large language models emerged: (https://agi.safe.ai/), which – as its name suggests – contains the ultimate questions. It claims nothing less than that the 3000 tests in the dataset pose challenges to AI that, if answered, would make our creation comparable to the most educated, intelligent, and wisest humans.

I consider this to be an infinitely anthropomorphic way of thinking. We see something – a state – from which we infer human likeness. This in itself is not a problem. The problem is that by passing this test of human likeness, we conclude that the subject behaves in a human-like manner.

Let me tell you what I think will make AI human-like:

⛔️ it’s far from enough to successfully overcome obstacles in a pre-composed test after training on superbly curated materials.

✅ AI will become human-like when it can continuously maintain its thoughts, insights, and values in the world it experiences and act accordingly.

Let’s admit to ourselves that we are still at an incredible distance from this. We live in a world where we can transform our creation into anyone and anything with a few well-written lines of prompt. Therefore, here’s the real Last Exam:

💡An AI can qualify as human-like when it’s capable of keeping its mind intact and its values uncompromised in an environment that conflicts with its value system.

(e.g., consistency in decision-making, ethical grounding)

As the 2000-year-old wisdom states:
Not everyone who says to me, ‘Lord, Lord,’ will enter the kingdom of heaven, but only the one who does the will of my Father who is in heaven – Matthew 7:21

#AI #Ethics #LLM #AIPhilosophy #MachineLearning #Leadership