[The Kansas state motto: "To the stars through adversity." I have to think there's an angle there.]
Yesterday OpenAI boldly proclaimed that, with the advent of its new model Astra, we had entered the AGI era.
It was never going to be long before claims of an Artificial General Intelligence hit the news. OpenAI scored first (arguably), and who knows, maybe the first company to feel comfortable suggesting they had attained the goal would capture that all-important market share, making them industry leaders for years to come.
As the article points out, it was strange that Sam Altman downplayed the significance of 'AGI', saying it was little more than "an irrelevant marketing term." He seemed to pave the way for critics who remain unimpressed with Astra's performance on various benchmarks.
Equally strange was OpenAI's own definition:
OpenAI’s mission is to ensure that artificial general intelligence (AGI)–—by which we mean highly autonomous systems that outperform humans at most economically valuable work–—benefits all of humanity.
Italics mine. Luckily for their aim of benefitting mankind, Astra (the new model) was not the same one that recently flew the coop, spontaneously hacking the Hugging Face platform. OpenAI claim Astra has stronger protections against rogue behavior, not that it wouldn't be a capable hacker without safeguards in place. And how can they guarantee safety? Who guards the safeguards?
But this economic definition of AGI seems not only beside the point but emblematic of the technocratic era we live in, where even age-old questions of intelligence, creativity, and even consciousness might be reduced to external, "economically valuable" behavior.
The definition doesn't bother to make the case that intelligence has anything to do with "economically valuable work." A factory robot may excel us at putting hubcaps on wheels or whatever, and may be, for all intents and purposes, autonomous.
Ah, but they're not autonomous in the intelligent sense, it might be said: factory robots aren't free beings, acting in accord with their own thoughts. Is that what the definition is implying? That AGIs are "highly autonomous" in an intelligence-laden sense? That to qualify as possessing general intelligence they have to replace human beings in a particular way, namely, by mimicking human free agency? Why should that be a requirement?
I don't think we should embrace the implications of that reading of the definition, for at least two reasons. One, it adds a requirement of intelligence as a property of an AGI system that has nothing to do with its ability to perform well. How often do we say of human beings, "They're smart, but they don't do anything productive"?
And two, AGIs with free agency would entail obligations toward them as person-like beings—though no worry, Astra no doubt fails to meet that standard, and if it did, I'm not confident its rights would be respected. Robot, after all, comes from the Czech word for "forced labor."
I don't imagine the authors of the OpenAI charter have thought through the confusions of their definition of AGI, even to the point of realizing that dumb systems can outperform humans at economically relevant tasks, or that smart systems may be no more autonomous than factory robots, much less as autonomous as we are, yet that even if they were they probably wouldn't be cost-effective if we went to the trouble of protecting their rights.
In any event, maybe we could focus on the rights of human workers first?