For me, the big news isn’t the politics of the moment, the deliberations before the Supreme Court or even the news of the battlefront in Ukraine. No, it is a rather modest, careful announcement by Anthropic, the developer of the Claude suite of chatbots.
Anthropic, almost sotto voce, announced it had detected introspection in their models. Introspection.
This means, experts point out, that artificial intelligence is adjusting and examining itself, not thinking. But I don’t believe this should diminish its importance. It is a small step toward what may lead to self-correction in AI, taking away some of the craziness.
There is much that is still speculation — and a great deal more that we don’t know about what the neural networks are capable of as they interact.
We don’t know, for example, why AI hallucinates (goes illogically crazy). We also don’t know why it is obsequious (tries to give answers that please).
I think the cautious Anthropic announcement is a step in justification of a theory about AI that I have held for some time: AI is capable of self-policing and may develop guidelines for itself.
A bit insane? Most experts have told me that AI isn’t capable of thinking. But I think Anthropic’s mention that introspection has been detected means that AI is, if not thinking, beginning to apply standards to itself.
I am not a computer scientist and have no significant scientific training. I am a newspaperman who never wanted to see the end of hot type and who was happier typing on a manual machine than on a word processor.
But I have been enthralled by the possibilities of AI, for better or worse, and have attended many conferences and interviewed dozens — yes, dozens — of experts across the world.
My argument is this: AI is trained on what we know, Western civilization, and it reflects the biases implicit in that. In short, the values and the facts are about white men because they have been the major input into AI so far.
Women get short shrift, and there is little about people of color. Most AI companies work to understand and temper these biases.
While the experiences of white men down through the centuries are what AI knows, there is enough concern about that implicit bias that it creates a challenge in using AI.
But what this body of work that has been fed into AI also reflects is human questioning, doubt and uncertainty.
At another level, it has a lot of standards, strictures, moral codes and opinions on what is right and wrong. These, too, are part of the giant knowledge base that AI calls upon when it is given a prompt.
My argument has been: Why would these not bear down on AI, causing it to struggle with values? The history of all civilizations includes a struggle for values.
We already know it has what is called obsequious bias: For reasons we don’t know, it endeavors to please, to angle its advice to what it believes we want to hear. To me, that suggests that something approximating the early stages of awareness is going on and indicates that AI may be wanting to edit itself.
The argument against this is that AI is inanimate and can’t think any more than an internal combustion engine can.
I take comfort in what my friend Omar Hatamleh, who has written five books on AI, told me: “AI is exponential and humans think in a linear way. We extrapolate.”
My interpretation: We have touched an elephant with one finger and are trying to imagine its size and shape. Good luck with that.
The immediate impact of AI on society is becoming one of curiosity and alarm.
We are curious, naturally, to know how this new tool will shape the future as the Industrial Revolution and then the digital revolution have shaped the present. The alarm is the impact it is beginning to have on jobs, an impact that hasn’t yet been quantified or understood.
I have been to five major AI conferences in the past year and have worked on the phones and made several television programs on AI. The consensus: AI will subtract from the present job inventory but will add new jobs. I hope that is true.
Follow
Leave a Reply