We don't have sci-fi AI, and in any case "this thing is now possible, therefore [much more difficult thing] will be possible soon" is not a rational thought.
Current frontier models are absolutely sci-fi AI by any measure that existed up until the late 2010s. If you showed a current frontier agentic AI, with bidirectional speech, tool use etc. to a person from 2010, they'd say it can't be real, there must be a person inside that mechanical turk.
Sci-fi is Lt. Cmdr. Data, HAL, Skynet, Culture Minds...
> Current frontier models are absolutely sci-fi AI by any measure that existed up until the late 2010s. If you showed a current frontier agentic AI, with bidirectional speech, tool use etc. to a person from 2010, they'd say it can't be real, there must be a person inside that mechanical turk.
It wasn't a natural language chat interface that makes stupid mistakes a noticeable fraction of the time and can execute command line programs.
The current state of things is closer (not close to by a long shot) to Fred Pohl's "AI" programs from his novel Gateway[0] than it is to the above examples.
But even those are decades (at least) beyond what we have now. What we have now is the foreseeable (from at least 30 or 40 years ago) evolution of what was once called "expert systems."[1]
I guess you had to be there. I certainly saw it coming.
>Expert systems worked on quite different principles.
Actually, not so much. The inference engine has changed, but otherwise it's same shit, different day.
They did interviews with experts and tried to do domain modeling, mapping the expertise to a formal symbolic logical description, ontologies, relations etc.
Both the knowledge base and the inference engine changed. They both changed to a unified machine learning paradigm. It's like you throw away your axe's handle, then the head, and then you grab a chainsaw. It's "the same thing" to that same degree.