How can we build human-like artificial intelligence, that is, an AI with human-like desires and intelligence (such as self-awareness, communication skills, sense of purpose or adaptability to different environments)?

See, for example, Diagram of Intelligence Network or System and Google a step closer to developing machines with human-like intelligence for more details.


That's a multifaceted question. The referenced articles are conjectures much like the hundreds of thousands of other interesting and carefully considered works on the net and in libraries. Whether any of them bring us down a dark ally or directly toward the objective the question implied is unknown.

The array of common human desires are suggested by derivatives of Abraham Maslow's hierarchy of needs. Self awareness is not among them. From a cognitive science perspective, self awareness is not a desire but rather trait of consciousness. The plausible explanations for its presence ranges between the extremes of a completely useless artifact of mental evolution to a deliberate infusion from some supreme entity, and everything in between.

The other desires, traits, and capabilities listed in the question are easier to discuss in a way that is backed by existing cognitive science, psychology, neuroscience, and self-evident human experience.

Regulating and prioritizing may an important element in intelligence beyond problem solving. For instance, a sense of purpose is only a desire as one moves up the hierarchlistedny of needs. People running from a Mammoth have no interest in purpose. Regarding competent communication skills, what is and what is not sufficiently competent would require a definition, and not all humans may qualify all of the time.

The sustainable ability to adapt is perhaps a super-set of intelligence in some ways, as in the case of the prioritization that exhibits the hierarchy of needs above and as in the case of DNA based adaptation. In other cases adaptability may be a subset of intelligence, in the case of applying taught facts and logical deduction to select an optimal choice without the benefit of empirical observation or even trial and error.

The intelligent agent approach may produce something useful, but it is not likely the path that led to a human brain and may not produce anything like it. We already understand what placing a number of such agents on a problem queue produces. A time sharing computer is exactly of that architecture. Parallel computing is not a significant deviation from that model.

None of the current extensions of automation architectures from Bell Labs, Harvard U, MIT, or other historical centers of electronic brain and robotic research and development have produced anything close to the varied integrated capabilities of a biologically healthy, socially aware, motivated human brain functioning as one of the controlling elements within a human body.

We do, however, have machines with human-like capabilities (simulations of task that were once manual), and such devices sometimes out-perform humans. But none of the existing automation is so integrated that they would blend in with humans. Nonetheless, many already have careers.

Single capability machines with job security:

  • Calculators
  • Automated hoppers
  • Anti-aircraft weaponry
  • Mail sorting machines
  • Speech recognizing devices
  • Speech synthesis

A general purpose computer, connected to a well designed, high end robot could do any combination of these and other things in a time shared way. Several networked computers could provide parallel execution of these tasks and even make basic decisions on how to prioritize the various tasks within the computer network.

However, the type of integration of sophisticated translation of arbitrary generalized goals to action leading to success and the imagination and prioritization of such sophisticated objectives is not yet available or at least deployed for general home, commercial, or industrial use.

Some available machines already accomplish well defined goals, but only within a very specific solution domain. The interaction of various capabilities mentioned is still largely undiscovered at anything close to the level of detail necessary to produce synthetic capabilities of the same caliber.

Whether it is possible to do so, Marvin Minsky voted yes, when he popularized his colleague's characterization of the human brain as a meat machine. Whether consciousness transcends neural electro-chemistry has yet to be proven or disproved in either the mathematical sense or the laboratory sense of the term Proof.

Interestingly, a form of self-consciousness, some layer of it, seems to require the ability to make a noise and hear it. Another layer beyond that seems to require the ability to move a muscle and view the results in a reflection. These and others form an interesting set of phenomena.

Perhaps intelligence requires embodiment. One of the episodes of Sarah Conner Chronicles implied this, and the theory is an old one recognized independently by several researchers. The question author implied this vaguely self-evident reality in the mention of action in the question.

  • $\begingroup$ FYI: "A New Approach to Understanding How Machines Think", quantamagazine.org/… $\endgroup$ – Dave Jan 11 at 19:47

Well, the low-hanging-fruit answer is that you simulate a human being - brain, hormones, everything. We should have the computing power for that to be feasible by 2040 or so.

Building up self-awareness from first principles on a different foundational technology platform could be a bit more difficult!


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.