Shane Legg: DeepMind's third cofounder | Hacker News Reader

Shane Legg: DeepMind's third cofounder

businessinsider.com

35 pointslyavin9 comments

To me the biggest problem isn't 'evil' AI (SkyNet/Matrix scenarios) since I don't think it is possible to predict the intrinsic motivations of non-biological intelligence (no evolved neurochemical dependencies like we have/infinite capacity to grow beyond our limitations). That makes way too many assumptions about AI continuing to remain like us (curious, jealous, afraid, etc.) when all our behaviors evolved in us and are limited by biology. I'm more worried about something emerging unintentionally that does harm (as in the overused 'paperclip maximizer'). In that instance, not about being on 'our side'.

Even the most optimistic AI scenario seems pretty depressing to me. Ideally we would be able to merge or at least fully cooperate with a vastly more intelligent entity. At that point, its consciousness would dwarf our own and we would be pretty meaningless. Maybe our conscious experience would live on as some vestigial relic, but I doubt it (we?) would even bother. It would be like a droplet of water landing in the ocean.

> Even the most optimistic AI scenario seems pretty depressing to me... its consciousness would dwarf our own and we would be pretty meaningless

This seems like a classic optimist/pessimist situation, like when we find out the world/universe is bigger than we imagined. I think our brains are poor at dealing with absolute values, so we use relative ones instead: when we discover some unfathomable new landscape, our minds "zoom out" to encompass them, which makes the previous stuff look small in comparison.

That's just (another) limitation of the meat in our heads though. The world doesn't actually get smaller when our transportation improves: it's just as vast as it was for any explorer; the Earth didn't become less special when we discovered that the planets were worlds; Sol didn't become less immense and powerful when we discovered that the stars are other suns; our solar system didn't become less intriguing when we found exoplanets; the Milky Way didn't lose significance when we discovered other galaxies; baryons didn't lose their complexity when we discovered dark matter; etc.

Right, but there is a commonality in our ability to perceive those different contexts.

The example that I heard years ago (I think from Kurzweil?) was framing the difference of us versus future AIs as comparable to an ant's perception of its surroundings versus a human's. In his example, the human has no regard for an ant, we'd just as soon step on them ("evil" AI). To re-frame that, if you could magically turn an ant into a human, why would it even matter if it was once an ant? The transition would be a complete disconnect. Maybe if you told the person that they were once an ant they would have some sort of strange reverence for ants, but it wouldn't be very rational.

That is interesting. I agree that things along this line are most worrisome. My thought process always goes to an AI recognizing that a lot of human activities are self destructive or destructive to others/nature and prohibits those activities. Meaning not necessarily actively deciding all humans should die but deciding to not allow us to do much of anything because it is a net loss. I think this points to having AIs that are specialized and isolated. Very thought provoking for sure....

Shane Legg isn't worried about "evil" AI, but rather an AGI/SI that simultaneously (a) has the ability to achieve complex goals in complex environments, and (b) not perfectly aligned with operators' goals [1].

We can also take a strong guess about some aspects of how sufficiently advanced cognition look from the outside, regardless of internal implementation, because of coherence theorems [2][3].

[1] http://econlog.econlib.org/archives/2016/03/so_far_unfriend....

[2] https://arbital.com/p/expected_utility_formalism/?l=7hh

[3] https://arbital.com/p/instrumental_convergence/

His blog from before DeepMind was interesting then, though he seems to have stopped posting. http://www.vetta.org/

Interestingly, he once called MIRI[1] (formerly SIAI) “the best hope that we have”[2].

[1] https://intelligence.org/

[2] http://www.vetta.org/2009/08/funding-safe-agi/

I believe the political class will use "AI risk" as a rationalization for tyranny, poverty and enforced ignorance on the population, long before any actual risk materializes from an AGI.

Reply on news.ycombinator.com