Latest Everything: Bot shows signs of consciousness | Hacker News Reader

Latest Everything: Bot shows signs of consciousness

sixwoffers.blogspot.com

97 pointsHusafan34 comments

It's an interesting article but I feel a bit bait-and-switched by the title; this is a definition of consciousness so weak that it's not really very interesting. I don't think I accept a definition of consciousness that appears to boil down to "shows similar perceptual illusions as the human perceptual system".

The idea of parallel unconscious processes selectively passing important info to a "broadcast" network of neurons was new to me, and if that's the way consciousness works in the brain I think that's absolutely fascinating.

That the first half of the article never talks about a bot actually was a pleasant surprise. The human brain is way more interesting than a bot.

The phrasing is new to me too. But it seemed to boil down to a circuitous way of saying "subconscious filter".

Something I'd imagine many married couples are all-too-familiar with.

Why not something more easily understood by everyone:

"You are now aware that you are breathing".

I think the implementation here is quite interesting; pigeonholing it into the idea of a "filter" is probably missing the point.

I think you can be pretty sure all such titles are bait-and-switch. As soon as one isn't, I'm sure we will have noticed the Singularity in some other way before seeing articles describing it.

say it will post an article about itself. Or a bunch of trolling comments. Lets hope for the former.

AI doesn't imply Singularity. Self-improving AI implies Singularity. It's totally possible to have e.g. a neuromorphic AI of roughly human intelligence that doesn't cause an intelligence explosion.

Technically, AI used as a tool by humans to develop better AI may not be considered self-improving AI, but would have possibly the same effect, would it not? Self-improving AI might be an implied step after AI as tool, though.

I don't see how. The only thing that prevents human intelligence explosion is limited intake rate of data (and possibly hardware, of which an AI's would presumably be more reliable).

A neuromorphic AI could be constructed without understanding how the brain actually works. You could run that faster, but even with a billion years subjective you still might not be able to come up with ways to improve your own mind if you aren't smart enough in the first place. Of course, the intelligence threshold for self-improvement is lower if your mind is simpler...but the clearest (and fastest?) route to something you could reasonably call AGI may be dumbly reimplementing a human brain.

That's kinda like the old AI canard: once it's obvious how it works, it's no longer considered AI.

If you believe in the physicalism of the human brain and mind, then any definition of consciousness boils down to "a bunch of chemo-electrical impulses."

Fascinating study with some fairly straight-forward follow on experiments. I worry about these things getting the same bad rap that plagued 'fuzzy' logic because the investigators try too hard to describe them in human terms. So rather than saying "A new selection mechanism for data assimilation based on similar mechanisms that have been proposed in neural processes." they say "Its almost conscious!" which sets off a whole bunch of distractions which don't help in the underlying algorithm development or understanding its applicability to things like machine learning or process control.

Strange; the "button test" says humans take 200ms to press a button once a green light flashes.

I set that experiment up on Engineering Days at my college. Kids who put their faces right up to the light could press the button in under 30ms. The human detector response seemed to depend upon the amplitude of the neurological signal!

Sounds about the right amount of time for network lag.

Kids who put their faces right up to the light could press the button in under 30ms.

That seems really fast. I can't do better than 200ms on the online reaction time tests I've tried, and zooming the screen or putting my face closer doesn't seem to help.

30 ms doesn't seem possible. Check the Human Benchmark for statistics, the fastest ones are around 100 ms: http://www.humanbenchmark.com/tests/reactiontime/stats.php

Maybe light delay was not randomized and people were just predicting when it'd light up? I remember this kind of trickery being used in Jeopardy to beat Watson.

Another example is the Olympic sprint, where any reaction under 100 ms is considered a false start (if I read correctly): http://www.condellpark.com/kd/reactiontime.htm

100 ms is for an adult... The kids have smaller arms, hence shorter axons. Assuming that their conduction speed are similar to that of young adults, a shorter reaction time isn't surprising for young children.

[1] for adults, they decline with age, I don't know how they evolve during development

Yes, presumably there is a difference, but it's not that much. The axons that conduct motor signals to the skeletal muscles have a conductance speed of 80-120m/s [1]. In 70ms (100ms - 30ms), at 80m/s, such a signal would travel 5.6m.

[1] http://en.wikipedia.org/wiki/Axon#Motor

Neat. It makes sense that you'd have a fast-path for deflecting objects that are suddenly in your face.

Lots of interesting stuff here on factors affecting reaction time: http://biae.clemson.edu/bpc/bp/Lab/110/reaction.htm#Type%20o...

Your experiment must have been set up differently though - the fastest time they mention is 180msec for teenagers.

Thats really fast. Can it be result of many hours shooter computer games practice? Newly wired fast path for pressing button.

Edit: just run some reaction tests with about 0.25 avg. I see, ie. mentally identify, the signal at ~1/5 - 1/10th of that time. The rest of the time is material muscle reaction and actual mouse button movement down. So, with hand trained for the movement (my fingers can't even take a simple accord on guitar - compare to virtuosos playing musical instruments ) and good quick button i see how it can be down to ~50ms at least.

Honest question : how did you measure how long it took you to 'mentally identify' something?

it just an estimate. The motor reaction in my case is taking most of the time, around 0.2s. So the rest is the time it takes to see and identify. It is also pretty compatible with basic estimates : 10-20ms for several layers of neurons in primary visual cortex signal processing and about the same for several layers in the upper layer.

My guess is that when the kids' faces were right up to the light, their subconscious mind was actually doing the triggering because the light change was noticed in their peripheral vision. At my kung fu school, we train ourselves to avoid focusing our vision on the person we're fighting because acute vision is translated/filtered by the conscious mind; the tiny delays caused by the conscious mind are enough to get you killed. Acute->Conscious reaction times are far slower than Peripheral->SubConscious reactions.

You'd probably see similar results in your experiment if you had the kids stand close to the light but focus their vision just outside the perimeter of the light.

In case someone's interested in why this doesn't tell us much about what's so ridiculously mysterious about consciousness and subjectivity:

http://consc.net/chalmers/ [David Chalmer's website]

http://instruct.westvalley.edu/lafave/nagel_nice.html [Classic paper by Thomas Nagel]

"Functional consciousness" is a fine goal if it helps the bot behave in a way that is useful. I don't know how we would even know if we ever achieved subjective consciousness in a computer, except as it says at the end of the article: 'We can't even prove that each of us isn't the only self-aware being in a world of zombies. "Philosophers have been dealing with that for more than 2000 years," says Franklin. Perhaps we will simply attribute subjectivity to computers once they become sufficiently intelligent and communicative, he says.'

On the other hand, it's a little flaky that they set a parameter to try to get human-like response times. Statistical fitting.

We built it to hit this number, we're surprised when it does!

Cool calculation process, though.

I'm under the impression they fitted only the most basic operations, and the surprise was that the bigger picture followed too, as if the program and the human brain were following the same steps.

It's that "as if" that is a little bit hand-wavy. Fitting/overfitting is insidious in that it corrupts both the procedure and the data.

I'm not saying that's exactly what happened here, but I am saying that it's a _hunch_ that human brains are following the same steps. It might be a good hunch, but it's still _not_ a scientific statement with the weight of statistical backing.

Where's some source code for this?

There are related papers here, http://ccrg.cs.memphis.edu/papers.html, but I'd like to see some code.

This is reminiscent of Funk2, http://www.funk2.org/about/, it would be interesting to compare and contrast these approaches.

It's pretty interesting that what we think of as consciousness seems to arise from a combination of processing delays and ignoring external stimuli.

"The User Illusion: Cutting Consciousness Down to Size" is a great read if you're interested in exploring consciousness. A significant portion is devoted to discussing just how much information our bodies are taking in, versus how much reaches our consciousness.

http://www.amazon.com/User-Illusion-Cutting-Consciousness-Pe...

In a similar vein is Thomas Metzinger's work.

http://www.philosophie.uni-mainz.de/metzinger/

Bots show no such thing...

Reply on news.ycombinator.com