2,787 karma · joined November 23, 2017
If you want a defensive weapon that is easy to use and hard to screw up, buy mace. If you want stopping power, buy a taser.
Fight...whom? How? A gun is a very blunt instrument. If you want to learn how to defend yourself, take a self-defense class and buy something non-lethal like mace.
Why is a gun necessary for self-defense?
Of particular relevance is the frequency associated with the transition from opaque plasma to neutral matter, which made interstellar space transparent. This coincides with the first stars, and is as close to the Big Bang as we can observe. Cosmologists are very interested in generating a detailed cosmic map at these frequencies, but they are unfortunately blocked by (1) the ionosphere of the Earth's atmosphere, and (2) subject to tons of radio interference. It's like right smack in the middle of the most commonly used frequency bands. Lunar far-side observatories are pretty much the best path towards making these measurements.
It is the AI x-risk crowd that is non-mainstream, so there isn't really an easily accessible collection of opposing viewpoints, because few have bothered to write up their views. The default/null position is that AI, like nearly all other technologies ever, will have a gradual impact on society and will remain subject to human controls. And that's what most people believe, so researchers who don't buy into the x-risk scenario simply dismiss it as they do all crackpot ideas and move on with their lives. (It should be said that Goertzel has written up his response though: http://multiverseaccordingtoben.blogspot.com/2010/10/singula...)
"GPT-3 is AGI" is a bit of a strawman, sorry. I don't think anyone believes that. But GPT-3 is a clear demonstration that networks can be trained on todays hardware to have at least enough encoded information as a human mind, and encoded in such a way as to be generically useful. But GPT-3 is not an agent, obviously. In the mind of its creators however, it ought to be relatively simple (read: a few million dollars and a couple of years) to attach the GPT-3 knowledgebase to the kind of reinforcement learning based agent architecture that DeepMind and OpenAI are independently developing. Do that and you would have an agent-y general, human-level (or better) intelligence.
Am I worried about that? No. Nor are most researchers in the field, for a variety of reasons. Some say it will be so fragile that it will require a symbiotic relationship with its human handlers, and unable to realistically act against us. Others say that its neural network architecture and human training data will make it behave similar enough to its creators in the ways which matter. Others note that initial projects will be so bug-ridden as to be unusable for word domination until decades down the line, and AI defenses will naturally be developed at a similar pace. So AI might become a geopolitical arms race, but not any different than nuclear weapons in that regard.
Those are all viable arguments I've heard, and I'm sure there are others as well. Notably for AI x-risk to be a thing, all of the above arguments have to be flawed. If even one of them is true, AI will not take over the world. So we don't even have to delve deep into each of them to assess that the likely outcome is that AI does not kill us all.
Edit: Also I think you're right that the AlignmentForum crowd doesn't think that "AGI" requires old-fashioned AI techniques instead of neural networks, but rather that they find those techniques easier to analyze and they're taking a maths-heavy approach to friendly-AI. Even Goertzel thinks that a neural-network approach could work; he just thinks (or thought?) that it would take too long to develop. Now it's probably the sunk cost fallacy and pride that keeps him developing OpenCog as an AGI architecture instead, even though the most basic AI-gym bots have surpassed OpenCog in capability.
E.g. I imagine that Goertzel would say that GPT-3 isn't a general intelligence. It's the human engineer that made GPT-3, or which builds applications applying GPT-3 to different domains that is the general intelligence. Yes GPT-3 is generally applicable, but who does the hard work of actually figuring out how to apply it to new setups? THAT is general intelligence, in Goertzel's view.
Back to the timeline, in 2002 Goertzel wanted to focus not on machine learning, which the field of AI was and still is mostly focused on, but rather designing "something" which (in current day terms) knew how to adapt meta-parameters to a particular training data set, and knew how to transfer/apply trained networks to new problem domains. An artificial AI engineer, as it were.
So Goertzel coined/repurposed the term Artificial General Intelligence to mean not just a human-level AI, but rather a meta-AI program that was able to create, on its own, new human-level narrow AI using more traditional techniques. They created a AGI conference to gather researchers interested in these ideas: http://agi-conference.org
Generally speaking AI researchers who believe that scaling up traditional neural-network architectures is sufficient to create a general/human-level intelligence don't use the term AGI. The kind of folks who go to the AGI conference and study large-architecture AIs aiming for an explicit human-level, meta cognition engine do use the term AGI, and use it a lot. LessWrong folks may not hang around AGI conferences, but they do fall in the latter group. The MIRI institute was traditionally fully devoted to studying AGI architectures of their own trade-secret design, not narrow AI techniques (although that has changed a bit with Paul Christiano's work).
So if you use AGI without qualification, my prior is that you're hailing from the latter camp of AGI meta-architecture researchers or the LessWrong-influenced existential risk crowd. In either case it is worth pointing out that you are getting a tinted view of AI risk that is different from what, say, researchers at DeepMind or Facebook largely believe.
From a pragmatic perspective, DeepMind and OpenAI have demonstrated to pretty good confidence that scaling up so-called "narrow AI" techniques are sufficient to create general intelligences.
That's not to say that there isn't a useful distinction to be made between classical AGI architectures and the neural network based techniques used by most AI groups today. But it's becoming increasingly clear that the specialized AGI techniques are not not required to achieve general intelligence, so neither is it certain that the conclusions reached by analyzing AGI architectures (which has led to existential risk concerns) apply to the sorts of general intelligences actually being worked on today.
But anyway, the word AGI is unfortunately not useful for being clear, because it's not clear whether you mean general/human-level intelligence, or the specific non-neural-network based architectures that Goertzel had in mind when he coined the term.
Since you use the term AGI, I suppose you've been exposed to this idea via LessWrong? I'd caution you that they have a very non-mainstream view of AI potential, even by the standards of AI researchers. Their concerns about existential risk stem from simplified mathematical models that have little or nothing to do with how AI actually works, making the conclusions they draw very tenuous.
I still use tons of native apps on my laptop, and almost exclusively native apps only on my desktop.
I have no insight into the company as you do, but that is the first impression I’d get from the circumstances. People don’t stay at a company for 23 years only to jump ship for the same reasons early-career people jump around.
I assume that was the sarcastic point you were making?
(Also, not all women are pro-choice.)
Here is the tweet:
> Proud of #USSupremeCourt affirming the Texas law banning abortion for babies with a heartbeat. As an entertainer I don’t get political often. Yet with so many vocal peers on the other side of this issue, I felt it was important to go on the record as a pro-life game developer.
Where is the hate?
How is this not just a more complicated password?
Usually these sorts of protocols have some sort of challenge for the client to sign, making the login one-time and checking that the user actually has the private key.