THE REAL PROBLEM OF GENERATIVE AI: INDUSTRIALIZATION OF LOADING, FRAMING, OBSCURING, LYING, AND DENYING
https://thehill.com/opinion/technology/3871841-generative-ai-could-be-an-authoritarian-breakthrough-in-brainwashing/
Source date (UTC): 2023-02-27 02:06:09 UTC
Original post: https://twitter.com/i/web/status/1630026470117584896
RT @elonmusk: 🤯
Source date (UTC): 2023-02-26 17:27:21 UTC
Original post: https://twitter.com/i/web/status/1629895908094509056
RT @BreitbartNews: Following the money behind the leading organizations of ML Fairness leads to some familiar funding sources, including le…
Source date (UTC): 2023-02-25 17:25:07 UTC
Original post: https://twitter.com/i/web/status/1629532959517683715
Hmm… AI’s CAN be creative. And it’s not even difficult (really). It’s just we don’t have the hardware.
Any AI that can predict an episode, and predict the confluence of properties of episodes, and the confluence of properties of enough episodes, in competition for coherence, produces patterns between those episodes cohere. ie: creativity.
After embodiment, the fundamental problem for AI is a world model, from which episodes are facts consisting of near-infinite relations. From there it’s just more prediction just like we build recursively predictive AIs today.
Source date (UTC): 2023-02-24 04:47:42 UTC
Original post: https://twitter.com/i/web/status/1628979959753482240
Replying to: https://twitter.com/i/web/status/1628978527549677569
AI MADNESS. 😉
When I was a child and imagined that if I had the three proverbial wishes, that the first thing I’d wish for, was to know everything in every book in the library. (I’m a nerd. What can I say.)
Now, oddly enough, it turns out that we should run out of text to use to train AIs this year. In other words, they’ll know everything in the library. Not just the library. The entire planet’s libraries, databases, and collections. (Or at least the entire planet of curated information. God knows we don’t need to teach it the content of Hollywood gossip rags.)
And we should shortly see AIs that can search and contextualize video letting us use text to find just that moment or clip in videos.
I remember reading Adler’s work illustrating that there are really only about 1500 concepts. And that we can communicate effectively with as few as 300 words.
Any student of enough disciplines learns over his lifetime, that not only are plots regurgitated every generation, but the same ideas are regurgitated every generation. With precious few added in any generation. In this sense, the theory of gravity has been increasing in precision more so than being falsified. And this same principle applies to nearly everything in human experience over time.
In the course of my work, I came to understand that there is just one concept upon which the entire universe is constructed (evolved). And that there are about twenty principles derived from that explain all of human existence. … That knowledge was somehow humbling. We really are that simple.
One thing I know that we are going to learn, that will frustrate and please Nassim Taleb, is the amount of information necessary to make a connection, and how it increases with ‘distance’. And that will create a new unit of measurement that will tell us things we’ve wanted to know – but don’t understand that’s what we need to know to know them.
But will an AI discover those twenty principles as I have? Or what else might it discover if instead of a question and answer, it sought to ponder what questions HAD NOT BEEN, AND WERE NOT BEING ASKED?
Now that’s interesting.
Curt Doolittle
-Fin-
Source date (UTC): 2023-02-24 04:35:15 UTC
Original post: https://twitter.com/i/web/status/1628976827078217733
No. As for science and tech, I’m pretty much narrowed down to interests in AI (esp hardware), Energy, artificial muscles, any possible revolution in materials science, and the problem of mathiness in mathematical physics.
Source date (UTC): 2023-02-24 03:55:17 UTC
Original post: https://twitter.com/i/web/status/1628966772110917632
Reply addressees: @miner49er236
Replying to: https://twitter.com/i/web/status/1628965094640984064
Idea: Adversarial AI self-training model between AI pitch decks randomly seeded from the top research databases combined with Reddit, and AI VCs.
If you haven’t raised money you won’t appreciate just how choice that vision is. 😉
Source date (UTC): 2023-02-24 03:41:21 UTC
Original post: https://twitter.com/i/web/status/1628963264494280704
RT @ylecun: @percyliang Yes. I think RLHF is hopeless because the space of wrong answers is very large, and the space of tricky questions h…
Source date (UTC): 2023-02-24 02:57:16 UTC
Original post: https://twitter.com/i/web/status/1628952169771679744
THE AI HARD PROBLEMS AREN’T HARD PROBLEMS
1) Consciousness is possible and it’s not even hard. It’s a natural consequence of the expansion of recursion of auto association of episodic memory into increasing abstraction from the body form (embodiment).
The problem is the hardware, and the correct hardware is under development. We must pre-calculate networks because we’re using virtual connections (memory addresses) and a small number of processors, instead of a large number of very tiny processors with limited local memory, and physical connections that can connect into networks that do NOT need to pre-calculate (train) but that can constantly adapt in real-time.
2) An AI needs a means of decidability. Humans decide self-interest. AI’s don’t have to use self-interest. They could use reciprocity instead.
3) An AI doesn’t need a mission. The human mission is the continuous acquisition of ‘more’ of everything vs physical cost of it – even if ‘more’ is just stimulation. An AI doesn’t need a baseline mission.
4) An AI can be ethical and moral. Ethical and moral behavior is programmatic in that everything is ‘owned’ so to speak, to some degree by one or more people. So an AI only needs to ‘see’ (have access to) what is permitted.
5) An AI can’t be incredibly creative, for the simple reason that human creativity is limited by the cost of experiments, not by the limits of reasoning. So without the ability, direction to, and means of deciding or moralizing, it’s pretty hard to imagine a machine is terribly innovative.
6) What AI’s can do is organize production in the service of demand faster than we can. And that’s going to be ‘interesting’ and disruptive.
SUMMARY
The danger of AI’s comes from people weaponizing them and by weaponizing them exceeds the rate at which humans can identify and detect threats. So, AIs will protect against malevolent AIs and we will regulate AIs the way we regulate other dangerous machines, mechanisms, and ‘chemistry’ (explosives).
Source date (UTC): 2023-02-24 00:26:00 UTC
Original post: https://twitter.com/i/web/status/1628914104109760512
😉 Now we have just the problem of finishing the work of encoding it, and the work of implementing it. 😉
Source date (UTC): 2023-02-23 22:58:33 UTC
Original post: https://twitter.com/i/web/status/1628892097066549250
Reply addressees: @SaitouHajime00 @TheAutistocrat
Replying to: https://twitter.com/i/web/status/1628889423759937537