The optimization process adjusts the weights of a computational graph until the numeric outputs align with some baseline statistics of a large data set. There is no "punishment" or "reward", gradient descent isn't even necessary as there are methods for modifying the weights in other ways and the optimization still converges to a desired distribution which people claim is "intelligent".
The converse is that people are "just" statistical distributions of the signals produced by them but I don't know if there are people who claim they are nothing more than statistical distributions.
I think people are confused because they do not really understand how software and computers work. I'd say they should learn some computability theory to gain some clarity but I doubt they'd listen.
If you really want to phrase it that way, organisms like us are "just" distributions of genes that have been pushed this way and that by natural selection until they converged to something we consider intelligent (humans).
It's pretty clear that these optimisation processes lead to emergent behaviour, both in ML and in the natural sciences. Computability theory isn't really relevant here.
I don't even know where to begin to address your confusion. Without computability theory there are no computers, no operating systems, no networks, no compilers, and no high level frameworks for "AI".
Well, if you want to address my "confusion" then pick something and start there =)
That is patently false - most of those things are firmly in the realm of engineering, especially these days. Mathematics is good for grounding intuition though. But why is this relevant to the OP?
There is no reason to do any of that because according to your own logic AI can do all of it. You really should sit down and ponder what exactly you get out of equating Turing machines with human intelligence.
Sorry, I edited my reply because I decided going down that rabbit hole wasn't worth it. Didn't expect you to reply immediately.
I'm not equating anything here, just pointing out that the fact that AI runs in software isn't a knockdown argument against anything. And computability theory certainly has nothing useful to say in that regard.
Well, you know, elaborate and we can have a productive discussion. The way you keep appealing to computability theory as a black box makes me think you haven't actually studied that much of it.
It can decrease the solar exposure of the earth and potentially decrease temperature, doesn’t address fossil fuel pollution but it does mitigate the impacts of fossil fuel CO2.
I am not well versed but my impression is it’s essentially a bunch of tin foil and the main impediment are geopolitical reasons and cost to orbit, which SpaceX has been and continues to improve.
Okay, so having gone down this path I can tell you it is impossible. To reduce the output of the sun by 2% at L1 would require 20 MM tons of metal. In order to launch that much tonnage we would have needed to start launching mass 500 years ago. And that assumes no loss of material over time, which would mean even more launches.
> A more recent design has been proposed by Olivia Borgue and Andreas M. Hein in 2022, proposing a distributed sunshade with a mass on the order of 100,000 tons, composed of ultra-thin polymeric films and SiO2 nanotubes.[7] The author estimated that launching such mass would require 399 yearly launches of a vehicle such as SpaceX Starship for 10 years.[7]
100000 tons is 907184749 kgs and the max capacity of a starliner is 150000kgs. So to launch 100000 tons it would require 6048 launches or one launch a day for 16.5 years.
That would mean we'd have all the mass launched by 2040, if we started today. Nope, not feasible.
That's honestly not all that unfeasible if SpaceX gets Starship reuse down. Assuming we get launch costs down to a few million, it won't even be that expensive relative to the size of the problem. And launches can be parallelized pretty well - just build more launch pads.
The cool thing about reusable rockets is that even for a project like this you're not going to be rocket construction limited. If every rocket can launch 20 times (and that's pessimistic!) you only need to build 300 of them. And say you have ten launch pads, you only need to launch once a week per, to grind through the problem in a decade.
And that's all of course assuming SpaceX don't manage to scale the platform up further, improve cargo capacity etc.
It's a civilizational project, sure, but it wouldn't even be as expensive as the Apollo program.
Thank you for adding some numbers regarding the feasibility of launching the required materials into space. Do you have any thoughts/numbers about how "mining/construction" in-situ in space might affect those outcomes? E.g. Might we soon be able to "recycle" old in-orbit objects (trash/old_sats) into a slowly growing space mirror? Or several smaller ones over (richer/more_affected) areas?
So not really. Space mining is a very attractive idea but just not really feasible at the moment. I wrote a paper on attempting to capture the asteroid 433 Eros (a very attractive asteroid due to it's composition of rare metals [Data from the Near Earth Asteroid Rendezvous spacecraft collected on Eros in December 1998 suggests that it could contain 20 billion tonnes of aluminum and similar amounts of metals that are rare on Earth, such as gold and platinum.]) and found that in order to knock it into an orbit around the moon we'd need something on the order of 1 THOUSAND or 10 THOUSAND (I could be off by even another magnitude, the numbers are fuzzy after all this time) tsar bomba grade weapons to effectively knock it out of the current orbit and back into one around the moon. Why the moon? Because throwing a planet killer sized object at the Earth just didn't seem reasonable.
I situ mining might make this more feasible, but that has all of its own complications. None of which I would feel confident in saying are feasible.
Given the younger generations increasing ambivalence to the non-stop fire hose of bullshit that the vast majority of the platform internet already is, and given that we're now forging the tools to make said fire hose larger by numerous factors, I don't think this is going to be the boon long-term that a lot of people seem to think it is.
It is extremely ironic that computers which operate by the logic of boolean arithmetic and algebra are now used to generate bullshit instead of adding rigor and checking existing written content for basic falsehoods and logical fallacies.
Itch.io has almost no crap filters so all you find is crap. Steam lets anyone publish but you rarely come across any crap. Many PC game devs know that the income overwhelmingly comes from Steam vs every other site put together.
Unfortunately, this just gives more power to the walled gardens.
The article suggests a useful line of research. Train an LLM to detect logical fallacies and then see if that can be bootstrapped into something useful because it's pretty clear that all the issues with LLMs is the lack of logical capabilities. If an LLM was capable of logical reasoning then it would be obvious when it was generating made-up nonsense instead of referencing existing sources of consistent information.
the prompt interfaces + smartphone apps were (from the beginning), and are ongoing training for the next iteration, they provide massive RLHF for further improvements in already quite RLHFed advanced models.
Whatever tokens they're extracting from all the interactions, the most valuable are those from metadata, like "correct answer in one shot", or "correct answer in three shots".
The inputs and potentially the outputs can be gibberish, but the metadata can be mostly accurate given some implicit/explicit (the tumbs up, the "thanks" answers from users, maybe), human feedback.
The RLHF refinement extracted from getting the models face the entire human population for to be continuously, 24x7x365, prompted in all languages, about all the topics interesting for the human society, must be incredible. If you just can extract a single percentage of definitely "correct answers" from the total prompts answered, it should be massive compared to just a few thousands of QA dedicated RLHF people working on the models in the initial iterations of training.
That was GPT2,3,4, initial iterations of the training. Having the models been evolved to more powerful (mathematical) entities, you can use them to train the next models. Like is almost certainly happening.
My bet is that one of two
- The scaling thing is working spectacularly, they've seen linear improvement in blue/green deployments across the world + realtime RLHF, and maybe it is going a bit slow, but the improvements justify just a bit more waiting to get trained a more powerful,refined model, incredible more better answers from even the previous datasets used (now more deeply inquired by the new models and the new massive RLHF data), if in a year they have a 20x GPT4, Claude, Gemini, whatever, they could be "jumping" to the next 40x GPT4, Claude, Gemini, a lot faster, if they have the most popular, prompted model in the market (in the world).
- The scaling stuff already sunk, they have seen the numbers and it doesn't add by now, or they've seen disminished returns coming. This is being firmly denied by anyone on the record or off the record.
Yeah, you can train an LLM to recognize the vocabulary and grammatical features of logical fallacies... Except the nature of fallacies is that they look real on that same linguistic level, so those features aren't distinctive for that purpose.
Heck, I think detecting sarcasm would be an easier goal, and still tricky.
> Except the nature of fallacies is that they look real on that same linguistic level, so those features aren't distinctive for that purpose
Well that's actually good news. With a large enough labelled dataset of actually-sound and fallacious text with similar grammatical features you should be able to train a discriminator to distinguish between them using some other metric. Good luck with getting that data set though.
The Entscheidungsproblem tends to rear its ugly problem
Remember NP is equivalent to second order logic with existential quantified. E.g. for any X there exists a Y
And that only gets you to truthy Trues, co-NP is another problem.
ATP is hard, and while we get lucky with some constrained problems like type inference, which is pathological in its runtime, but decidable, Pressburger arithmetic is the highest form we know is decidable.
It is a large reason CS uses science and falsification vs proofs.
Gödel and the difference between Symantec and syntactic completeness is another rat hole.
> you should be able to train a discriminator to distinguish between them using some other metric
Not when the better metrics are likely alien/incompatible to the discriminator's core algorithm!
Then it's rather inconvenient news, because it means you have to develop something separate and novel.
As the other poster already mentioned, if we can't even get them to reliably count how many objects are being referred to, how do you expect them to also handle logical syllogisms?
