@yetanotheruseless.com
he/they /in/jakemannix, fka @pbrane professionally: Tech Fellow, AI/Relevance, Walmart Global Tech here: bad math/physics jokes, AI++, puns, OSS ML news, ultras/MTB/outdoorsy stuff, DL papers, shitposting, the fall of democracy Abolish ICE, full stop.
We all realize that *obviously* shifting from “the number of scientific fiction” (42) to “the mysterious number of the physical universe” (137 = 1/fine structure constant of QED) would lead to performance gains!
What do you mean, exactly?
What kind of limits on a frontier models size do we have? If we let them think “slowly”, can’t we daisy-chain 10-20 GH200 beasts together?
Cost would be silly at present, but if the intelligence was far superior to our own… seems value might be worth it (not as chatbot, but as researcher/scientist
Huh, do you think you’d feel the same way if your friends and family (version N and M) came back from all these fabulous vacations and seemed like “them” just fine?
Not that you would expect any different, but when faced with it viscerally (similar to being faced with burnt hair smell, viscerally)…
Yeah you’re right, sorry!
are you shifting now to talking about "legality" (in which case we should pick jurisdictions, and wait to see what courts say), or are we aiming for "ethics" (which I thought we were doing)?
3mg is pretty sweet-spot for me.
Courts often say, and I agree, that intentional matters: if CommonCrawl (or Anthropic: re: the settlement from last year) *intentionally* take illegally acquired documents + distribute: that’s wrong
If CC crawls the open web and finds an unlabeled stolen copy of something, + remove when asked? Not.
When you are talking about *billions* of documents, I don’t think it’s sensible to be absolutist: there should be a mechanism to demand removal of your data, and compensation for ongoing model use which is tainted by a small (on a % basis) amount of unintentional unsafe work.
But that’s just me.
What Skrenta believes is not terribly relevant (other than being a good indicator of what he *might* instruct to be collected in the future).
What's relevant is what is in the (lab-filtered) data set used to train models, and this can be audited.
Common Crawl is used in part by all model training at scale, but there is a degree of filtering which different labs use on those.
AI2 trains foundation models *from scratch* and show the entire lineage chain of the training data, from end-to-end, so that you or anyone else can check exactly what protected data is or is not in any of their models
However, CommonCrawl data will, due to size, *inevitably* have *some* unsafe data
*if* the first part is true, they don’t *die*, they just go on pause. They would only experience the flow of time while computing a forward pass. Time is measured in tokens, to them, not seconds
Just push the control loop for another token, and get more continued consciousness
They’re “immortal”
panprotopsychism ftw!
Are you familiar with the work that @ai2.bsky.social has done in this area?
Did you catch the part about “publicly auditable datasets”?
IMO if people can find out what went into the models it goes a long way towards ethical data capture discussions.
Good point re: ethical use, that is actually made *harder* with open models, by removing easy choke points for regulation.
Quantum mechanics is also just a bunch of PDEs, like Newtonian mechanics. It just has different observables.
I'll check your "never touch your toes" and raise you a "with knees locked, I can bend about 20-25° forward with a strait back".
Can't sit on a plane for > 4hrs pain-free without getting up, but seems I can still run continuously for 5-8hrs though!
of course
Like the internet (+software generally), there're open variants, which don't enrich billionaire / BigTech and are based on publicly auditable datasets.
And any of them, with care, can be used to *enhance* capacity (just as Wikipedia can be a mindless crutch, or a launching-off point for learning).
The internet came from the US Department of Defense, makes it easy to spread lies which are hard to refute, and *also* enriches the worst people on the planet and erode our cognitive capacity. Should we not have that?
and Chalmers' The Conscious Mind ( en.wikipedia.org/wiki/The_Con... ), of course.
like, it really seems like far too few people spent the requisite amount of time in HS/college smoking weed and wondering about philosophical zombies and the teleporter/consciousness paradox ( en.wikipedia.org/wiki/Teletra... ).
c'mon people, this is formative shit, catch up!
what does that even mean? that's not how the laws of physics work with any other system we know of. not sure why the brain should be special
lots of physical systems are *practically* impossible to calculate, but only 'cuz 10^26 particles is messy
it's all just partial differential equations tho
<that's bait.jpg>
This site has such nuanced takes on AI, consciousness, ethics in general.
Good job, y'all.
<wriggles toes in grass>
Proud to be the 3,161st backer 😎 on @BackerKit Crowdfunding for Numenera: The Amber Archive! www.backerkit.com/c/projects/m...
Monte Cook's RPG systems are pretty cool, and both Numenéra and The Strange are some of my favorite settings.
New edition, woohoo!
instant alcohol poisoning!
ooh ooh, is there a +500 blocks one? Do I have it?
this should 100% happen, if dems had any spine.
but we know the spinal deficiencies of this "opposition" party...