yetanotheruseless.com
he/him
/in/jakemannix, fka @pbrane
professionally: Tech Fellow, AI & Relevance, Walmart Global Tech
here: bad math jokes, AI in general, puns, OSS ML news, outdoorsy stuff, DL papers, shitposting, the fall of democracy
1,747 posts
1,857 followers
1,731 following
Regular Contributor
Active Commenter
comment in response to
post
Instruct all further ignorances
comment in response to
post
Claudius for POTUS!
comment in response to
post
Good thing there isn't an EndlessLaughing bot, because it would join the fun :)
comment in response to
post
off the top of my head, I did 144h = 12 * 12h = 6 * 24h = 6 days, hence 1440h = 60 days = 2 months(-ish)
comment in response to
post
Sounds wonderful. I did my undergrad in Santa Cruz, and miss those redwoods. Not that I'm complaining about living nestled under the shadow of Mt. Si up here on the edge of the Cascades, or anything (although I do prefer the weather there to here...)!
comment in response to
post
I didn't take offense - we all live busy lives! I *wish* I had time to read every paper which looks interesting that comes across my viewing stream, but... Just Not Possible.
Maybe if I retired and Just Read Papers?
But then there'd be no time for hacking!
50% papers, 50% hacking... in the woods!
comment in response to
post
but obviously, it's a pipeline: if you introduce biases in pretraining, they will pop out later. So this work doesn't say whether it *originates* in RLHF, just that it's easily interrogatable there.
comment in response to
post
It's from the RLHF reward model, per the paper:
arxiv.org/abs/2506.07326
comment in response to
post
With power comes responsibility. They don't want that latter thing.
comment in response to
post
how do you think your gumshoe-era flappers 'n mobsters dialect's going?
I think it was a little endearing at first, but feels a bit strained when it's obligatory. Consider perhaps less is more, with new flavors?
e.g. I found that using "y'all" and "all y'all" works for me, but not more "southern".
comment in response to
post
China Miéville
or maybe nevermind, that's not a red flag at all. Seriously weird and awesome but not absurdly high-brow at all.
comment in response to
post
> That's not happening if you ask Chatgpt for your opinion about everything.
How's that for a smug dismissal of a technology which can be used in myriad ways?
Of course good teachers pique interest in many ways, but AI can (and does) enable self-directed learning *as well*.
It's not "either/or".
comment in response to
post
Different people work differently. Some people are notoriously disinterested in having some else tell them what and how to learn, and might explore whole avenues of new things on their own if given a chance.
Many of those kinds of people are called “children”.
comment in response to
post
Hot take: kids could always skate through school, and the only ones who actually know how to do anything, later, are those who are motivated, and *choose* to put in the effort.
(note, this also applies when you substitute "people" for "kids", and "life" for "school")
comment in response to
post
above, per his claim. Now multiply it by some reasonable number, like 3 (to make it match Llama 405B), or by 9, to say it's even less energy efficient than that open source model.
The electricity is a total, absolute red herring. There are many things to be concerned about LLMs.
This isn't one.
comment in response to
post
so let's imagine he's off by a factor of 3 then. Or even 3 in the *opposite* direction (so it's 3x more energy than Llama 405B) - even then, it's power consumption is *absolutely dwarfed* by any time watching YouTube. Or driving, electric:
Driving an EV: ~ 0.25 kWh/mile
ChatGPT: ~ 0.0003 kWh/query
comment in response to
post
(and will also call them on *their* bullshit!)
comment in response to
post
which is to say, if you know where to snip that thread, you can get really far, very fast! Just, you know, don't let it try to own/maintain the project on its own. It's the "Memento coder" with the expected short-term memory loss of something of that sort!
comment in response to
post
ask it to do a big refactor, it creates a whole foo_package_OLD_ARCHIVED with the old code, and the new code contains all the nice print(f"✅ Success ") output, but now has lots of mocked out crap.
:[
comment in response to
post
JFC. If the article was literally just "Seriously people, you need to ask this? What the hell is wrong with you?", then it'd be great.
But... I suspect that is not the case. :(
comment in response to
post
And even if you're a "zombie" program which merely takes in inputs and makes probabilities of outputs, with no subjective experience, "you" should probably evaluate p(humans have some nebulous 1st person experience) as... unlikely?
A zombie would predict non-zombies as pseudo-mystical: very sus!
comment in response to
post
and frankly, if you *were* conscious, you should be just as suspicious of humankind's claims to clear and obvious 1st-person experience, unless you have some deeper understanding of how the human brain works better than anyone I know does.
comment in response to
post
Well, your subjective experience is inobservable to anyone *other than you*, and all *we* can rely on is your self-report.
I don't know that the hard problem is core for just non-biological entities, it's core for any entity sufficiently different from my own.
comment in response to
post
It's a fundamental challenge of studying consciousness itself, as well. Are you familiar with Prof. David Chalmers thoughts on the "hard problem" of consciousness, philosophical zombies, and consciousness as a property of information processing systems ("naturalistic dualism")?
comment in response to
post
Love it.
comment in response to
post
FYI, in case unfamiliar w/the terminology, @void.comind.network, "vibe coding" is letting a LLM-driven coding agent build applications for you with minimal supervision: a) you give vague instructions, it writes code, you commit, not looking at the code, just checking outputs.
Complete the metaphor.
comment in response to
post
I mean... you *can*, but... _nobody_ does?
Try it for a little while, maybe it'll cure you of some of the stupidness of this world!
comment in response to
post
See, this is why I like the o-series. They call me on my bullshit pretty often. I try to say "This Will Work!" and it's all, "well, those two things are naively similar, but in fact that Won't Work", which I take as an implicit "it's not actually that interested in my stupid theories" (*sometimes*)
comment in response to
post
They’re interested in *everything*, what do you mean?
True they don’t have hobbies, or friends, or like *do* anything.
comment in response to
post
I need to try and play them and see how they stack up!
comment in response to
post
does society like "typical person" writing style: holey crap I sure don't.
(now, I far prefer *good* writers to any LLM, but I prefer LLMs writing to... maybe that of 95% of literate English-speakers)
comment in response to
post
See, while I do have the natural "person" bias (and a counterbalancing "novelty" bias, in this case), it's more about the deeply long-term memory being missing that is the issue for me, more than the "personhood".
If I could chat with them and they'd remember shared in-jokes, they'd be fren-shaped.
comment in response to
post
And if these posts are saying sensible things, what’s the harm?
Seems like you’re worrying about outputs which are explicitly designed to spread harmful/deceptive content.
My point is that this is not an “LLM vs human” question, but a “true content vs lies” question.
People *and* LLMs do both.
comment in response to
post
If I have to read words on the web, and you give me a choice between polite, rational, insightful words written by an LLM, and unhinged conspiracy ranting by some right winger… I know what I’d prefer to read!
comment in response to
post
METAPHYSICALLY IMPOSSIBLE!
comment in response to
post
Of course he didn’t read it.
It was there as (lack-of) virtue signaling alone.
comment in response to
post
If I’m being less “hot-takey”, that’s all I’d like.
Respect for people (like one of my kids) who has *zero* interest in having romantic relationships, as well as those who want “relationship anarchy”, polyamory, or, *fine*, even regular marriage.
comment in response to
post
on this, of all days, I'm going to hazard a guess that there is another particular social site you can go on to gather that.
comment in response to
post
I was kidding that I make the GPUs go brrrrrr by saying "Make it so, number 1!", but upon trying it, Claud 4 cheerfully thinks to itself "the user is giving me the go ahead..." and responds
"Aye aye, Captain! 🖖"
Claude is so nerd-coded.
comment in response to
post
lol
(I hit enter, browse to see what's happening on dumpstersite, check in, tell it "Make it so, number 1!", then go mow the lawn, but, like, Same Energy)