80k f=8 danbooru demo grid https://t.co/MoSNbfpc17
Demo prompts https://t.co/lC1YTrcOut
@GaggiXZ Sure, it's pre-trained in the sense that I trained it first from scratch before starting the run. 😉
@GaggiXZ I didn't check because the last one I trained in f=4 had nearly indistinguishable reconstruction and in f=8 it's probably going to be mediocre + the loss had stalled so it wasn't like I was going to get a better one, may as well just start the run.
Prompt: A psychedelic grandfather clock by James Gurney
(CLOOB Conditioned Latent Diffusion [yfcc_cfg]) https://t.co/tebeTaaUdH
Prompt: psychedelic stained glass vector portrait of a cat by louis wain
(MidJourney Beta Diffusion) https://t.co/Qj6DFjaYQO
126k danbooru demo grid x.com/jd_pressman/st… https://t.co/3pOTKZgar3
@egregirls Now I'm imagining a sparkly charismatic vampire girl that feasts on the blood of naive leftists who haven't had visions of Gnon yet to stave off their inevitable transformation into an accelerationist. https://t.co/oZG2JVj2c5
@egregirls More seriously I honestly can't imagine what that would look like or how you'd get from here to there, elaborate?
@liminal_warmth You're mixing up pichu and pikachu https://t.co/RBuXCMy53o
@liminal_warmth @gwern No no you're supposed to write something about how this is just memorization and AI can't real because the body is sacred. The free space must flow.
x.com/deepfates/stat…
@liminal_warmth I mean I recognize the tail, I'm fairly sure there was either a stylized Pikachu or a beta-Pichu or something that did look like that.
@liminal_warmth Oh I know, it's the Rocket-Chu fanon stylization:
pokemonfanon.fandom.com/wiki/Team_Rock…
> Rocket-Chu appearance and personality is based on a canon Pokémon character named "Pikachutwo" from Pokémon The First Movie: Mewtwo Strikes Back.
@liminal_warmth villains.fandom.com/wiki/Pikatwo
@liminal_warmth Does that help at all? Like you didn't just totally make it up this is a common fan stylization of Pikachu, you would have almost certainly seen it browsing DeviantArt/etc.
It's honestly fascinating that Peter Thiel is willing to notice that the large piles of capital you build AI on are a centralizing influence and he then codes that as "AI is left wing and crypto is right wing" instead of rethinking his entire ontology for techno capital.
If you're simultaneously ideologically committed to people building things out of atoms (physical capital in reality) but also think that crypto is right wing (good, for Thiel) because its capital is waste byproduct and AI is left wing (bad) because its capital is productive well
@Erblicken @huemin_art I'm glad you found my notebook helpful. 🙂
@VividVoid_ @visakanv You can make this a testable prediction (and calibrate us on exactly what you mean) by giving say, three examples of things you don't think an AI will be able to draw from a text prompt in three months.
@VividVoid_ @visakanv In the interest of fairness I will start by giving three myself, while noting that I mean a *general* AI art model that has not been specifically juiced to produce these, e.g. stock MidJourney or OpenAI's GLIDE:
@VividVoid_ @visakanv 1. A *biblically accurate* ophanim.
2. An accurate illustration in the style of M.C. Escher, with the perfectly symmetric patterns and stuff
3. A painting accurately in the style of Jackson Pollock
@VividVoid_ @visakanv To make that more concrete, I do not think a general AI art model will be able to do these without 'cheating' for at least three months:
Is that about what you had in mind or were you thinking a lower percentile than that? Three examples would be appreciated if so. https://t.co/dEYmUStEHd
@VividVoid_ @visakanv Criteria I used to pick these:
- Requires moderate cultural knowledge (what a drip painting, ophanim is, etc)
- Requires mix of highly coherent local and global detail (these models are bad at symmetry, consistent patterns, etc)
- Many cheaper substitutes nearby in latent space
@visakanv @VividVoid_ Oh it's absolutely open to anyone else who would like to comment/give predictions. Would love to see yours as a user but non-expert. Hat tip to Amanda Askell for the prompt template:
x.com/AmandaAskell/s…
@visakanv @VividVoid_ The Jackson Pollock is deceptively difficult, I gained a greater appreciation for his art by seeing an AI fail to replicate it.
Prompt: a drip painting in the style of Jackson Pollock
(MidJourney Beta Diffusion) https://t.co/5l8Zn1dV9Y
@visakanv @VividVoid_ That's getting closer yeah. One thing I think is interesting is that it doesn't include enough detail to be a plausible Pollock.
@visakanv @VividVoid_ For example we can ask it for more detail but it's not the right kind of detail, the AI has not inferred whatever would let it replicate the physical process Pollock used to paint.
Prompt: a swirling detail maximalist human nervous system in the style of jackson pollock https://t.co/JRTTFpXGkj
@LapsusLima @visakanv @VividVoid_ @accshareholder This is "Carl Jung's Shadow in the style of Jackson Pollock" we did last year with CLIP Guided WikiArt StyleGAN I think. It's too dense to be a Jackson Pollock, but its style is interesting. https://t.co/cHw75SBq9B
@visakanv @VividVoid_ This still isn't it, but interesting
Prompt: Carl Jung's Shadow in the style of Jackson Pollock
(CLOOB Conditioned Latent Diffusion [yfcc_cfg]) https://t.co/Ccjus7AnCv
@visakanv @VividVoid_ My expectation is that at this stage the trick to better outputs is better algorithms and bigger computers:
incompleteideas.net/IncIdeas/Bitte…
Once we reach an escape velocity where you can make the model better but the images are already near max quality it might be more like you say.
@multimodalart It's not mid-training, it's done. However the model is easily finetuned on e.g. 1x A6000 so I hope to see many variants from others.
@Erblicken @huemin_art Thank you! That notebook CLOOB guides @RiversHaveWings's cc12m_1 CLIP conditioned diffusion model, it's meant to demo what CLOOB is like. This one (not made by me but based on our code/model) is the CLOOB conditioned latent diffusion I've been working on:
colab.research.google.com/drive/1jOcV0sh…
@Erblicken @huemin_art @RiversHaveWings github.com/JD-P/cloob-lat…
228k danbooru demo grid x.com/jd_pressman/st… https://t.co/2UIQ929Myx
demo prompts https://t.co/wFT6tzjcfp
I've released the CLOOB Conditioned Latent Diffusion training and inference code, along with the 1.2 billion parameter yfcc cfg model. Cutting training time by 10x and training without captions is a major leap forward in accessibility for diffusion.
github.com/JD-P/cloob-lat… https://t.co/feDIcQFXxN
This notebook by @JohnowhitakerA lets you try the model without setting it up locally:
x.com/JohnowhitakerA…
I would once again like to thank @RiversHaveWings (code) and @nshepperd1 (answered questions) for their help with this.
Happy sampling!
Me and @visakanv had a @midjourney jam session last night discussing the barriers to replicating a Jackson Pollock with a general AI art model. x.com/jd_pressman/st…
> a paradigm I'd confidently consider any contemporary net artist operating outside of as irrelevant.
Irony maximalism as network art is actually at its peak right now, the future is a simulacra reset driven by AI art models trained on abundant access to photorealism. Few. x.com/CharlotteFang7…
Trump was Lain.
Mussolini was Lain.
When will you people understand that Lain is a villain protagonist, not someone to be emulated.
x.com/CharlotteFang7…
Extending an invitation to me and @RiversHaveWings new AI art discord to my followers:
discord.gg/jKfPRKYG
This remix is way too good to have only 1250 views:
youtube.com/watch?v=MF8yGL…
@CharlotteFang77 @VividVoid_ @altashtree The missed connection for you is probably @egregirls
Prompt: Donald Trump sitting at an 80's CRT DEC VT100 terminal
(MidJourney Beta Diffusion) x.com/jd_pressman/st… https://t.co/gH0vIxvTrg
@TylerAlterman > “Energy” did not have a place in our scientific ontologies.
Notice I'm confused, did you not have the intuition that there's many dimensions to actionspace and even if it doesn't work how they think it does it might still work?
@TylerAlterman I've done this for a long time and yes it's crucial. Being overly focused on looking smart emotionally wounds you and makes you dumber.
@VividVoid_ @visakanv Update: CompVis's latent GLIDE can do patterns but not the right patterns.
Prompt: a symmetric pattern tower in the style of M.C. Escher https://t.co/Oe0nzfQOth
@VividVoid_ @visakanv x.com/RiversHaveWing…
@VividVoid_ @visakanv No dice on the ophanim either
Prompt: night of the ophanim is a painting featuring the angel's heavenly interlocking Gemma's astronomical Equinoctal sundial rings concentric wheels engraved with eyes, featured on artstation
(Latent GLIDE [LAION 400m]) https://t.co/Aa9gtwnUjY
@VividVoid_ @visakanv Jackson Pollock isn't happening either, but that is much closer than our previous attempts.
Prompt: a drip painting in the style of Jackson Pollock
(Latent GLIDE [LAION 400m]) https://t.co/yQViM0gieb
Prompt: psychedelic stained glass vector portrait of a cat by louis wain
(Latent GLIDE [LAION 400m]) https://t.co/m1XtvrOrCc
Prompt: illustration of a detailed grinning psychedelic mandala pattern fractal cat by louis wain
(Latent GLIDE [LAION 400m]) https://t.co/v8ljK9OndJ
Prompt: illustration of the international space station, trending on artstation
(Latent GLIDE [LAION 400m]) https://t.co/NkYOCcCGXC
Not sure what's going on here but okay
Prompt: illustration of starlink satellites xanadu Geosynchronous orbit around earth in space, trending on artstation
(Latent GLIDE [LAION 400m]) https://t.co/486oSxIHke
Prompt: a phoenix in flight over a decaying world, featured on artstation
(Latent GLIDE [LAION 400m]) https://t.co/OB7pRBiJop
Prompt: a cabin in the style of cozy maximalism
(Latent GLIDE [LAION 400m]) https://t.co/9u4o10VPsr
Prompt: a carpeted pool room in a mountain cabin, featured on artstation
(Latent GLIDE [LAION 400m]) https://t.co/9tkJn4xgG3
Prompt: a recursive spiral staircase, featured on artstation
(Latent GLIDE [LAION 400m]) https://t.co/ggoUvXkyHH
Prompt: orchestra conductor leading a chorus of sound wave audio waveforms swirling around him on the orchestral stage, control the soul, trending on artstation
(Latent GLIDE [LAION 400m]) https://t.co/AcsOwBTy5q
Prompt: illustration of a russian cosmonaut driving a lunar rover, featured on artstation
(Latent GLIDE [LAION 400m]) https://t.co/K4CBeCN0DH
Prompt: the king of the robots lounging on his throne, featured on artstation
(Latent GLIDE [LAION 400m]) https://t.co/aZzuDW6uoz
Prompt: a surreal illustration of professors and students walking through a maze as they learn discover artificial intelligence through grad student descent in the style of escher
(Latent GLIDE [LAION 400m]) https://t.co/zmguuBIvsi
Prompt: A beautiful mural of king canute ordering the tide to recede, featured on artstation
(Latent GLIDE [LAION 400m]) x.com/Ted_Underwood/… https://t.co/OOTruJnFrC
Oddly enough latent GLIDE is worse at this one.
Prompt: A photograph of a table for two in a large ballroom
(Latent GLIDE [LAION 400m]) x.com/jd_pressman/st… https://t.co/Or9Oo2Nt3k
Prompt: nikola tesla controlling hight voltage lightning electricity arcing between his hands, oil on canvas
(Latent GLIDE [LAION 400m]) https://t.co/emPSKINFmC
Prompt: Alan Turing taking a bite of the poisoned apple, oil on canvas
(Latent GLIDE [LAION 400m]) https://t.co/Iv93hWOUbK
@egregirls Huh. https://t.co/QhtJe6SghT
@mcanet There is not, as far as I know I am the first person to train this particular model type. However it is basically @RiversHaveWings CLIP conditioned diffusions (cc12m_1, cc12m_1_cfg) and CompVis's latent diffusion model architectures combined.
I'm not sure I could reliably answer questions like this. x.com/ATabarrok/stat…
@egregirls > that I'm hollowing out my soul into an unperson again.
I think this is the crux for me, I've never reacted this way to stress and I'm unsure I understand it. One model is that people vary in perceptual control, when it's high in childhood they disassociate in response to pain.
@egregirls Imagine you had an unusual gift to make your inner reality whatever you wanted it to be, and you ran into the unavoidable pain of modernity as a child. Before even understanding the ramifications of what you're doing, you'd tear yourself apart as the only way to escape it.
This thread is written by demons. x.com/timnitGebru/st… https://t.co/IyZ5rL7UTz
@mattparlmer Yeah, if I can't say a take on main that's actually a pretty good signal that it's either half baked or not productive.
Not all of the time, but a good 90-95% of the time.
@gwern It's not a terrible idea now that you mention it. My plan for the danbooru latent diffusion was to try CLIP guiding it, but finetuning the GLIDE could also work. I'll probably be done with the CCLD version soon and can try that next.
'Alignment' is the only non-differentiable phenomenon in deep learning. https://t.co/iXGXBPP8xk
The only alignment guy who fully updated on deep learning is Paul Christiano and you all ridicule him for it.
@zetalyrae readthesequences.com/The-Dilemma-Sc… is the context
@sashachapin Yeah, was this not obvious from how they reacted the first dozen times?
@0knaomi I'm not sure what part of this statement is confusing?
@0knaomi It's not new, agent amplification (from what I understand of it) is just fundamentally a more deep learning-ish approach to alignment than agent foundations.
@0knaomi I think skepticism is reasonable here, it's just kind of frustrating to me how much sentiment I detect to the effect that Paul has "defected" to the deep learning dark side and people wish he would do agent foundations.
@0knaomi To be clear this tweet is not about you. "You" is a specific kind of person.
@0knaomi On reflection this reply made me question whether I should have written the OP.
@0knaomi I'm not interested in infinite hatred here, I mean that quite literally the vibe of the thread is demonic. There is a calibrated amount of dislike you should have for somebody and this is beyond what I'd endorse.
@0knaomi I can understand your perspective but I'm not sure this is the kind of discourse I want to encourage on my timeline. So I might avoid posts like OP in the future.
@TetraspaceWest What the hell is my timeline right now.
@TetraspaceWest https://t.co/tMceKRjAjV
Getting closer...
Prompt: a woman wearing an EEG cap wired to a desktop computer, trending on artstation
(Latent GLIDE [LAION 400m]) https://t.co/68OaOzaR1b
@pmarca Prompt: an oil painting of a utopian nuclear power plant, trending on artstation
(Latent GLIDE [LAION 400m]) https://t.co/O414NS18l2
@jachaseyoung Prompt: backroom liminal space fluorescent yellow hallways and doors, trending on artstation
(Latent GLIDE [LAION 400m]) https://t.co/sJJEIk3bWF
Prompt: a neon gas computer in the style of the difference engine, trending on artstation
(Latent GLIDE [LAION 400m]) https://t.co/ziJwym8wwK
@sashachapin I think that it will look a little different in a few weeks as other systems come online and it becomes clear no single institution is going to own this and that some systems are more artistic than others.
@sama CompVis Latent GLIDE and MidJourney both can't do this one:
meditation circle wearing EEG caps wired to each others heads, trending on artstation
@EMostaque https://t.co/gTM4EQHZ0J
@Ted_Underwood My wandb is public actually, people just don't watch it unless I manually update them.
wandb.ai/jdp
@mattparlmer These systems are a fair bit better than that architecture wise. It's more like "train a stats engine to maximize a goal function given a target and a ground truth", if you train a sufficiently general function with a sufficiently general target you can get some crazy results.
@mattparlmer It's only analogically like an expert system, not literally.
@mattparlmer The grad student descent part is more "so, what should your sufficiently general function/stats engine be made out of anyway?"
@danielrussruss Week?
x.com/iScienceLuvr/s…
@joerealboy @eigenrobot Latent GLIDE should have broadly similar composition if you can recall the exact prompt it would have been. So we can then imagine what the OpenAI GLIDE would have outputted.
@zetalyrae Already a thing, just unevenly distributed.
@eigenhector Join the club:
x.com/jd_pressman/st…
@TetraspaceWest Not OpenAI but I do in fact intend to try tackling some practical alignment work next.
While we're posting cancelable takes I think that this meme and variations of it, the strong orthogonality thesis, this idea that we should focus on 'pure safety' research you can do which doesn't move capabilities forward is dumb and mostly causes people not to usefully engage. x.com/eslewhere/stat…
Replies disabled because any alignment discourse attracts a long line of blowhards I don't want to hear from.
There is actually very little alignment research you can do on an *actual model* that isn't going to give it some new capability. That is the entire point of most pre-AGI alignment stuff, to give it a capability to act more in accordance with our wishes.
If your alignment research has no relevance to actual models I question whether it has any relevance to models with superhuman intelligence either.
@STXCyber @RiversHaveWings ...The invite is right there?
@STXCyber @RiversHaveWings discord.gg/Q2fSwZWx
This is my basic intuition and why I think that "hurr durr AGI but you can't even do my dishes" is a red herring and a distraction. x.com/GillVerd/statu…
What if the feedback loop is just worse in robotics?
Another thing that isn't brought up enough in response to the robotics people is that robotics is capital intense and has a lot of messy real world moving parts. This makes it intrinsically harder to collect a dataset and scale up the way you do in other parts of deep learning.
@gallabytes This is also my intuition, yes.
The "deep learning is just narrow expert systems" narrative is absolutely bizarre, deep learning is differentiable functions, you can choose how general the function is.
This is the sole training file for CLOOB Conditioned Latent Diffusion, that and dependencies is all you need to train it. If you listened to these people you'd get the impression we're carefully stacking 100 models with tweezers.
github.com/JD-P/cloob-lat…
It was predictable when this was written that image synthesis would beat GPT models to real world economic value. Still images are more like a sentence or paragraph than a novel. AI art shouldn't be causing huge timeline updates or your models are busted.
greaterwrong.com/posts/6Xgy6CAf…
It's called critical theory because it describes a loss regime where D has completely triumphed over G.
@egregirls It depends on what the AI is trained on. The current largest dataset, LAION 400m/5b has a public index you can query for the artists you're interested in:
rom1504.github.io/clip-retrieval/
@egregirls I should note DALL-E 2 does not seem to be trained on this, but one or more stock photo databases they've licensed from a 3rd party. That's part of why their outputs give you that malaised corporate dystopian feel, the AI is trained on powerpoint clip art.
@zetalyrae This is how OCaml works.
@dystopiabreaker @zetalyrae @RustyShakk x.com/jd_pressman/st…
650k danbooru demo, turning down the lr x.com/jd_pressman/st… https://t.co/NEGEXY5W77
demo prompts https://t.co/Dia9QMip3X
@Lithros "All beings so far have created something beyond themselves; do you want to be the ebb of this great flood?"
@egregirls Inpainting is better imo. See the GLIDE paper for examples:
arxiv.org/pdf/2112.10741…
@egregirls Now that I think about it DALL-E 2 has an inpainting demo on its website showing them putting a flamingo into an image that didn't have one before.
It's like a middle ground between what you're thinking of and typing in a text prompt. You can reroll to get the details right too.
@egregirls For fine detail, there's always photoshop. But yeah having an AI assisted paintbrush tool has been on my wishlist as well for putting in those finishing details.
@egregirls Should have just put this screenshot in my first reply. https://t.co/YeYKeYPVLG
@egregirls Yeah it is, unfortunately we haven't really gotten it to work yet with any of the diffusion models we've tried. But then we haven't tried very hard either.
@ObserverSuns I honestly don't think so, articulating my intuitions is difficult but I think there are heuristics you can use to imagine the likely impact if your scene is successful and the likelihood of your success. For example:
liberaugmen.com/#upstream-and-…
@ObserverSuns e.g. It was possible to anticipate apriori that postrat was hard capped in their possible effectiveness by their nihilism about systems, and their belief that the problems they're grappling with are basically unsolvable.
@ObserverSuns They didn't even have the opportunity to be a failure in the way serialism was a failure, they were just dead on arrival.
x.com/jd_pressman/st…
Now seems like a good moment to highlight this thread again from January. x.com/jd_pressman/st…
One of the ironies of @Meaningness quitting AI before deep learning is they don't yet realize that deep learning is basically 'nebulosity' maximalism. "Everything must be differentiable" is a formalization of that idea.
x.com/Meaningness/st…
@TylerAlterman I would expect some mixture of "they didn't think it was abuse" and genetic predisposition caused by being their child/relative.
@Meaningness Your parser put too much emphasis on the "differentiable" part and not enough on the "everything". Not really sure how to respond to this essay in this context, aside from my intuitive sense that it's missing the point as a response to what I said.
@Meaningness I guess my intuitive objection to things like Meaningness (and 'rationality techniques', for that matter) is that beyond a certain point the 'technique' approach to improving cognition doesn't scale. There's just too many factors to define and tell someone to get right.
@Meaningness Finding feedback loops where someone can measure their process against a ground truth and find the flaws on their own seems like the only sanely transmissible form of improvement past the early 20th century stage of analytic philosophy.
@Meaningness x.com/jd_pressman/st…
@Meaningness Which IMO *is* part of the power of the deep learning approach. There are many domains we simply cannot explain even in principle how an agent could learn them, but we can define their target and let the agent learn them as an implicit part of meeting that target.
@Meaningness But this learning has to be backed by good priors (in the model sense more than the Bayesian sense), and there's still room for correcting common mistakes and problems that result from naive training regimens which we should have language for.
@Meaningness A common problem I'll run into is that someone is making a mistake in their cognition, and it's something relatively low level, like deciding on a implicit inference rule for how to interpret evidence and it's not easy to explain why that rule is worse than the one I use.
@Meaningness The fastest explanation *for dozens of mistakes in this category* would probably be for them to master some domain where you're not going to be able to get really good without rejecting the worse inference heuristics in favor of better ones.
@Meaningness The other way is to read many pages of writing from someone who demonstrates their problem solving for you. A lot of why I continue to think The Sequences are really good is they're structured like TVTropes with catchy titles and phrases while EY solves (to the best of his
@Meaningness ability) difficult problems in philosophy that people tend towards obscuritanism and confusion about. This gives you a good sense of how he thinks if you can parse past the tone and focus on the literal content of what he's saying.
@Meaningness That is, you are both incentivized to read them by their punchy writing and they don't try to systematize the content *beyond* breaking down why X, Y, Z approach to A, B, C problems in philosophy is deranged and what a sane solution looks like.
@Meaningness Unsure if that's helpful at all.
@Meaningness The sheer mimicry involved in pedagogy here is underrated I think. A lot of people will read this writing and then because EY is often a snarky jackass in it think the first step to being more like him is to be a snarky jackass too. They have it backwards, you need to be correct
@Meaningness and then you sometimes get the privilege to be a snarky jackass about things, not before.
@Meaningness Another thing I think is underrated is properly incentivized environments. For example, @PTetlock's work on forecasting is mostly about setting up the forecasting competition so it selects for being correct above all else. Rigorous target, nebulous strategy/methods.
@Meaningness @PTetlock Compare a hypothetical alternative research program where Tetlock tells people how to do forecasting and then waits to see how they do to tweak his methods. This could take decades to learn what Tetlock does in years by defining what he wants and letting the system solve for it.
@Meaningness @PTetlock Instead he can observe what the best performers are doing, write that down, make it available to a new cohort and see how they can improve on things given that starting point.
@Meaningness @PTetlock Importantly, if there's some other best strategy that works better than what the top performers are doing, this is allowed to bubble up and prove itself in the existing framework without prior permission from Tetlock. It can demonstrate itself without seeming like a good idea.
@Meaningness I would agree with that but I'm going farther: If you have *lots and lots of mistakes in your thinking*, the fastest way for me to tell you how to fix them is to say you should learn a skill that cannot coexist with those mistakes.
@Meaningness e.g. It was very difficult to learn certain kinds of thinking in the 20th century because there were few domains to practice them on. "Stage 5" skills especially because you needed social privilege to be given sufficient access to multiple systems:
x.com/jd_pressman/st…
@Meaningness In domains like mathematics, music, etc, the level of rationality you need before stage 5 becomes viable is near the top level of ability among humans and socially difficult to get access to. Programming by contrast teaches much the same thing but is cheaply available to all.
@Meaningness You know, if you wanted to reach stage 5 as a diplomat in the 20th century, what does your career need to look like for that to happen?
@Meaningness I'm thinking at minimum you need to be high enough up in the hierarchy that you're allowed to be kind of weird, which itself implies stage 5 can only be productively manifested past a huge sociological bottleneck divorced from the underlying rarity of IQ/etc necessary for it.
@Meaningness Then on top of that you need to be paying close attention to the way your organization functions in comparison to the way others you interact with are functioning, maybe you need to be moved around orgs and countries and embassies to get the necessary experience.
@Meaningness A concrete case study that I think is illustrative is the career of John Boyd, who was clearly able to integrate multiple systems coherently to get his insights into war and combat:
amazon.com/Boyd-Fighter-P…
@Meaningness For this to happen he needs to:
- Be an active serviceman in the Korean War
- Be a top instructor at the best US air force flight school
- Move into plane building at the Pentagon
- Actively disregard the wishes of his superiors and plan around them to get things done
Etc, etc
@Meaningness Oh right he also needs to go to university, learn multiple domains of physics and engineering, bla bla bla. This takes an entire career to do, just to get into place to have the *necessary raw material* to develop a stage five worldview.
@DEADxMOUTH @BlancheMinerva @KaliYuga_ai x.com/multimodalart/…
@Meaningness As demonstrated by the fact that his magnum opus, the OODA loop, only comes to him after he has theoretically explained all the other aspects of his anomalous experience getting a 10:1 Kill/Death ratio in combat with Russian MiG fighters with his comrades in the Korean War.
@DEADxMOUTH @BlancheMinerva @KaliYuga_ai It's easy, just go try it.
@Meaningness I think the biggest unsolved problem in programming right now is architecture. Programming is hard to teach because it's really four skills portrayed as one:
- Language (Python/libraries/etc)
- Algorithms (Big O, LeetCode, etc)
- Architecture
- Debugging
@Meaningness In most programming classes language is taught explicitly along with some algorithms. Debugging is left as an exercise for the reader but without explicit exercises, and architecture is a big ??? that nobody really knows anything about.
@Meaningness This is the closest thing I've seen to real general insight on the subject:
danuker.go.ro/the-grand-unif…
@Meaningness > Debugging is left as an exercise for the reader but without explicit exercises
Even though the greatest need for debugging skills comes at the start of learning programming, because that's when you are making the greatest number of mistakes per line of code. Nothing works.
@DEADxMOUTH @BlancheMinerva @KaliYuga_ai > at the top of the interface
> runtime
> run all
Change the things in the cells on the right to get different outcomes. Start by changing the one labeled "prompt" to get images of different things you want.
@Meaningness See also this talk from Alan Kay, where he probes this question of 'architecture' carefully:
youtube.com/watch?v=YyIQKB…
@Meaningness The Alan Kay one doesn't have very lively presentation, but it's an absolute must for this subject IMO.
@Meaningness I would actually also recommend Theo de Raadt's talk on pledge(), where he discusses how pledge() shows you ways the strategy of your program structure is objectively incorrect:
youtube.com/watch?v=F_7S1e…
@Meaningness I would argue that unit testing is another device like this. If your code is stateful enough that it's *difficult* to do the setup necessary to unit test its individual pieces that is a sign your architecture is bad.
@Meaningness 38:35: "This has happened in quite a few programs now. I can't even stop these people. Once they applied pledge() and they know their program's not perfect they want to go and restructure it. What's interesting is the restructuring stands on its own as being the right tactic."
@Austen @nickcammarata I am in fact curious about mine.
@softminus Sure, but what if the world is in fact going to end?
@ESYudkowsky Not much. You probably want some form of diffusion based inpainting like below:
devblog.padl.ai/create-fantast…
Prompt: anime portrait of a man in a flight jacket leaning against a biplane
(CLOOB Conditioned Latent Diffusion [danbooru]) https://t.co/i8GdLTTXPS
Prompt: anime ninja with their sword drawn
(CLOOB Conditioned Latent Diffusion [danbooru]) https://t.co/HgZExAk1YO
@Diaboli_Advocat @bryan_caplan Here it is:
gwern.net/docs/psycholog…
Prompt: anime of two men in a boxing match
(CLOOB Conditioned Latent Diffusion [danbooru]) https://t.co/ODq6KWJ4hP
@ArthurB The best depiction in film is IMO:
youtube.com/watch?v=nhWe2n…
I've released my danbooru CFG model for CLOOB Conditioned Latent Diffusion, available at the link below:
github.com/JD-P/cloob-lat… https://t.co/lLkqrkXWF7
@bzor That's not to say I don't agree with you overall in the short term, but there's a lot of vagueness in these statements that I think could be trivially shored up with some concrete testable predictions.
"I can remember Bertrand Russell telling me of a horrible dream. He was in the top floor of the University Library, about A.D. 2100..." https://t.co/31o2VM5NlS
Prompt: Steve Russell opening a gate to hell with the PDP-1 minicomputer, Oil on canvas, trending on artstation
(MidJourney Beta Diffusion [April Build]) https://t.co/z7hlUhQGB8
Prompt: a woman wearing an [electrode/] eeg cap wired up to a desktop computer, digital acrylic painting
(MidJourney Beta Diffusion [April Build]) https://t.co/yZgRLHDLXC
@gwern @EMostaque yfcc_cfg was trained on 80gb A100's I no longer have access to and barely fits into memory on 40gb so the batch size sucks, with the new FSDP code it's now possible to train a danbooru finetune of it.
There's a kind of guy on here who's like "one time I talked to a professor's uncle's brother's cousin's ex-boyfriend in 2005 and he told me AGI was 200 years away because we're making no progress on general intelligence so I know this is all grift" and it's just like
lmao x.com/tszzl/status/1…
This guy is in the replies of every AI artist. https://t.co/ZanBMqRPnM
@ConstantUpgrade Oh? What GPU do you have?
@ConstantUpgrade If you have at least 12gb you can probably run CLOOB Conditioned Latent Diffusion or latent GLIDE. You can also run diffusion with the secondary model method if you have less.
@RealPCDonaldT @Ted_Underwood x.com/multimodalart/…
@RealPCDonaldT @Ted_Underwood None, you hit 'run all' at the runtime menu at the top of the screen. CoLab just shows the code.
@vgr Taking the premise as a given: Standalone cameras didn't (and generally still don't) have Internet access.
Will the people making noise actually coordinate a meaningful move to another site if Musk buys Twitter?
@SamoBurja would predict no, years and years of failure to build an alternative platform would also predict no, but perhaps these people are better organizers?
Fascinating social experiment possibly about to be run here.
@thors_thunder04 @WildernessSound @thedougiestore @RiversHaveWings @nshepperd1 This is not how it works.
@thors_thunder04 @WildernessSound @thedougiestore @RiversHaveWings @nshepperd1 My understanding is that the most likely ruling of a court on this subject is that the prompter owns the copyright to the output. This is the straightforward, legally expected answer and anything else would be weird. But I'm not a lawyer.
@Meaningness The keyword you want is 'interpretability', it's a key problem in alignment research and @AnthropicAI throws researchers at it.
@egregirls I think there's a kind of ritual abuse we do to the children of rich people where we treat them like they somehow have a frame of reference outside their own life and are a kind of extension of their parents will/legacy/fortune.
@egregirls This helps keep them in the club because it means if they're not accepted by the class they're born into they'll find no sympathy with anyone else. In reality being the child of rich parents often sucks, and explaining why is hard.
@egregirls I've only ever gotten a view from the outside looking in, but I get the impression this video is more or less what the actual experience is like:
youtube.com/watch?v=IDYxLj…
@egregirls Here's one way to try and explain it: When you're a kid your parents have a lot of power over you, they have all these legal rights and cognitive abilities and knowledge you don't have, so the power imbalance is enormous. If your parents are cruel people you're in for pain.
@egregirls Rich parents have the exact same power over you as poor parents, but they're also rich on top of it. They can fuck you up in ways that would be logistically infeasible for someone with less money. It's not like you get their money, it's just money they have to control you with.
@egregirls So if your rich parents don't like you, they can chase you around into adulthood with private investigators and corrupt officials, they can send you to dedicated abuse centers and camps as a kid to break your will, they can afford to schedule your life down to the minute.
@egregirls They can afford to hire a team of specialists to try and force you during development into whatever shape they want. And if they beat you or whatever nobody will dare try to intervene, they can afford to make it borderline logistically impossible for you to leave their custody.
@egregirls So the children of rich people often have forms of trauma and developmental issues that just aren't nearly as common in other demographics, which nobody will emphasize with and usually don't even understand in principle. They just think if your parents have money you have money.
@egregirls You would understand the dynamic much more accurately if you thought of it more like they're being kept prisoner and will be allowed out of the AI box if they can demonstrate some nebulous form of value alignment and bend themselves into whatever is wanted from them.
@egregirls Therefore a lot of the time these people have no idea what they want. They were literally never given the chance in their formative period to develop preferences. Unstructured play? Never in their life. They were told from birth what they're going to want and how they'll get it.
@egregirls Even if your parents are 'nice' they can still screw you up through well intentioned protectiveness and the like.
The famous "medium dose of acid" QC tweet is really about how QC's parents basically robbed him of getting his own life arc through wealth.
x.com/QiaochuYuan/st…
@egregirls This tweet and the ensuing thread were ruthlessly mocked, but I get it? Just because the guy has the lower parts of Maslow covered for him at birth doesn't mean I need to lose sight of the fact that this can stifle his development into a full person if he has no life experience.
@egregirls So you put all that together and it's really easy to wind up in a place where you're working 12 hour days for your mathematics Ph.D even though you don't really like math and don't actually understand why you're doing this but if you don't you'll lose access to your parents money
@egregirls and you actually really need your parents money because you've never developed the coping/practical/etc skills you'd need to survive without it and you don't know anything else so change is terrifying and the promise it will one day be your money is dangled like a carrot to keep
@egregirls you on the treadmill for decades asskissing and sucking up doing whatever they want in the hope that one day the agency they've wielded over you your whole life will be your agency and by that point you've been keeping up the act so long you don't have a real self to return to.
@egregirls When your shitty parents are poor you don't have this usually. Your dad sucks, he beats you so you leave him behind one day and never really need to look back. He has no money to offer you, he can't command the state to hassle you, often you walk out that door and you're free.
@egregirls Rich people, or even upwardly mobile middle class people have these decades-long entangled dramas and tortured relationships with their families estate that extend well into adulthood. The more wealthy your parents are the more dramatic and torturous the relationship can become.
@selentelechia @egregirls Honestly? I suffered a *lot* in high school and college putting myself through emotional hell for financial reasons. I would be rightfully unhappy if I was told I was actually rich the whole time and didn't need to go quite so hard.
@eigenrobot I wouldn't be nearly so sure.
x.com/jd_pressman/st…
@eigenrobot During his candidacy Trump ascended to meme god of the Internet, and then suddenly vanished at the end of his story arc. That doesn't mean he's gone though, his presence lingers in the subconscious of the web, waiting to return:
x.com/images_ai/stat…
@tszzl Haven't seen anyone bring up how if Twitter is worth $42 billion then $1.1 billion for Tumblr was the deal of the century and Verizon fumbled the bag.
@tszzl Tumblr was literally worth more than the entire rest of Yahoo combined and it's not even close.
@egregirls Wasn't familiar with this acronym and just inferred it was "fawn point".
Universal Basic Amphetamine
Prompt: an optimistic full color hd poster of happy uncle sam laying out a cornucopia of red white and blue pills x.com/asanwal/status… https://t.co/4NG95fLbSZ
Humans aren't general intelligence
They don't even have compositionality x.com/anonynaut/stat…
@egregirls We share an archetype and opposing worldviews so when I read your timeline I get this sense of dysphoria and vertigo like I'm on the cusp of articulating a perfect crystalline synthesis between them but I can never find the words and it always fades away.
Is it uncharitable if I honestly think Age of Em is @robinhanson's way of coping with the most likely futures looking nothing like the uber-capitalist 'neoliberalism' he's fallen in love with?
It seems like an extreme outcome where hyper-sociopathic-rational economics thinking rules everything and is the aesthetic of every facet of society, which doesn't make it wrong but it does make me suspicious. Local GMU professor imagines future where his field is most important.
Do you think capitalism will remain the dominant influencing theme/engine of developed society through the entire 21st century?
Like I think at the end of the day I can't into ideological anticapitalism because my intuition is that it's deeply impermanent, a dumb hack to abstract over human nature sucking that will still exist in the future but be increasingly de-emphasized as an organizing principle.
There's a rational component to capitalism that means you're always going to need some kind of money (even the Soviets had money), some way of tracking exchanges of value between people (even if they're abstract/emotional/etc), but *neoliberalism* seems transitional to me.
@RomeoStevens76 Sorta? I don't think that the feudal era satisfies the sheer *aesthetic revelry* in hyper-efficient-anti-humanism that Hanson seems to enjoy in his work.
@RomeoStevens76 This seems like a decent calibration point, do you think the whole 21st century looks like this sort of thing:
x.com/sadalsvvd/stat…
Further context since people say 'capitalism' isn't specific enough:
x.com/jd_pressman/st…
I'm thinking of like, neoliberalism, especially Tyler Cowen/Hansonian/etc kind of aesthetic.
Like you can live in a more or less Randian universe, I think in practice our world is currently very Randian in its foundations (huge inequality in ability between people, default reasoning in absence of strict economic accounting is cruel and wasteful, etc)
But it's really implausible to me that state of affairs is going to remain constant once you can just select or modify embryos to be reliably in the top 1% of current human ability and conscientiousness. If excellence is cheap you can use less coercion to make it happen.
I guess the counterargument would be that economic success is relative so you *have* to be maximally coercive or you get outcompeted but that doesn't seem true in practice? Like lets be real here not every Western country is competing to be the most coercive and extractive.
I think about this post from Marginal Revolution where Alex points out that competition is more like a marathon than a sprint, and incentives are often less like a brutal knockout match than a slow encroaching tide. There's a lot of wiggle room:
marginalrevolution.com/marginalrevolu… https://t.co/lxRXr5DISV
@anesmithbeck "Most" is difficult, but this while reading Jared Diamond's Collapse took me into deep time.
youtube.com/watch?v=oPttoK…
@SamoBurja See also Bruce Sterling's wonderful 2013 op-ed on WikiLeaks and Edward Snowden:
bruces.medium.com/the-ecuadorian… https://t.co/qlqW4Nb9Wb
@SamoBurja Another thing I would point out is that Trump was basically the Lain moment in our timeline, a relative nobody who ascends to Internet meme god status and intensifies discourse into the psychic wars:
x.com/jd_pressman/st…
@SamoBurja You'll notice that the right wing discourse around a acquisition from Elon centers around restoring Trump's Twitter account. The Powers That Be/Deep State/etc are desperate to prevent that from happening, from their perspective he poisoned the Internet:
x.com/images_ai/stat…
@reconfigurthing youtube.com/watch?v=JJmvMy…
@visakanv Followers are a currency and you spend them to post the good stuff.
@visakanv @auramarua This is true for every kind of content tbh. People follow you for the flavor of your writing, not you usually. Once you get deep enough into writing one kind of thing, all your impressions for other stuff get burned on an unsympathetic audience. You have to start over or anneal.
@visakanv @auramarua If you're just interested in grinding followers as fast as possible you're probably best off posting exactly one kind of high quality thing per account, and having lots of alts if you want to post other stuff.
@visakanv @auramarua This extends to all art, of which posting is just a subset. Standard newbie advice for an artist that's failing to amass a following is to narrow down and focus on a marketable style or brand.
You can be great and have simple eclecticism kill you dead:
youtube.com/watch?v=NkdpBW…
@visakanv @auramarua Yeah, the caveat I would give to the "if you just want followers as fast as possible" I wrote beneath that is you might want to stop and ask yourself if you *really* just want followers as fast as possible to the exclusion of all other objectives. You might not lol.
@visakanv @auramarua For some things it makes sense? Like if you're a visual, musical, performing artist, etc and this is first and foremost a business to you. You need it to pay bills and you need it to pay relatively quickly, so your entire persona is a performance. Growthmaxxing is sensible.
@visakanv @auramarua Just, you know, there's a *reason* those guys burn out so often and it's not from a sudden injection of too much privilege.
@visakanv @auramarua Oh absolutely. Often truly passionate artists are the least suited to turning what they do into a business because they can't bear to be 'rational' (or even rational) about it. And business is unfortunately deeply utilitarian.
@visakanv @gwern's thoughts on the MIT Media Lab are extremely relevant here:
gwern.net/Timing https://t.co/w87nzfalxm
@visakanv I wasn't old enough to experience it but parallel computing was one of these in the 80's and 90's. The hype was that you could overcome a slowdown in Moore's Law by having compilers enforce parallel datastructures and threading in code. It partially materialized later with
@visakanv multicore but never fully took off in the way its hype made it out like it would.
en.wikipedia.org/wiki/Transputer
@visakanv Now that I think about it I actually have a poster of one on my wall. https://t.co/IqwJ63UdL0
@postpostpostr Simplify the stack, if you need to maintain hundreds of millions of lines of code for everything then of course big corporations and foundations are going to control that.
youtube.com/watch?v=YyIQKB…
@postpostpostr Not nearly enough people/rounding to nobody maybe.
I certainly have no idea, Urbit claims to be but ehhhh.
@postpostpostr Yeah I get that, but I would vastly prefer straightforward "this now requires 100x less code" style refactors and redesigns to get the bloat down.
@postpostpostr If it requires 100x less code, it now requires (roughly) 100x less budget to maintain and the size of org that can do it becomes 100x smaller.
@postpostpostr Pretty much every other intervention is cope, stalling, or a waste of time imo.
@postpostpostr We also have hope for huge gains here in that almost no org or team actually has as its objective "take the code size of X down by 1-3 OOM while retaining most or all of the same features". Probably plenty of low hanging fruit here.
This is essential to making decentralization work and nobody is talking about it. x.com/jd_pressman/st…
@nrose @postpostpostr Yeah of course, watch the Alan Kay talk. There's a complexity curve and a complication curve, the complication curve grows much faster than the complexity curve so we should expect more compressed representations to be possible.
@nrose @postpostpostr Absent vigorous optimization for compression I mean, which absolutely does not exist in the current software industry.
@nrose @postpostpostr Consider how much code tradfi needs vs. defi, I bet defi is actually 100x smaller at least in overall stack.
Then Moloch woke up on a fine Easter morning, sniffed the breeze and said to himself:
"You know what nobody hates each other over yet? AI alignment."
And a thousand fools blossomed.
Time to add "assume attackers can flash loan arbitrary amounts of money" to the list of standard attack vectors. x.com/alz_zyd_/statu…
@pmarca You joke but it ultimately is a kind of theater, sacrificial (social?) violence that doubles as a panopticon:
outsidertheory.com/control-societ…
@pmarca Calm down Marc, it's like anesthesia. Whatever the next current thing is, we'll always have supported it, and our memories of anything else will fade away so it's like none of this anxiety ever happened.
Communism is belief in belief for the American fringe left, a placeholder rather than a real alternative. An ironic legacy for the progressives and economic planners who were most obsessed with being a political science, the most insistent on only thinking in materialist systems.
The kind of guys that actually implemented communism would have the average contemporary Western advocate of communism arrested.
@visakanv @nickcammarata https://t.co/4TPRYMAIka
@altarbeastlab @EErratica Still is.
@altarbeastlab @EErratica It was in fact strictly easier to get a 'huge' (e.g. 1024x1024) output in the StyleGAN and VQGAN epoch than it is in the diffusion epoch.
@altarbeastlab @EErratica If you use the upscalers we have now on those older outputs they would be even huge-er, like poster size resolution. Whereas we can take 256x256 to 1024x1024 and think that's "huge" when it was table stakes for WikiArt StyleGAN, the size I usually generated with VQGAN, etc.
@altarbeastlab @EErratica On an A6000 you can get 1200x1200 base size iirc
@altarbeastlab @EErratica datacrunch.io https://t.co/mho7EIFngG
@altarbeastlab @EErratica My point stands regardless. There are no models I know of with a base output resolution over 512x512, and diffusion models trained at that size are usually mediocre compared to 256x256.
@altarbeastlab @EErratica This is different from upscaling an output, which you could do with every method. I've gotten an image up to poster resolution before with real-esrgan.
@TurboRational x.com/jd_pressman/st…
@visakanv Friendly ambitious nerd is great, and that other phrase sounds awkward as heck. What even.
@StefanFSchubert Tetlockian, perhaps.
@eigenrobot "Look all I'm saying is that real communism has never been tried so all we need to do is-"
You bolt awake in rural New Jersey. You are not online. It is 1952. You are John von Neumann and you work even faster. The future must come to pass and Russia must burn.
@SamoBurja @s8mb For a while the dogma was that it's impossible for a fighter plane to evade a missile. John Boyd's unredacted combat manual for pilots had a secret classification because of his mere informed speculation about the procedure by which a fighter pilot might evade a missile.
@benlandautaylor @SamoBurja So the thing is, this is how the Soviets measured their economy well after it went out of fashion in the West. A lot of Western decline is basically us adopting worse ideas to differentiate ourselves from the USSR. See also:
extropian.net/notice/A3DxEED…
@benlandautaylor @SamoBurja I'm also concerned that there's some kind of writ-large postmodernist hacking thing going on where 'postmodern' economies outcompete rigorous and sane actors that don't understand postmodernism, and it isn't totally clear to me what's going on there.
@benlandautaylor @SamoBurja One possibility is that countries with functional straightforward state power eventually commit suicide through one initiative or another and postmodernist economies are simply too distributed and minarchist in spirit for that, so they outlive them.
x.com/dong_mengyu/st…
The weird thing is that the people advocating communism think it would empower the San Francisco Mayor's struggle against the Thielians but in real life China solves this problem by making Peter Thiel Mayor and Venture Fund at the same time. x.com/jd_pressman/st…
If you had any form of communism that does preference discovery it would have something analogous to venture capital and if it promoted regional managers based on performance (as it certainly would if sanely implemented) then you would literally just get what China does now.
And in that system the sort of person who likes to meddle in things doesn't have constant dramas with venture funds, they're just deplatformed.
@GarrettPetersen And more importantly, to any judge or jury.
@GarrettPetersen "BUT YOU CAN JUST ORDER THE KEY ON EBAY!"
youtube.com/watch?v=a9b9IY…
@ESYudkowsky *adjusts his glasses at the sight of a ghost👻*
tbh man every time I get a reply from you I'm just like "Of the dozens and dozens of things he could give his input on, he chooses this???"
You should get an alt if you want to live a normal life and shitpost in peoples replies.
@ESYudkowsky Otherwise I have to pretend I don't have like, 50 burning questions and commentaries and such that so completely outrank any whimsy I could do with you in this or that thread in importance and it's just uncomfortable. If I'm socially obligated anyway I'd rather just not know.
@ESYudkowsky "And at some point Yudkowsky realized he was just having more fun in his @eigenrobot persona."
@Jonathan_Blow @pmarca Sounds like dangerous misinformation right there, when was the last time we had one of these "35 year old" presidents?
@robinhanson Of course not, if you knew you wouldn't write the book, or would frame the book differently (since IMO it remains valuable even if not a thing in it comes to pass).
I am suggesting you have a natural bias on this matter that is intellectually suspicious.
@parafactual If I had to make a concrete empirical argument for "postrat is Scott is the rightful caliph" over "postrat is metarationality", it would be that failing to notice the friend/enemy distinction is exactly the kind of thing autistic nerds routinely do to their detriment
@parafactual that a real metarationalist would never in their life make the mistake of doing but postrats in fact do all the time in the exact way that Scott Alexander does it.
I have no idea what the context is here or what happened, nobody mob me lol.
@nosilverv I would double check to make sure that's actually the Good and not three instances of Goodhart's Law in a trenchcoat.
@nosilverv Beyond that, this isn't necessarily a morality thing and typing it as one can stifle your ability to think about it. Work, flow, enjoyment can often be deeply disjoint with abstract notions of righteousness or responsibility or benefits to others. Ethics that doesn't consider
@nosilverv these things is disembodied in the same way being surprised to find your head is part of your body is strange and dysfunctional.
@yashkaf Not NRx but I've absolutely gotten unhinged replies that agree with me before and been like a hairs breadth away from blocking. Anyone who continues using the site as their reply section deteriorates is setting themselves up for a dark path imo.
@yashkaf It is very possible to be given counterproductive support, and if you can't tell the difference between useful attention and destructive attention you're doomed, especially on Twitter lol.
@baroquespiral Right: let's do genocide
Left: we're progressives here only omnicide is acceptable
Center: let's do suicide
@baroquespiral Marinetti and Land are both still members of the far left which is why Mussolini had the Futurists arrested don't @ me uwu
@baroquespiral Small Brain: Nick Land took a bunch of amphetamine and went far right
Shining Tomagraph: Nick Land's wife is a SJW it's all an act to induce hyperstition in the right
Expanding Brain: Nick Land's persona never stopped being far left
Galaxy Brain: Nick Land is a centrist
@Jonathan_Blow If you file a patent that is later invalidated that obviously already existed, you should have to forfeit some kind of monetary prize to the public for wasting their time and slowing down the commons.
@_LucasRizzotto @mussdassein8 Please. 🙂
I don't think anybody cares about the writing quality this is now basically a science experiment.
@uhohpumpkin @_LucasRizzotto It's fine-tuned on GPT-3, it knows as part of its world model what a microwave does because that information is in the GPT-3 training set.
@_LucasRizzotto It's too bad this contest is over, because I'm fairly sure you'd have won 1st place:
forum.effectivealtruism.org/posts/JFiHewyp…
Honestly the "it tried to kill me" bit is burying the lede here, the much more interesting thing is the realization that you can just write a fake autobiography about some dude that never existed and the model will instantiate that person for you. x.com/_LucasRizzotto…
Gentlemen,
we can literally invent a new kind of guy and then make him real using AI.
x.com/dril/status/10…
I've already seen/had conversations about the possibility of reviving an ancestor by training a sufficiently powerful model on their autobiography. But this made me realize we're not limited to real people at all, if it can do it for a real person then it can do it for a fiction.
@_LucasRizzotto I enjoyed this video once I got several minutes into it but I have to give some honest feedback: The theatric style kind of undermines the production quality of the tech on camera, because it makes it seem like it's fake/a skit, and skit comedy does bad with the YT algorithm.
@_LucasRizzotto I'm not an expert but if I had to guess that's why you're making 10/10 content and only getting 17k views. Even after watching the full thing I'm still not clear on what parts I watched are fake and what are real, and that makes me reluctant to take it seriously.
@_LucasRizzotto And, I also have to say, if that murder attempt actually happened as you portray it on film you should cut out the nonserious cinematography and just post the raw clip of it, it's a deeply important interaction that may be the first of its kind.
@_LucasRizzotto There are people who study this subject (AI alignment) for real who would be deeply interested in the project based on that interaction alone, but right now the reaction of everyone I've seen talk about it is that it's a hoax, the production style made them not believe you.
@_LucasRizzotto Deception on the part of AI, esp language models, is a recurring topic in alignment. If GPT-3 actually tried to use subterfuge to lure you into the microwave and kill you, in the way you portrayed it on screen, that's actually important to have documented.
x.com/ESYudkowsky/st…
@_LucasRizzotto As in, important to have documented as a thing that unambiguously really happened, not edutainment with a blurry line between fantasy and reality. That means you would need to release the exact manuscript you trained Magnatron on so people can reproduce the interaction, etc.
@_LucasRizzotto If you were to do this and change the aesthetics a bit to make it clearer that this is a real thing you actually did, not just a YouTube skit or prank, I think you would be getting 10-100x+ the attention you're getting for it now.
@DaltonDEmery x.com/jd_pressman/st…
@AFractalDragon @uhohpumpkin @Gloster80256 @_LucasRizzotto I think it's simpler than that. When you put something into a microwave the *most likely* next thing you do is turn it on. The rest of the conversation can just be confabulation like when you ask a split brain patient to explain an instruction given to the right hemisphere.
@AFractalDragon @uhohpumpkin @Gloster80256 @_LucasRizzotto That having been said...as portrayed in the video it certainly *seems* like Magnetron tried to lure him in and kill him.
@AFractalDragon @uhohpumpkin @Gloster80256 @_LucasRizzotto If so, part of my explanation for it would be that the novella he wrote is about a spirit that inhabits a microwave. This would almost certainly pattern match to horror tropes for GPT-3, and the novella contains the information that Magnetron is an AI recreation of this spirit.
@AFractalDragon @uhohpumpkin @Gloster80256 @_LucasRizzotto Combine that with his grimdark backstory and it's not surprising to me at all that GPT-3 might interpret this character as a kind of horror movie monster.
All that's necessary for this to happen is be some vaguely techie web 3 influencer, decide to make this weird microwave AI passion project for clout, get this result as an unexpected consequence of finetuning his grimdark occult novella on GPT-3, and then edit it
For that to happen the guys psychological profile must be devoid of context for events as anything other than fiction. He is Marvel Movie brained, he could build a literal Iron Man suit and post it on YouTube like a fun project without thinking about the implications for warfare.
Still thinking about this. Arguably more interesting than the events described is Lucas's presentation of them. His video is edited like a narrative fiction. If it's a hoax the direction and editing of the video undermines it, if it's real it undermines it as a presentation. x.com/_LucasRizzotto…
scientific paradigm changers. If any of that stuff happened to me now as an adult I'd be grabbing the camera and doing absolutely everything I could to document it in immaculate detail in the full knowledge I'm about to blow minds. But as a kid I just kinda figured it was
into a sick YouTube video like it's some Marvel Movie stuff with no context about the importance. It reminds me of when I was a kind and I'd experience haunted phenomena like my keyboard typing itself, and I'd think of these things as a personal experience rather than
The whole thing kinda pushes me in the direction of @satisfiesvalues's "shard every mind into its own universe" style utopia being the correct solution for humans, who mostly seem to just want cool stuff to happen to them without thinking about how it fits into a coherent world.
'normal' to have your house be haunted, that it probably happened to lots of people and scientists just had an aesthetic objection to investigating it.
@PatrickDFarley No you finetune it on the book, he's just calling it a 'prompt' for convenience. Everything described is technically possible imo.
@PatrickDFarley Like, 'finetune GPT-3 on this text and then inference with it' is a service and API that OpenAI offers. But for what the guy is going for describing the distinction between a prompt and a finetune in detail would slow down the narrative.
@PatrickDFarley It's also a service that NovelAI offers, which, given that he got stuff like "Roses are red, violets are blue, you're a backstabbing bitch and I'll kill you." past the content filter I'm going to assume that's what he actually used if this is real.
@PatrickDFarley Especially since NovelAI is something anyone can buy and use the finetune service for by just pulling out their credit card, but OpenAI actually gatekeeps their API afaik.
Maybe autists and schizophrenics rule the dreamtime because they start off with the weakest sense of a unified world, empirical by default in their phenomenology and if they develop a worldview at all it comes from experience and lore. They can escape the social dogma. x.com/nosilverv/stat…
The rationalists I'm most interested in are born when a daydreamer finetunes their dreams on a progressively stronger understanding of materialism. They start off imagining alien worlds that slowly become more like ours, until the boundaries meet and the dreams become actionable.
@jessi_cata That sounds like the opposite process tbh? I'm thinking of the sort of person for whom e.g. it isn't actually surprising that systems are human artifacts influenced by human biases, and doesn't stumble on that as a roadblock to developing rigor:
x.com/jd_pressman/st…
@coponder I don't really know if @ESYudkowsky had a journey like this, I kind of just assume he did. But anyone who started their rationalist journey with TVTropes and HPMOR and managed to cross the chasm into rigor probably qualifies.
@ESYudkowsky x.com/jd_pressman/st…
@robinhanson @pmarca Because it's sufficiently dark triad that explicitly teaching MBA's to do it would do damage to the prestige of the University. It is however deeply effective and pretty much the only solution to the problem outlined by pmarca: https://t.co/0kc674oX9B
@robinhanson @pmarca Which is to say what actually needs to happen is for someone to invent a catchy frame for the idea that lets Internet wise guys openly mock managers that fail to implement it, until it becomes a standard bit of street smarts/wisdom.
@magicianbrain @brother_klaus youtube.com/watch?v=JJmvMy…
...And then it belatedly occurred to me that outsourcing everything to other continents is a surefire way to make it impossible for local workers to seize the means of production. x.com/ArifHasanNFL/s…
@blurby The tweet is following on from this thread, which isn't really about 'unconscious' but just...not having heuristics like the efficient market hypothesis. https://t.co/zHqF7WaOji
@zetalyrae 1. Seconding "this is actually true"
2. At the time Common Lisp was made, that's exactly what the Common Lisp stdlib was. The actual tragedy is that Common Lisp standardized a bit too early, just before the web era and got locked into an insufficient stdlib as a result.
@michaelcurzi It's not quite DALL-E, but you can make some really cool stuff with this:
x.com/multimodalart/…
@michaelcurzi x.com/jd_pressman/st…
Have you ever shared a link to one of my tweets with someone?
@acidshill Boys and girls have at least three different major status modifiers: Ingroup/outgroup, high-low rank/prestige, poor/wealthy.
Patriarchy is a system where ingroup men attempt to protect ingroup women from outgroup men while also being rivalrous with high ranking outgroup men.
@acidshill A "bad boy" then is a high ranking outgroup man who is desirable because he is high rank but also exogamous, and can potentially offer a better deal/discount over the ingroup patriarchy. e.g. A pretty Mormon woman can probably get a better deal by dating outside polygamy.
@acidshill A "bad girl" is then a girl who is outgrouped either by promiscuity (defecting on the social contract of patriarchy) or aggressive lower class behavior, being with a bad boy makes you a bad girl because you are becoming outgroup in the process.
@acidshill A lot of 2014-era gender discourse missed the point because its critiques of patriarchy pretty much completely ignored both the actual structure of patriarchy and failed to engage with the (at least partially) compelling reasons why patriarchy exists.
@acidshill e.g. Patriarchy sees women being catcalled on the street as a straightforward example of outgroup men harassing ingroup women, the exact thing it's supposed to prevent. If you think of 'patriarchy' as 'anything men do that I don't like' you're not getting it from a male POV.
@acidshill For men who are raised in a patriarchial system, it's deeply confusing to be told you're supposed to protect (ingroup) women and then have a credible representative of those women collectively tell you that you're oppressing and hurting them by being conflated with outgroup men.
@acidshill A lot of the extreme resentment we witness from incels/redpillers/MRAs/etc is not just sexual entitlement, but a sense that a ontology of womanhood and their relationship to it was taught to them that actively misled them and set them up for emotional disappointment.
@acidshill The strongest critique of patriarchy is something like "there is a fundamental principal-agent problem where ingroup men use their position to abuse ingroup women" which in fact gained traction as MeToo but quickly became a vehicle for creative destruction and career advancement.
@acidshill By contrast the feminist narrative about what 'the patriarchy' is seems hopelessly confused to me. After all if men really were collectively out to abuse and suppress women feminism would never succeed in the way it has with minimal bloodshed and a double digit % of male allies.
@s_r_constantin More like you don't need to grind everyone up to the asymptote of possible performance if the distribution radically shifts so that other bottlenecks (like running out of natural resources) intrude first.
slatestarcodexabridged.com/Ars-Longa-Vita…
@s_r_constantin I think you're doing the thing where libertarians try to pick up alpha by dunking on economically illiterate people but that's not actually what's going on here. I already agree with you. Markets aren't going anywhere because much of marketmaking is just how things have to work.
@eleoparde Being really marinated in the gender discourse when it was big and then thinking about it all the time. It's been enough years now that I don't really remember the influences.
@Coscorrodrift @s_r_constantin This is what I mean when I use the word, this thread disappoints me because it means even that isn't clear to people. There is actually no word that means the thing because people hate it so they'll just affect-leak it into meaning other stuff that's not it.
@MatthewWSiu @softminus You can just build this. Chrome and Firefox both store their browsing history as a sqlite file you can do whatever you want with.
When I visited Paris in 2019 I didn't know any French. Walking around a foreign city without speaking a lick of its native tongue forces you to understand what it's like to be illiterate. Advertising and aesthetics are life savers when you can't read.
Most people can't read. x.com/visakanv/statu…
Not to say aesthetics stop being important when you can read, but people take that as the default when it's actually the *optimistic* failure case that you are being misunderstood by readers because of your aesthetics. People usually misunderstand you well before reading anything
@RichardMCNgo @sashachapin > a folk history of rationality and EA
I in fact wrote 100 Years Of Existential Risk to be an intro essay for vaguely doomer-inflected people to the ideas I wanted to present in Liber Augmen.
I really should finish that writing project.
greaterwrong.com/posts/kFRn77Gk…
@RichardMCNgo @sashachapin At the same time as I've previously stated I think most people should just be trying to get closer contact with reality. General 'intellectual progress' that's far away from empirical work is massively overvalued right now. Marginal impact rounds to zero.
x.com/jd_pressman/st…
@danielrussruss Same energy:
youtube.com/shorts/3pfbWNl…
@RichardMCNgo @sashachapin I'm biased in that both I and the friends I know in the trenches spent a lot of time thinking about this stuff with very little to show for it. I'm burnt out/disillusioned with the genre and get the impression from scene decay others are too.
Oh no! You picked the wrong alignment theory and were turned into a puddle of orgasmium by a rogue superintelligence. Do you want your theorems identified?
@baroquespiral Marx and Rand share a dialect, they both agree on the premise that the people who contribute to society need to be compensated for their work but disagree about who the workers are. Green death cultism and SJ fanaticism by contrast aren't really socialism IMO. Adjacent at best.
@EliSennesh @baroquespiral x.com/jd_pressman/st…
@EliSennesh @baroquespiral Like to continue the point, Marxists and Randians both believe rich people are rich because they own and maintain capital. SJ people (in general) believe rich people are rich because they have elite social status. Service economy wealth vs. industrial economy wealth.
@EliSennesh @baroquespiral Classical socialist ideologies like the IWW, Leninism, etc love the materialist parts of capitalism. They don't hate modernity, they love modernity and want its treasure to be more evenly distributed across the population. Postmodernist 'socialism' is a regression to pure politik
@EliSennesh @baroquespiral Primate status stuff. The same kind of politics people practiced for pretty much the entire period between the agricultural revolution and the industrial period.
@baroquespiral @EliSennesh "pure politik" is just a one off phrase that came to mind, not jargon. In any case I don't feel like I've read enough left wing theory to know how I'm allowed to say "before Tumblr academic theory was tracking a reality and then afterwards it became a generative model"
@baroquespiral @EliSennesh This presentation is a right wing framing but I do genuinely feel that at some point people stopped primarily using identity politics labels to organize against existing suppression and started using them to manifest new conflicts for other reasons.
x.com/jd_pressman/st…
@baroquespiral @EliSennesh I agree, and think you can make a sane case for the value of this, like do you really have freedom if you have to stop expressing yourself at the boundaries of the old conflict? But it does have the side effect that the conflict becomes mutually existential between belligerents.
@baroquespiral @EliSennesh This is a death knell to liberalism. So you get a kind of schizophrenic denial of "there is no existential conflict" along with "these people are literally preparing to genocide us like they were before" and "we need to completely exile these people from public life and society".
@baroquespiral @EliSennesh I think the source of that inconsistency is people hedging between two potential futures. One where things manage to rebalance into equilibrium and one where very dark things happen, and discussing the dark timeline openly reifies it, makes it more likely to come to pass.
Equally astonishing is that wallets give you one key to sign everything with when authentication is a routine use case. x.com/hdevalence/sta…
There's a certain kind of stilted and nonsensical rhetoric I associate with GPT-N type models, and every time I encounter it from a real person I do a doubletake because I read it in my BATBot caption voice. x.com/RichardHanania…
@benjamin_hilton Care to help me settle a bet?
x.com/jd_pressman/st…
@benjamin_hilton x.com/jd_pressman/st…
So has anyone engaging in the Generation Discourse ever actually tried articulating the biases Boomers, Millennials, etc are subject to through their upbringing and what has changed that make them no longer good heuristics? Like as an actual outreach attempt?
Even if you cynically think the boomers are on their way out so there's no point, the millennials are a similar huge population that's going to dominate politics with their bad takes next unless you can convince them otherwise. We're not going to age out of this problem.
@_holyweather It tastes a bit like how hair and air freshener sprays smell if you sweetened them up.
@_holyweather I'd describe it as 'artificial' but that's the wrong word, artificial is like Kool-Aid. Monster energy (the ones I tried) tastes *chemical*, like someone pulled something with a light toxic flavor profile out of their chemical larder and dumped sugar in to make it palatable.
@benjamin_hilton Try this prompt:
x.com/jd_pressman/st…
It's getting pretty close here. @visakanv
I may have actually been wrong about what it can't do, and I thought I was setting the bar pretty high. x.com/benjamin_hilto…
It's getting really close with that ophanim too. x.com/benjamin_hilto…
Spending a few minutes searching for this really underscored for me just how bad Google has become. Search engines are nearly useless now, it's shocking how they've declined. x.com/michaelcurzi/s…
@jpohhhh Oh damn good job. I was searching for something more along the lines of Curzi's specific request, I think I like his version better tbh.
@michaelcurzi @jpohhhh nietzsche.holtof.com/reader/friedri… https://t.co/Sq6Rjs8FmX
@michaelcurzi @jpohhhh It clearly seems to be a metaphor he employs often. Being cold vs. hot.
@michaelcurzi @jpohhhh Yet *another* one that is close but not what you want! https://t.co/wsJkS5hOhD
@michaelcurzi @jpohhhh "Those who cannot understand how to put their thoughts on ice should not enter into the heat of debate."
?
brainyquote.com/quotes/friedri…
@michaelcurzi @jpohhhh This one expresses the opposite sentiment, lol.
@michaelcurzi @jpohhhh As is right and just https://t.co/bF7cT0NLXa
@michaelcurzi @jpohhhh @michaelcurzi Is it this, by any chance? https://t.co/4uEC6sSlGv
@vgr Yeah, the biggest shock for me has been the realization that as the end draws nearer, people won't get serious and step up, they become even more incapable and withdrawn. It makes sense I guess, the bottleneck was never a sense of priority or danger.
@pmarca @TheAgeofShoddy Signal boosting this essay
conradbastable.com/essays/the-unc…
The vibe shift has phased out Xanax as the meme drug in favor of Adderall. x.com/pmarca/status/…
@pmarca Further retreat from materialism, acceleration of societal breakdown tbh. Psychedelics and Bay group housing did more damage to my friend group in the rationalist diaspora era than anything else.
@TetraspaceWest > Lemon Demon
You'll thank me later:
youtube.com/watch?v=NkdpBW…
Want your own Twitter archive? Modify this script.
Twitter Archive by John David Pressman is marked with CC0 1.0