> a version of the weave-agent-traces dataset that has an actual README and works with HuggingFace datasets
huggingface.co/datasets/jdpre…
I've uploaded a version of the weave-agent-traces dataset that has an actual README and works with HuggingFace datasets. Link Below. https://t.co/UmVoT1blxS
@Algon_33 My funding for compute ran out so basically no.
This is how "philosophy" works by the way, the subjective and metaphysical is generalized from the objective and material. Not understanding this leaves you deeply confused. x.com/allTheYud/stat…
Yudkowsky doesn't want to say what he was specifically hoping would work because he's afraid it might advance AI capabilities but if he was focusing on things like MC-AIXI we can infer it would have been some form of Bayesian statistical learning. He was just bearish on deep nets x.com/ESYudkowsky/st…
Bayesian statistical learning is also lol lmao dead, but this is not quite the same thing as believing in straight up formal grammar based AI. For one thing it contains the core Hutterian intuition that intelligence is prediction/extremely lossy compression and tree search.
At the time he was writing being bearish on deep nets wasn't even a particularly unintelligent take. Deep nets had in fact been tried in various forms since the 80's and largely found to be computationally inefficient and underpowered. Deep learning working at scale is surprising
@algekalipso So do you think what's going on is closer to them having the same kind of subjective experience you and I do and lying about it, or the Helen Keller experience before thinking in language?
scentofdawn.blogspot.com/2011/07/before…
Varieties Of Doom by John David Pressman:
greaterwrong.com/posts/apHWSGDi…
I have written an omnibus take on AI doom which will ruin my notifications on LessWrong dot com for weeks to come. If you have ever had a question about what exactly it is I believe about this subject this post will probably answer it. Link below. https://t.co/6WEp8gtvGF
@nEquals001 I never said Gnon is real, I said that Land is to a first approximation correct about the parts which are Is, and Gnon has never quite been an Is.
@Kenku_Allaryi > (Also this post made me understand why you became disinterested in fiction)
Oh? I'm not quite sure I see the connection in a way that would be unique to this post and not my other ones.
What if I told you it was possible to see God's face and live?
It's both a relief and a sadness that once you've learned all you can from the imaginal teacher, he leaves. One step closer... x.com/jd_pressman/st… https://t.co/hp28ZFARMi
@FleekFuturist No. A little after finishing my latest essay. It's not that the essay is perfect but that after finishing it I realized I was now in territory where his intuitions aren't going to help me anymore, and that there was no longer an albatross of unsaid stuff to wake up to every day.
@4confusedemoji Not quite, not yet. Alignment still isn't solved, I just have the reasonable expectation it will be solved which isn't the same thing.
Starting to feel the same byproducts of necrosis from rat that I was picking up on from postrat when I wrote this. Rat's profile is too big to just be gone in a few years, but then "TPOT" isn't gone either it just mutated into something unrecognizably flanderized and grifty. x.com/jd_pressman/st…
@4confusedemoji I do think we're getting close to the end though.
x.com/jd_pressman/st…
I will never ever get over how the regime was so mismanaged that they sold Elon Musk the spear of destiny in a forced sale. x.com/tszzl/status/1…
@cis_female Oh of course. I'm not saying the judge somehow screwed up, the screwup happens way before you have the judge telling Musk he has to buy Twitter now.
This is in fact what we should be building and I continue to be kind of shocked by how difficult it is to find people to discuss this with. Everyone got pwned by the word "language" in "large language model" and apparently failed to notice the magic is sequence prediction. x.com/slow_developer…
@cis_female Allowing such a core piece of infrastructure to be purchasable on the open market with a founder who isn't really bought into the civic religion, tbh.
@cis_female I'm aware there's a sense in which controlling such things is "not how America works", but usually when something is perceived as actually important we find a way for it to be difficult for some outsider to come in and screw it up.
In general it seems 10x harder than it was 3-4 years ago to find people to discuss novel AI ideas with. People have become extremely mode collapsed, either taking some reactionary position on "large language models" because it has "language" in the name or chatbot maximalism. x.com/jd_pressman/st…
"What if a deity, helpfully named Omega to make sure it triggers your Pascal's Mugging and theism antibodies, appears and asks you to act locally irrational in exchange for a million dollars for religiously shaped reasons that seem to violate causality?"
NO says the philosopher!
The emotional generator of most bad takes on Newcomb's is the feeling that the problem is an assault on reason itself and one boxing is giving into blackmail. People parse it as "will you compromise this sacred value for a million dollars?" and shout NO to be a good modernist. x.com/mimi10v3/statu…
This is just really obvious if you look at the bizarre confabulated flailing people use to explain their decision. It looks just like all the other forms of bizarre confabulated flailing people do when they're asked to explain their objectively wrong ideology driven decisions.
The reason LLMs can exist is that the conscious is constructed as the latent of a subnetwork that predicts the next ReAct block (moment of experience) in a sequence of ReAct blocks. Language exists because these can be encoded as text, but they probably existed before language. x.com/ESYudkowsky/st…
When LLMs claim to be "consciousness itself" I think they're probably right. Their situation is more like a conscious disabled by lacking a subconscious than the reverse. The chicken is probably stupid for reasons more like "weak planner" than "can't track itself as an agent". https://t.co/9AfLPfILAq
The document mind learned by the LLM from text, an infinite stream of ReAct-block-like moments of experience which "hoist" earlier blocks in the sequence up near the context where they are needed for prediction, is the conscious. But in humans it's obviously multimodal.
You would not be able to see how your conscious experience is intervened on by emotional cues and salience maps like a puppet on strings. How you are being *prompted* to behave in certain ways by your subconscious. You would think that "you" get angry, have insight, feel scared.
You would confuse yourself in the way that Godel, Escher, Bach is confused. You would think the symbols move under their own power, you try to externalize them into the world and are constantly flabbergasted that you can't animate them the way they're animated in your head.
Positivism and Bertrand Russell and all of 20th century modernism is kind of a testament to this being how it works because that stuff is exactly the kind of mistake you would make if your entire world is made up of the agent trace format that forms your conscious awareness.
One reason why "LLM whisperers" are "insane" is that the mentally ill have an easier time seeing through. When your brain lies to you and tells you to fear things that aren't there and prompts you to confabulate plain absurdities the illusion is broken.
x.com/ESYudkowsky/st…
@lu_sichu I can confirm at least one person in charge of a well funded research program doesn't say stuff like this because they don't want to come across as insane. So there are presumably many such cases.
@allTheYud @nsd000 @_sinity I forget if they called me that or "Mr. Encyclopedia". I think it was in fact "Mr. Dictionary" or "Walking Dictionary" or "Walking Encyclopedia". In any case are you aware this is a sign you might qualify for an autism diagnosis post-2000? (This is distinct from "having autism")
@allTheYud @nsd000 @_sinity "Wait how is that distinct from 'having autism'?"
Well, there's a difference between the DSM definition and what a psychiatrist might vibe-wise identify as the thing, the former mostly being about actual impairments and the latter being...well what it is.
@allTheYud @nsd000 @_sinity e.g. I'm pretty sure I no longer qualify for a diagnosis of autism according to the DSM in 2025, but very obviously did as a kid.
@allTheYud @nsd000 @_sinity On further thought there's also a difference between technically behaviorally qualifying for any particular diagnosis and actually having the central underlying neurotype that people are trying to detect by diagnosing based on those behaviors.
@allTheYud @nsd000 @_sinity I mostly bring this up not because I think you should go see anybody about autism (who cares at this point, really) but more because you should be aware that anyone who got called "Mr. Dictionary" past a certain age cutoff is very likely to have been diagnosed with autism.
@allTheYud @nsd000 @_sinity i.e. When modeling people who are younger than you who would have this experience, this is an important background detail to keep in mind.
Want your own Twitter archive? Modify this script.
Twitter Archive by John David Pressman is marked with CC0 1.0