Big Yud
Eliezer “Fat Yud” Yudkowsky is a hack fraud leading a sweaty, smelly sex cult of dysgenic nerds and low-agency tools. Unfortunately, I continue to hear his name among people with inflated salaries and senses of self importance. My objections to his brand of rationalism are too many to be summed up in a few social media posts, so this page exists for the categorizing of them. Hopefully, this post will protect me from tiresome conversations by being so vitriolic and repugnant that, upon being linked to it, any Yudkowskian interlocutors will slink away.
The Man
Eliezer has been a hanger-on of AI research for over two decades now, initially captured by the optimizer. He is also fat. He’s known for founding MIRI, formerly the Singularity Institute. During SI’s early years, he aped research happening in universities in an attempt to help accelerate AGI development - however, with a grifter and LARPer at the helm, the Institute’s research output was understandably sparse. Feeling the burn of imposter syndrome, and desperate to secure Thiel funding, Yud grasped for a niche. He turned to the same refuge of all substandard entryists - Human Resources, and Public Relations. In other words, having carved a modicum of reputation for himself based on little more than handshaking, he began to use it to lecture other, more talented researchers on what they ought not to do.
Yud’s first paper in 2001 contains his only idea:
The term “Friendly AI” refers to the production of human-benefiting, non-human-harming actions in Artificial Intelligence systems that have advanced to the point of making real-world plans in pursuit of goals.
It also contains his entire sales pitch:
What is at stake in Friendly AI is, simply, the future of humanity.
AI needs to help humans. AI won’t help humans by default. If anyone does this wrong, the world might end.
Yud spent the next 20+ years elaborating more and more complex ideas of what “helping humans” means and especially what “the world might end” means. I won’t pick on his 2001 paper anymore, since changes in the field have made everything from that era seem quaint. I will only ask you to do me one favor: scroll through it briefly, scan it if you will, and see if any algorithms jump out at you. Any pseudocode, any tables of results, any state diagrams - is that a flowchart? ah, it’s only a theoretical example of a flowchart, never mind. Any mathematical equations? None of those, but it does contain some book reviews. Do a search for the string “hypothe”, to capture all its derivatives.
This will be a repeating pattern.
The Myth
Yudkowsky, sometimes referred to as “Yud the Dud” or, less often, “Blub-cow-ski”, is good at constructing hypotheticals. I’m also good at constructing hypotheticals. I once imagined a future skin cream so good it could regrow flayed skin in seconds, and strapping Kim Kardashian to a giant spindle and skinning her alive, regrowing the skin as it was peeled off her body, unwinding a giant hundred foot long unbroken sheet of Kardashian-hide, which was tanned and made into exclusive luxury goods. But my being able to do a creative writing exercise doesn’t mean the world should pay me to sit on my ass and doomsay about the dangers of futuristic steroid cream.
To step beyond the sweaty cargo-cult leader for a second, and examine his even sweatier harem - let’s grab a more recent paper from his org, and give it a scan. Some actual state diagrams this time, so some progress has been made in the past 19 years! Details are still few and far between, especially when it comes to practical implementation of any of the 11 algorithms ostensibly covered in this survey.
Actually, Hubinger (who is not only sweatier, but somehow stupider) does kind of explain how these safety methods are implemented - by somebody talking. All 11 of his ideas appear to be variations of sitting someone down in a chair, and asking them hypothetical questions (I can’t say for certain, as some time through skimming section 5 I was interrupted by muscle cramps from the constant compulsive wanking motions my left hand was making). He does not discuss whether this has something to do with data pre-processing, or is only applied during fine-tuning.
Specifically, we want Amp(M) to look for deceptive or otherwise catastrophic behavior in M by utilizing things like transparency tools and adversarial attacks.
Thanks for the specificity, Evan! His only concrete reference for “transparency tools” is the Inception project - OpenAI’s work making a fully explainable neural net. How this is possibly practical or relevant to the issue of, say, deploying GPT-3’s 175B parameters, Evan is too modest to say. “Adversarial attacks” receives no reference number at all - presumably the research this term is based on is so ubiquitous among Big Yud’s Ugly Sex Cult MIRI Fellows that it doesn’t even merit a footnote.
[…]it seems unlikely to me that it would successfully catch deceptive pseudo-alignment, as it could be very difficult to make transparency tools that are robust to a deceptive model actively trying to trick them.
Aha! A Bayesian claim - I recognize these. Go on then, Evan, provide any justification whatsoever that “it seems unlikely” to- oh, woops, never mind, that’s the end of the section.
Yud and the MIRI Morons deal in unjustified claims constantly. It’s a defining feature of their work - open any one of their papers or blog posts (oh god their blog posts) and you’ll see a stream of big sciencey sounding words that, upon closer examination, are nothing more than baseless postulates on subjects they seem to only fuzzily understand. Everything is high level. Everything is “to be elaborated on later”, “to be expanded” - by someone less important (and more mathematically capable) than any of the important sweaty cult nerds. They are much too busy proclaiming the end of the world and running giant money laundering operations to explain exactly how “Amp(M)” will “look for deceptive behavior” “utilizing things like transparency tools”.
Now, look - I get it, not every important paper has to include math formulas or practical implementation to be scientific. Here’s OpenAI publishing essentially a blog post on arxiv, just like MIRI fellows do. And here are the reasons their opinions matter. Fuck it, this paper is OpenAI doing MIRI’s entire gimmick, except semi-competently. It’s not one of their best, but look at that! It’s concise, elaborates on claims made, and names specific tools, methods, and applications for each area of concern! Moreover, these are all issues that map onto something that exists in the real world. Hardware security is a real thing. Differential privacy is a (kind of) real thing. “Alignment” is not.
MIRI has virtually no hard science to their name. In fact, the only paper of theirs my review turned up with any rigor whatsoever was from a grad student they poached out of UC Berkeley. He went back to Berkeley within two years.
The Legend
If it feels like I’m jumping around too much, I apologize. I’m trying to break down a cadre of lunatics who have been vomiting words onto the internet for two decades - it’s a lot to encompass in a short blog post. Let’s bring things back to Yud.
Yud spins hypotheticals. Take, for example, this recent podcast appearance. (Fair warning, actually clicking that link and playing the video will display Eliezer’s hideous gremlin visage on your screen.) He starts by fluffing his ego. Next, Yugly malappropriates the concept of generality (an ML research term meaning how successfully a single model can be applied to multiple tasks) to biology. ChatGPT, he says, is not a scary general intelligence.
But can you imagine - what if there were???
That’s as concrete as he gets.
The unfounded assumption that generality applies equally between systems of linear equations and animal brains. The unfounded assumption that human consciousness is an artifact of the brain exclusively and deterministically, and so there is nothing separating silicon from carbon. The unfounded assumption that “self-improving AI” is possible or desirable. Yud builds science fiction narratives on a hotbed of unfounded assumptions, the sort that “sound reasonable” to a certain kind of self-important Bay Area atheist only because they already believe them. Eliezer “Face Only a Mother Gopher Could Love” Yudkowsky exploits these naive and usually unexamined beliefs to full effect, taking them to their absurd conclusions with a straight face. He constructs elaborate glossaries of self-defined technical terms - a common sign of religious fraud a la Scientology, or else just a garden variety nutter.
But Yud has had too many followers for far too long to be considered a harmless kook. He clearly knows how to manage a cargo cult. Some members he cows with complex incantations of self-referential liturgy, completely and intentionally at odds with any mainstream research it might be compared to. Some members come hoping to cloak themselves in the same pseudoscientific gibberish, seeking a shield for their stupidity and lack of mainstream success. Some are no doubt only casually interested, and not very bright. All of them have far too much money. And they smell bad, and look hideous.
Artificial Intelligence is an exciting subject to think about. It might even be real some day (ML models like ChatGPT, which perform well defined operations only when commanded, are closer to Notepad than Skynet). But if an artifical general intelligence singularity does happen, Ugly Yud’s only material contribution in any regard will have been to slightly accelerate it.
eliezer has IMO done more to accelerate AGI than anyone else.
— Sam Altman (@sama) February 3, 2023
certainly he got many of us interested in AGI, helped deepmind get funded at a time when AGI was extremely outside the overton window, was critical in the decision to start openai, etc.