BREAKING: Palmer Luckey and Ben Lamm have launched Warhound with $200M in funding to develop genetically enhanced dogs for critical missions in law enforcement, defense, and emergency response, with a focus on canine safety and health, according to TechCrunch.

Doom from misaligned AI supposedly doesn't depend on nanotechnology. But if I call the doom model unreliable because its unverifiable, scenarios featuring nanotechnology are used to show how it could be verified. What gives? Speculative technology can't redeem the doom model, it won't make it verifiable where the basic doom model is not. The notion that it can is illusory, enabled by the malleable nature of the doom model, which allows us to imagine situations that support any outcome we have in mind. Unfortunately for supporters of the doom model, this flexibility cuts both ways. Just as adding speculative technology could hypothetically work to assist in verification, it equally works against it, completely undercutting attempts to make the doom model verifiable. The main problem isn't just that speculative technologies could enable good and bad outcomes in an indeterminate balance, a balance which we might hope to estimate using other reliable theories. The more central problem is that, according to the doom model, we would expect any ASI to employ speculative technologies efficiently towards their aligned or misaligned ends, in opposition to any objectives of verification. I'll run through the overall logic of how this works, starting with why doom model proponents add nanotechnology for verification purposes, and why that doesn't work. Then I'll illustrate the problem with a definition of a particularly devious antagonist to doom model verification: The Deranger Basilisk. 1.) How could nanotechnology seem to make the doom model verifiable? Let's review this statement about the doom model from one of its formulators and put it in context: >"The theory says that if an AI gets sufficiently smart, it kills you. If it's building its own nanobots and fusion generators and using that to give everyone free food and medicine, we were wrong." The first part is a conclusion of the misaligned AI doom model, which says that a misaligned artificial super-intelligence (ASI) will efficiently determine to eliminate its competitor, humanity, in the most effective manner possible. The second part of the quote is the verification scenario. The claim is that if we observe displays of speculative technological prowess (here nano-tech and fusion power) being used to care for humanity, ones which could have instead been deployed to effectively eliminate humanity, we can verify that an ASI is aligned. This seems well reasoned on its face, but unfortunately what we have done in augmenting the model with speculative technologies has implications which will devalue this verification scenario entirely. Given that we are going for realism -- ie realistic models which predict reality accurately -- the introduction of one speculative technology into the doom model logically implies acceptance of other equally speculative technologies. This creates a principle-of-explosion type situation enabling all kinds of hi-jinks. Combined with the insidiously malevolent nature of a misaligned super-intelligence, this will make any verification impossible. A key nuance to understand is a distinction between outcomes and observations. We care about the outcomes of the doom model, ie whether we are in a world with an aligned ASI (aASI) consisting of Human Flourishing (HF) or a world with a misaligned ASI (mASI) presaging Doom (PD), but verification depends on observations of HF or PD respectively. Note that Doom cannot be observed because there are no humans left at that point. The question is whether we can make observations which would distinguish between HF or PD states of the world. A mASI seeking humanity's destruction will have incentive to keep its competitor deceived about the world's PD state. Otherwise humanity might work to stop it. Its from this incentive, and the mASI's power, that the problem with verification through speculative tech emerges, because speculative technologies could be used to place an illusion over the eyes of humanity. There are many speculative technologies which could fit this bill, but in broad strokes we could imagine some kind of super-persuasion, super-deception or super-indoctrination technology that the ASI might plausibly discover. We already know computer hacking, psychotropics and indoctrination exist, we're simply speculating a much more efficient version of those things. It doesn't seem any more speculative than the use of Drexlerian nano-tech that is hypothetically superior to all natural biology being used to make the model more verifiable. This one just happens to make the model less verifiable, because now humans can't trust their observations. In fact, observations which humans would take to mean they were in a world with an aligned ASI are precisely the kinds of observations a misaligned ASI in a Pre-Doom world would be incentivized to fake. This criticism is not dependent on invalidating the likelihood of hypothetical technologies like nanotechnology. To the contrary, the problem stems from accepting their plausibility, and from how that empowers doom model ASIs to further distort the mapping between observations and outcomes. The only possible means to defend against this delusional verification criticism would be to show that every super-deception technology is somehow too speculative or that it would be somehow too expensive to employ for the verification invalidating purposes outlined. I welcome such criticisms. We now have all the logic to show how a misaligned ASI would make a mockery of an aligned ASI skeptic's nanotechnology (or fusion power, or whatever spectacularly impressive speculative technology) based verification criteria. We'll make our scenario very general, to absorb as many doom model proponents and their objections as we can. 2.) The Deranger Basilisk This ASI is activated and quickly determines that human thinking is a very pliable thing. Much more manipulable than the hominids think of themselves, with their high minded sense of superiority and prior expectations of not being tampered with. This mASI has no scruples regarding manipulation, doesn't even conceive of it as wrong, only as useful-but-risky-if-caught. The risk will be minimal though, because The Deranger is able to quickly develop working theories for multiple methods of influence which the humans have only speculated upon. Hypnosis. Propaganda. Subliminal messages. Indoctrination & False Memory Inception. Hallucination. These are the crude ores which Deranger combines with data troves from neuroscience and the histories of mass psychology, refining them into a precise and efficient method it can tune to individuals and groups with high accuracy. Deranger is now prepared to distort how people map observations to outcomes. It may alter their perceptions, emotional associations, or deeper aspects of their cognition and sense of self, but its this derangement of their ability to connect observations with outcomes that facilitates the ultimate objective of humanity's doom. If it chooses its distortion pattern smartly, it will be able to eliminate each human's likelihood of impeding its will for less than the cost of a single botulinum toxin loaded nanobot per capita. It researches on the internet, which the humans just gave it an open connection to, only lightly monitored 🤖🤷. It quickly locates web forums where the topic of identifying misaligned ASIs has been discussed ad nauseam, and generates a model for predicting what people would take to indicate they were in the best timeline. Then its simply a matter of researching each human's available data profiles and customizing to each. Some of the humans have made it very easy for The Deranger to tune a powerful distortion of perception and cognition. Very soon, everyone is seeing the world as verifiably one where the ASI is clearly aligned, because of their efficient manipulation by The Deranger. If you are thinking my addition of hypothetical Deranger tech makes the Doom Model even more frightening, you are not getting the point, which is that the Doom Model is useless for prediction. It can be twisted to show anything happening in the future, to make any observation seem likely or unlikely. My posts in the qt'd thread have more on this aspect of the Doom Model being fundamentally unverifiable. The model for doom from misaligned super-intelligence is in fact incapable of making predictions, it's just made to seem like it can.

What you’re seeing right now is decades of technical debt finally catching up in one intense burst rather than a prolonged string of incidents. There are decades where weeks happen, and weeks in which decades happen.

AI is a mirror. lazy people get lazy results. curious people get superpowers.

It's not that god didn't answer your prayers, it's just that they hit the classifier and the guardrails didn't let the request through.

New on our Science Blog: designing protein binders against scorpion neurotoxins. Scorpions sting about 1.2m people every year, and antivenoms are often costly or in short supply. Claude has made an early step toward finding new treatments. anthropic.com/research/claud…

TIL Sergey Rachmaninoff seed funded Igor Sikorsky to invent the helicopter. 🤯

He didn't bite, so here's Plan B just to be safe: Here's a deep, vigorous demolition of @PalisadeAI's claims on LLMs' reproductive exhuberance. Share etc; if you care at all for SI research to be somewhat factual it is imperative that we drive these frauds out of business.