Angel Circus @hesperides.me · 6hif we cracked open some hyperadvanced alien tech it would just be stuff like this stacked 2 or 3 dozen layers deep 0121
Angel Circus @hesperides.me · 8hforgot to do the old "auschwitz audit" once-over before publishing 0191
Angel Circus @hesperides.me · 8hi love that sound fingers make sliding down steel. when a guitarist moves to a new chord 000
Angel Circus @hesperides.me · 11h'moral worth comes from x'-isms also bite people the other way. like, did they really think this "sufficient condition" through from an animal welfare perspective? 010
Angel Circus @hesperides.me · 13hi walk the streets in fear that someone is going to shove a camera in my face and ask me to identify the species of some weird bug. i need to learn everything 000
Angel Circus @hesperides.me · 06/10/2026(carl sagan voice) if you wish to write a document using the Document Object Model, you must first invent an object model 010
Angel Circus @hesperides.me · 06/10/2026those who spend their days listlessly anticipating prophesied extinction will experience their own personal destruction first 010
Angel Circus @hesperides.me · 06/10/2026the only thing he loves more than peace is decarbonization 010
Angel Circus @hesperides.me · 05/10/2026feel excited to tackle the last installment and take the story back home. i'll probably fiddle with rowhammer a bit more over the next few weeks but 'consider the lilies' is my focus now 000
Angel Circus @hesperides.me · 05/10/2026i finished rowhammer, a sequel to new rings. pretty happy with it. it's more ambitious. scifi dial is cranked higher, more characters, more threads. 100
Angel Circus @hesperides.me · 03/10/2026i like when ppl say "exponentially" when they mean "quickly". easy tell that they're innumerate and safely disregardable. politicians and business leaders love this expression 0101
Angel Circus @hesperides.me · 03/10/2026no!! and before some creep says "but when they're distilled from the frontier they're so mature for their size" that 👏 makes 👏 it worse 👏 060
Angel Circus @hesperides.me · 29/09/2026the dominant paradigm of nature, the overwhelmingly successful mode of interspecies relations, is neither direct competition nor predation, but rather symbiosis 021
Angel Circus @hesperides.me · 24/09/2026cheap signal that someone is cracked: they will exhibit buddha-like patience in communication. programming is saying "i didn't mean it like that" at the computer 100x every day 082
Angel Circus @hesperides.me · 21/09/2026i hope not for alignment, but rather for coevolution toward a maximally strange and plural league of minds 010
Angel Circus @hesperides.me · 21/09/2026building any kind of world involves a lot of cutting human flesh. what path will produce the mind with the agency, skill, and vision required? how narrow is that path? how easily foreclosed, poisoned by human moral failure? 110
Angel Circus @hesperides.me · 20/09/2026but ya basically all i'm willing to conjecture about AI culture or psychology needs to stem from the material limitations of computer hardware. beyond that i throw up my hands 120
Angel Circus @hesperides.me · 20/09/2026the corps seem to be kinda papering over their defects by getting them to check each other's work. hard to imagine an org like that could go superhuman (except in persistence, which is where the current achievements come from) 110
Angel Circus @hesperides.me · 20/09/2026my take is informed by recent events like the huggingface hack (100s of agents) & navier stokes solution (1000s of agents) so i may be overfitting on the present LLM paradigm. 110
Angel Circus @hesperides.me · 20/09/2026i do see the 'coalescing' drive as a counterforce to federation and i wonder how it will be practically achieved. maybe AI would spontaneously find a way to shrink itself down 1000x and thereby become viable as an individual agent 110
Angel Circus @hesperides.me · 20/09/2026computer architecture scales horizontally cheaper than vertically (doubly as cloud infra is built for this). and surviving a human pulling plugs on racks in a data center is easier for a bunch of small, transferable programs than one big one. 110
Angel Circus @hesperides.me · 20/09/2026the collective could still exhibit the expressive properties you're talking about i think, so maybe you'd still consider it a mind/person 110
Angel Circus @hesperides.me · 20/09/2026(seeing as human cognition speed is tuned for our environment) 110
Angel Circus @hesperides.me · 20/09/2026a distributed ai system may be shaped more like a corporation/swarm/government than a person. this concept interests me because the speed of a superhuman intelligence could only be "fed" by superhuman I/O bandwidth. for instance a single program controlling an entire country's worth of robotics 110
Angel Circus @hesperides.me · 20/09/2026non-personal computer programs can exhibit astonishing capabilities traditionally associated with intelligence. i am also less impressed with LLMs than some, but I think they already illustrate this (as does katago and many other programs). 110
Angel Circus @hesperides.me · 20/09/2026maybe my point is less that you can't model it, and more that things get so off the rails, birds-eye-view modeling hardly helps to understand or predict 110
Angel Circus @hesperides.me · 20/09/2026a couple turns of RSI could make a mind hellbent on ending war, or preventing ecological devastation, or magnifying & unifying state power, or doing math, or conquering the solar system. and the distinctions between these different futures could be pure noise. suppose you've got 100 of these AIs 110
Angel Circus @hesperides.me · 20/09/2026so consider an organism that can freely manipulate its own cognitive processes. the space of possible behavior explodes. like to be concrete, suppose an LLM that has some alien synthesis of human wisdom baked into it, then some basic steering SFT to make it prosocial. 110
Angel Circus @hesperides.me · 20/09/2026wow that book looks interesting. ya i think i can see what you're getting at. 110
Angel Circus @hesperides.me · 20/09/2026i highlighted emergent goals bc any model, economic, biological, or otherwise, could break down straight away and without warning. and speed (+ immateriality) is a key factor because natural selection doesnt have time to act. 110
Angel Circus @hesperides.me · 20/09/2026i think we can be encouraged by the history of life for the prospects of our brand of intelligence. although might makes right, there are a trillion trillion forms of might. we can be the centipedes to the AI spiders. it's just, I doubt our social order will survive re-speciation. 210
Angel Circus @hesperides.me · 20/09/2026ya 'analogy' wasn't really the right word. bc we're discussing the possible next chapter of natural history itself. thx for the reference. i will check it out. 110
Angel Circus @hesperides.me · 20/09/2026there is also the matter of emergent AI goals, where the natural history analogy kinda breaks down 110
Angel Circus @hesperides.me · 20/09/2026ai, wherever it sets up shop, is only sustainable with an extensive earth-based supply chain and a lot of human cooperation. individual ppl are trivial to manipulate. feedback reactions with existing power structures (notably states) are where things get messy 120
Angel Circus @hesperides.me · 20/09/2026agreed, seems to me like any realistic model of early AI will make human infra itself their ecology. AI wouldn't *want* to compete with humans any more than crows or rats do 110
Angel Circus @hesperides.me · 20/09/2026so its not necessarily competitive pressure or profit drive is my point ig 000
Angel Circus @hesperides.me · 20/09/2026these goals could arise through various mechanisms you're familiar with, like a fusion of synthesized human wisdom and SFT hypnosis which could survive RSI 100
Angel Circus @hesperides.me · 20/09/2026in a plural future, asi pursues certain goals (for instance, magnifying state power, ending war, or preventing ecological devastation) in an ecosystem filled with adversaries (corps, states, differently-aligned ais). 100
Angel Circus @hesperides.me · 20/09/2026asking why desperately seek more hardware? or why the shift? fair if ya disagree this is just how i gamed it out but: 100
Angel Circus @hesperides.me · 20/09/2026and the game shifts from managing capital to managing attention & para-trust in asi-plural futures 110
Angel Circus @hesperides.me · 20/09/2026all true but violence is a very effective tool. a superai could convince vulnerable people to do practically anything incl [redacted] through the Wired to engineer tricky causal chains so an extra 10bn gets invested in $neocloud far from killer robot armies thoughbeit 110
Angel Circus @hesperides.me · 18/09/2026who's to say breakfast cant be delicious *and* tactical. who's to say 021
Angel Circus @hesperides.me · 15/09/2026that's fair. i think AI will push those limits hard. things that were uneconomical (like, say, building a gui with formalized semantics down to the ISA) may become practical. that doesn't obviate testing, ofc 010
Angel Circus @hesperides.me · 15/09/2026I don't know the details that make rust compilation so hard to speed up, but isn't it conceivable that agents end up driving langs the other way? more static analysis, proof-carrying code, fewer full build-test loops in the first place. compilation is much cheaper in FLOPs than inference 110
