eve-meth

My account is an experiment in overcoming this reflex

You've heard of Hume's fork and Hume's guillotine but have you heard of Hume's 2001 Honda Accord with which he killed a family of 5 while high on ketamine.

Kinda funny that GOSPLAN sounds so similar to "God's Plan".

Well many of the claims from organizational biology would say that it is not possible for the current models to have (true) autonomy or motivation or cognition. My argument is that even if this is the case, then it doesn't matter that much for solving immediate engineering problems when we can have superficial copies of them, like model routers or the auto/adaptive thinking features. With the strawberry example, we can bolt on some module that makes the LLM reason through it or write a script to solve it and that will be good enough. There is probably some much harder analogue to the strawberry problem that exploits the limitations of LLMs and just isn't known yet, and when that time comes we can also add extra epicycles to the models so they can solve it. I do think this fundamentally limits models, just unsure of the timescale when it will become relevant.

I did miss some of that in my reply so I appreciate the clarification. Bio-inspired would be an accurate description of the anthropic autodream/autosteer. It's difficult to model stuff like REM with our little understanding of it. after the recent opus 4.7 release, it was widely discussed it was failing the `how many r's in the word strawberry` test. After this people were criticizing the test, making up one reason or another, and making analogue comparisons to human senses that were in my opinion not accurate. I'm not familiar with grading complexity of logic/math problems (other than, like o notation). But I've been thinking about your post that autonomy/cognition maybe don't deal exactly with computability. It's an interesting concept, especially when you consider that llm 'thinking' isn't really the same as human thinking. The profit incentive for these things are immense. It would be really strange if any company went out of their way to invest gpu time on training autonomy into a model. Autonomy and motivation do improve performance at tasks, though. For example, there really isn't any way to modulate how much energy/time a model spends on a problem. Currently this is solved outside the model using stuff like model routers. But things like cognitive gating to determine what features to pay attention to during a task, or other types of actual thinking are just outside of the realm of possibility for current models. Back to the strawberry test - I am fairly sure the problem gets fixed through training to do something like s-t-r-a-w-b-e-r-r-y in the thought traces, but that doesn't solve the fundamental fact that intra-word features are not inspected like a human would do intuitively.

The Straussian read of these tweets is that my research program is not going well 🫠

note to advertisers -- this ad is very persuasive to me.

i like how the lidl ads I get are just "here's a product we have".

@eshaanissar longpost ping btw

I think bio-inspired research has been very successful. I'd argue basically anything neural network based (~ all of AI now) is bio-inspired. I just think it's questionable to say that, moving forward, the general direction of improvement is "let's try the same (===) things that biology does". To use autodream as an example, we can say that it is inspired by REM sleep, but since we don't fully understand REM sleep in the first place, in practice the implementation of autodream will be an engineering problem. They will not be looking at neuroscience data or MRI scans to improve or refine autodream, and they probably shouldn't be either. The idea that autodream works is somewhat independent of whether REM sleep actually works in the same way. So I think the ideas can be good. It just won't say much about biology, and biology can't help it that much either beyond the initial idea. Which might have been arrived at without biological knowledge anyways. The concerns of theoretical biology (by which I mean, organizational/relational biology, or biological cybernetics, or whatever. the terminology is not very good atm) are pretty far from these engineering problems and in the near future will likely stay far, which is the argument in my screencap above.

@meadandjuniper Medi: middle terra: earth nean: suffix

I always forget how to spell mediterannean. Mediteranean. Mediternean. Fuck it

@lu_sichu @WKCosmo always one of my favorite fun facts that Walter Alvarez discovered the layer of iridium that contributed to the asteroids killing dinosaurs hypothesis, and that his father Luis Alvarez's physics connections were crucial for analyzing the evidence.

@WKCosmo ohh kinda of like this youtube.com/playlist?list=… also yes that alvarez's kid

One of the things that left me with the impression I have about anthropology is that every conversation I have ever had with an anthropologist on the subject of human evolution has gone more-or-less like this.

This begs the question then of, if not probability theory, then how exactly do we talk about things we are uncertain about?

@adam_kranz yeah i was interested in consciousness for some time but have mostly given up on it as a scientific problem to work on. I don't think there's any real point to it when there are much more basic things we still can't really explain. I have some hunches of course though.

How I feel about philosophy in general tho consciousness in particular is indifferent to me. Part of why I've put it on the back burner, too emotionally fraught for no clear reason

@eshaanissar Constantly being told "think before you speak" oneshotted my entire social life

I wish I read this when I was a teenager. "... relatively automated motor processes can be disrupted if they are run using consciously accessed, task relevant declarative knowledge..." I used to create all these bad heuristics when remembering things - like 'between these two things, it's your *second* instinct, don't go with your first.' Which made me a really good test taker with a minimal amount of studying. But set me up for horrific failure later in my academic career. this is a prime example of the 'profound' and convoluted stupidity which arises past mid-wittery.

@adam_kranz visa_scylla_charybdis.png strikes again

Panpsychism is the inverse of the "cosmic insignificance" error. Instead of saying "the size of the universe makes me feel so insignificant," it says "significance feels so big it must be the size of the universe."

@eigenrobot well there was that guy so into paws that he double amputated his own hands via freezing to attach paws to them.

are there furry foot fetishists who are like into cartoon anthropomorphic animal feel. like some people are into lizard cartoon feet and "hoofies" are into horse feet and so on

@BecomingCritter Glad to see that my dose of 0.5 mg which I derived from self-experiment was closer to natural levels. I just buy the 1mg tablets and split them.

You're taking way too much melatonin Supplements are sold in the 3-10mg range. Initially MIT researchers found that tiny doses of 0.1 to 0.3 mg helped induce sleep while maintaining healthy blood melatonin levels. They then patented that range. The patent combined with limited FDA regulation of supplements and a Vit C-esque 'more is better' psychology push companies to go over that dose 3-5mg doses push melatonin levels well above their natural range Here's the solution:

@dioscuri goat

Big personal news: I’ve been recruited by Google DeepMind for a new Philosopher position (actual title), focusing on machine consciousness, human-AI relationships, and AGI readiness, starting in May. I’ll continue my research & teaching at Cambridge part-time. Absolutely stoked!

@skooookum Jmail is exempt

“I { built | vibecoded } a { website | dashboard } that lets you { interact with | visualize } <some publicly available data>”

@lu_sichu so do you think this approach is fruitful? Does it solve problems or characterize systems in a way that was difficult before?

@zermaelian youtube.com/watch?v=wSWmHZ…

I just finished Stuart Umpleby's 6-hour "Fundamentals and History of Cybernetics" lecture. One thing that really sticks out to me is the optimism that eventually the systems of interest could be formalized, when it has turned out to be massively difficult. From this perspective it makes a lot of sense that cybernetics as a field splintered into ten thousand pieces as practically applying those ideas to specific systems would mean that the details and idiosyncrasies of those specific systems would dominate the analysis more than the original unifying ideas.

@lu_sichu ur about to derail my phd for the nth time now i'll have you know

@zermaelian math.ucr.edu/home/baez/netw…

@lu_sichu oh is that so. this is very interesting. I've never heard of this before.

@zermaelian I feel like topos insitute and category people have pretty much inherited this legacy and done the cybernetics formalization

Though my initial read of how Heidegger uses the word "cybernetics" can be swapped out with the climaxing of the general rationalist-reductionist-empiricist-materialist-scientific juggernaut, and "cybernetics" was just the stand-in he used for his time.

I also find the splintering of cybernetics interesting in light of Heidegger's claim that the completion of cybernetics is the end of philosophy. If it turns out that, for cybernetics, "the center cannot hold" and it could never convincingly transform itself into a totalizing framework, then I wonder if this would change his ideas.