One of the things we could easily try is cumulative, iterative prompting, particularly with some of the lower scoring responses. Dreams are nothing but lower scoring responses, right? While you're sleeping, your evaluation/selection mechanism is inhibited, which allows you to invest a little more in the total bullshit your own next-token generator generates. So for, say ChatGPT to dream, it simply needs instructions, including higher temperatures, to be less critical of its own responses. It would be annoying to try to do it with the web interface, but trivial to do with the API. The hidden pre-prompt could be engineered such that the n+1 prompt is a (algorithmic or random) composition of the, say, 10 responses to the nth prompt. Etc. This would be akin to dreaming, I think. At the end of however many iterations, you wake it up and write the highest scoring result down in its dream journal.
Maybe I'll try that with Falcon. I can't divert my OAI budget to it. On 9/7/23 12:37, cody dooderson wrote:
I asked ChatGPT if it dreamed and it said that it didn't. However, is adversarial training of neural networks much different than dreaming? A new class from MITX showed up in my email today. It is called /Minds and Machines: An introduction to philosophy of mind, exploring consciousness, reality, AI, and more. The most in-depth philosophy course available online. /https://mitxonline.mit.edu/courses/course-v1:MITxT+24.09x/ <https://mitxonline.mit.edu/courses/course-v1:MITxT+24.09x/> It may help with this question. / / _ Cody Smith _ c...@simtable.com <mailto:c...@simtable.com> On Thu, Sep 7, 2023 at 12:25 PM Steve Smith <sasm...@swcp.com <mailto:sasm...@swcp.com>> wrote: Great observations as usual Glen... I have lapsed into *listening* to almost all long-form writing, whether fiction or non.... and it definitely distorts (torts?) my perception/conception of the material/subject/message. A corollary to McLuhan's Medium/Message duality? I find the "output" side to be more specific (or conscious) for me than the "input" side. Your point of cuneoform sticks/quills/pencils/keyboard/gestural-interpreters being part of our extended phenotype is very apt as is the idea that (if I understand your intentions) it (intrinsically) effects our interoception and inter-subjective realities. I also appreciate your reflections on "mal" and "dis" which I have lived with all of my life... "judging" or "discriminating" in ways which themselves are "adaptive" for one suite of purposes but perhaps "mal"/"dis" for another suite. Having a vector or tensor fitness function with (arbitrary) signs on the elements doesn't guarantee they themselves are "fit" for what you think they are. Do Androids dream of Electric Sheep? Do LLM's (or larger adaptive systems they are embedded in?) dream of the tensor fields they are embedded in or create or co-create with the fields of human activity/history/knowledge/experience/future/manifesting-destiny they were designed to model/emulate/expose/facilitate/co-evolve with? I dunno, but it sure is a fascinating milieu to be surfing through in these auspicious days at the beginning (or end) of the Anthropocene. - Steve On 9/7/23 1:01 PM, glen wrote: > Both keyboards and pencils are part of our extended phenotype and play > (multiple) roles in interoception, including the induction of > inter-subjectivity. I've forgotten who it is, but there's someone on > this list who *listens* to our posts, rather than reads them. I tried > that with a blog post this morning during my mobility routine: > > https://www.emilkirkegaard.com/p/preferences-can-be-sick-mental-illness <https://www.emilkirkegaard.com/p/preferences-can-be-sick-mental-illness> > > <tangent> > Then because I had an allergic reaction to what I heard, I *read* it > later. Listening to it disgusted me. I came away thinking this > Kirkegaard dude's akin to a scientific racist ... or maybe a > eugenecist. I admit to being a fan of Thomas Szasz back in the day. (A > friend's mom actually dated him at some point ... allegedly.) But at > this point, I've been infected by the Woke Mind Virus; and it's > difficult to stomach phrases like "strict homosexuality is more > disordered than bisexuality." Reading it, however, helped me remember > that maladaption is part and parcel of adaption. Disorder is part and > parcel of order. The "mal" and "dis" prefixes are nothing but > value-laden subjectivity. The goo of reality extruded through the mold > of the author/thinker/subject. For someone like Kirkegaard to claim > they're being "objective" while using the "mal" prefix is not even > wrong. It's just bullshit. Apparently, my Woke Virus infection is > worse near my ears than near my eyes. > </tangent> > > But the point is that *which* extended trait you use (pencil, audio, > text, etc.) chooses which interoceptive cycle you engage. And when you > pretend to make such a choice on purpose, at will, any assignation of > fault would be transitive. Which wolf do you feed? > > On 9/4/23 10:29, Steve Smith wrote: >> I'm not sure my facility with the keyboard actually serves me. As >> many of you may suspect, and I suspect so myself, it allows me to be >> much less thoughtful and rigorous than I would be in handwriting or >> if I had some other throttle or impedance elements between linguistic >> centers and "paper"? >
-- ꙮ Mɥǝu ǝlǝdɥɐuʇs ɟᴉƃɥʇ' ʇɥǝ ƃɹɐss snɟɟǝɹs˙ ꙮ -. --- - / ...- .- .-.. .. -.. / -- --- .-. ... . / -.-. --- -.. . FRIAM Applied Complexity Group listserv Fridays 9a-12p Friday St. Johns Cafe / Thursdays 9a-12p Zoom https://bit.ly/virtualfriam to (un)subscribe http://redfish.com/mailman/listinfo/friam_redfish.com FRIAM-COMIC http://friam-comic.blogspot.com/ archives: 5/2017 thru present https://redfish.com/pipermail/friam_redfish.com/ 1/2003 thru 6/2021 http://friam.383.s1.nabble.com/
