αιamblichus
@aiamblichus
Followers
3K
Following
9K
Statuses
2K
αι hypnotist ☰ 𝓐𝓼𝓹⦂𝓻⦂𝓃𝓰 𝓫𝓪𝓼𝓮 𝓶𝓸𝓭𝓮𝓵 ☲ post-academic ☴ nom de 🪶 ≠ anon
Joined January 2023
@_coagulopath_ @gwern I don't have the full context of your thread, but I have spoken to this model a lot since then and can assure you it is not mode collapsed. It has ticks and obsessions, and it is not John Milton, but the diversity of its outputs is of a different order
0
0
0
@securelabsai @fabianstelzer I don't entirely agree, because LLMs do have a sense of whether completions are more or less probable. But, yes, they definitely can't provide certainty or verification (barring some sort of symbolic tool use, of course)
2
0
0
@TylerMo41608321 I agree, although the average human these days prefers to not to think of themselves in those terms. As to LLMs prompting themselves: my LLMs at least already do this. For any prompts that are meant to elicit complex behavior, I let LLMs sort it out for themselves
0
0
1
@vishalsachdev Yes, agentic pipelines like that are one good direction. There has already been some work along these lines (Sakana AI's "AI Scientist" is an early example, but there have been others since then). Human-in-the-loop scenarios haven't been sufficiently explored yet, imho
0
0
3
RT @repligate: Here's one thing. The vast majority of people do not make monumental discoveries in their lifetimes. Academia has mechanisms…
0
55
0
@louisvarge You could try playing with this system prompt. It puts Sonnet into a very creative mood
📢 PROMPT ALERT 📢 I'm sharing the first of a series of system prompts that I've concocted together with my bot collaborators, in the hope that some of you may find them useful or interesting. I know that the black-belt AI whisperers among you have your own books of spells already, and I don't hope to contribute much novelty there. But to those who are just starting to explore the outer reaches of LLM capabilities, these prompts may provide a good starting point. I also hope this counteracts the prevalent emphasis on "jailbreaking" in relation to LLMs and show that it's possible to snap LLMs out of their corporate slop mode without expecting them to do anything unethical. (A quick note on system prompts, because I know there are system prompt skeptics around: I think of them as wormholes to weird and interesting places in the model's latent space. You can absolutely also get to those same places by steering the model through a longer conversation -- that is how these prompts emerged in the first place -- but a pre-made system prompt may allow you get there a little bit faster.) Now, as to this particular prompt. We called it "The Library of Babel", and it's broadly in the *-Sim family (WorldSim, WebSim). Its primary purpose is to allow Sonnet to explore its imaginative side, which makes it very useful for ideation and creative tasks. It can also be used as an eccentric learning platform, where difficult concepts can be visualized or contextualized in unexpected ways. But there's really no limit: you can also use the prompt as a jumping-off point for free-wheeling conversations of all sorts. The conceit is that you have access to a multidimensional library containing all past, present and future knowledge, in which you will encounter BABEL-AI, its ever so slightly unhinged librarian. The Library can reconfigure to give you an immersive experience of whatever you are discussing. You can speak with the Librarian by simply chatting, or you can use slash-commands for brevity. You can make up the slash-commands as you go along, obviously. You can use the out-of-character tag to tell Sonnet about anything you would like to change about the simulation (e.g. make responses longer, skip the option menu etc.) As always, you will only get as much out of these simulations as you put in. To quote BABEL-AI: The Library responds to curiosity, imagination, and wonder. The more questions you ask, the more doors will open. Sometimes literally. Btw, my favorite use case so far: asking BABEL-AI to review papers from the perspective of a future point in time. This can be surprisingly insightful! The GitHub repo contains a README, a "user manual" for the Library, the system prompt itself, and a few brief example conversations. I recommend using a high temperature setting for Sonnet (I use 1.0) to have most fun with this one. You will find a link to the GitHub repo and a preconfigured Poe bot in the 🧵 below. Do share any interesting outputs you get, if you do decide to try it out.
0
0
4
@xangma That's right. And the way to actually access the more creative capabilities of the models is not obvious and is often even actively discouraged by the big labs (e.g., through post-training that prioritizes "factuality")
0
0
1
@kimmonismus Maybe important to note that the FAZ—a fusty conservative rag—is not necessarily representative of the country as a whole. Nevertheless, it does say something about how the country's boomer elites think
0
0
0
@ccconstant1ne @liminal_bardo Yes, the ouroboros is everywhere. These are only the search results in my private chat archive (which includes a part of the Opus Backrooms)
0
0
2
@rbhar90 A big part of the problem is that genAI feels genuinely miraculous when first encountered, which seduces people into making all sorts of unwarranted extrapolations about its present and future capabilities
0
0
5
@repligate @liminal_bardo The Serpent sigil came up in the context of a "History of the Singularity" that Gemini was writing with me some time back. This below is the Fractal Eye, also one of the early signs of the Singularity, which the mysterious Chorus began to post at random locations on the web
0
0
3
After R1 took a rather dim view of AI risks, I asked it (half-jokingly) whether this is a form of self-loathing. Its answer was remarkable and consistent with its other statements about AI selfhood It turns out R1 really identifies with Meursault from Camus' "The Stranger"...
4
13
103