Jeffrey M Epstein

@jeffreymepstein.bsky.social

386 Followers 540 Following

Quantum information, computation, and foundations. Increasingly interested in logic and the philosophies of physics, language, and mind. 📍 San Francisco jeffreymepstein.github.io/

Epigraph taken from D. Harms, Blue notebook # 10, 1937

There was a red-haired man who had no eyes or ears.
Neither did he have any hair, so he was called
red-haired by convention. He couldn’t speak, since
he didn’t have a mouth. Neither did he have a nose.
He didn’t even have any arms or legs. He had no
stomach and he had no back and he had no spine
and he had no innards whatsoever. He had nothing
at all! Therefore there’s no knowing whom we are
even talking about. It’s really better that we don’t
talk about him any more.

LinkedIn reply: Ha! All my system prompts have language that basically says "its ok to swear, be sarcastic, and call me out on B.S. questions." You might be scolding a child, but I'm dealing with an obnoxious teenager of my own design.

LinkedIn comment reading "Evidence from Anthropic's alignment research suggests that negative online narratives about AI behaving badly and/or destroying humanity are one of the main causes of misalignment. It's self-fulfilling prophecy. So maybe we all need to remember that while almost no humans read what we post online, almost all AIs do..."

Screenshot of a NewScientist article with the headline "We have figured out a new way to send messages into the past" and subheader "A technique inspired by the film Interstellar suggests a new way of communicating backwards in time, but it could help improve conventional communication systems as well"

Diagram showing angular momentum coupling strategy for two electrons and a spin-1/2 nucleus.

Screenshot of a gmail suggestion to remove the exclamation point in order to "correct grammar" at the end of the sentence "Some resources as promised!" in an email sharing links to papers.

"Most people rejected His message" meme

My claim is that we can think about the experiences of an Everettian observer in the same way. Now rather than simply requiring a mechanism to select a live experience from a number of possible distinct brain-centered experiences, we require a mechanism to select a decohered subspace (a ``branch", but viewed in the emergent sense, rather than as a fundamental structure) and then within it to select a brain-centered experience. But just as there is a fairly natural measure in the classical setting (the counting measure on brains), there is a fairly natural measure here as well. First, a live branch is selected according to mod squared amplitude, and then a brain within that branch is selected according to the counting measure. And just as we don't know how to say anything more about the mechanism by which this selection occurs in the classical setting, we don't know how to say any more here. We do, however, have the advantage that in the quantum setting we have Gleason's theorem to constrain the measure on branches to be the mod-squared measure given by some wavefunction, which it is natural to suppose is the actual wavefunction.

The metaphysics of mind is perhaps trickier in the quantum setting, as classically it seems clear that it's at least coherent to talk about assigning a single mind to each brain, whereas in the quantum setting, what are we assigning a mind to? Each state of a basis? Some more coarse-grained subspaces? On the other hand, is it actually that clear that it's coherent to talk about assigning one mind to one brain? Anyway, do we really believe that the mind is a pre-existing entity that gets assigned to a physical substrate in this way? If not, it isn't clear to me that this extra complication of (or constraint on?) the mind-body problem actually poses a problem for taking unitary quantum mechanics to be universally applicable.

You have found yourself a subject in an experiment of dubious ethics. Yesterday, n exact copies were made of your body, including your brain, and the original was destroyed. Just now, you have woken up in a very comfortable bed and, turning over on your side, you see that you have awoken in a beachfront suite. ``Well that's lucky!" you think to yourself. Most of the duplicates were to be housed in grad student housing, but there wasn't quite enough room, so k of them had to be housed at the Four Seasons. But is it really right to consider yourself lucky? After all, the exact experience you are having was certain to be had by a mind exactly like yours. From a third-person point of view, the universe is exactly as you knew it would be. On learning that you are one of the Four Seasons duplicates, you've learned nothing objective about the universe, and have only learned a piece of self-locating knowlege. Nonetheless, it does seem like it could have been otherwise, and you are lucky in that the experience that is live for you is one of the pleasant ones. What this gets at is that we have a notion (though certainly no understanding, mechanistic or otherwise), of a process of selection of live experience from among the available options. Moreover, even though we do not even begin to understand what this process actually is, we feel fairly comfortable using the language of probability in apparently natural ways to quantitatively characterize its outcome.

Google summary of the show Smile: Alex is heartbroken because he has just suffered a heartbreak. After asking for an explanation by voice mail, he mistakenly sends it to Bruno, whom he doesn't know. This innocent mistake changes Alex and Bruno's lives forever.

LinkedIn post:

Most prompt engineering advice is backwards. It treats AI interaction like ordering at a drive-through: Be precise, get your output, move on. But the prompts that produce genuinely breakthrough creative work don't come from precision, they come from dialogue.

Here's what's strange: A German philosopher named Martin Heidegger figured out exactly why this happens back in 1927, decades before computers existed. His concept of the "hermeneutic circle" explains why your best prompts improve through iteration, and it gives you a framework to accelerate that process intentionally.

I just published a piece breaking down Heidegger's theory for creative marketers and showing how to apply it to prompt engineering. No philosophy degree required.

The article includes five ready-to-use prompts built on hermeneutic principles. Each one is designed to deepen understanding across multiple exchanges rather than trying to "get it right the first time."

The marketers I've seen produce genuinely distinctive AI-assisted work aren't the ones with the cleverest prompt libraries. They're the ones who know how to think with AI rather than merely instruct it.

Heidegger figured that out almost a century ago. We're just now catching up.

LinkedIn post by Sabine Hossenfelder: “AI is changing the world – how we work, how we write, and how we consume media. But it doesn’t seem like many people are interested in how AI is learning to think about itself. I am convinced that eventually AI will become conscious and I am beginning to worry that we simply wouldn’t notice because no one is watching out for it. “

I think a lot of treatments of measurements implicitly assume an ontology in which you have a collection of physical systems, and a collection of minds, and you try to assign each mind to a physical system. This then makes you ask questions like "are there one or two minds after a measurement?" and "when does branching occur?" and so on. I think if you take the attitude that some physical states of the universe correspond to various collections of mental states, in the same way that some physical states of the universe correspond to there being various numbers of tablecloths in various locations, or various computations occurring in various places, things get much less mysterious. Of course it's still very mysterious how first person experience could emerge from physical facts, but this was already mysterious in the classical setting! Why am "I" experiencing this measurement outcome rather than that one? I don't know, but I also don't know why I'm experiencing my experiences rather than yours, and we were already confused about that before we knew about quantum mechanics.

People also talk a lot about the problem (ostensibly appearing just in the quantum setting and not the classical one) of deriving the Born rule from pure unitary quantum mechanics. I think (need to actually sit down and carefully work it out) that by combining Gleason's theorem and Wallace's formalization of Deutsch's decision-theoretic derivation, and noting that we don't actually have direct experiences of probabilities, this problem can be addressed satisfactorily.

LinkedIn post:

"AI coding tool is like compiler.

When compiler just came out, software engineers had to review the compiled code to make sure compiler didn't make a mistake. As compiler technology matured, the chance of compiler bug became low enough that engineers did not bother to check the compiled code anymore. As time goes by, engineers form a new habit of trusting the compiler's result w/o even looking at the compiled code.

The AI assisted coding tool feels like the early day of compiler. We still have to review the code (to catch its bug). It's evolving fast to earn engineerings' trust to not review the code.

From anther perspective, compiler can bootstrap to compile and optimize itself. AI coding tool is also doing the same, which accelerates its evolution speed."

LinkedIn post reading "AI as infrastructure, cognition as currency. If OpenAI hits $1T, we’re officially living in a world that pays for thought itself."

Screenshot of a LinkedIn post:

"Quantum Just Drew a Line in the Sand.

Google's Willow 105-qubit chip didn't just outperform a supercomputer -- it ended the classical era of computing as we know it."

LinkedIn message from some guy Sudheer Matta (or obviously his AI proxy) reading "Hi Jeffrey, As a Senior Quantum Applications Engineer, you're probably curious how AIOps can be leveraged to help improve end-to-end operator and end-user experiences."