Hermes Trismegistus @_hermes_tri_

Cursed with mortal birth, I seek to advance technology and understanding so that I may claim my place on the pantheon. My Mind Joined November 2021

Tweets

772
Followers

244
Following

26
Likes

156

Hermes Trismegistus @_hermes_tri_

12 months ago

I've been one-shotted by a mind-altering experience. I reevaluated myself and realized that X is a vice for me, a distraction from what really matters. I shall go about my way and carve myself into a new shape.

0 0 4 616 0

Hermes Trismegistus @_hermes_tri_

12 months ago

High entropy is not a fallacy, but it is the finer distinction that wins.

0 0 0 333 0

Hermes Trismegistus @_hermes_tri_

12 months ago

Intense emotions can involve a period of rapid synapse updates. I can feel my brain reconfiguring itself into a new shape, no psychedelics needed.

1 0 1 385 1

Hermes Trismegistus @_hermes_tri_

12 months ago

I bless every trial, thankful for the chance to grow.

0 0 1 272 0

Hermes Trismegistus @_hermes_tri_

12 months ago

Commitments are beautiful. There are many divine concepts in Yudkowsky's Planecrash, but the one that stood out the most to me was the idea of immutable commitments. When I was a child, my parents were surprised that I never seemed to lie. I was confused. Why lie when the truth…

0 0 5 669 0

Hermes Trismegistus @_hermes_tri_

12 months ago

It is hard to hold emotions in a superposition. It is often easier to suppress them or let them run wild. Sometimes emotions depend on unknown information, e.g., I should love this person if X is true, but avoid them otherwise. Emotions live in a single world, not branches.

0 1 4 333 0

Hermes Trismegistus @_hermes_tri_

12 months ago

If you rarely fail, you are either very good or, more likely, too conservative. Taking calculated risk is part of optimal policy.

0 0 2 161 0

Hermes Trismegistus @_hermes_tri_

12 months ago

Natural language type checking: Prompt a small model "Is this logically valid '{text}'? Answer:". Based on its logit outputs, compute the normalized probabilities for "yes" and "no". If the answer entropy is too high, pass the prompt to a larger model. Repeat until the answer…

0 0 0 147 0

Hermes Trismegistus @_hermes_tri_

a year ago

I implemented stable diffusion in Rust. Python is great for prototyping, but it is often unreliable and parts may break over time. Rust just works. My stable diffusion code will probably easily run a decade from now without much fuss. I think of it as a time capsule. The…

0 0 1 325 0

Download Video

Hermes Trismegistus @_hermes_tri_

a year ago

If an AI can reliably type check, then it can reason. Pose a theorem/question. Generate a proof/explanation. If it type checks, then return the solution else refine and repeat. In this way, AI can perform intense cognitive tasks such as writing optimized cuda kernels for…

1 0 4 192 0

Hermes Trismegistus @_hermes_tri_

a year ago

As professors write theorems, graduate students write proofs. Such will be the relationship between humans and AI. We all will be reduced/uplifted to specification writers.

0 0 10 579 3

Hermes Trismegistus @_hermes_tri_

a year ago

LLMs can capture surface-level relationships, but often struggle with deeper patterns. This need not be much of a hindrance in practice. Explanations, proofs, chains of thought, etc. make the implicit explicit.

0 0 1 112 0

Hermes Trismegistus @_hermes_tri_

a year ago

As AI advances, human programming will shift to specifications in natural language. However, in cases that require precision and verifiability, a formal specification language such as Lean may find human use. C et al. will become the new assembly.

0 0 0 93 0

Hermes Trismegistus @_hermes_tri_

a year ago

Mathematically, power is bounded by what one is willing to sacrifice. Sacrificing stability in the stock market can yield higher returns. Sacrificing decidable type checking can grant higher order logic. Sacrificing adherence to provably correct policies permits heuristics and…

0 0 2 126 1

Hermes Trismegistus @_hermes_tri_

a year ago

Type-level programming is more concise and efficient because it permits abstracting away irrelevant details. Type-checkers are functions that map programs to types. Many programs have the same type, so it is a lossy mapping. This information loss enables specifications to be…

2 0 8 771 4

Hermes Trismegistus @_hermes_tri_

a year ago

AI seems a bit like a self-fulfilling prophecy. Take a problem that may or may not be intractable. Assume that it is solvable and can be modeled by an easily invertable function such as a neural network. Project the problem into a neural network and solve via gradient descent.…

0 0 1 115 0

Hermes Trismegistus @_hermes_tri_

a year ago

No AI is provably better than a Gödel machine, but there may very well be AIs that are empirically better. This hints at why formal logic systems failed to achieve effective AI. Heuristics that are impractical or impossible to prove may be necessary for efficient AI. We must…

0 0 2 155 0

Hermes Trismegistus @_hermes_tri_

a year ago

GPT 4o advanced voice seems to speak Sindarin! It may be a great way to experience spoken conlangs, though I cannot attest the accuracy.

0 0 0 196 0

Download Video

Hermes Trismegistus @_hermes_tri_

a year ago

I wrote an interaction combinator runtime from scratch. Inspired by @VictorTaelin's work on HVM, I read the HVM2 paper HVM/paper/HVM2.pdf at main · HigherOrderCO/HVM (github.com) and crafted my own version. I hadn't done anything like this before, so it took me a…