Skip to content

Instantly share code, notes, and snippets.

View p3nGu1nZz's full-sized avatar
🌎
w o r k i n g • f r o m • s p a c e

p3ngu1nZz p3nGu1nZz

🌎
w o r k i n g • f r o m • s p a c e
87 followers · 100 following
View GitHub Profile
@p3nGu1nZz
p3nGu1nZz / gist:3acdf783cf14ec21d678db6fc04cc8ee
Created November 14, 2024 08:41
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>Multiple Wobbly Plasma Bubbles</title>
</head>
<body>
@p3nGu1nZz
p3nGu1nZz / .py
Created December 2, 2024 14:14
EWC for fine-tuning models.
import torch
import torch.nn as nn
import torch.optim as optim
# Example model
class SimpleNN(nn.Module):
def __init__(self, input_size, hidden_size, output_size):
super(SimpleNN, self).__init__()
self.fc1 = nn.Linear(input_size, hidden_size)
self.fc2 = nn.Linear(hidden_size, output_size)
@p3nGu1nZz
p3nGu1nZz / gist:4f0108fb137c0cd1a34fffe9fb205252
Created July 15, 2025 16:41
{"raw": "X-SPANFORMER\nSPAN-AwARE ENCODER\n5.4 Qualitative Span Interpretability\nTo assess the plausibility and semantic alignment of X-Spanformer's induced spans, we perform side-by-side comparisons against syntactic and semantic reference structures. Using single-sentence prompts drawn from the validation sets of WikiText and Stream-Mix, we visualize the top-K spans selected at various layers and entropy regimes. We benchmark span boundaries against:\nSyntactic parses: Constituents produced by Berkeley Neural Parser", "type": "mixed", "id": {"id": "a0409606-f532-4dd2-b02e-2a0bae5bfeee"}, "meta": {"status": "keep", "tags": [], "doc_language": "en", "extracted_by": "pdf2seg", "confidence": 0.78, "source_file": "XSpanformer_TokenizerFree_SpanAwareEncoder_RawsonChrzanowski_Preprint_v1.0_2025-06-26.pdf", "notes": "The segment contains a mix of technical terms and structured information that can be segmented into meaningful spans, such as \"X-SPANFORMER,\" \"SPAN-AwARE ENCODER,\" specific versions like \"5.4 Qua
@p3nGu1nZz
p3nGu1nZz / gist:ada1d467e74d5505780e46937a867236
Created October 1, 2025 01:13
Fantastic work, Claude! Love what we came up with—that’s exactly what people want. Let’s start by updating #file:ACTION_PLAN.md and deleting #file:M3_COMPLETION_TODO.md. Once that’s done, create a new file in the project root called TODO.md with an unordered [ ] to-do list containing short descriptions for each item (no nesting). Update the instructions in #file:copilot-instructions.md and #file:AGENTS.md to reflect this process. After updating the documentation, commit the changes to the current branch, push them, and then begin working on the tasks in TODO.md.
@p3nGu1nZz
p3nGu1nZz / gpt5.2-meta
Created December 19, 2025 11:07
A “non-physical, mathematical, hyper-dimensional space” that *contains* all information is a coherent idea in a few different senses—but those senses often get blended together in a way that makes the claim sound more scientifically settled than it is. The most useful way to think about it is to separate (1) **mathematical frameworks that model phenomena**, (2) **philosophical positions about what is ultimately real**, and (3) **spiritual/metaphysical doctrines** that make additional ontological claims.
## 1) The idea is mathematically natural, but “contains all knowledge” is an extra step
It’s completely standard in math/physics to represent a system as a point in a high‑dimensional space:
- **State spaces** in dynamical systems: the “space” is a bookkeeping device for all degrees of freedom.
- **Hilbert space** in quantum theory: a mathematically precise space of possible states.
- **Information geometry** / statistical manifolds: spaces whose points are probability distributions.
- **Markov chains**: evo