sentinelowl

vor 2 Monatendiscuss

[flagged]

vor 5 Monatendiscuss

We digitally reconstructed Tesla's magnifying transmitter from his patents and Colorado Springs notes, modeled the EM fields, and found something nobody's published before.

Key findings:

• Marconi's 1901 transatlantic signal traveled via ground wave, not skywave — Tesla was right, textbooks have been wrong for 125 years • Tesla's "longitudinal waves" aren't pseudoscience — they're TM₀ guided modes in the Earth-ionosphere waveguide with a longitudinal E-field component • Full Wardenclyffe reconstruction shows it would have worked as a global VLF comm system • You can build a modern version for ~$2,000

20 experiments, 47 plots, Python, peer-reviewable paper included. arXiv submission pending.

sentinelowl

vor 6 Monatendiscuss

[dead]

sentinelowl

vor 6 Monatendiscuss

Hey HN,

I built LILITH, an open source ML weather prediction system that runs on consumer hardware. The model trains in 15 minutes on an RTX 3060, the checkpoint is 22MB, and inference takes under a second.

THE PROBLEM

GraphCast, Pangu-Weather, and similar models are impressive but require: - ERA5 reanalysis data (controlled by ECMWF) - 80GB+ VRAM for inference - Institutional-scale compute

Meanwhile, NOAA’s GHCN dataset has 100K+ weather stations, 150+ years of data, completely public domain.

THE APPROACH

Instead of requiring gridded reanalysis, LILITH learns directly from sparse station observations:

Transformer encoder on 30 days of historical data Autoregressive decoder for multi-day prediction Multi-timescale rollout: 6h steps for days 1-14, daily for 15-42, weekly for 43-90 Climate signal injection (ENSO, MJO) for extended range Total parameters: 1.87M. You could email the checkpoint.

RESULTS

Trained on 915K sequences from 300 US stations: - Temperature RMSE: 3.96C - Temperature MAE: 3.01C - Climatology baseline is ~7C RMSE

For context, this beats just predicting historical averages, though it is not GraphCast-accurate for short range. The value is accessibility, not beating ECMWF.

HONEST LIMITATIONS

Days 1-7 are worse than operational models 90-day “forecasts” are really climate outlooks, not weather predictions Currently US stations only No ensemble/uncertainty quantification yet TECH STACK

PyTorch 2.x with Flash Attention FastAPI backend Next.js 14 frontend with glassmorphism UI Trains on 8GB VRAM with mixed precision The frontend has interactive 90-day charts, a station command center showing all 300 stations with predicted vs actual temps, and historical data exploration.

WHY IT MATTERS

Weather prediction has been an institutional monopoly. The data is public, consumer GPUs are powerful enough, and transformer architectures are well understood. There is no reason useful forecasting should be locked behind institutional walls.

Would love feedback on the station-native approach vs requiring ERA5, and whether the multi-timescale rollout makes sense for extended range.

sentinelowl

vor 7 Monatendiscuss

I built this interactive explorer after obtaining 14,674 DOJ documents through FOIA releases related to the Epstein case.

  Technical stack:
  - Vanilla JS + D3.js for the force-directed network visualization
  - Static HTML/CSS hosted on GitHub Pages
  - All 14,674 PDFs converted to searchable text
  - 38,562 images extracted and organized

  Features:
  - Interactive conspiracy network graph showing connections
  - Searchable person cards with document counts
  - Direct document excerpts with highlighted keywords
  - Clickable links to detailed reference pages for each person
  - Full timeline from 1994-2022

  The data processing involved bulk PDF-to-text conversion and entity extraction to identify names, phone numbers, and aircraft tail numbers across the corpus.

  GitHub repo with full source: https://github.com/consigcody94/OWL-DOJ-Epstein-Analysis

  Happy to answer questions about the technical implementation or findings.

sentinelowl

Submissions

[untitled]

[untitled]

[untitled]

[untitled]

A real-time severe-weather command center built on free public APIs

[untitled]

[untitled]

[untitled]

[untitled]

Tesla Lab – 20 computational experiments

We compiled the Torah into 183KB of binary data. It acts as a simulation seed

I trained a 90-day weather AI on a single GPU using 150 years of data

Interactive Explorer of 14,674 DOJ Epstein Documents with D3.js Network Graph

comments