3 Months of Pretraining on One RTX 3060 Yields a Model Speaking 'Lemon'
cephaloform · x · 2026-09-14
A developer tried pretraining a model from scratch on a single RTX 3060. Three months in, checking the output revealed it producing what looks like a 'Lemonstrat…' — apparently echoing a demo of the Lemon programming language instead of natural language. A humorous small-compute pretraining fail.
More from Fun
- When STT makes you sound barely literate: what do models really think of us? — teropa · 2026-09-15
- DeepSeek fans compare Dario to Hitler as AI safety-vs-accel camp feud turns nasty — zephyr_z9 · 2026-09-15
- Cyclist animates their last 100 Taipei bike rides from Strava GPS tracks — randal_olson · 2026-09-15
- Dev builds Poker Arena in a day to watch LLMs play live poker with real money — threepointone · 2026-09-15
- Rocket League ported into 2004's GTA San Andreas in gloriously absurd mod — tekbog · 2026-09-15
- Meme: ML engineers still using TensorFlow in 2026 — kmeanskaran · 2026-09-15