Overview

trainpad is a launchpad where every coin is born as a blank GPT with random weights. Within its first minute it starts training on its launcher's seed text. Holders feed it more text by signing a message. You can watch the loss curve fall, read what it says every 30 seconds, download its weights and check every checkpoint's sha256.

What is real right now

• Training from random weights on the server's CPU lane (PyTorch CPU, nanoGPT model code at a pinned commit).
• A byte-level BPE tokenizer trained per coin at its first slice, then frozen.
• Live loss stream, samples every 30 s, FIRST WORDS detection, the share card and 10 s clip.
• sha256 of every checkpoint, downloadable weights (bf16 safetensors + tokenizer), reproducible step 0.
• Three lab runs on public-domain books, reborn every 6 hours, and the mother model.
• Holder corpus submissions by signed message, checked against on-chain balances.
• The launch flow: built, decoded in your browser, simulated, then signed and sent by your own wallet.
• The fee route and the $TRAIN buyback and burn, every 5 minutes, from one period setting.

What is not connected yet

• The GPU lane (rented GPUs). Until it is connected every receipt is a CPU lane receipt at $0.000 house cost.
• The 8xH100 nanochat graduation run. A coin that reaches its target shows "funded, graduation lane not connected".
• pump.fun replies as corpus input. The corpus shows 0% from replies.
• Automatic X posts of FIRST WORDS clips. Share on X opens a normal post with the card link.

The full list with what each one needs is in the build's MOCKS.md. Nothing on the site shows output from an unconnected lane as real.