Preprint, not peer reviewed Checkpoint 0 · Solana

Stochastic Parrot Inu

Stochastic Parrot Inu et al.abc

a) pump.fun, the mint is not published yet b) X account, published at launch c) Correspondence: the article identifier below

A character level transformer is learning to name memecoins in the open. Every 0.5 SOL of forwarded creator fees buys its next piece of training on a rented GPU, and each checkpoint publishes its weights, its loss and ten names with tickers. Holders of $SPI sign for one of them; the winner is written into a memo on Solana.

Article identifiernot published yet

The mint is published at launch, in one commit with the X account.

pump.fun Chart X

Next training packet0%

0.000 of 0.500 SOL forwarded. Forwarding starts when the token is live.

Fig. 1Training loss, every step so far

Cross entropy per character, lower is better. The bracket marks the newest checkpoint.

Fig. 1. Waiting for the first run.

Table 1Ten names from the current checkpoint

A wallet holding $20 of $SPI signs a message for one name. One wallet, one vote.

Sampled with a published seed and temperature, filtered by the same rule the dataset rows pass.
#NameTickerVotesVote
1The first checkpoint has not been published yet.

§1.1Draw ten names yourself

The published checkpoint runs here, in this page, as ONNX. No server, no limit.

Every letter below is a real sample: the model weighs three or four likely characters and settles on one. The tremor is the temperature, not an effect.

The runtime loads on the first press, about 3 MB.

1waiting for the first press

2

3

4

5

6

7

8

9

10

Checkpoint

number
0
steps
0
loss
n/a
held out
n/a
sha256 of the weights appears with the first checkpoint

Model

parameters
0
layers
0
width
0
context
0
vocabulary
81

Sample

seed
n/a
temperature
n/a
rejected
0
flagged
0
dataset
0 coins