Just grabbed 20 Slonks on OpenSea over the past couple days. This thing launched May 1st and the floor's already sitting at 0.062 ETH after just 5 days—318 ETH moving in the last 24 hours alone. Let me break down what's actually happening here because it's pretty wild.

So the basic idea: take 10,000 CryptoPunks, feed them into a tiny transformer model embedded directly in a smart contract on Ethereum, and let it redraw them. Yeah, you read that right—there's literally an AI model running on-chain. The genius part? AI art usually gets clowned for being obviously AI-generated, right? The creator leaned into that. The model imitates the Punks but makes mistakes on purpose. More errors equals higher value. "Slop is the art"—they actually made that the whole vibe of the project.

Each NFT has three key attributes that drive everything. First is the original punk number (1-10,000). Second is the slop value—basically how many pixels out of 576 total did the AI mess up? That's the real stat that matters; higher slop means higher quality in this inverted logic. Third is level, which starts at 0 and increases with each merge.

The gameplay loops are where it gets interesting. You can merge two cards of the same level, burn one, and level up the other. When you do, the AI averages the "impressions" of both cards and redraws. The result? Way less similar to the original punk, which means higher slop, which means more token value. It's irreversible though—the burned card is gone.

Then there's the token split mechanic coming tomorrow. You can disassemble a card for $SLOP tokens—the number of slops on your card equals the number of tokens you get. A 287-slop card gives you 287 $SLOP. If that token bundle sells for more ETH on Uniswap than the card itself, you arbitrage it. This is where the nft and token mechanics really start blending together in a way I haven't seen before.

The third action is spinning for new cards using tokens—think lottery ticket vibes. You burn tokens in a Dutch auction (starts at 576, drops by 1 every few minutes, minimum 100). Most of the time you get something mid. But there's a 1% chance the AI goes completely unhinged and creates a near-perfect slop masterpiece. That's your moon shot.

But here's what really got me: the technical execution. They actually embedded a 22.7 KB transformer on-chain. The weights are split across 9 contracts using SSTORE2, and every mint or merge runs forward inference directly in the EVM, drawing the SVG in real-time. Vocabulary of 10,000, 10-dimensional embeddings, 18 attention heads. It's a tiny model but it's genuinely running.

The economics are elegant too. Total $SLOP supply is exactly 5,760,000—that's 10,000 punks times 576 pixels. Burn an nft, get slop tokens equivalent to its defects. Burn tokens to generate a new card from the void, redrawn with added noise. I've never seen pixel-level nft decomposition turned into a token system quite like this.

The design philosophy is consistent across the board. Collectors can chase rare original punk types or hybrid mutants from multiple merges. Gamefi players can calculate arbitrage angles and enjoy the merge strategy. It completely eliminates IPFS dependency since everything's generated on-chain. Different player types, different value propositions.

Sure, it's not the easiest thing to understand at first glance, but that's kind of the point with good games—there's a learning curve. Once it clicks though, you've got something genuinely innovative. This is the kind of on-chain engineering that deserves attention. The nft token economy here isn't just slapped together; it's actually thought through.