Home

A sentence keeps showing up in conversations about AI, usually with optimism: “When memorization becomes cheap, understanding becomes valuable.” Sometimes it shows up in technical clothing: “When code production becomes cheap, verification becomes valuable.” The claim sounds like simple economics: when one thing gets cheaper, the sca...

If you want to understand a machine, don’t start with the pitch deck. Put it on a lift. Trace the pipes. Find the parts that glow when you push them too hard. “Equity perpetuals” keep resurfacing in crypto as live markets: TSLA, NVDA, sometimes broad indices. The promise is simple to say out loud: trade equity price exposure like BTC-24/7, wit...

TL;DR: Who’s right about the ADL trilemma? A lot of people are tired already, so here’s the scoreboard up front. Dan is right that: The “ADL trilemma” in Tarun’s paper does not apply to all perp venues in all regimes. The trilemma only bites when the exchange is in a structural deficit regime: expected tail shortfalls...

Tarun Chitra’s ADL paper triggered a big argument around Hyperliquid’s October 10–11 event1234. Who’s right about what, and what actually needs fixing? The ADL debate around Hyperliquid’s October 10–11 event has three main characters: Hyperliquid, whose ADL is implemented in contracts and described only briefly in docs5. Tarun’s paper, w...

When a Transformer controls a $100M DeFi vault, can it be jailbroken into draining funds? When an AI oracle publishes “anonymized” embeddings on-chain, can adversaries recover the private data? These aren’t hypothetical questions, they’re architectural constraints imposed by two recent theorems about the functional properties of deep neural netw...

The Arbitrum post proposes replacing a single, exponential “backlog” pricer with a multi-constraint version and, further, a constraint ladder that inserts intermediate time scales between long-horizon and short-horizon resource constraints1. Short-horizon constraints (keeping executors from falling behind) should prompt fast fee responses, while...

Richard Sutton’s weariness was palpable in his conversation with Dwarkesh Patel. After fifty years in AI, the lesson he delivered was a “bitter” one: intelligence built on human intuition consistently fails, while systems that learn from experience ultimately prevail. Computation and feedback win; imitation and intuition lose. Aristotle, who gr...

Overview Monad’s base-fee controller, inspired by adaptive optimizers like RMSprop, aims for responsiveness without twitchiness by making the fee step size inversely proportional to short-horizon variance in gas usage. However, in a permissionless blockchain, block producers control the very signal, gas usage $g_k$, that drives this variance es...