A log is all your need. Turns out you can just build a log db with hooks on CRUD and maintain arbitrary secondary indexes. It is as good as thor's hammer with extensions.

Append-only (authenticated) databases are inevitable. Huge thanks to @ryanzarick, @IsaacN0tNewton, and @Thomas_S_Kim for taking the time to unpack the insights behind the pioneering QMDB and explore how its scalability can drive SOTA TPS via FAFO.

An overly simplified version of work-incentive system in the future has three pools: no job, real job, Steve jobs. They correspond to proof-of-no-work, proof-of-work and proof-of-change, respectively.

Turns out you can just stack MMRs (and MMR-shaped Merkle trees) on top of each other to add functionality (amortizing proof cost in the process) 👀

Just realized that the word “fast” is very serial. When you can do one thing at a time, the latency (how fast you are) determines the throughput. But with parallelism, throughput is decoupled from latency. For example, a business that sells 12-year-old whiskey has high latency but high throughput that you can make infinite batches per year. And if you make a batch every year, then after 12 years you get to sell a lot every year. And this is called pipelining.

Suppose we have 100 uniswap pools with high contention. FAFO can make each pool run at 10K TPS.

1M TPS, merklized, EVM, single node with state persisted. One really needs to do a lot of things right to make it possible.

With infinite knowledge at your finger tips, the best quality a man can have is curiosity and persistence.

feature request: before we can normalize all models into one, lets just create a hot key like shift+enter triggers thinking models like O3. otherwise, a cheaper mode. @sama it will save a lot of user time