Good times with @albert2001pp

Tokenized deposits are having their matcha moment They are presented as the alternative to coffee (stablecoins). The problem is that, just as with coffee, a premium, high-quality, imported ceremonial matcha is not the same as a matcha loaded with additives Therefore, it is the task of the tokenizing bank to fully understand its technological stack, assess its capacity to modernize its core, and then do (if you allow me to describe it this way) the easiest part, which is issuing a token on a blockchain. The difficult work is determining what the record will be, when balances become economically available to the client, how transfers achieve definitive finality, and how that liability converts into another bank's money I also believe the coffee versus matcha approach overlooks the reality that both can occupy different layers of the same monetary stack. A tokenized deposit is a direct liability of a bank and can provide bank-native money for treasury, payments, and settlement within a controlled and defined institutional perimeter A stablecoin is a separate liability designed to travel beyond that perimeter across more open networks. The former preserves the banking relationship and, potentially, all associated rights, while the latter provides portability and distribution In fact, the most interesting architecture, and the one we will see the most in the future, could place one underneath the other. Tokenized demand deposits will form the operational liquidity layer of a payment stablecoin's reserves The holder would own a claim against the stablecoin issuer, the issuer would hold a deposit claim against the bank, and the bank would record the corresponding deposit liability on its balance sheet, either natively on the blockchain (Deposit Token) or as a representation (Tokenized Deposit) This could synchronize the receipt of reserves with minting, and burning with the release of redemption funds. Although we are already starting to see some stablecoin issuers offer real-time on-demand minting and burning to certain systemic players, it remains a synchronization with systems that require orchestration, whereas here we could achieve true atomicity that would work 24/7 anywhere in the world (a world that has clear regulation, infrastructure, and technological development, of course...)

AI capex is projected to reach $765B in 2026, passing oil and gas for the first time. Compute is now getting its first futures contract. Here is what has to be true for it to work…

understanding reasoning from pretraining to post-training standard LLM pipelines eval RL post-training largely in isolation from the massive pretraining phase. Instead, the authors quantify pretraining-to-RL interface by using chess as a controlled testbed to isolate how pretraining scale shapes RL performance. Training is performed sequentially across a complete pipeline. At each stage, the model is first pretrained on human chess games, then SFT on synthetic reasoning traces, and finally uses RL on chess puzzles with verifiable rewards They test this across 36 compute combos with models ranging from 5M to 1B parameters. Some findings are: 1) pretraining loss strongly predicts post-RL performance; the rate of RL improvement grows approximately linearly with number of pretraining tokens. 2) Mechanistically, RL does not simply sharpen existing SFT preferences; on hard puzzles, it surfaces correct moves that were nearly absent, a predictive pattern that also transfers to a 1B language model trained on math text!