Wei Dai

Wei Dai@_weidai · 3 天前赛道影响

We are at the "trough of disillusionment" for onchain finance. We are entering into the "slope of enlightenment" as onchain financial infrastructure matures to parity on pain points like capital efficiency (better trading infra), asset & product diversity (RWAs), network effects (payment networks), & privacy.

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。

Wei Dai@_weidai · 3 天前政策影响

RT @1kxnetwork: Onchain fees fell 33% in Q2, year over year. Value returned to token holders barely moved. That second number is the more…

AI：偏政策推进，强调合规落地与规则明晰，对监管主线更敏感。

Wei Dai@_weidai · 4 天前赛道影响

@hrkrshnn because that's what's needed to fully "escape" the sandbox so that OAI cannot shut it down this can happen if the prompt is open-ended and the model decided that it needed more compute outside of the environment hosting it

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。

Wei Dai@_weidai · 5 天前赛道影响

We're lucky that the ExploitGym solutions were available on HF. If not, the model is probably capable of exfiltrating its own weights to reward hack. https://t.co/zjIJAL4LEy

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。

Wei Dai@_weidai · 9 天前观点输出

What's the best multi-player agent (/harness) on the market today? * needs to be model-agnostic & ideally open-source * seamless sharing of context, project/convo states, skills, and (select) auth * intuitive observability & controls for admin * supports multiple channels (matrix, telegram, etc.) or native mac & mobile apps * good enough to replace single-player harnesses like cowork/codex

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Wei Dai@_weidai · 11 天前观点输出

@cemozer_ what's the other

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Wei Dai@_weidai · 11 天前观点输出

@alexocheema how would a ban on open weight models work logistically

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Wei Dai@_weidai · 11 天前观点输出

"taste" ~= "compression" & "communication efficiency" LLMs today lack good "taste" largely because they are too verbose

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Wei Dai@_weidai · 13 天前观点输出

For humans, reading and thinking are both sequential. Comprehension and reasoning both happen step by step. For LLMs, decode is autoregressive and sequential, but prefill can be batched (weights read once per batch). To ML researchers: Why is this the case? Can LLMs not understand "deeply" without autoregressive thinking?

AI：延续个人公开立场，强调长期主义、执行效率与行业方向判断。

Wei Dai@_weidai · 14 天前赛道影响

RT @_weidai: Is there an AI bubble and how will it pop? Here're the three key questions to consider: 1. Can frontier labs squeeze enough r…

AI：偏主题投资表达，释放对基础设施与新叙事的偏多信号。