Gram 價格GRAM

➜ Some things don’t look valuable until you try to take them. A tweet. A sentence. A code snippet. A dataset. Alone, they're just fragments. But inside an AI system, they become invisible fuel. ▸ You train a model. ▸ It absorbs the fragments. ▸ And then… they vanish. No credits. No receipts. No clue where the outputs came from. That's not innovation, that’s amnesia. But what if AI had memory? Not just of what it says, but of who gave it the spark to say it? ➜ Imagine if Google paid every StackOverflow user it learned from. Or if ChatGPT tipped you every time your repo shaped an answer. Sounds like science fiction? @OpenledgerHQ is building that exact future. It’s called Infini-gram. ➜ Here’s how it flips the rules: ✦ You post data → it enters a datanet ✦ A model uses it → Infini-gram matches the span ✦ You get rewarded → instantly, onchain Even if the model paraphrases you? Still counts. Even if it’s a 4-layer agent chain? Still counts. This isn’t just NLP, it’s influence tracking. ➜ Most people still think “AI attribution” means slapping a link under a blogpost. Infini-gram turns it into a live reward loop. You don’t need to hunt citations. The system finds them for you. And settles them in tokens. Automatically. ➜ It's not just about credits, it’s about creating economic gravity. ▸ The more your data shapes AI output ▸ The more tokens you earn ▸ The more you’re incentivized to upload better data ▸ The smarter the models become ▸ The more trust the outputs hold That’s a positive-sum loop. ➜ The real twist? @OpenledgerHQ's Infini-gram isn’t tied to a single model or chatbot. It’s modular. It plugs into agents, datasets, protocols. It doesn't care what model said it. It only cares where it came from. In short: Infini-gram turns AI outputs into receipts. And receipts into rewards. ➜ So next time you drop knowledge into a system, ask: ▸ Will this vanish into the model’s memory hole? ▸ Or will it echo back, traceable and rewarded? Because the AI game isn’t just about answers anymore. It’s about attribution. And attribution is the new oil.

很多人现在都在说“AI要可验证”、“要能追溯来源”，但说实话，大多数模型根本做不到。前两天看到 @OpenledgerHQ 发布了通过 Infini-gram 支持归因证明的逻辑，仔细研究了一下感觉还是挺有意思的。 在使用AI的过程中，我们去问它一个问题，它回答得头头是道，可我们没法知道这句话到底是怎么来的，是“猜”出来的，还是真的在哪段训练数据里看过。这就好像你问一个人问题，他告诉你了答案，每句话都说“我觉得是对的”，但从来不给出处。 OpenLedger 最近提出的 Infini-gram 系统，就在试图解决这个问题 —— 把模型每个 token 的来源“标出来”。 简单理解，传统语言模型用的是 n-gram 技术 1️⃣uni-gram 是看单个词 2️⃣bi-gram 是两个词连在一起 3️⃣tri-gram 是三个词组合 以上的语言逻辑会提供一部分上下文，但内容十分局限，只看现有的问题，根据小语句关联来回答，但忽视当前问题在整个对话的逻辑。 而 Infini-gram 是另一个思路。它不仅看现有问题，而是用一种类似“符号匹配”的方式，把模型输出的每个片段都拿去对照训练集里所有可能出现的“语句”，看它到底是从哪里学来的、和谁的贡献有关。 比如你问模型：“怎么判断一个钱包是Bot？” 一般的模型会告诉你：“这种地址通常在极短时间内高频交易多个DEX合约。” Infini-gram 能告诉你：判断依据是告诉相似于某个公开数据研究者在某个 Dune dashboard 或 Github repo 里写的注释。甚至能定位到是哪一行。 它背后的技术其实挺硬核的，用的是基于 suffix-array 的 ∞-gram 框架 —— 本质上，它把训练集里所有片段都提前建好索引，输出的时候直接比对，不需要重新跑模型，也不靠梯度计算。这意味着快、稳、可复现。 对用户来说，你可以知道模型回答是不是“原创”还是“搬运” 对数据贡献者来说，你可以获得应有的“署名权”甚至“经济激励” 对监管机构来说，这提供了一个“可解释”的接口 OpenLedger 正在做的，不是让模型更“聪明”，而是更“负责任” —— 回答每句话时都能说清楚：“我为啥这么说，我是从哪学来的”。 在我看来，OpenLedger 提出的 Proof of Attribution 系统是“可信 AI”迈出的关键一步，也可能是构建数据所有权和贡献溯源的核心基建。

OpenLedger 刚发布了它最核心的一块技术底座：Infini-gram。简单讲，它能追踪语言模型输出的每一个词，精确地找到背后用到的是哪段训练数据 这意味着，AI 生成的每一句话，都能在链上溯源到具体的数据贡献者，做到谁贡献谁得分，谁影响谁收益 它用的是类似“无限长度 n-gram”的机制配上后缀数组，能处理长上下文，又能在超大数据量下跑得很快，不看模型参数，也不影响效率 这套机制不仅让模型透明，也让数据有价，不会让我们贡献一次数据拿一次收益，整体上长期 building @OpenledgerHQ 会更加有动力了！

The reason that purple has traditionally been associated with royalty was because, in Ancient Rome, the only source of purple was milking and fermenting the liquid from a snail. It took 12,000 snails to produce 1 gram of dye

Now at Rank #65 on @OpenledgerHQ’s Kaito 30D LB and #9 on Cookie LB. I’m doubling down because OpenLedger isn’t just putting AI on-chain, it’s building AI as blockchain. From Infini-gram to Proof of Attribution to the $25M SeedLab, they’re unlocking real agent and data liquidity for Web3. Who’s with me?