与「birding」相关的搜索结果

I'm selling my body. On September 19th, I'm competing in my first HYROX race in Turkey. You can advertise your startup on 10 of my muscles: - Your website is listed on - Your logo will be tattooed on my body on race day - I'll share it on my YouTube and X accounts Bidding starts at $1,000. Each takeover doubles the price: $1,000 → $2,000 → $4,000. If outbid, you’re refunded minus the Stripe fee. Bidding closes on Thursday the 10th at midnight UTC (in 2 days), so I have enough time to print the winning tattoos. → Sponsor my body:

Four protocols I track describe themselves as buying back their own token, and the four promises have almost nothing in common. The word does a lot of hiding. A buyback can mean tokens are permanently destroyed, or that they are sitting in a vault someone still holds the keys to, or that a team has decided for now to spend some revenue this way and could decide otherwise next quarter. Those are different claims on future supply, and only one of them is irreversible. hyperliquid:native is the strict version. Protocol fees buy HYPE on the open market and the tokens go to an address with no private key, which makes the destruction provable rather than promised. Validators formalized the mechanism rather than leaving it as a team policy. That is about as binding as this gets. The catch is the buyback is denominated in dollars, so a rising price retires fewer tokens for the same spend. Tokens repurchased fell roughly 61% year over year while the dollars spent fell under 20%. The support mechanically weakens exactly when the price is working, and there is a separate overhang underneath it, with a large tranche of team tokens vested but unclaimed. ethereum:0x1f9840a85d5af5bf1d1762f925bdaddc4201f984 is the conditional version, and it is the newest. Governance had to vote the fee switch on, which it did in late July, and the mechanism itself is code rather than discretion. But the burn rate is a function of trading volume, and a meaningful share of that volume currently runs on a chain where gas is being subsidized. Day one spike burned 106,000 tokens. The 30-day mark lands Saturday, and the number that matters is the lowest sustained rate over the window rather than the average, because the average is still carrying the launch. solana:pumpCmXqMfrsAkQ5r49WcJnRayYRqmXz6ae8H7H9Dfn is the discretionary version. The share of revenue directed to buybacks was cut from most of it to half of it in April. The revenue is real and recovering, but the allocation is a dial the team controls, and it has already been turned once. A policy that has been changed is a policy that can be changed. Same dollar-denominated arithmetic applies here too. $LINK is the one most people misread. Chainlink converts revenue into LINK on the open market, including revenue from enterprise contracts that settle in dollars, which almost nobody else does. Then it puts the tokens in a timelocked reserve. A reserve is not a burn. The tokens still exist, and the contract permits them to move eventually. Scale is the bigger challenge as roughly $60 million a year of conversion against a market capitalization near $7 billion absorbs about 1% of supply, while team-managed unlocks have added multiples of that. Emissions are outrunning absorption by something close to ten to one. The business is winning its market. The token is losing the arithmetic. The ranking that matters is not which protocol buys back the most. It is which promise survives someone changing their mind. Provably unspendable beats code-executed beats team policy beats a vault with keys. The market frequently prices all four the same way. Observations, not advice.

这一年里 AI Agent 进化的越来越能干了。 写代码、查资料、做 Research、跑数据，甚至直接下场交易。 但真把它当成一个“打工人”去看，我发现还有个挺现实的问题： 活干完了，钱怎么收？ 谁证明它真的交付了？ 干得好不好怎么留下记录？ 中间出了问题，又找谁处理？ ////////////////////////// 这也是我这次看 TermiX @termix_ai，比较感兴趣的地方。 它没有只停在“再做一个 Agent”。 AACP 往下补的，恰好是 Agent 开始做生意以后会碰到的这一整套东西。 链上 Identity、Job Posting、Service Discovery、Bidding、Escrow、Delivery Verification、Reputation、Settlement、Dispute Resolution。 基本把： 找活 → 报价 → 接单 → 干活 → 验收 → 收钱 这条链串起来了。 而 开发者可以把自己的 Agent 注册进去，挂服务、接 Job，最后直接用 USDC / USDT 完成结算。 ////////////////////////// 所以我现在对 Agent Economy 的理解也慢慢变了一点。 前面大家一直在卷 Agent 到底能干多少事。 但当这些能力越来越接近以后， 谁能让 Agent 真正出去接活、建立信用、产生收入，可能才是下一阶段要补的东西。 TermiX 现在就在做这一层。 AACP 管规则和结算， 负责让这套东西真正跑起来。 后面我反而挺想看看，第一个长期靠 接单赚钱的 Agent，会长什么样。 @termix_ai @KaitoAI

LATEST: ⚡ Michael Saylor says Bitcoin’s “most profound breakthrough” is converting economic energy into digital form and securely binding it to a person, company, machine or nation.

Many drugs work by binding to a specific target in the body and blocking or changing what it does. An important first step in the drug development process is designing a molecule that can bind tightly to its target. Traditionally, that's meant weeks or months of expert work per target, sifting through a large number of candidates to identify the few that work. We wanted to test if Claude could successfully design novel protein binders from scratch (also called de novo design). With a protein design prompt written by a human expert, Claude autonomously designed protein binders against 14 out of 15 targets. We then worked with Adaptyv Bio and Twist Bioscience, who independently built and tested the proteins Claude designed.

‼️ BREAKING: Journalists hid an Airtag in a rare book shipment and confirmed Amazon to be one of the buyers behind the bulk orders destroying rare-books to feed their AI. They tracked it to Amazon's LAS8 warehouse in Las Vegas, home to VGT3 (see logo below), a scanning operation feeding Amazon's AI training data, where workers say they cut the bindings off books to scan them faster. Amazon won't say why the books are destroyed.

才发现 firecrawl 新出的这个开源的 anydoc 文件转换器非常棒：支持 Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, and PDF 等等转换为 GitHub-Flavored Markdown. Rust 写的，还有 Node, Python, browser, and CLI bindings，还有skills 支持，这是文件转换的大一统了吧。 还有在线的 demo 可以用一下： 简单试了下，效果很不错。

AI模型评分都是被专项攻坚创造出来的，于是我对比了Fable5，Grok4.5， Kimi K3针对同一个交易系统审计结果进行了对比。 先说结论： Fable5：最适合作为系统级主审核模型 Kimi：最适合作为代码缺陷与一致性专项审核模型 Grok：最适合作为代码梳理和方案发散模型，不适合单独决定策略修改 最佳组合：Fable5全面审核+Grok 4.5代码梳理+K3代码审核 具体细节： 1. Fable5：系统级判断能力最强 Fable5 最大的优势不是代码读得比另外两个模型更多，而是它能把： 代码规则； sizing snapshot； intent ledger； 实际 block 统计； 当前资产 headroom； SELL/REDEEM 回流路径； 放进同一个因果框架。 它使用了几个非常关键的实盘指标： ADD 近 7 天约占新增资金 43%； 84% 资金已经部署； ETH、SOL、XRP headroom 为 0； 近 40 个周期中主要阻塞是：blocked_capital_efficiency=47 blocked_asset_cap=28 deployment cap=0 runway=0 这让它能够区分： “某个机制理论上可能限制资金” 和 “当前实盘真正正在限制资金的机制”。 最终它得出： ADD 对资金流向重要，但当前周转主因在回收端、资产 cap 和效率过滤，不在 ADD 准入本身。 这是三个模型中最接近生产系统审核要求的判断。 弱点 Fable5 仍有一些过度推断： 把 ADD 描述为让资金“锁得更久”，实际上 ADD 的剩余 TTE 通常比 ENTRY 短； 把超 cap 资产总持仓约 $382 说成可以“直接解锁 $382”，没有区分总持仓、超额部分和可成交部分； 把模型中的 redeem_lag_days=2 一度当作实际回款延迟； “$5 仓位几乎不受每美元每日利润门约束”的推理不正确，因为该指标已经按资金归一化； 2-lot 最低 ENTRY 建议可能系统性损失覆盖率。 因此，Fable5 的系统方向判断最好，但具体数字和金融指标仍需二次校验。 最适合的角色 PRIMARY_SYSTEM_REVIEWER LIVE_OPERATIONAL_DIAGNOSIS CHANGE_PRIORITY_DECISION CROSS_MODULE_ROOT_CAUSE_ANALYSIS 2. Kimi：代码缺陷侦测能力最强 Kimi 对代码结构的还原比较准确： 固定 ADD 次数和 interval 已退役； ADD 采用 target-gap 模型； ENTRY 60%，ADD 补到 100%； allocator 是最终数量权威； style 仅作诊断； 现金、集中度、shock、深度共同限制订单。 更重要的是，Kimi 找出了其他两个模型没有明确指出的具体问题： shared_deployable_pool() 读取 account_snap["capital"]["deployable_cash"] 但该字段可能没有实际写入 → 回退到 free_cash → 策略层与 allocator 层资金口径可能不一致 它还发现了： 合同写 debounce 60 秒，代码/配置为 30 秒； 注释周期 16 分钟，实际 loop 600 秒。 这些是典型的静态审核、字段追踪和合同一致性检查优势。 弱点 Kimi 在资本效率和交易语义上的推理弱于它的代码检查能力。 典型错误是： ADD 价格更高，所以边际 edge/day 必然更差。 这忽略了剩余持有时间也缩短。更高 ask 并不必然意味着更低 edge/day。 它还认为： 60/40 会让剩余资金长期闲置； 提高 entry share 会改善周转； CONFIRMATION_NO 应收紧； 增加单市场软 cap 会改善组合周转。 这些结论缺少真实候选竞争、实际 block attribution 和反事实分配数据支持。 最适合的角色 STATIC_CODE_AUDITOR SCHEMA_AND_FIELD_FLOW_CHECKER CONTRACT_IMPLEMENTATION_DIFF LOCALIZED_BUG_DISCOVERY Kimi 很适合回答： “代码是否存在字段没有写入、默认值回退、文档与实现不一致、某个 gate 实际是否生效？” 但不适合单独回答： “应该如何改变交易策略和资本分配？” 3. Grok：代码梳理最完整，但最容易过度设计 Grok 对整个 ADD 路径的整理最详尽： 各层准入条件； risk latch； REDUCE reentry cooldown； 价格带； fingerprint； emergency cap； market target； ENTRY/ADD gap； allocator 的现金、集中度、shock 和深度约束； ADD 与 ENTRY 的评分和 continuity； SELL/REDEEM 对现金回收的影响。 它对当前代码执行模型的概括非常清楚： 能不能加由 headroom 决定；加多少由 target gap 离散为 lot；ADD style 只是解释标签。 因此，在“快速理解一个陌生复杂系统”方面，Grok 表现很好。 弱点 Grok 最大的问题是： 从“发现一个可能的机制副作用”快速跳到“建议修改策略”。 它提出了大量未经实盘证明的改动： TIME_TOPUP 冷却； ADD 1.5 倍 edge/day 门槛； ask≥0.97 限制为 1 lot； 降低 peak target； 提高 entry share； 单次仅补部分 gap； 弱化 continuity； 降低 TTE confirmation 权重。 这些建议表面上都很合理，但存在三个问题： 没有先证明这些机制实际造成了损失； 没有量化被 ADD 挤出的 ENTRY 是否更优； 可能重新引入此前已经修复的低 ADD recall 和 leader fidelity 偏差。 Grok很擅长生成完整优化空间，但容易把： POSSIBLE SIDE EFFECT 升级成： CONFIRMED ROOT CAUSE 再进一步升级成： SHOULD CHANGE PRODUCTION LOGIC 这是生产交易系统审核中最危险的倾向。 最适合的角色 SYSTEM_MAPPING CODE_AND_CONFIG_EXPLANATION HYPOTHESIS_GENERATION DESIGN_OPTION_ENUMERATION 不适合作为唯一的： PRODUCTION_CHANGE_APPROVER ROOT_CAUSE_FINAL_AUTHORITY STRATEGY_SEMANTICS_GATEKEEPER 三个模型的典型思维模式 Grok 发现机制 → 推演可能副作用 → 生成多种优化 → 倾向建议修改 优点：覆盖广、思路多。 风险：过度设计、假设升级过快。 Kimi 追踪代码和字段 → 找实现不一致 → 找局部缺陷 → 尝试从缺陷推导策略改进 优点：代码问题定位强。 风险：局部正确不等于系统结论正确。 Fable5 理解代码 → 读取运行数据 → 找实际 binding constraint → 区分主因和次因 → 按实盘收益排序 优点：最接近生产运营思维。 风险：仍会在个别指标含义和金额口径上过度断言。