Simon @linsizhen

用 ChatGPT 审查本地代码，把 Codex Token 留给开发。通过只读 MCP 桥接本地 Git 仓库，限制写入与命令执行。 Review local code with ChatGPT and save Codex tokens for development. A read-only MCP bridge keeps access scoped and secure. github.com/simonlin1212/l…

我的投研Agent，希望能打造成贾维斯的感觉

Vibe Research v1.1.0，完整演示来了。 这次最想展示的两个变化：WorkBuddy 也能接入了，而且有了独立的 Mac 客户端。 打开首页直接聊，简单问题不用跑 Agent；需要深入研究时，再开启Agent，进行联网搜索、取数和计算。 从接入 AI，到查看市场与产业线索，再到公司研究、研报管理，视频里把整个工作台走一遍。 Codex、Claude Code 订阅和模型 API 接入也都保留，按已有账号选择即可。 项目已开源，源码、Mac 安装包和使用说明都在 GitHub： github.com/simonlin1212/V…

Vibe Research v1.1.0 更新，把 AI 对话、市场数据和研究工具放进一个本地投研工作台。 这次重点升级： • 支持 WorkBuddy / CodeBuddy、Codex、Claude Code 订阅，也支持模型 API。 • 新增 Mac 独立客户端，M 系列电脑无需另外安装 Node、Python 或 Codex。 • 打开首页就能聊天。简单问题用普通对话，需要联网查资料、取数和多步研究时，再开启 Agent。 功能也集中到了一起：每日复盘、资讯雷达、产业信号、板块分析、六阶段公司研究、多空辩论、回测，以及研报和研究记录管理。 研究不只留下一篇报告，还保留来源、证据和计算过程，方便回头核对。 模型接入使用自己的订阅或 API，研究结论仍需独立判断。 项目已开源，源码、Mac 安装包和使用说明都在 GitHub： github.com/simonlin1212/V…

Vibe-Research 是一套基于开源 Codex Harness 打造的本地金融研究 Agent，目前已获得 GitHub 2,200+ Stars。 它把每日复盘、资讯雷达、产业信号、标的研究、多空辩论、回测和研究归档，放进同一个浏览器工作台。 支持 A股、美股和港股，可使用本机已有的 AI 订阅，也可以配置 MiMo、DeepSeek、Qwen、Kimi、OpenAI 等模型 API。 不同于普通的 API 聊天框，它能维护研究上下文，调用 117 个信息源与专业工具，持续推进任务并保存完整报告。 所有研究数据和记录均保存在本地，让 Agent 真正成为一套可以长期使用和持续升级的个人投研系统。

Vibe-Research V1.0.0 is live. Built on the open-source Codex Harness, it is now a local financial research agent with 117 callable data sources, market reviews, equity research, bull/bear debates, and A/HK/US backtesting. Use an AI subscription or your own API. github.com/simonlin1212/V…

My open-source project Vibe-Research just crossed 2.1k stars ⭐ New in v0.4.0 — "Industry Signals", starting with GPU rent: • 1-year daily rent history for B200 / H100 / A100 • Latest spot price + how much of each fleet is rented out • Forward expectations for the next 13 months, from public event contracts All from zero-auth public data. A snapshot ships with the repo, so a fresh clone opens with full history. It shows the price facts and the caveats on how to read them — it does not tell you whether compute is "oversupplied". That call is yours. github.com/simonlin1212/V…

美国拟禁中国光模块。 【来源与时间】 路透社独家，2026年8月4日 14:08 GMT（北京时间 22:08）， 援引四名知情人士。起草方是 FCC（美国联邦通信委员会）， 不是商务部。官员希望年内发布，发布即生效。 路透同时强调：FCC 仍可能修改或搁置。 先厘清一个时间差：A股8月4日15:00收盘，这条稿件22:08才发。 当天光模块的大涨（中际旭创+13.2%、新易盛+13.7%、天孚+17.5%） 驱动是英伟达 Vera Rubin 与 Spectrum-6 发布 ＋ 野村上调预测， 与这条新闻无关。 美股是在稿件发出后涨的：应用光电+19.4%、Fabrinet+16.5%、 Coherent+12.4%、Lumentum+8.9%，而英伟达只有+2.6%。 这不是板块普涨，是光模块供应链的定向换仓。 【机制：不是"禁中国"，是"禁非美国产 + 逐个豁免"】 路透明确说，这次沿用 FCC 对无人机、路由器、逆变器、机器人 的同一套模板。查 FCC 官方对7月28日机器人/逆变器那次的说明， 架构是： · "foreign-produced" 定义为任何不构成美国 domestic end product 的物品（48 CFR §25.101(a)） · FCC 原文：生产实体的国籍"不相关"；该行动"country neutral" · 新型号拿不到设备认证 = 不得进口销售； 已获认证的存量型号，继续进口、销售、使用不受影响 · 唯一出口：向美国国防部（DoW）申请 Conditional Approval， 可按单一型号，也可按整类或整个实体批准 也就是说：基线是"所有非美国产的新型号一律禁"， 歧视不写在规则里，落在豁免审批环节。 路透的措辞是"豁免许多非中国供应商"——是许多，不是全部。 【由此，流传最广的那条安慰是错的】 昨夜多份解读的核心论点之一是： "头部厂商早已布局东南亚产能，北美出货基本来自海外产能， 因此即使极端情况也有充分应对空间。" 这在这套规则下不成立。 泰国产 = 非美国产 = 仍在禁令基线之内。 搬去东南亚规避的是关税，规避不了设备认证这道闸。 两套逻辑的适用条件完全不同。 真正的出路只有两条： 一是美国本土产能——2025年前后才起步，太早； 二是拿到国防部豁免——而中际旭创、新易盛已于2026年6月8日 被列入五角大楼1260H清单（同批共65个实体）。 同理，另一侧也被高估了： Coherent 的马来西亚/越南/菲律宾工厂、Lumentum 与 Fabrinet 的泰国工厂，在规则文本里同样是"非美国产"，同样需要申请豁免。 买它们，押的是国防部的自由裁量，不是条文。 【伤害的真实形状：不砍存量，砍代际】 存量已认证型号照常出货。被禁的是新型号—— 而1.6T与2.4T恰恰是2026-2027的全部增长来源。 野村上调预测的理由正是"800G、1.6T为主要收入来源， 2027年起2.4T与NPO进入商业化"。 这条规则打的不是当期利润， 是估值倍数所依赖的那条增长曲线。 所以"在手订单已覆盖2026全年"这个安慰，恰好答不了这个问题。 全部问题因此收敛为一个可查的事实： 主要厂商的1.6T型号，在规则发布之前拿到认证了没有。 规则"年内发布、发布即生效"，意味着从现在到发布日， 理性行为是抢认证。 【另一侧的力量同样真实】 · 替代能力有物理上限。美方智库 Foundation for American Innovation 自己的报告承认，Coherent 与 Lumentum "缺乏替代中国供应商的规模"。中际旭创全球数通光模块 份额27%（Counterpoint），90%收入来自中国以外。 · 路透明言这会推高 AWS 等美国云厂的成本。 · 美国商务部今年2月已把数据中心设备进口限制作为 贸易缓和的一部分搁置，FCC 是绕过缓和的另一条通道； 但这也意味着，放松同样可以只在豁免环节完成，不必改规则。 · 中国在这条链上握有直接反制杠杆：磷化铟（InP）自2025年2月 起纳入中国出口管制，中国控制全球约七成铟储量与产能， 而海外厂商要扩产替代，吃的正是 InP。管制后6吋 InP 价格一度上涨约250%。美国要卡成品，中国卡的是它 替代所需的原料。 【结论】 一、这是一条用国家中立的写法、通过豁免环节精准指向中国、 并且在文本层面封死了东南亚退路的规则。 三点缺一不可，漏掉任何一点都会读错这条新闻。 二、它不砍当期收入，砍的是1.6T之后的整条增长曲线， 也就是高估值里最贵的那一段。当期订单饱满不构成对冲。 三、它仍然只是草案。落地与否的观察点不在新闻，在三处： 规则发布时对"新型号"的口径定义、国防部豁免名单的实际构成、 以及中国在磷化铟上的反应。 四、短期定价的是情绪，中期定价的是规则口径， 长期定价的仍是良率与交付。 但这一次与2020年以来历次传闻有一个关键区别： 过去中国厂商能靠转移产能化解，这一次那条路 在规则文本层面被堵住了。 以上为公开信息整理与产业分析，不构成投资建议。

I shipped an AI multi-agent dashboard this week. Then I cut the AI down to about 10% of what's on the screen. On purpose. It's a daily-review board for short-term A-share traders. Every evening it pulls the limit-up pool, the consecutive-limit ladder, block-trade prints, sector money flow and theme attribution, and collapses an hour of manual work into one screen. Five analyst agents, one judge agent. Standard setup. But every number you actually read is computed, not generated. The reason is boring: those readings are arithmetic. Median move of yesterday's strong names, follow-through rate, ladder distribution, market breadth. Each has exactly one right answer. Hand that to a model and it will probably get it right, or it will get it plausibly wrong. Nobody questions a number that's well formatted and looks reasonable. So the split is: ▸ Hard metrics: pure computation, never touches the model, there the second you open it ▸ Five analysts + judge: the only job the LLM has is turning eight data sources into something a human can read Unplanned benefit. When your key expires or the model rate-limits, 90% of the product still works. Built the other way round, a model outage is a blank page. 341 tests. Most of them aren't guarding against crashes. Crashes announce themselves. They guard against the other kind: renders fine, numbers look reasonable, conclusion is wrong. Runs fully local. No API key needed if you already have a Claude or Codex subscription, it just drives the CLI you're logged into. It doesn't pick stocks. Reviewing what already happened is factual work, and that never needed a recommendation. Apache-2.0: github.com/simonlin1212/v…

Meta: free cash flow $784M, down 91%. Capex took 97.6% of operating cash flow. Cash halved to $15.5B. It still guided 2026 capex up. Microsoft: capex +110%, free cash flow still $19.6B, capex at 65% of operating cash flow. Revenue isn't the problem — Meta grew 28%, faster than Microsoft's 18%. Its operating income fell 8%. You can fund +100% capex from profit, from cash, or from lenders. Only one of those compounds. So the number that breaks first isn't capex, it's the growth rate of capex. Not investment advice. Company filings, 2026-07-29.

Microsoft and Meta both reported after the close. Both are spending record money on AI. Microsoft rose 3% after hours, Meta fell 8%. The gap is not about how much they spend. Microsoft: revenue $90.0B (+18%), operating income $40.6B (+18%), adjusted EPS $4.74 against $4.24 expected. Azure grew 43% and passed $100B annually. Backlog rose 84% to $678B. Capex doubled to $35.8B for the quarter, $115.9B for the year, and free cash flow was still $19.6B. Capex took 65% of operating cash flow. Meta: revenue $60.80B (+28%), which is faster growth than Microsoft. Operating income fell 8% to $18.78B. Margin went from 43% to 31%. Costs rose 55%, depreciation rose 46%. Capex $31.08B against operating cash flow of $31.86B, leaving free cash flow of $784M, down 91%. Cash fell from $35.9B to $15.5B. One ratio explains both reactions. Capex as a share of operating cash flow: Microsoft 65%, Meta 98%, Alphabet 115%. Alphabet's free cash flow was negative $5.86B last quarter, and in June it priced an $84.75B equity raise to fund AI infrastructure. AI capex has moved from being paid out of profit to being paid out of cash cushions and shareholder equity. Depreciation is the lagging part, and it does not negotiate. None of them is cutting. Meta raised the floor of its 2026 guidance to $130–145B. As long as financing stays available, the orders hold. Amazon reports tonight and AWS is the cleanest test, since it sells compute without another business attached. Not investment advice. Figures from company filings, 2026-07-29.

Nvidia -5% Monday. China's optics and PCB names fell 10-17% in a single session, now 35-50% off their highs. Every chart screams "dip." I'm passing. Here's the full reasoning — including the data that argues against me. ▸ THE BULL DATA IS REAL I spent two days pulling every price I could find: · B200 listed rents: $5.94/hr median — roughly 2x a rough breakeven estimate · Rental forward curve: August priced HIGHER than July · China grey market: a DGX B300 server doubled in six months to ~2.75x US retail · DRAM contracts +13-18% QoQ; refurbished H100s sell at just 15% below new Physical demand for 2026 is not dead. That build is contracted, and it will happen. ▸ WHAT ACTUALLY BROKE The WSJ reported Nvidia is in talks to backstop ~$250B of financing tied to OpenAI's datacenter buildout. Read that twice: the seller guaranteeing the customer's ability to buy the seller's product. Telecom veterans have seen this movie. Lucent and Nortel financed their customers' purchases in 1999. Revenue looked great — until it didn't. Nvidia's CDS just had its biggest one-day move on record. The market is repricing the guarantor. ▸ THE LOGIC THAT MADE ME WALK Every price I can track — rents, grey market, memory — is a LAGGING indicator. It measures today's scarcity, not tomorrow's returns. What broke last week is the financing layer, and credit is a LEADING indicator. And one simple thought experiment: if any big buyer cuts capex, compute gets cheaper for everyone else — so the others don't need to add. The "everybody adds forever" assumption fails in both directions. I have no edge on 2027 credit quality. And when you have no edge, you don't need a position. There's an old Chinese saying: a wise man does not stand under a wall that is being load-tested. ▸ WHERE MY ATTENTION GOES INSTEAD Humanoid robotics. Not because it's cheap — it isn't — but because its driver is a product cycle, not a credit cycle: · During Monday's compute crash, humanoid supply-chain names fell ~2% while compute proxies fell 10-17%. One day proves little, but the market drew a line. · Tesla's Optimus V3 reveal window is Jul-Aug; the Fremont line targets ~70k units/yr annualized · Unitree cleared IPO registration in 104 days — about to become the first PROFITABLE humanoid maker to go public · 2025 global shipments: 13,318 units, +479% YoY The layer I study is the bottleneck components — harmonic reducers, planetary roller screws, dexterous hands — where irreplaceability lives. ▸ METHOD Same discipline as always: predefined entry bands, no chasing, no averaging into falling knives. Watching is also a position. Disclosure: I hold a position in a Chinese humanoid supply-chain company (harmonic drives). No position in any compute name. Not investment advice.

Everyone paywalls options Greeks. I built a free one. Full Greeks + IV + same-day 0DTE flow, straight from CBOE's official feed. yfinance doesn't have Greeks. OpenBB's free tier doesn't either. Also in there: daily short volume for all 12,112 US symbols (FINRA), a market-wide screener across 5,309 companies (SEC EDGAR), and same-day insider + institutional filings. 13 layers, 30+ endpoints, zero API keys. One Python file, MIT licensed. Built with Claude Code. Repo link below.

━━━━━━━━━━━━━━━━━━━━ CXMT LISTS TOMORROW 20260727 · WHAT TWO MARKETS PRICE ━━━━━━━━━━━━━━━━━━━━ China's largest DRAM maker — world #4 — lists on the STAR Market Monday. IPO price 8.66 CNY, or 579B CNY (~$85B). Only 6.73% of the shares float on day one. ▸ TWO MECHANISMS, ONE NUMBER Hyperliquid pre-IPO perp (xyz:CXMT) $6.09 → 2.77T CNY (~$408B) Polymarket day-1 closing-mcap ladder median ~2.76T CNY A leveraged perp and a set of binary event contracts. Different traders, different mechanics, no shared order book. They land 0.4% apart. The convergence is the signal, not either number by itself. ▸ THE LADDER ≥2.0T CNY 88.5% ≥2.5T 65.5% ≥3.0T 35.5% ≥3.5T 21.5% Look at the tail. The 3.0–3.5T bucket holds only 14%, but everything above 3.5T holds 21.5%. The book is pricing a break of 3T as a break to somewhere much higher. Also: the perp has de-rated 18% in eleven days, from 3.37T on pricing day to 2.77T tonight. ▸ THE WEEKEND CUT BOTH WAYS Friday — Micron -7.0%, SK Hynix -8.3%, Samsung -7.6%, SOXX -4.4%. Saturday — Korea and the US announced $950B of memory MOUs. SK Hynix: $750B of long-term supply to US buyers, including a $500B+ Nvidia partnership and a 2GW Vera Rubin + HBM4 datacenter for 2027. Samsung: $200B with Broadcom, covering sub-2nm foundry and packaging. SK's chairman relayed that Broadcom's CEO told him memory demand would likely surpass available supply. The perp went 2.81T → 2.77T across both. Professional money read an 8% sector drawdown and a $950B demand signal as roughly offsetting. ▸ THE PART NOBODY IS PRICING 2.77T is ~26x 2026 earnings — H1 guidance doubled, then stripped of ~25% minority-interest leakage. The IPO price was ~5x. Retail is anchoring on Micron's ~$1T cap instead. But the prospectus says the H1 surge is not sustainable. 2025 depreciation ran 4.6x adjusted net income. And the 29.5B CNY use-of-proceeds contains zero HBM — I grepped all 401 pages. HBM is the option. DRAM contract pricing is the earnings. ▸ METHOD Pulled tonight: Hyperliquid /info, Polymarket gamma API, the listing filing on CNINFO (Jul 24), Yahoo closes. FX 6.79. Share count 66.881B. Event-contract depth is thin (~$160k total) — read the ladder as a sentiment scale, not a price. No position. I applied for the IPO and was not allocated. Not investment advice.

━━━━━━━━━━━━━━━━━━━━ 📡 HBM HIRING RADAR 20260726 · BASELINE ━━━━━━━━━━━━━━━━━━━━ SK HYNIX vs MICRON ▸ SNAPSHOT SK hynix America 42 openings Micron (HBM roles) 193 openings First edition = baseline. From here, every issue ships the delta. ▸ THE TREND The leader is moving toward the customer. The challenger is moving toward margin. ▸ THE DETAIL SK hynix — 76% of US roles sit in San Jose. And they aren't memory jobs: • 3D Stacked DRAM-on-Logic Design Engineer • AI/HPC System Architect (x2) • AI Memory Solution Architect • Director, System-in-Package Architecture DRAM-on-logic is base-die work. You don't hire four system architects in Silicon Valley to sell commodity die. That's a memory vendor refusing to be commoditized. Micron — 193 HBM roles across six countries: • Singapore (Fab 10A) — 54 • Richardson, TX — 47 • Taichung — 22 · Hyderabad — 17 • Jalisco, MX — 16 · Boise HQ — 8 Design in Texas, packaging in Singapore, 28 layout seats largely in Mexico. That's a cost curve being engineered. ▸ METHOD Public ATS endpoints (Greenhouse + Workday), keyword-matched. Baseline edition — no month-over-month claim made.

━━━━━━━━━━━━━━━━━━━━ 📡 HUMANOID HIRING RADAR 20260726 · 30-DAY DIFF ━━━━━━━━━━━━━━━━━━━━ TESLA OPTIMUS vs FIGURE ▸ SNAPSHOT (Jun 25 → Jul 26) Tesla Optimus 231 → 282 (+22%) Figure 101 → 122 (+21%) Tesla opened 80 new reqs, net +51 — meaning ~29 closed or filled. This board is moving, not parked. Figure posted 25 new reqs in the window. ▸ THE TREND Same growth rate. Opposite bets. Tesla is buying capacity. Figure is buying deployment. ▸ THE DETAIL Tesla's 80 new reqs: • Manufacturing — 34 (42%) • Tesla AI — 31 • Austin, TX — 24 of them Production-hell playbook: build the line first. Figure's 25 new reqs: • Data Collection — 9 • BotQ Manufacturing — 3 • Commercial Operations — 3 But the titles are the real signal: "Humanoid Robot Operator — Commercial Site Team" "Apprentice Robot Service Technician" "Fleet Coordinator" · "TeleOps Quality Support" You don't hire site operators for robots that aren't on sites yet. ▸ ONE MORE THING Figure opened AI Data Operations Manager roles in the UAE and Mexico City this month, plus its first Asia supply-chain seats. Data ops is going offshore before the robots ship at volume. ▸ METHOD Public ATS endpoints, identical query both months. Tesla exposes no post dates — new reqs inferred from ascending job IDs.

Most "free" US market data APIs stop at quotes. No options Greeks. No short volume. No SEC filings. So I built one that has them — 13 layers, 30+ endpoints, zero auth, no API key. Every source labeled with its license tier. github.com/simonlin1212/g…

I built a-stock-data — an open-source A-share data toolkit for AI agents (Claude Code, Codex). Launched May 2026, now 6.7k⭐ on GitHub. Zero auth, 100% free: quotes, K-lines, research reports, fund flows, filings, limit-up boards, ETF options & more. github.com/simonlin1212/a…