Personal Private Programmable: Balaji on Claude Code + Obsidian + Crypto¶
Core Thesis¶
Balaji Srinivasan identifies an emerging tech stack he calls "personal private programmable" — the intersection of Claude Code, Obsidian, open-weight AI models, and crypto wallets (ENS/SNS). He argues that within months, open-weight AI models will match Claude Code's capabilities and handle Google Gemini's context window. When local AI can process all personal data (markdown, git repos, emails) while crypto wallets enable encryption and ENS/SNS names enable secure transmission, "redecentralization and keeping data local becomes advantageous" — local data is easier to encrypt (crypto) and compute on (local AI). Conclusion: the personal becomes private and programmable.
WHY: Cloud-based AI requires sending data to external servers (privacy risk). Local AI + crypto enables both privacy (encryption) and power (programmable).
WHAT: Convergence of open-weight AI models, diverse file types (Obsidian, git, email), and crypto identity (ENS/SNS).
SO WHAT: Reverses centralization trend — "it's now better to redecentralize" because local data unlocks both encryption (crypto) and computation (AI).
Three Converging Technology Trends¶
1. Open-Weight AI Models (Local Computation)¶
Current State: - Claude Code works well today - Google Gemini handles large context windows
Projection (Next Few Months):
"If you squint ahead a few months, we will likely have open weight AI models that work as well as Claude Code does today, and that can handle the context that Google Gemini does."
Implications: - No need to send data to Anthropic/Google servers - Full AI capabilities running locally - Privacy + power in one package
Examples of Open-Weight Models: - LLaMA 3, Mistral, DeepSeek - Trend: Models shrinking while capabilities grow - Hardware: Apple Silicon, NVIDIA consumer GPUs increasingly sufficient
2. Diverse Personal Data File Types¶
What's Available: - Markdown files — Obsidian notes, documentation - Git repos — Code, version history, commits - Mbox files — Email archives - Miscellaneous file types — PDFs, spreadsheets, databases, photos, documents stored online or on disk
Key Insight:
"The open weight AI models can do a lot if they can see all these different data types at once."
Obsidian as Reference Implementation: - Plain markdown files (portable, future-proof) - Local-first (data on your machine) - Graph view (connections between notes) - Claude Code integration — AI can read, write, organize your notes
Advantage of Local File Access: - AI sees everything in unified context - No API rate limits - No upload/download delays - Full filesystem access (read git history, parse emails, cross-reference documents)
3. Crypto Wallets & Name Systems (Identity + Encryption)¶
Infrastructure Already Exists: - Hundreds of millions of local crypto wallets — MetaMask, Phantom, Rainbow, Coinbase Wallet - Scaled crypto name systems: - ENS (Ethereum Name System) — balaji.eth - SNS (Solana Name System) — balaji.sol
What These Provide: 1. Encrypted data handling — Public/private key cryptography 2. Identity — Human-readable names (balaji.eth) map to addresses 3. Secure transmission — Encrypt to recipient's public key
Workflow Example:
1. You have local data (Obsidian vault, git repos)
2. Local AI processes data (Claude Code summarizes, analyzes)
3. Encrypt output with crypto wallet (to friend.eth's public key)
4. Transmit to friend.eth via ENS/SNS
5. They decrypt locally with their private key
6. End-to-end encrypted collaboration without central server
The Synthesis: Personal Private Programmable¶
Three Capabilities Combined:
| Capability | Technology | Benefit |
|---|---|---|
| Personal | Local file access (Obsidian, git, email) | AI sees all your data in context |
| Private | Crypto wallets (encrypt with keys) | Data never leaves your machine unencrypted |
| Programmable | Open-weight AI (Claude Code locally) | Full AI capabilities without cloud |
Why This Changes Everything:
Before (Cloud-Based AI): - Send data to Anthropic/OpenAI → privacy risk - Limited context (API constraints) → capability limitation - Subscription fees → ongoing cost
After (Local AI + Crypto): - Process data locally → privacy preserved - Full context (all files accessible) → maximum capability - One-time hardware cost → no subscriptions
Balaji's Key Phrase:
"The personal data becomes far more programmable (with local AI) and yet also more private (because it's being computed on locally)."
Collaboration via ENS/SNS¶
Problem: How do you collaborate with others while keeping data private?
Solution: End-to-end encryption via crypto name systems.
Workflow:
- Your Setup:
- Local AI (Claude Code) processes your Obsidian vault
- AI generates summary/analysis
-
Crypto wallet encrypts output to
collaborator.eth -
Transmission:
- Encrypted data sent to
collaborator.ethaddress -
Could use IPFS, Arweave, or even email (data is encrypted)
-
Recipient Decrypts:
collaborator.ethreceives encrypted data- Their crypto wallet decrypts with private key
-
Their local AI processes decrypted data
-
No Central Server Sees Plaintext:
- Unlike Google Docs, Notion, or Slack
- True end-to-end encryption
- Identity via ENS/SNS (not phone numbers or emails)
Advantages Over Current Collaboration Tools: - vs Google Docs: Google can read everything; ENS/SNS encrypted collaboration = Google sees nothing - vs Slack: Centralized server stores all messages; ENS/SNS = peer-to-peer encrypted - vs Email: Email is plaintext by default; crypto enforces encryption
Technical Challenges (Acknowledged)¶
Balaji admits this vision requires work:
"There are many networking details to be worked out regarding secure synchronization of packets between different ENS/SNS names across different machines."
Specific Challenges:
1. Synchronization Across Machines¶
Problem: If you have laptop + desktop + phone, how do they stay in sync?
Current Solutions (Insufficient): - Obsidian Sync (paid, centralized) - Git (requires manual push/pull) - Dropbox/iCloud (centralized, not encrypted)
Crypto-Native Solution Needed: - Decentralized sync protocol - Automatic conflict resolution - End-to-end encrypted - Works across devices with same ENS/SNS identity
2. Pure P2P Is Hard¶
"You might need some kind of private cloud like Gitlab to really make it work depending on the application (because pure p2p is hard)."
Why P2P Is Hard: - NAT traversal (firewalls, routers) - Offline devices (if recipient is offline, where does data wait?) - Discovery (how do you find collaborator.eth's IP address?)
Hybrid Solution: - Private GitLab instance as relay - Data is still encrypted end-to-end - GitLab just stores encrypted blobs, can't read contents - Similar to how Signal uses servers for message delivery but can't read messages
3. Key Management¶
Problem: Losing private key = losing all encrypted data.
Solutions Emerging: - Social recovery (Argent, Gnosis Safe) - MPC wallets (multi-party computation, key shards) - Hardware wallets (Ledger, Trezor)
The Redecentralization Thesis¶
Core Argument:
"The Claude/Obsidian trend points at a powerful emerging concept where it's now better to redecentralize and keep all your data local."
Historical Context:
1990s-2000s: Decentralized (Desktop Software) - Microsoft Office on local machine - Email downloaded via POP3/IMAP - Files on local disk
2010s: Centralized (Cloud Software) - Google Docs, Office 365 - Gmail web interface - Dropbox, iCloud
Why Centralization Won Then: - Collaboration (real-time co-editing) - Sync across devices - No local software installation - Backups handled by provider
2020s: Redecentralization (Local AI + Crypto) - Local AI = power without sending data to cloud - Crypto = collaboration without central server - Open-weight models = no vendor lock-in
Why Redecentralization Wins Now: - Privacy: Data never leaves your machine unencrypted - Capability: Local AI can see all your data at once (no API limits) - Ownership: You control the data and the compute - Cost: One-time hardware vs ongoing subscriptions
Balaji's Punchline:
"Because local data is easier to encrypt with crypto and compute on with local AI. Thus: the personal becomes private and programmable."
Obsidian as Proof of Concept¶
Why Obsidian Matters: - Local-first: Plain markdown files on your disk - Portable: No vendor lock-in (files are just markdown) - Extensible: Plugins, APIs, scripting - Graph view: Visualizes connections between notes
Claude Code + Obsidian = Killer Combo: - Claude can read your entire vault (all notes in context) - Claude can write new notes, update existing ones - Claude can follow links, build connections - All computation happens locally
Example Workflow: 1. You have 10,000 notes in Obsidian vault 2. Ask Claude Code: "Summarize everything I've written about AI safety" 3. Claude reads all relevant notes (local file access, no upload) 4. Claude generates comprehensive summary 5. Claude creates new note with cross-references 6. All without sending data to Anthropic
This was impossible with cloud AI: - Upload 10,000 notes to ChatGPT? (Slow, expensive, privacy risk) - API limits? (Context window too small) - But local Claude Code? Easy.
Implications & Predictions¶
For Individuals¶
Near-Term (2026-2027): - Open-weight models reach Claude Code quality - Local AI becomes standard for knowledge workers - Obsidian-like tools proliferate (Logseq, Foam, Dendron)
Medium-Term (2028-2030): - ENS/SNS name ownership becomes standard (like email today) - End-to-end encrypted collaboration becomes default - "Cloud storage" sounds as archaic as "mainframe"
Long-Term (2030+): - Personal AI trained on your entire life's data (all local) - Zero-knowledge proofs enable selective disclosure - "I sent it to alice.eth" = normal phrase
For Companies¶
Threat to Incumbent SaaS: - Notion, Roam Research, Evernote → Why pay subscription if local AI is better? - Google Docs → Why share data with Google if encrypted collaboration works? - Slack, Discord → ENS/SNS encrypted chat = same UX, better privacy
Opportunities: - Hardware (local AI acceleration) - Tooling (Obsidian, GitLab, etc.) - Crypto infrastructure (ENS/SNS wallets, key management)
For Web3¶
Validation of Crypto Thesis: - Crypto isn't just "magic internet money" - Real utility: encryption, identity, secure collaboration - ENS/SNS becomes the "email of web3"
Infrastructure Needs: - Better UX for wallets (most people can't manage keys) - Sync protocols (decentralized Dropbox) - Discovery mechanisms (how to find friend.eth's devices)
Critiques & Open Questions¶
Critique 1: "Open-Weight Models Aren't There Yet"¶
Counterargument: - LLaMA 3 70B already competitive with GPT-3.5 - DeepSeek V3 approaching GPT-4 quality - Trajectory is clear (exponential improvement) - Balaji says "a few months" — optimistic but not unreasonable
Critique 2: "Most People Won't Run Local AI"¶
Counterargument: - Apple Silicon Macs can already run 7B-13B models locally - Next-gen hardware (M4, M5) will handle 70B+ models - Anthropic could ship "Claude Local" (like Photoshop vs Photoshop.com) - Once someone experiences "AI that works offline and sees all my files," hard to go back
Critique 3: "Crypto UX Is Terrible"¶
Acknowledgment: - Key management is hard - Most people lose seed phrases - ENS names cost money (registration, renewal)
Counterargument: - Social recovery wallets solve key loss - ENS UX improving (Coinbase integration, etc.) - Once killer app exists (encrypted AI collaboration), UX improves fast
Critique 4: "GitLab as Relay = Not Decentralized"¶
Balaji's Admission: - Pure P2P is hard - Hybrid approach (private GitLab) is pragmatic
Nuance: - GitLab only stores encrypted blobs (can't read data) - Still better than Google Docs (Google can read everything) - Long-term: Decentralized alternatives (IPFS, Arweave, Filecoin)
Open Question: What About Mobile?¶
Challenge: Phones have limited compute for local AI.
Possible Solutions: - Smaller models (7B-13B) on phones - Heavy compute on laptop/desktop, sync encrypted results to phone - Future: Mobile chips improve (Apple A-series, Snapdragon)
Connection to Other Trends¶
1. Local-First Software Movement¶
Key Text: Martin Kleppmann's "Local-First Software" (Ink & Switch)
Principles: - No spinners (data is local, instant) - Your work is not trapped on one device - Network optional - Collaborate with others - Data lasts forever (no startup shutdowns)
Balaji's Vision = Local-First + AI + Crypto
2. Zero-Knowledge Proofs¶
Future Extension: - Not just encrypt data to friend.eth - Prove facts about data without revealing data - Example: "I have 10,000 notes about AI" (provable, but notes stay private)
3. Sovereign Identity¶
Self-Sovereign Identity (SSI) Movement: - Control your own identity (not Facebook, Google, government) - ENS/SNS as building blocks - Local AI + crypto wallet = digital autonomy
Actionable Takeaways¶
For Individuals (Now)¶
- Start using Obsidian (or similar local-first tool)
- Export notes from Notion, Evernote, Roam
- Get comfortable with markdown
-
Build your "second brain" locally
-
Get a crypto wallet + ENS/SNS name
- MetaMask + ENS name (yourname.eth)
-
Experiment with encrypted messaging (Status, XMTP)
-
Try Claude Code with local files
- Point Claude at your Obsidian vault
- Experience "AI that sees all your data at once"
- Understand why this is different from ChatGPT
For Developers (Now)¶
- Build local-first tools
- Obsidian plugins
- Local AI wrappers (llama.cpp, Ollama)
-
Encrypted collaboration layers
-
Experiment with ENS/SNS identity
- XMTP (messaging via ENS)
- Ceramic Network (decentralized data)
-
IPFS + ENS (distributed storage)
-
Prepare for open-weight model parity
- Fine-tune local models on domain data
- Build tooling that assumes local AI = default
For Companies (Strategic)¶
- SaaS companies: Prepare for "local mode"
- How do you monetize if users run AI locally?
-
Hardware sales? Support contracts? Freemium local + paid cloud?
-
Crypto companies: Focus on UX
- Key management is the blocker
-
Solve seed phrase problem = unlock adoption
-
Hardware companies: Optimize for local AI
- Apple, NVIDIA, AMD → AI acceleration chips
- Memory bandwidth matters (transformer models are memory-bound)
Balaji「個人、私密、可程式化」:Claude Code + Obsidian + 加密(繁體中文詳細版)¶
來源: https://x.com/balajis/status/2016437837013660071
作者: Balaji Srinivasan (@balajis)
日期: 2026-01-28
收藏日期: 2026-01-28
標籤: #去中心化 #本地AI #隱私 #加密貨幣 #obsidian #claude-code
核心論述¶
Balaji Srinivasan 提出「個人、私密、可程式化」(Personal Private Programmable)的新興科技堆疊 — Claude Code、Obsidian、開放權重 AI 模型、加密錢包(ENS/SNS)的交會點。他主張未來幾個月內,開放權重 AI 模型將達到 Claude Code 水準並處理 Google Gemini 的上下文窗口。當本地 AI 能處理所有個人資料(markdown、git repos、emails)的同時,加密錢包能加密資料、ENS/SNS 名稱能安全傳輸時,「重新去中心化、保持資料在本地變得有利」 — 本地資料更易加密(crypto)、更易運算(local AI)。結論:個人變得私密且可程式化。
WHY(為什麼重要): 雲端 AI 需要發送資料到外部伺服器(隱私風險)。本地 AI + 加密同時實現隱私(加密)和能力(可程式化)。
WHAT(核心機制): 開放權重 AI 模型、多樣檔案類型(Obsidian、git、email)、加密身份(ENS/SNS)的匯聚。
SO WHAT(影響): 逆轉中心化趨勢 —「現在重新去中心化變得更好」因為本地資料同時解鎖加密(crypto)和運算(AI)。
三個匯聚的技術趨勢¶
1. 開放權重 AI 模型(本地運算)¶
現況: - Claude Code 目前表現良好 - Google Gemini 處理大型上下文窗口
預測(未來幾個月):
「如果你往前看幾個月,我們很可能會有開放權重 AI 模型,其表現與今天的 Claude Code 一樣好,並且能處理 Google Gemini 的上下文。」
意義: - 無需發送資料到 Anthropic/Google 伺服器 - 完整 AI 能力在本地運行 - 隱私 + 能力合而為一
開放權重模型範例: - LLaMA 3、Mistral、DeepSeek - 趨勢:模型變小、能力增強 - 硬體:Apple Silicon、NVIDIA 消費級 GPU 越來越足夠
2. 多樣的個人資料檔案類型¶
可用的資料: - Markdown 檔案 — Obsidian 筆記、文件 - Git repos — 程式碼、版本歷史、commits - Mbox 檔案 — Email 檔案 - 雜項檔案類型 — PDFs、試算表、資料庫、照片、線上或硬碟上的文件
關鍵洞察:
「開放權重 AI 模型如果能同時看到所有這些不同資料類型,就能做很多事。」
Obsidian 作為參考實作: - 純 markdown 檔案(可移植、未來證明) - Local-first(資料在你機器上) - 圖形視圖(筆記間連結) - Claude Code 整合 — AI 能讀、寫、組織你的筆記
本地檔案存取的優勢: - AI 在統一上下文中看到一切 - 無 API 速率限制 - 無上傳/下載延遲 - 完整檔案系統存取(讀 git 歷史、解析 emails、交叉參考文件)
3. 加密錢包與名稱系統(身份 + 加密)¶
基礎設施已存在: - 數億個本地加密錢包 — MetaMask、Phantom、Rainbow、Coinbase Wallet - 規模化的加密名稱系統: - ENS(Ethereum Name System) — balaji.eth - SNS(Solana Name System) — balaji.sol
這些提供什麼: 1. 加密資料處理 — 公鑰/私鑰密碼學 2. 身份 — 人類可讀名稱(balaji.eth)對應到地址 3. 安全傳輸 — 用接收者公鑰加密
工作流範例:
1. 你有本地資料(Obsidian vault、git repos)
2. 本地 AI 處理資料(Claude Code 總結、分析)
3. 用加密錢包加密輸出(到 friend.eth 的公鑰)
4. 透過 ENS/SNS 傳輸給 friend.eth
5. 他們用私鑰本地解密
6. 端到端加密協作,無中央伺服器
綜合:個人、私密、可程式化¶
三個能力結合:
| 能力 | 技術 | 好處 |
|---|---|---|
| 個人 | 本地檔案存取(Obsidian、git、email) | AI 在上下文中看到你所有資料 |
| 私密 | 加密錢包(用密鑰加密) | 資料永不以未加密形式離開你機器 |
| 可程式化 | 開放權重 AI(本地 Claude Code) | 完整 AI 能力,無需雲端 |
為什麼這改變一切:
以前(雲端 AI): - 發送資料到 Anthropic/OpenAI → 隱私風險 - 有限上下文(API 限制)→ 能力限制 - 訂閱費用 → 持續成本
現在(本地 AI + 加密): - 本地處理資料 → 隱私保護 - 完整上下文(所有檔案可存取)→ 最大能力 - 一次性硬體成本 → 無訂閱
Balaji 的關鍵語句:
「個人資料變得更可程式化(透過本地 AI),卻也更私密(因為在本地運算)。」
透過 ENS/SNS 協作¶
問題: 如何在保持資料私密的同時與他人協作?
解決方案: 透過加密名稱系統的端到端加密。
工作流:
- 你的設定:
- 本地 AI(Claude Code)處理你的 Obsidian vault
- AI 產生總結/分析
-
加密錢包將輸出加密給
collaborator.eth -
傳輸:
- 加密資料發送到
collaborator.eth地址 -
可使用 IPFS、Arweave、甚至 email(資料已加密)
-
接收者解密:
collaborator.eth收到加密資料- 他們的加密錢包用私鑰解密
-
他們的本地 AI 處理解密後的資料
-
無中央伺服器看到明文:
- 不像 Google Docs、Notion、Slack
- 真正的端到端加密
- 透過 ENS/SNS 身份(不是電話號碼或 emails)
相較於現有協作工具的優勢: - vs Google Docs: Google 能讀一切;ENS/SNS 加密協作 = Google 看不到任何東西 - vs Slack: 中央伺服器儲存所有訊息;ENS/SNS = 點對點加密 - vs Email: Email 預設是明文;加密強制加密
技術挑戰(已承認)¶
Balaji 承認這個願景需要工作:
「關於不同機器上不同 ENS/SNS 名稱之間封包的安全同步,有許多網路細節需要解決。」
具體挑戰:
1. 跨機器同步¶
問題: 如果你有筆電 + 桌機 + 手機,如何保持同步?
現有解決方案(不足): - Obsidian Sync(付費、中心化) - Git(需要手動 push/pull) - Dropbox/iCloud(中心化、未加密)
需要加密原生解決方案: - 去中心化同步協議 - 自動衝突解決 - 端到端加密 - 跨裝置運作,同一 ENS/SNS 身份
2. 純 P2P 很難¶
「你可能需要某種私有雲,像 GitLab,才能真正讓它運作,取決於應用(因為純 P2P 很難)。」
為什麼 P2P 很難: - NAT 穿透(防火牆、路由器) - 離線裝置(如果接收者離線,資料在哪裡等待?) - 發現(如何找到 collaborator.eth 的 IP 位址?)
混合解決方案: - 私有 GitLab 實例作為中繼 - 資料仍然端到端加密 - GitLab 只儲存加密 blobs,無法讀取內容 - 類似 Signal 使用伺服器傳遞訊息但無法讀取
3. 密鑰管理¶
問題: 遺失私鑰 = 失去所有加密資料
新興解決方案: - 社交恢復(Argent、Gnosis Safe) - MPC 錢包(多方計算、密鑰碎片) - 硬體錢包(Ledger、Trezor)
重新去中心化論述¶
核心論點:
「Claude/Obsidian 趨勢指向一個強大的新興概念,現在重新去中心化並保持所有資料在本地變得更好。」
歷史背景:
1990s-2000s:去中心化(桌面軟體) - Microsoft Office 在本地機器 - Email 透過 POP3/IMAP 下載 - 檔案在本地硬碟
2010s:中心化(雲端軟體) - Google Docs、Office 365 - Gmail 網頁介面 - Dropbox、iCloud
為什麼當時中心化贏了: - 協作(即時共同編輯) - 跨裝置同步 - 無需本地軟體安裝 - 供應商處理備份
2020s:重新去中心化(本地 AI + 加密) - 本地 AI = 能力,無需發送資料到雲端 - 加密 = 協作,無需中央伺服器 - 開放權重模型 = 無供應商鎖定
為什麼現在重新去中心化贏了: - 隱私: 資料永不以未加密形式離開你機器 - 能力: 本地 AI 能同時看到你所有資料(無 API 限制) - 所有權: 你控制資料和運算 - 成本: 一次性硬體 vs 持續訂閱
Balaji 的結論:
「因為本地資料更易用加密加密、更易用本地 AI 運算。因此:個人變得私密且可程式化。」
Obsidian 作為概念驗證¶
為什麼 Obsidian 重要: - Local-first: 純 markdown 檔案在你硬碟上 - 可移植: 無供應商鎖定(檔案只是 markdown) - 可擴展: 插件、APIs、腳本 - 圖形視圖: 視覺化筆記間連結
Claude Code + Obsidian = 殺手組合: - Claude 能讀你整個 vault(所有筆記在上下文中) - Claude 能寫新筆記、更新現有筆記 - Claude 能追蹤連結、建立關聯 - 所有運算在本地發生
範例工作流: 1. 你在 Obsidian vault 中有 10,000 則筆記 2. 問 Claude Code:「總結我寫過關於 AI 安全的所有東西」 3. Claude 讀取所有相關筆記(本地檔案存取,無上傳) 4. Claude 產生綜合總結 5. Claude 建立新筆記並交叉參考 6. 全程不發送資料到 Anthropic
這用雲端 AI 不可能: - 上傳 10,000 則筆記到 ChatGPT?(慢、貴、隱私風險) - API 限制?(上下文窗口太小) - 但本地 Claude Code?輕鬆。
意義與預測¶
對個人¶
近期(2026-2027): - 開放權重模型達到 Claude Code 品質 - 本地 AI 成為知識工作者標準 - Obsidian 類工具擴散(Logseq、Foam、Dendron)
中期(2028-2030): - ENS/SNS 名稱所有權成為標準(像今天的 email) - 端到端加密協作成為預設 - 「雲端儲存」聽起來像「大型主機」一樣古老
長期(2030+): - 個人 AI 在你所有人生資料上訓練(全部本地) - 零知識證明實現選擇性披露 - 「我發送到 alice.eth」= 正常用語
對公司¶
現有 SaaS 的威脅: - Notion、Roam Research、Evernote → 如果本地 AI 更好,為什麼付訂閱費? - Google Docs → 如果加密協作有效,為什麼與 Google 分享資料? - Slack、Discord → ENS/SNS 加密聊天 = 同樣 UX,更好隱私
機會: - 硬體(本地 AI 加速) - 工具(Obsidian、GitLab 等) - 加密基礎設施(ENS/SNS 錢包、密鑰管理)
對 Web3¶
加密論述的驗證: - 加密不只是「魔法網路貨幣」 - 真實用途:加密、身份、安全協作 - ENS/SNS 成為「web3 的 email」
基礎設施需求: - 錢包更好的 UX(大多數人無法管理密鑰) - 同步協議(去中心化 Dropbox) - 發現機制(如何找到 friend.eth 的裝置)
批評與開放問題¶
批評 1:「開放權重模型還不行」¶
反駁: - LLaMA 3 70B 已經與 GPT-3.5 競爭 - DeepSeek V3 接近 GPT-4 品質 - 軌跡清楚(指數改進) - Balaji 說「幾個月」— 樂觀但非不合理
批評 2:「大多數人不會運行本地 AI」¶
反駁: - Apple Silicon Macs 已經能本地運行 7B-13B 模型 - 下一代硬體(M4、M5)將處理 70B+ 模型 - Anthropic 可能推出「Claude Local」(像 Photoshop vs Photoshop.com) - 一旦體驗過「離線運作並看到我所有檔案的 AI」,很難回頭
批評 3:「加密 UX 很糟」¶
承認: - 密鑰管理很難 - 大多數人遺失助記詞 - ENS 名稱需要錢(註冊、續費)
反駁: - 社交恢復錢包解決密鑰遺失 - ENS UX 改進中(Coinbase 整合等) - 一旦存在殺手應用(加密 AI 協作),UX 快速改進
批評 4:「GitLab 作為中繼 = 不去中心化」¶
Balaji 的承認: - 純 P2P 很難 - 混合方法(私有 GitLab)務實
細微差別: - GitLab 只儲存加密 blobs(無法讀取資料) - 仍然比 Google Docs 好(Google 能讀一切) - 長期:去中心化替代品(IPFS、Arweave、Filecoin)
開放問題:手機怎麼辦?¶
挑戰: 手機對本地 AI 運算有限。
可能解決方案: - 手機上較小模型(7B-13B) - 筆電/桌機做重運算,同步加密結果到手機 - 未來:行動晶片改進(Apple A-series、Snapdragon)
與其他趨勢的連結¶
1. Local-First 軟體運動¶
關鍵文本: Martin Kleppmann 的「Local-First Software」(Ink & Switch)
原則: - 無轉圈圈(資料在本地,即時) - 你的工作不被困在一個裝置 - 網路可選 - 與他人協作 - 資料永久保存(無新創公司關閉)
Balaji 願景 = Local-First + AI + 加密
2. 零知識證明¶
未來擴展: - 不只加密資料給 friend.eth - 證明關於資料的事實,不揭露資料 - 範例:「我有 10,000 則關於 AI 的筆記」(可證明,但筆記保持私密)
3. 主權身份¶
自主身份(SSI)運動: - 控制你自己的身份(不是 Facebook、Google、政府) - ENS/SNS 作為基石 - 本地 AI + 加密錢包 = 數位自主
可行動的建議¶
對個人(現在)¶
- 開始使用 Obsidian(或類似 local-first 工具)
- 從 Notion、Evernote、Roam 匯出筆記
- 熟悉 markdown
-
本地建立你的「第二大腦」
-
取得加密錢包 + ENS/SNS 名稱
- MetaMask + ENS 名稱(yourname.eth)
-
實驗加密訊息(Status、XMTP)
-
用本地檔案試試 Claude Code
- 讓 Claude 指向你的 Obsidian vault
- 體驗「AI 一次看到你所有資料」
- 理解為什麼這與 ChatGPT 不同
對開發者(現在)¶
- 建立 local-first 工具
- Obsidian 插件
- 本地 AI 包裝器(llama.cpp、Ollama)
-
加密協作層
-
實驗 ENS/SNS 身份
- XMTP(透過 ENS 訊息)
- Ceramic Network(去中心化資料)
-
IPFS + ENS(分散式儲存)
-
為開放權重模型平等做準備
- 在領域資料上微調本地模型
- 建立假設本地 AI = 預設的工具
對公司(策略)¶
- SaaS 公司: 準備「本地模式」
- 如果使用者本地運行 AI,如何獲利?
-
硬體銷售?支援合約?Freemium 本地 + 付費雲端?
-
加密公司: 專注 UX
- 密鑰管理是瓶頸
-
解決助記詞問題 = 解鎖採用
-
硬體公司: 為本地 AI 優化
- Apple、NVIDIA、AMD → AI 加速晶片
- 記憶體頻寬重要(transformer 模型受記憶體限制)
關鍵金句¶
「個人資料變得更可程式化(透過本地 AI),卻也更私密(因為在本地運算)。」
— Balaji 的核心論述「現在重新去中心化並保持所有資料在本地變得更好。」
— 逆轉中心化趨勢「因為本地資料更易用加密加密、更易用本地 AI 運算。因此:個人變得私密且可程式化。」
— 結論「開放權重 AI 模型如果能同時看到所有這些不同資料類型,就能做很多事。」
— 關於本地檔案存取的價值「純 P2P 很難。」
— 誠實承認技術挑戰
相關資源¶
- 原推文: https://x.com/balajis/status/2016437837013660071
- 引用推文: Kepano(Obsidian CEO)關於 Claude Code + Obsidian 工作流的提問
- 相關概念:
- Martin Kleppmann "Local-First Software"
- ENS (Ethereum Name System) - ens.domains
- Ceramic Network (decentralized data)
- XMTP (messaging via ENS)
- 實踐工具:
- Obsidian.md
- Claude Code
- MetaMask + ENS
- llama.cpp / Ollama (local model hosting)
狀態: ✅ 收藏完成
應用價值: 高 — 指出 local-first + AI + crypto 的戰略匯聚點
行動建議:
1. 開始用 Obsidian(如果還沒有)
2. 取得 ENS 名稱(實驗加密身份)
3. 嘗試 Claude Code + 本地檔案(體驗差異)
4. 關注開放權重模型進展(LLaMA、Mistral、DeepSeek)