新模型出招,大神焦慮,AI 進入價格戰
W16

新模型出招,大神焦慮,AI 進入價格戰

Anthropic 4/19 推出 Opus 4.7,HN 同步出現「Claude Design」系統設計討論串;Karpathy 在 Robot Brains podcast 公開談 PhD 級焦慮,與 NYT「Jagged Intelligence」框架並置;Ben Horowitz 寫「開源 AI 將決定美國國運」;a16z 連發 PMF 案例 — Why We Need Continual Learning + Surviving AI Price Wars + Frontier Systems for Physical World;Tokenizer cost 分析在 HN 上 trending

256篇文章
162+來源
Y

AI 模型 & 產品更新

Anthropic:Claude Opus 4.7 正式發布(4/19)

本週主秀。Anthropic Introducing Claude Opus 4.7 4/19 上線。重點規格 + 對接點:

升級點跟 4.6 比
Vision2,576px(4.6 是 1,152px) — 設計稿、Figma、白板照都能直接讀真實度
長 reasoning內建 xhigh effort tier,介於 high / max 之間
Auto modeMax plan 用戶的 /auto 自動 permission decision,30 min+ 任務不打斷
Memory跨 session 自動 memory 強化

實測場景(我自己用):vision 升級對醫療影像專案 / 設計稿審查實質 useful — 之前 1,152px 的圖看不清字,現在能直接讀 Figma 截圖

觀察: Opus 4.7 是「長跑任務」模型 — 不是 Sonnet 那種快速 turn 用的。如果你在做研究 / 重構 / 計畫書寫作,4.7 會讓 turn 數降下來。但日常 coding 還是 Sonnet 4.6 划算

Claude Design — HN 同步爆紅

Anthropic 4/19 還推 Claude Design,HN top 兩篇:「Claude Design」+「Thoughts and feelings around Claude Design」

Claude Design = Anthropic 把「設計系統思考」內建到 Claude — 你給一張截圖或描述,Claude 直接產 Figma-friendly 的設計 spec + tokens + 可實作的 React component

觀察: Claude Design 對接 ui-designer / shadcn-component-discovery 那個 workflow,是更上游的設計決策幫手。如果你在用 design pipeline,這是 ➊ 階段(設計決策)的新工具

Measuring Claude 4.7's tokenizer costs HN top

社群開始用 OpenAI tiktoken / Anthropic tokenizer 對比 Claude 4.7 的成本曲線。重點:long context mode 的 token 消耗比 short context mode 高 40%(不是線性),因為 reasoning 階段會 backtrack


AI 開發工具 & Agent

"I'm spending months coding the old way" — HN 反潮流貼

I'm spending months coding the old way HN top

一個資深開發者寫他刻意不用 AI 寫 code 三個月,重新練手感。HN 留言兩極 — 一半「這是 buggy whip 思維」,一半「我也這麼做,AI 讓我變笨」

觀察: AI coding 過度依賴的反彈正在浮現。Cherny 自己最近也說「自 2025-11 我沒手寫過 code」 — 兩個極端各有人選。重點是「你知道自己在哪個位置」

Are the costs of AI agents also rising exponentially?

Are the costs of AI agents also rising exponentially? (2025) HN top

回顧 2025 年下半的 agent 經濟學:每個 agent 任務的 token 成本,從 $0.01 級往 $1-10 級走。多 agent orchestration(Claude Code spawn 5 個 subagent 平行跑)把成本推到 single chat 的 50x


大神觀點

Karpathy
Karpathy — Robot Brains podcast:公開談 PhD 級焦慮,跟讀 PhD 同感。Tesla AI / FSD 採訪
Karpathy
Karpathy — "Jagged Intelligence"(NYT)— AI 在某些領域超人某些領域智障,這個 jaggedness 才是 AI 的真實狀態
Boris Cherny
Boris Cherny — Claude Code 新版:plan mode + auto-accept edits + CLAUDE.md mistakes log,他個人從 2025-11 沒手寫 code
The Batch (Andrew Ng)
The Batch (Andrew Ng) — The Batch:Meta Pivots From Open Weights / Big Pharma Bets On AI / Regulatory Patchwork / Simulating Human Cohorts

創投 & 市場

Ben Horowitz:開源 AI 將決定美國國運

Why Open Source AI Will Determine America's Future — Ben Horowitz a16z

論點:中國正用 Qwen / DeepSeek / Kimi 把 AI 變成全球公共財;美國如果只走 closed-source 路線,會像 Cold War 美國只給盟友 mainframe 結果 Linux 在歐洲長出來那種戰略誤判

觀察: 這是 a16z 把「投開源 AI」框成「愛國」的論述準備。Horowitz 的政治影響力 + Trump 親 a16z 路線,下一波美國 AI 政策可能會看到 federal 補助開源模型訓練的招

Surviving AI Price Wars Without Destroying Your Business

Surviving AI Price Wars Without Destroying Your Business a16z

主題:當 token 成本一年降 90%,怎麼定價才不會被甩開。重點:別綁死 token 成本到客戶價格,要在「價值 metric」上定價(成功完成任務的次數、節省的時間)

Why We Need Continual Learning(早於 W17 的伏筆)

Why We Need Continual Learning a16z

關注「模型上線後怎麼學新知識」這個 unsolved problem。RLHF / fine-tuning / RAG 都是 workaround。a16z 在押 continual learning startups

Frontier Systems for the Physical World

a16z 投資 Hilbert — 物理世界 frontier systems。配合 Anduril / Stipple Bio 路線:a16z 在物理 AI(biology / robotics / defense / hardware)連發

Prediction Markets:They Grow Up So Fast

a16z 看 Polymarket / Kalshi 規模化路線。重點:prediction market 從「政治賭盤」變成「金融衍生品」的速度比想像快


商業 & 產品策略

Anthropic 估值 + 新一輪傳聞

W16 多個融資 / 估值傳聞:Anthropic 跟 Treasury / Fed 開銀行 CEO 會議談「Anthropic Mythos」(金融市場 narrative 影響)。下一輪估值市場猜 $400B+

First Round 進入 advice 階段

W16 First Round Review 的 article 偏向「給早期創辦人」:comp rules / FDE 招募 / company name 選擇 / discovery toolkit。比 W15 的 PMF 案例集更 actionable


行動建議

  1. 升 Claude Opus 4.7 試 vision 升級:如果你在做設計稿 / 圖表分析 / Med Vision / 補助案文件審稿,2,576px vision 對應 Figma 截圖 + 白板照
  2. /auto mode 評估:Max plan 用戶可以開長跑任務(>30min)測 auto mode 自動 permission decision。風險動作(push / deploy)一般 mode 還是要在
  3. Tokenizer cost 重算:如果你重度跑 long context(CLAUDE.md 大 + RAG),4.7 的 token 成本曲線跟 4.6 不一樣,先 measure
  4. Continual learning 觀察:a16z 押 continual learning startups。如果你的接案有「模型上線後維運」需求,這條路線會快速成熟

參考來源

RSS Digest: 256 articles from 11 sources(Anthropic Opus 4.7 + Claude Design 連發、Meta AI 持續、a16z 5 篇 essays、Karpathy 媒體 mentions、HN 16 篇)

主要訊號分布:

  • AI Companies: 9 篇(Anthropic Opus 4.7 主秀 + Meta MTIA/SAM/Muse Spark continued)
  • a16z Blog: 6 篇(Continual Learning、Open Source AI、Price Wars、Hilbert、Prediction Markets、Frontier Physical Systems)
  • Hacker News Top: 16 篇(Claude Design / Tokenizer costs / "I'm spending months coding the old way" / Opus 4.7)
  • Thought Leaders: 4 篇(Karpathy NYT + Robot Brains podcast + Tesla mention)

Generated from /research-scan W16 digest

Claude Opus 4.7AnthropicKarpathyJagged Intelligencea16zBen HorowitzOpen Source AIContinual Learning