AI Native Article — Podcast

賢さが安くなるほど、価値はモデルそのものから離れていきます。Netflixが専門家よりシステム思考者を採る理由、財務エージェントの成否を分けるもの、そしてAI市場の主役交代——一見ばらばらの三つの事実を、「余った資源は値を失い、希少な"配線の判断"に価格が集まる」という一つの原理でつなぎ直します。自社のAI投資がモデル選びに寄りすぎていないか、点検しながら聴ける一本です。

Netflix CPTO on AI and the future of product & engineering — Lenny's Newsletter — https://www.lennysnewsletter.com/p/netflix-cpto-on-ai-and-the-future

Agentic AI in finance and accounting: the prerequisites — Financial Times — https://www.ft.com/content/1930c048-dbca-4c14-a6f5-c3a3b9468764

The real AI race may no longer be at the frontier: open models on Hugging Face — TechCrunch — https://techcrunch.com/2026/07/14/the-real-ai-race-may-no-longer-be-at-the-frontier-open-models-hugging-face/

Source article (L3 analysis): 知能が余るほど、配線の判断が希少になる — kohuehara.xyz — https://kohuehara.xyz/ai-native-article/article/3a5d0f0be61e

「どのモデルが最も賢いか」を競う時代は、静かに終わりつつあります。大規模なコード移行の実例、出典のない市場規模の物語、オープンモデルが上位モデルと品質で統計的に同点となった評価——三つを並べると、価値の源泉がモデル単体から、それを回すプロセスと自前の測定へ移ったことが見えてきます。ベンダーが差し出す見出しの数字ではなく、自社の実タスクで測ることの意味を話します。

@claudedevs (X) — Anthropic code-migration economics: Bun's Zig→Rust migration, ~1M lines in under two weeks at ~$165k API-equivalent cost; a 165,000-line Python→TypeScript migration over a weekend; the six-step loop method and the "worst case is deleting a branch" decision-cost framing — https://x.com/claudedevs/status/2079654423828304282?s=46&t=JLj57N67SbZn6GQzDkxKag

The Kobeissi Letter (X), 8 July 2026 — agentic AI reframed from chatbot to enterprise software layer / "digital employee", market sized above $155bn by 2030; the post itself is disclosed as a paid partnership — https://x.com/kobeissiletter/status/2074999422249787885?s=46&t=JLj57N67SbZn6GQzDkxKag

Databricks — "Benchmarking coding agents on Databricks' multi-million-line codebase": GLM 5.2 statistically tied with Opus 4.8 on quality at $1.28 vs $1.94 per task; harness choice alone moved per-task cost by more than 2x — https://www.databricks.com/blog/benchmarking-coding-agents-databricks-multi-million-line-codebase

Source article (L3 analysis, 2026-07-22): 「価値の源泉はモデルの外側へ移った」 — https://kohuehara.xyz/ai-native-article/posts/3a5d0f0be61e

実行が並列化されて安くなるほど、価値は個々の実行ではなく、それらを束ねる一段上の層へ移ります。AIが人を束ねた熟議の実験、人がAIを束ねた運用、そして委任するほど楽観的になってしまう自動化のパラドックス。束ねる主体がAIか人かは価値の所在を決めません。いま自分が実行層にいるのか、束ねる層にいるのか——そこが分かれ目だという回です。

VentureBeat — "How America's 250th birthday became a test of AI-powered collective intelligence" (Unanimous AI "Thinkscape"; 277 participants in parallel 4–5 person groups; 94 initial ideas converged to 3; Louis Rosenberg's "100% human intelligence" framing) — https://venturebeat.com/technology/how-americas-250th-birthday-became-a-test-of-ai-powered-collective-intelligence

Anthropic Economic Index, June 2026 report ("Cadences") — Claude usage logs joined to ~9,700 self-reported survey responses; the automation paradox; self-reported capability exceeding observed exposure — https://www.anthropic.com/research/economic-index-june-2026-report

Roan (X) — quant alpha-research swarm design: six specialised agents across the research pipeline, 24-hour parallel operation, maker-checker separation; cites Boris Cherny and promotes the "Slate" tool — https://x.com/rohonchain/status/2074134246784921977?s=46&t=JLj57N67SbZn6GQzDkxKag

Source article (L3 analysis, 2026-07-21): 「価値は「実行」から「束ねる層」へ移る」 — https://kohuehara.xyz/ai-native-article/posts/3a4d0f0be61e

エージェントの信頼性は、最終的な答えが合っていたかではなく、実行の過程を追跡できるかで決まりつつあります。実行トレースを多次元で捉え直す研究と、「人間がソフトウェアより安くなった」時代の浪費をどう診断するかという議論を突き合わせ、価値の律速要因がどこへ移ったのかを整理します。過程の検証可能性が新しい参入障壁になる——反証できる形で提示する回です。

"From Agent Traces to Trust" (arXiv:2606.04990v4) — survey proposing six trace dimensions and typed provenance relations (Support / Derive / Depend-on / Contradict / Invalidate / Trigger / Update); names the process-level accountability gap; cites InjecAgent, AgentDojo, AgentPoison — https://arxiv.org/html/2606.04990v4

George Sivulka (a16z), "You Just Hired a Million Bad Employees", 14 July 2026 — "humans are cheaper than software for the first time", token-spending loops, the seven org analogies, Eval-suite-as-OKR — https://a16z.news/p/the-next-ai-goldrush-tokens-loops

a16z — "The Next AI Goldrush: Tokens & Loops" (same piece, alternate URL) — https://www.a16z.news/p/the-next-ai-goldrush-tokens-loops

Source article (L3 analysis, 2026-07-20): 「過程の検証可能性が価値を律速する」 — https://kohuehara.xyz/ai-native-article/posts/3a3d0f0be61e

生成のコストがゼロに近づくほど、希少になるのは「作る力」ではなく「その出力を保証する力」です。上場ソフトウェア株の選別的な売り、自律エージェントの運用、広告に付き始めた「どう作られたか」の表示——別々の話に見える三つを、同じ移動を別の高さから撮った写真として並べます。検証が価格を決める世界で、自分の仕事がどちら側にあるのかを問う回です。

a16z — "Charts of the Week: Software's Selective Selloff" (AI-defensibility screening of listed software equities; Jevons demand dynamics, AI-related hiring, datacenter/power-price inverse correlation) — https://www.a16z.news/p/charts-of-the-week-softwares-selective

CyrilXBT (X) — 60 overnight autonomous-agent workflows framed as trigger / agent / verification — https://x.com/cyrilxbt/status/2076351205043744877?s=46

Google Ads — "How this ad was made" AI transparency panel, announced 9 July 2026 (SynthID watermarking; 2023 election-ad disclosure requirements) — https://blog.google/products/ads-commerce/google-ads-ai-transparency-labels/

Source article (L3 analysis, 2026-07-19): 「生成がタダになる世界で、検証が価格を決める」 — https://kohuehara.xyz/ai-native-article/posts/3a2d0f0be61e

広告の開示ラベル、国防費を超えるAI設備投資、社債発行の急増。無関係に見える三つの動きを、ひとつの転換として読み解きます。これまで技術は既存の制度に合わせて導入されてきましたが、いまは制度の側がAIに合わせて作り替わり始めた——適応の矢印が逆転しています。次にどの制度が書き換わるのかを、この原理から考えます。

Google Ads — "How this ad was made" AI transparency labels (Google blog) — https://blog.google/products/ads-commerce/google-ads-ai-transparency-labels/

The Kobeissi Letter — big-tech investment-grade bond issuance, $182bn in 2026 vs ~$13bn prior-year period — https://x.com/kobeissiletter/status/2073777167691845818?s=46&t=JLj57N67SbZn6GQzDkxKag

The Kobeissi Letter — AI capex reaching ~3.2% of US GDP in 2027, surpassing defence spending — https://x.com/kobeissiletter/status/2075286130941931676?s=12

Source article (L3 analysis, 2026-07-19): 「AIが制度を作り替える側に回った」 — https://kohuehara.xyz/ai-native-article/posts/3a2d0f0be61e

CyrilXBT on X — 60 Fable 5 overnight agent workflows, each built from trigger + agent + verification step (2026) — https://x.com/cyrilxbt/status/2076351205043744877

Databricks — Benchmarking coding agents on Databricks' multi-million-line codebase (2026-07-08): GLM 5.2 statistically tied with Opus 4.8 on quality at $1.28 vs $1.94 per task; harness alone moves per-task cost more than 2x — https://www.databricks.com/blog/benchmarking-coding-agents-databricks-multi-million-line-codebase

TechCrunch — Apple sues OpenAI over alleged trade secret theft (filed 2026-07-10, N.D. Cal.; metal-finishing process knowledge and unreleased product information via executive hiring) — https://techcrunch.com/2026/07/10/apple-sues-openai-over-alleged-trade-secret-theft/

Jerry Liu (LlamaIndex co-founder) on X — "The Best Model Routing is Task Specific" (2026-07-15), incl. OpenRouter "Fusion" and Cognition "Devin Fusion" — https://x.com/jerryjliu0/status/2077537847951945742

OpenAI — GPT-5.6 general availability: the Sol / Terra / Luna tiers, per-token and per-dollar outcome framing, and the parallel-agent "ultra" setting — https://openai.com/index/gpt-5-6/

The Kobeissi Letter on X — agentic AI as "digital employees", $155B market by 2030 (2026-07-08; disclosed paid partnership, no stated sourcing or methodology) — https://x.com/kobeissiletter/status/2074999422249787885

「いちばんいいものを一つ買えば解決する」が、そろそろ効かなくなってきました。「evalは新しいOKR」というシヴルカの主張、Anthropicのチップ多様化、そしてマスク発言への「測定不能ゆえ無意味」という反論。並べると、価値が能力そのものから「測る軸」と「束ねる構成」へ移っていることが見えてきます。能力を買う判断から、測り方と組み合わせ方の判断へ。

George Sivulka (Hebbia) on X — "You just hired a million bad employees." essay, "eval is the new OKR" (2026-07-14) — https://x.com/gsivulka/status/2077070925154161101

TechCrunch — Anthropic is discussing a new custom chip with Samsung (reporting The Information, 2026-07-02) (2026-07-02) — https://techcrunch.com/2026/07/02/anthropic-is-discussing-a-new-custom-chip-with-samsung/

MIT Sloan Management Review Middle East — Elon Musk predicts AI will eclipse human intelligence in 5 years (Musk post of 2026-06-22 and the measurability rebuttal) — https://www.mitsloanme.com/article/elon-musk-predicts-ai-will-eclipse-human-intelligence-in-5-years/

労働、推論コスト、そして営業秘密の訴訟。バラバラに見える三つのニュースが、ひとつの原理で説明できます。豊富になった資源からは価値が逃げ、隣接する希少な補完財へ吸い寄せられる——キャリアと事業のどちらを考えるときにも、そのまま使える見取り図です。

a16z — George Sivulka, "The Next AI Goldrush: Tokens and Loops" / "You just hired a million bad employees." (2026-07-14) — https://a16z.news/p/the-next-ai-goldrush-tokens-loops

TechCrunch — Why the rise of open source AI isn't hurting Anthropic yet (Decagon CEO Jesse Zhang; Vercel and OpenRouter data) (2026-07-07) — https://techcrunch.com/2026/07/07/why-the-rise-of-open-source-ai-isnt-hurting-anthropic-yet

TechCrunch — Apple sues OpenAI over alleged trade secret theft (filed 2026-07-10, N.D. Cal.) (2026-07-10) — https://techcrunch.com/2026/07/10/apple-sues-openai-over-alleged-trade-secret-theft/