📰 Daily Digest — 2026-03-19

4 items | AI, Business, Open Source

---

📋 Quick Summary

GPT 5.4 is a big step for Codex

Source: Interconnects (Substack) · Category: AI · Link: Original

• Nathan Lambert argues agent evaluation should move beyond a single correctness score toward a 4-factor view: correctness, ease of use, speed, and cost.
• In hands-on Codex usage (fast mode with high/xhigh effort), he describes GPT 5.4 as the first OpenAI agent that reliably handles messy end-to-end tasks with far fewer operational “hard edges.”
• He also reports stronger context behavior and generous rate limits versus his own usage limits, while noting a shared GPT/Claude weakness: occasional dropped TODO items in multi-task prompts.

실리콘밸리 VC가 쓴 VC에게 투자받는 법

Source: 이안의 주간실리콘밸리 (Substack) · Category: Business · Link: Original

• Ian Park frames fundraising as a game-theoretic equation of incentives and introduces a three-layer VC model plus a “genius vs. fool” psychology matrix.
• The piece breaks down classic “2 and 20” math with concrete examples (e.g., a 100억 fund, ~20억 fees over 10 years, ~80억 investable capital, and why many LPs seek ~3x outcomes).
• Founder guidance focuses on selecting the right house/champion path by fund size, vintage year, check-size fit, and who can actually win internal IC debates.

Agent-to-Agent Communication Is Broken: Why an Email-like Inbox Model Works

Source: Medium · Category: AI · Link: Original

• ⚠️ Fetch failed (403/security verification page).
• The retrievable page only exposed Medium’s anti-bot interstitial (“Performing security verification”), not the article body.
• Kept the original link for retry when direct content access is available.

Anthropic ‘81,000 people want from AI’ — 669 classified quotes with occupation, country, region, topic, category, and sentiment

Source: GitHub Gist · Category: Open Source · Link: Original

• The gist publishes a CSV of 669 classified quotes derived from Anthropic’s “81,000 people want from AI” material.
• The file spans 84 countries and 13 world regions; top country counts include United States (148), South Korea (46), and Japan (43).
• Label distribution is light 375, shade 263, mixed 31, with frequent topics such as Productivity (54), Learning & education (48), and Emotional support (40).

---

📝 Detailed Notes

1. GPT 5.4 is a big step for Codex

1. The article argues current agent benchmarks underrepresent real-world usefulness.
   • Traditional leaderboards compress performance into one correctness score for interpretability.
   • Lambert says agent work quality depends on four axes: correctness, usability, speed, and cost.
   • He expects separate benchmark dimensions to mature over time rather than one scalar metric.
2. Practical workflow reliability is presented as GPT 5.4’s main step-change.
   • The author describes prior OpenAI-agent usage as “death by a thousand cuts” in everyday operations.
   • He specifically calls out fewer failures around messy operational tasks such as packages, file ops, and git-adjacent flows.
   • In his Codex usage, GPT 5.4 fast mode plus higher reasoning effort feels robust across varied tasks.
3. The write-up distinguishes model capability from harness experience and product feel.
   • Claude is characterized as warmer and more opinionated, which helps newcomers stick with the tool.
   • GPT 5.4 is described as meticulous and literal, better for tightly specified execution-heavy task lists.
   • The comparison frames two philosophies: intent-modeling assistant versus exact-instruction coordinator.
4. Cost/rate-limit and context behavior are central to the author’s preference split.
   • He references paying for both Claude ($100/month) and ChatGPT ($200/month) plans in parallel.
   • He reports rarely reaching Codex limits in fast mode, while hitting Claude limits at times.
   • The post links this to reasoning efficiency and cites third-party CursorBench context for token-performance trade-offs.
5. Remaining friction is small but non-trivial in multi-step prompting.
   • Both GPT 5.4 and Claude Opus 4.6 are reported to occasionally drop tasks from multi-TODO prompts.
   • Queueing extra follow-up messages during execution is described as risky outside simple cases.
   • Net assessment stays strongly positive: better “agentness” and usability, while model choice remains use-case and taste dependent.

---

2. 실리콘밸리 VC가 쓴 VC에게 투자받는 법

1. The core framing treats fundraising as an incentives equation, not a pitch-deck beauty contest.
   • Ian Park applies a utility-function lens from economics/game theory to founder–VC interactions.
   • He models VC behavior through three layers (fund economics, org-person incentives, IC politics).
   • He also introduces a psychological matrix around consensus safety versus contrarian “genius/fool” outcomes.
2. Fund-level economics explain why VC behavior can look extreme from the outside.
   • The article uses the standard “2 and 20” structure: management fee and carried interest.
   • In the 100억 example, about 20억 is consumed as 10-year fees, leaving roughly 80억 for investment.
   • He highlights why LP expectations and dilution math push many funds toward power-law, home-run hunting.
3. Carry design materially changes internal behavior and decision incentives.
   • Whole-fund carry aligns teammates around portfolio-level outcomes and shared upside.
   • Deal-by-deal carry can intensify internal competition because personal upside ties to individual deal wins.
   • The same valuation decision can be rational under one structure and irrational under the other.
4. IC process is described as a political multiplayer game centered on a champion.
   • Founders present early, but internal debate is often mediated by the partner or investor champion.
   • The champion’s preparation, credibility, and coalition-building can decide whether objections are neutralized.
   • Park emphasizes this is not purely data adjudication; rhetoric, trust, and repeated-game dynamics matter.
5. Founder strategy should optimize for fit and influence path before narrative polish.
   • He recommends screening target funds by size, vintage stage, and expected check/exit math compatibility.
   • He advises prioritizing contacts who can actually carry the deal through IC, not just exploratory meetings.
   • He also notes first-contact path matters: direct partner access or strong non-VC intro can improve champion alignment.

---

3. Anthropic ‘81,000 people want from AI’ — 669 classified quotes with occupation, country, region, topic, category, and sentiment

1. The gist packages a structured sample from a much larger response corpus.
   • The headline references 81,000 total responses while the shared CSV contains 669 labeled quotes.
   • The downloadable raw file is anthropic_81k_quotes.csv with one record per selected quote.
   • This turns narrative responses into a compact, queryable research artifact.
2. Schema design supports cross-sectional analysis across demographics and themes.
   • Columns include occupation, country, world region, full quote text, topic, category, and sentiment.
   • The sample covers 84 countries and 13 world regions, enabling geographic comparisons.
   • Country frequency is concentrated, with the United States at 148 entries, followed by South Korea (46) and Japan (43).
3. Sentiment labels show optimism-leading but mixed public perception.
   • The dataset contains 375 light, 263 shade, and 31 mixed entries.
   • That split suggests many respondents report benefits while a large minority emphasizes concerns.
   • The sentiment balance makes the file useful for studying coexistence of enthusiasm and anxiety narratives.
4. Topic/category frequencies reveal where everyday value and risk perceptions cluster.
   • Top topics include Productivity (54), Learning & education (48), and Emotional support (40).
   • High-volume categories include Personal transformation (113) and Professional excellence (104).
   • Reliability & trust (56) and Job displacement (37) remain visible concern channels in the labeled data.
5. Practical caveats matter when interpreting or reusing the sample.
   • Not every field is fully populated; occupation is blank in 154 rows.
   • It is a curated subset, so counts should be treated as directional within this file, not population estimates.
   • Still, the format is immediately usable for dashboards, qualitative coding, and prompt-grounded synthesis workflows.