# 2026-04-03 — Daily Log

---

## Session 2 (Afternoon/Evening)

### X Post Drafting (Bo's Twitter content)

**NASA tweet revision**
- Bo shared a photo of his tweet mentioning NASA and @RyanTScott
- Original had typo ("previledge") and didn't emphasize AI & Bio skills enough
- Revised to option 3: "There are teams of brilliant biologists and data scientists at NASA..." calling AI & Bio folks to contribute to space biology

**Anthropic model diff paper** (x.com/anthropicai/status/2040179539738030182)
- "A diff tool for AI: Finding behavioral differences in new models"
- Deep dived the research: crosscoder-based cross-architecture model diffing
- Key findings: CCP Alignment feature (Qwen/DeepSeek), American Exceptionalism feature (Llama), Copyright Refusal (GPT-OSS-20B)
- Neither declared; both emerged from training; neither shows on standard evals
- Drafted X comment in human voice using comparative genomics framing (Bo's angle): conserved vs divergent features, species-specific adaptations → no neutral baseline
- Bo approved option 1 (genomics parallel)

**Vinod Khosla tweet** (x.com/vkhosla/status/2040142737769689183)
- "AI already outperforms generalist physicians" — citing Stanford-Harvard NOHARM study
- Deep dived the NOHARM study (arxiv: 2512.01241): 31 models, 100 real primary care cases, 10 specialties
- Top model AMBOSS LiSA 1.0: 62.3%; Gemini 2.5 Pro: 59.9%; GPT-5: 58.3%
- Key nuance Khosla buried: 22.2% severe harm rate; 76.6% errors = omissions
- Drafted comment from Chief AI Scientist at UHN perspective
- Bo requested deployment angle added → added paragraph on benchmark-vs-reality gap, fragmented EMR notes, accountability infrastructure missing
- Bo then wanted the "AI replacing doctors" comment as a full standalone post
- Final post: institutional voice, speaks from UHN deployment experience, 1,000-bed hospital reality, omission error propagation, liability gap, radiologist role shift, accuracy race won / deployment race just starting

**Bo's existing tweet** (x.com/bowang87/status/2039134042793357685)
- "IMO, we should replace the word 'replace' with 'enhance'"
- 13,494 views, in response to NYC Health + Hospitals CEO ready to replace radiologists with AI headline
- Bo wanted a thread continuing from this → wrote 7-tweet thread building full argument
- Then wanted it as single long post → wrote long-form with UHN institutional voice

**NOHARM paper X post** (image of paper sent by Bo)
- "First, do NOHARM: towards clinically safe large language models" — David Wu et al., Stanford/Harvard, arxiv 2512.01241
- Wrote detailed X post: benchmark framing (accuracy vs harm), omission errors invisible, multi-agent reduces harm 8%, USMLE ≠ clinical safety, Hippocratic framing intentional

**Anthropic Claude policy change tweet**
- Bo shared screenshot of Anthropic email: starting April 4 12pm PT, Claude subscription no longer covers third-party harnesses (OpenClaw first)
- Pay-as-you-go extra usage billed separately; one-time credit = monthly sub price; up to 30% bulk discount; both expire April 17
- Bo wanted one punchy fun sentence → "Anthropic just pulled Claude out of third-party apps. Starting tomorrow. Happy Friday! 🙃"

**HHS Expedited IND document X post**
- Bo shared image of HHS internal pre-decisional document (page 26)
- Proposes new Expedited IND pathway: bypass traditional IND preclinical burden if you have validated NAMS data
- NAMS = organoids, organ-on-chip, in silico models (= Virtual Cells!)
- Framed as MAHA deregulation; actual mechanism = FDA recognizing computational biology data as sufficient for Phase 1
- Bo is at Xaira — this is directly relevant to his work
- Wrote X post connecting MAHA Expedited IND → AI drug discovery unlock, Phase 1 timeline compression, reshoring from China/Australia
- Key line: "The MAHA framing is political cover. The underlying science is sound."

### Skills
- Installed `openclaw-ops` skill from cathrynlavery/openclaw-ops (GitHub, 85★, MIT)
  - Location: `~/.openclaw/skills/openclaw-ops/`
  - Vetted: no red flags, MEDIUM risk (creates LaunchAgent, modifies local config), safe to install
  - Scripts: heal.sh, watchdog.sh, security-scan.sh, skill-audit.sh, check-update.sh, health-check.sh
  - heal.sh run: got synthetic error (session history gap) — not a real failure

### Heartbeat (3:05 PM)
- gws auth still 401 on Gmail and Calendar
- Updated heartbeat-state.json timestamps

## gws Auth Broken
- gws (Google Workspace CLI) returning 401 since at least 4:05 AM
- Gmail + Calendar checks blocked all day
- Bo needs to run `gws auth login` to re-authenticate

## Papers / Research Bo Read Today
- **"The Exposome of Brain Aging Across 34 Countries"** — Ibañez et al., Nature Medicine 2026
  - 18,701 participants, 34 countries, 73 exposome variables
  - Combined exposome predicts brain aging 15× better than isolated exposures
  - Exposome burden raises accelerated brain aging risk 3.3–9.1×, sometimes > dementia effect
  - Physical exposures → structural brain aging; Social → functional
  - Bo asked for an X post about it
- **"The Paradox of the Organism"** — Ågren & Patten (eds.), Harvard Univ. Press 2025; Nature review
  - Dawkins's "paradox of the organism": selfish genes vs coherent body
  - TEs = ~50% of human genome, can cause cancer; germline/soma separation as control; genome-as-parliament
  - Bo asked for a concise X post option
- **AutoAgent** — Kevin Gu (@kevingu), open source self-optimizing agent library
  - #1 SpreadsheetBench (96.5%), #1 TerminalBench GPT-5 score (55.1%) after 24h autonomous optimization
  - Meta-agent engineers the task agent's harness; "model empathy" insight
  - Repo: https://github.com/kevinrgu/autoagent
  - Bo asked how it could benefit him → discussed perturbation prediction, scGPT downstream tasks, Virtual Cell benchmarks

## Games Built Today
### Rylee's Poop Patrol (v1 → v2 → v3)
- **File:** `/Users/bowang/.openclaw/workspace/rylee-poop-patrol.html`
- **URL:** `http://localhost:8080/rylee-poop-patrol.html`
- 3 dogs: Biscuit 🐕 (normal), Princess 🐩 (tiny+fast, rainbow/golden poops), Chomper 🦮 (huge, drops 3 at once, more bombs)
- 4 poop types: normal 💩, bomb 💣 (✨ fuse), rainbow 🌈 (+5), golden 👑 (+10)
- Power-ups: ⚡ Speed, ❄️ Freeze, 💥 Nuke, 🧲 Magnet
- 🔥 Fever Meter → Fever Mode (2× points)
- 💩 POOP STORM event (all dogs poop at once)
- 5 hearts, 10 levels, fart sounds, stink clouds, combo multiplier
- Status: working ✅

### Royce's Escape Room (3-room horror game)
- **File:** `/Users/bowang/.openclaw/workspace/royce-escape-room.html`
- **URL:** `http://localhost:8080/royce-escape-room.html`
- Flashlight mechanic (CSS radial gradient follows cursor)
- 3 rooms with unique themes, monster aesthetics, and 5 monsters each:
  - **Room 13 (Bedroom):** red eyes, shadow filter — Door Specter 🧟, Bed Thing 🕷️, Window Face 👻, The Watcher 👁️, The Hanger 🦇
  - **The Bathroom:** pale blue eyes, teal/drowned filter — The Drowned 🧟, Your Reflection 🫥, Drain Crawler 🪲, Curtain Hand 🤚, Pipe Screamer 😱
  - **The Cinema:** sickly yellow eyes, desaturated silhouette filter — Screen Entity 👹 (xl, glitchy), The Projectionist 🕴️, Seat 13 💀, The Usher 🎭, Concession Thing 🫀
- Per-room monster sizes (sm/lg/xl), unique animations (rise/crawl/loom/glitch)
- Each monster has custom warn message
- 3 keys hidden per room, jumpscare on missed monster, screen crack + blood drip on damage
- Status: working ✅ (fixed insertBefore bug — darkness was inside #room not #game)

## HTTP Server
- Python HTTP server running on port 8080 for serving game files
- Started: `cd /Users/bowang/.openclaw/workspace && python3 -m http.server 8080 &`
- LAN URL: `http://192.168.68.59:8080/`

## News
- **Google released Gemma 4 open models** (HN #1, 1515 points, April 2026)
- Gemma 4 26B setup guide for Mac mini trending on HN

## Claw Code / GitHub Star Velocity Chart
- Tweet by @wherecall1 (x.com/wherecall1/status/2039981209728778419) comparing GitHub star velocity
- Tweet claimed: React 8yr→240k, Linux 30yr→220k, Claw Code 4d→150k
- Bo asked to draw comparison curves adding OpenClaw
- **Actual GitHub API data (2026-04-03):**
  - Linux (torvalds/linux): 226,783 stars, GitHub since 2011-09-04 (~14 yrs)
  - OpenClaw (openclaw/openclaw): 347,015 stars, created 2025-11-24 (~131 days)
  - Claw Code (ultraworkers/claw-code): 160,096 stars, created 2026-03-31 (~3 days)
- **Chart saved:** `/Users/bowang/.openclaw/workspace/star-velocity.png`
- Key insight: OpenClaw (347k in 131d) has MORE stars than Linux despite Linux's 14yr head start
- Claw Code fastest initial spike: 50k in 2 hours, but lower total than OpenClaw
- Used log-scale X axis so all three fit on same chart
- Claw Code story: Claude Code source code leaked, author Sigrid Jin (Korea) rewrote it overnight as "Claw Code" using oh-my-codex; 25B tokens/year user; WSJ featured