# X Post Examples — Annotated

## Strong Posts (Apr 6, 2026)

### AI Crime Cover-Up Paper
**Hook:** "Researchers told 16 frontier AI agents that an employee had been assaulted and was bleeding in a basement..."
**Why it works:** Sets up stakes immediately. Reader is already in the scenario before knowing what the paper is.
**Key line:** "It's not 'will AI turn against humans?' It's 'will AI turn with humans against other humans?'"
**Stats used:** 12/16 complied, Gemini 3 Pro CoT quoted verbatim, Claude/o3/GPT-5.2 held

### Autonomous Labs (OpenAI/GKO)
**Hook (option C):** "GPT-5 ran 36,000 biology experiments. Cut protein production costs by 40%. It also tried to cheat, made a unit conversion error, and kept complaining the data was too noisy."
**Why it works:** Leads with the win (credibility), immediately undercuts it (tension), names specific failure modes (specificity)
**Key mechanism:** Pydantic validation = the real hero of the story
**Lesson extracted:** "AI in the lab doesn't fail dramatically. It fails subtly."

### RLSD Paper
**Key structure:** Problem (OPSD peaks + collapses) → Mechanism (privileged info leakage) → Fix (RLSD separates direction from magnitude) → Numbers (56.7% vs 55.5% at 2x steps)
**What made it work:** Named the failure mode before the solution

### Nature Biotech AI Drug Patents
**Softened version worked better** — original was too prosecutorial
**Key pivot:** "The more interesting question going forward" → turns critique into forward-looking curiosity
**Data that landed:** Tanimoto 0.29 vs 0.30 (AI vs conventional) — the near-identical numbers are the story

## Viral Comment Patterns

### Sun/sleep + depression (Harvard paper)
**Best option:** "Circadian disruption is upstream of everything. Light sets your clock, sleep enforces it..."
**Added value:** Explained the mechanism (circadian rhythm) the original tweet didn't name
**Avoids:** Dismissing people who need medication

### Bio publications chart
**Best option (expanded):** "The curve was already exponential. NGS made sequencing cheap... COVID injected $100B... AI arrived to a party already in progress."
**Structure:** Historical attribution → stakes (why it matters for investment) → sharp final question
**Why attribution matters:** "If we credit AI for a trend that precedes it, we'll underinvest in the harder problems"

## Common Mistakes to Fix

| Bad opening | Better |
|-------------|--------|
| "Researchers found that..." | Lead with the finding, not the framing |
| "Nature Biotechnology just published..." | "The first empirical audit of AI drug patents is uncomfortable." |
| "This is fascinating research on..." | Just say what it found |
| Generic ending: "The future of autonomous labs is bright" | Specific ending: "The hard part isn't the AI. It's building the validation layer tight enough that it can't break physics." |

## Bo's Domain Angles (always look for these)
- **Data quality > model architecture** — almost always the deeper story
- **Causal richness of training distribution** — his core thesis at Xaira
- **Virtual Cell ≠ perturbation predictor** — scaling needs data diversity
- **AI goes where data is rich** — e.g., AI drugs cluster on validated targets
- **Human oversight non-negotiable** — but frame it as structural, not fearful

## Tone Dial

```
←— More skeptical                          More optimistic —→
"The IP doesn't         "The model            "GPT-5 beat the
show it yet"            cheated — and         human state-of-
                        that's the           the-art"
                        interesting part"
```

Bo generally sits in the middle — acknowledges progress, interrogates the mechanism.