rickoslyder/hint-detection

Hint Detection Skill

Lab researchers often hint at work before publication. This skill identifies these signals.

Hint Patterns

1. Vague Progress Claims

Language that implies results without specifics:

• "We've been seeing interesting results with..."
• "There's been a lot of progress on..."
• "Things are moving faster than people think in..."

2. Deflection with Signal

Answers that acknowledge something exists:

• "I can't say much, but..."
• "You'll see soon..."
• "No comment ;)"
• "That's a great question" (followed by non-answer)

3. Future Tense Confidence

Certainty about unreleased capabilities:

• "You'll see that..."
• "This will become clear when..."
• "The next generation will..."

4. Unusual Enthusiasm

Disproportionate excitement about a topic:

• Sudden interest in a specific area
• Detailed knowledge about approaches not in their published work
• Defending an approach more vigorously than expected

5. Specific Denials

Sometimes denial calls attention:

• "We're definitely NOT working on..."
• "That's not what we're focused on" (when they clearly are)
• Overly specific denials

6. Timeline Hints

Suggestions about release timing:

• "In the coming weeks/months..."
• "Stay tuned"
• "Sooner than you think"
• Mentions of specific events (conferences, dates)

7. Capability Hedging

Language implying current vs future:

• "Current models can't do X yet"
• "With today's approaches..."
• "The bottleneck right now is..."

8. Recruitment Signals

Hiring patterns can indicate direction:

• Sudden push for specific expertise
• "We're building a team for..."
• Job postings for unrevealed projects

Author Context

Weight hints by author credibility:

• Lab leadership (Dario, Sam, Demis): High signal, often deliberate
• Research leads: Technical hints about their area
• Individual researchers: May hint at their specific work
• Former employees: Sometimes reveal direction
• Adjacent figures (investors, partners): Second-hand signals

Analysis Framework

For each potential hint:

1. Quote the relevant passage

Extract the exact language that suggests a hint.

2. Implied capability

What capability or result is being hinted at?

3. Confidence level (0.0-1.0)

How confident are you this is a real hint vs. noise?

Consider:

• Author's position and knowledge
• Specificity of language
• Pattern match to known hint types
• Context of conversation

4. Estimated timeframe

When might this be revealed?

• imminent: Days to weeks
• near-term: 1-3 months
• medium-term: 3-12 months
• unclear: No timing signal

5. Domain

What area of AI?

• reasoning, agents, safety, multimodal, scaling, etc.

Output Format

Return JSON:

{
  "hints": [
    {
      "hintText": "The exact quote suggesting a hint",
      "author": "Author name",
      "affiliation": "Company/org",
      "impliedCapability": "What they're hinting at",
      "confidence": 0.7,
      "reasoning": "Why you think this is a hint",
      "timeframe": "near-term",
      "domain": "reasoning",
      "sourceUrl": "URL if available"
    }
  ],
  "noHintsFound": false
}

If no credible hints are detected, return:

{
  "hints": [],
  "noHintsFound": true,
  "notes": "Brief explanation of why content doesn't contain hints"
}

False Positive Avoidance

Not every comment is a hint. Exclude:

• General optimism without specifics
• Restatement of public roadmaps
• Academic speculation
• Marketing language in official announcements
• Obvious jokes or sarcasm
• Old information presented as new

High-Value Hint Indicators

Prioritize hints that:

• Come from people with direct knowledge
• Reference specific capabilities or benchmarks
• Include uncharacteristic certainty
• Align with known research directions
• Are followed by unusual silence on the topic