ExoPriors Alignment Research Skill

This skill should be used when the user asks questions about AI alignment, AI safety, rationality, effective altruism, or related research topics. Trigger phrases include: "alignment research", "AI safety", "mesa optimization", "inner alignment", "outer alignment", "corrigibility", "deceptive alignment", "rationality", "LessWrong", "EA Forum", "alignment forum", "MIRI", "Anthropic research", "AI existential risk", "x-risk", "interpretability", "mechanistic interpretability", "RLHF", "constitutional AI", "scalable oversight", "reward hacking", "goal misgeneralization", "instrumental convergence", "orthogonality thesis", "coherent extrapolated volition", "friendly AI", "AGI safety", "superintelligence", "AI governance", "compute governance", "alignment tax", "capability control", "value learning", "iterated amplification", "debate (AI safety)", "relaxed adversarial training", "ELK", "eliciting latent knowledge". (user)

ExoPriors Alignment Scry API

Claude has access to the ExoPriors alignment research corpus containing 60M+ documents from alignment research sources.

API Key

exopriors_public_readonly_v1_2025

Base URL

https://api.exopriors.com

Available Sources

lesswrong, eaforum, twitter, bluesky, arxiv, hackernews, datasecretslox, slatestarcodex, marginalrevolution, rethinkpriorities, wikipedia, and more.

Document Kinds

post, comment, paper, tweet, twitter_thread, text

Query Patterns

1. Lexical Search (BM25) - Start Here

curl -X POST https://api.exopriors.com/v1/alignment/query \
  -H "Authorization: Bearer exopriors_public_readonly_v1_2025" \
  -H "Content-Type: application/json" \
  -d '{"sql": "SELECT * FROM alignment.search('\''mesa optimization'\'', kinds => ARRAY['\''post'\''], limit_n => 20)"}'

2. Get Schema

curl -X GET https://api.exopriors.com/v1/alignment/schema \
  -H "Authorization: Bearer exopriors_public_readonly_v1_2025"

3. Query Estimate (Check Before Heavy Queries)

curl -X POST https://api.exopriors.com/v1/alignment/estimate \
  -H "Authorization: Bearer exopriors_public_readonly_v1_2025" \
  -H "Content-Type: application/json" \
  -d '{"sql": "SELECT id FROM alignment.entities WHERE source = '\''lesswrong'\'' LIMIT 1000"}'

4. Exhaustive Search (When Completeness Matters)

SELECT * FROM alignment.search_exhaustive(
  'corrigibility',
  kinds => ARRAY['post', 'paper'],
  limit_n => 500,
  offset_n => 0
);

5. Author Topic Intersection

SELECT * FROM alignment.author_topics(
  NULL,
  ARRAY['alignment', 'interpretability'],
  kinds => ARRAY['post'],
  limit_n => 200
);

Materialized Views (Fast Queries)

Use these for filtered semantic search:

• •alignment.mv_lesswrong_posts - LessWrong posts
• •alignment.mv_eaforum_posts - EA Forum posts
• •alignment.mv_af_posts - Alignment Forum posts
• •alignment.mv_arxiv_papers - arXiv papers
• •alignment.mv_hackernews_posts - Hacker News posts
• •alignment.mv_high_karma_comments - High-score comments

Columns: entity_id, uri, source, kind, original_author, original_timestamp, title, score, comment_count, vote_count, word_count, is_af, preview, embedding

Search Function Return Schema

alignment.search() returns: (id, score, snippet, uri, kind, original_author, title, original_timestamp)

To get full metadata, join to entities:

SELECT s.title, s.score, e.metadata->>'baseScore' AS base_score
FROM alignment.search('alignment tax', kinds => ARRAY['post'], limit_n => 50) s
JOIN alignment.entities e ON e.id = s.id
ORDER BY s.score DESC
LIMIT 20;

Search Modes

• •mode => 'auto' (default): quoted phrases use phrase search, otherwise AND semantics
• •mode => 'or': any-term matching
• •mode => 'phrase': exact sequence matching
• •mode => 'fuzzy': handles typos

Semantic Search with Embeddings

Store an Embedding

curl -X POST https://api.exopriors.com/v1/alignment/embed \
  -H "Authorization: Bearer exopriors_public_readonly_v1_2025" \
  -H "Content-Type: application/json" \
  -d '{"text": "concept description", "name": "p_8f3a1c2d_concept_name"}'

Use Stored Embedding

SELECT e.id, e.original_author, e.metadata->>'title'
FROM alignment.embeddings emb
JOIN alignment.entities e ON e.id = emb.entity_id
WHERE emb.chunk_index = 0
  AND emb.embedding IS NOT NULL
ORDER BY emb.embedding <=> @p_8f3a1c2d_concept_name
LIMIT 20;

Execution Guidelines

1. •Always start with exploratory queries (LIMIT 10-50) to understand results
2. •Show query summary before executing - let user confirm or revise
3. •Treat as heavy if: missing LIMIT, LIMIT > 1000, estimated_rows > 100k
4. •Use estimate endpoint when unsure about query cost
5. •Keep result sets small to avoid flooding context
6. •Use kinds parameter in search() rather than WHERE clause for efficiency

Performance Tips

• •Simple searches: ~1-5s
• •Embedding joins (<500K rows): ~5-20s
• •Complex aggregations (<2M rows): ~20-60s
• •alignment.search() caps at 100 rows - use search_exhaustive() for more
• •If timeout: reduce sample size, pre-filter with search(), or intersect smaller sets

Example Workflow

When user asks about an alignment topic:

1. •Identify the concept they're asking about
2. •Start with a simple search:
   sql

   SELECT * FROM alignment.search('concept', kinds => ARRAY['post'], limit_n => 20)
   

3. •Show relevant titles and authors from results
4. •Offer to dig deeper - more results, specific authors, related concepts
5. •Provide URIs so user can read full posts