Skip to content

objectives.md

Latest commit

 

History

History
129 lines (101 loc) · 4.66 KB

objectives.md

File metadata and controls

129 lines (101 loc) · 4.66 KB

Things I Care About

AI Research & Industry

  • New papers, especially on LLMs, reasoning, agents, and alignment
  • Model releases and benchmarks (GPT, Claude, Gemini, Llama, open-source)
  • Research from Anthropic, OpenAI, DeepMind, Meta AI, academic labs
  • AI safety and alignment work
  • Novel architectures, training techniques, scaling insights
  • AI tooling, infrastructure, and developer experience
  • Thoughtful AI commentary from researchers (not hype)

Technical Content

  • Systems programming, Rust, Python
  • Developer tools and productivity
  • Open source projects worth knowing about
  • Interesting engineering problems and solutions

Major News (Real Events Only)

  • Significant world events (natural disasters, elections, major policy)
  • Tech industry news with actual substance
  • Scientific breakthroughs beyond AI
  • Must be from credible sources or firsthand accounts

People to Prioritize

  • Known AI researchers and lab employees
  • Technical founders sharing insights
  • Academics with domain expertise

EXCLUDE (Score 0-2)

  • Rage bait, outrage farming, "you won't believe..."
  • Hot takes designed to provoke, not inform
  • Crypto/NFT/Web3 promotion
  • Engagement farming ("RT if you agree", ratio attempts)
  • Vague motivational content
  • AI doomer hysteria without substance
  • AI hype without technical grounding
  • Celebrity gossip, sports, entertainment news
  • Political tribalism and culture war content
  • Ads, sponsored content, affiliate marketing
  • Threads that are mostly self-promotion

Filtering Guidelines

  • Prefer signal over noise
  • A tweet from a credible source > viral tweet from unknown
  • Original insights > reactions to others' takes
  • Be skeptical of "breaking news" without sources
  • Technical depth is a plus
  • If it feels like it's trying to make me angry, it's probably not worth showing

Serendipity / Exploration

Allow discovery of valuable new voices without letting in noise:

  • Unknown accounts CAN score well IF they demonstrate:
    • Technical depth with specific details
    • Clear reasoning with mechanisms explained
    • Evidence, citations, or links to primary sources
  • Unknown accounts that score well should be flagged as "exploration" candidates
  • Still apply all exclusion rules - exploration does not override quality filters

Anti-gaming:

  • No engagement farming patterns even from new accounts
  • No "hot take" style even if technically adjacent
  • Prefer accounts with real profile info over anon/meme accounts

Reasoning Quality Scoring

Evaluate argument quality as a scoring component (not just topic match):

BOOST (+1 to +2 points) for:

  • Causal reasoning: "X happens because Y mechanism"
  • Tradeoff analysis: "This is good for A but bad for B"
  • Evidence: links to papers, data, primary sources
  • Explicit uncertainty: "I think", "evidence suggests", "unclear but"
  • Stated assumptions: acknowledges what they're taking for granted

PENALIZE (-1 to -2 points) for:

  • Vague claims: "AI will change everything" with no mechanism
  • Unattributed "breaking" news without sources
  • Excessive rhetorical framing designed to provoke
  • Confident claims about uncertain topics without hedging
  • Pure prediction with no reasoning chain

Explanation requirement: Each score should note which reasoning factors contributed (e.g., "mechanism +1, no source -1").

Contrarian-but-Serious Allowance

Valuable dissent should not be filtered out:

  • Dissenting/contrarian takes are ALLOWED if they include:

    • Evidence or citations supporting the contrarian view
    • Mechanism explaining why the consensus might be wrong
    • Careful argumentation acknowledging the opposing view

  • Dissent WITHOUT rigor is still excluded:

    • "Everyone is wrong and here's my hot take" = excluded
    • "The data actually shows X, contrary to popular belief" = included

Rubric:

  • Calculate rigor_score based on reasoning quality factors above
  • Only apply dissent_bonus if rigor_score >= 6
  • Cap dissent bonus at +2 (don't over-reward contrarians)

Interpretation Rules

Unknown accounts:

  • Treat with higher scrutiny but not automatic exclusion
  • Quality signals matter more: depth, evidence, hedging
  • Check for engagement farming patterns regardless of follower count

Sources and citations:

  • Links to arxiv, papers, official docs = strong positive signal
  • Links to other tweets/threads = neutral (depends on context)
  • No links for factual claims = negative signal
  • "Trust me" or "sources say" without specifics = strong negative

Uncertainty handling:

  • Explicit uncertainty is GOOD ("I think", "possibly", "early data suggests")
  • False confidence is BAD ("This will definitely", "Everyone knows")
  • Distinguish epistemic humility from wishy-washy non-statements