AI summary by TwinMind
"Not with my data" many people say to AI companies
Sabine Hossenfelder
7m 37s
AI Research Current State
- AI field experiencing "roller coaster" period with conflicting perspectives
- Research insiders warn of uncontrollable "new species" being bred
- Skeptics compare AI to failed 3D TV trend
- Major current issue: access to high-quality training data
Data Commons Shrinking Crisis
- Data Providence Initiative report highlights rapidly shrinking data availability
- Generative AI advancement previously fueled by vast free public data
- Website owners/organizations restricting content access
- Motivation: prevent free data use for others' profit
- AI companies struggling to find new data, especially:
- Current events content
- High-quality material
Legal Cases and Content Reproduction
- New York Times vs OpenAI:
- Allegations: Early GPT versions reproduced NYT paywall content word-for-word
- Issue doesn't occur with newer GPT versions
- Raises questions about grammar modification vs plagiarism
- Perplexity vs Forbes:
- Perplexity reproduced Forbes articles without attribution
- Content stealing predates AI (existed since internet dawn)
- New element: automation via bots
- 99% of journalism involves "stealing stuff without attribution"
Web Crawling and Blocking Mechanisms
- AI companies use web crawlers (software, not "robot spiders")
- Google has crawled internet since ~250 BC for indexing
- Blocking method: robots.txt document with crawler bot names
- Problem: need to know specific crawler names for identification
Restriction Implementation Timeline
- Sudden increase in restrictions starting ~1 year ago
- Triggered by data theft headlines and bot name distribution
- Website responses:
- Complete crawler blocking
- Partial content restrictions
- Predicted continued increase in blocks over next couple years
Competitive Imbalance
- Restrictions don't affect all companies equally
- More blocks against OpenAI crawlers than Meta's
- Creates AI competition imbalance
- Highest quality and newest data most likely restricted
- Makes AI models less useful for certain purposes
- Companies could rename crawlers but would create bad publicity
Proposed Solutions and Analysis
- Report suggests:
- More transparency about data usage
- Better licensing agreements
- Information about data handling
- Author's perspective: Issue more about output than input
- LLMs trained for language production, not information content
- Most common use: grammatically correct text/code generation
Scientific Data Challenge
- Different AI data crunch concern: lack of scientific data
- Difficulty making scientific data computer-readable
- Example: teaching computers to read contour/corner plots across 20 papers with 80 different conventions
- Prediction: AI companies will hire scientists for training
- Most valuable training data "locked away in human brains"
Sponsor Content
- NordVPN promotion focusing on AI-related internet security concerns
- Features: secure connection, threat protection, content access
- 5,000+ servers worldwide
- Special offer: nordvpn.com/sabina or code "sabina"
