Convergence
The Price of Power Is Blindness
Brains, markets, proteins, laws, and AI all discovered the same terrible bargain: peak performance requires abandoning the ability to explain yourself
5key numbers
AI SYSTEMS & MECHANISTIC.
175B
GPT-3 parameters enabling emergent abilities absent in smaller models.
Google DeepMind deprioritized sparse autoencoders in March 2025; SAEs underperformed baselines on harmful intent detection (AI Frontiers, 2025).
GPT-3 (175B params) showed emergent zero-shot learning; smaller models failed similar tasks (Quanta Magazine, 2023).
Emergent abilities appear suddenly at pre-training loss thresholds in tasks like MMLU, GSM8K (arXiv survey, 2025).
MOLECULAR BIOLOGY & PROTEIN.
90%
AlphaFold2's prediction accuracy for protein structures in CASP 2020.
Hydrophobic forces limit foldable protein domains to ~200 amino acids, matching 1,200-protein experimental distribution (2012).
AlphaFold highest-confidence predictions have ~2x errors of experimental structures; 10% have substantial errors (2024).
AlphaFold2 predicts structures of all 200 million known proteins, used by 2M people in 190 countries since 2020.
NEUROSCIENCE & BRAIN FUNCTION
100 trillion
Neural connections in human brain resist individual interpretation despite enabling coherent function
Human brain contains 86 billion neurons forming 100 trillion connections; connectomics field aims to map every neuron-to-neuron link, yet comprehensive understanding remains fundamentally limited by.
Brain exhibits high-complexity signals ~80% of acquisition time, interrupted by spontaneous complexity drops; this dynamic opacity means brain state cannot be predicted from static architectural.
Cortical hubs are dispersed throughout brain despite functional integration role; 30% of cortical pairs connected directly, 60% separated by only two processing steps, creating opacity between.
FINANCIAL MARKETS &.
75%
Market participants concerned about algorithmic trading opacity.
Liquidity peaks at 10% algorithmic trading share, then declines (WJAETS 2024).
1/3 of liquidity provider algorithms are market makers, improving spread 86% of eager orders (JFEC 2025).
68% ±7.4% confidence interval for US stock companies using algorithmic trading (RARS 2024).
LEGAL SYSTEMS & CASE LAW
38%
Recall rate for AI legal passage retrieval on top 10K precedents
LePaRD dataset contains millions of US federal judge citations to precedent, yet best classification-based retrieval models achieve only 38% recall on 10,000 most-cited passages, demonstrating.
Case order effects in legal decision-making show asymmetric patterns where responses to cases become less similar to preceding cases rather than more similar, contradicting consistency theories and.
US Supreme Court case law is more than 3x as disruptive as lower court rulings, with case law volume growing at 1.74x rate faster than population growth, indicating accelerating complexity in.
AI SYSTEMS & MECHANISTIC.
175B
GPT-3 parameters enabling emergent abilities absent in smaller models.
Google DeepMind deprioritized sparse autoencoders in March 2025; SAEs underperformed baselines on harmful intent detection (AI Frontiers, 2025).
GPT-3 (175B params) showed emergent zero-shot learning; smaller models failed similar tasks (Quanta Magazine, 2023).
Emergent abilities appear suddenly at pre-training loss thresholds in tasks like MMLU, GSM8K (arXiv survey, 2025).
MOLECULAR BIOLOGY & PROTEIN.
90%
AlphaFold2's prediction accuracy for protein structures in CASP 2020.
Hydrophobic forces limit foldable protein domains to ~200 amino acids, matching 1,200-protein experimental distribution (2012).
AlphaFold highest-confidence predictions have ~2x errors of experimental structures; 10% have substantial errors (2024).
AlphaFold2 predicts structures of all 200 million known proteins, used by 2M people in 190 countries since 2020.
NEUROSCIENCE & BRAIN FUNCTION
100 trillion
Neural connections in human brain resist individual interpretation despite enabling coherent function
Human brain contains 86 billion neurons forming 100 trillion connections; connectomics field aims to map every neuron-to-neuron link, yet comprehensive understanding remains fundamentally limited by.
Brain exhibits high-complexity signals ~80% of acquisition time, interrupted by spontaneous complexity drops; this dynamic opacity means brain state cannot be predicted from static architectural.
Cortical hubs are dispersed throughout brain despite functional integration role; 30% of cortical pairs connected directly, 60% separated by only two processing steps, creating opacity between.
FINANCIAL MARKETS &.
75%
Market participants concerned about algorithmic trading opacity.
Liquidity peaks at 10% algorithmic trading share, then declines (WJAETS 2024).
1/3 of liquidity provider algorithms are market makers, improving spread 86% of eager orders (JFEC 2025).
68% ±7.4% confidence interval for US stock companies using algorithmic trading (RARS 2024).
LEGAL SYSTEMS & CASE LAW
38%
Recall rate for AI legal passage retrieval on top 10K precedents
LePaRD dataset contains millions of US federal judge citations to precedent, yet best classification-based retrieval models achieve only 38% recall on 10,000 most-cited passages, demonstrating.
Case order effects in legal decision-making show asymmetric patterns where responses to cases become less similar to preceding cases rather than more similar, contradicting consistency theories and.
US Supreme Court case law is more than 3x as disruptive as lower court rulings, with case law volume growing at 1.74x rate faster than population growth, indicating accelerating complexity in.
Enjoyed this?
Get new stories in your inbox.