Polygenics and Machine SuperIntelligence; Billionaires, Philo-semitism, and Chosen Embryos – #102

Discussion of Chinese billionaire Xubo who reportedly has 100+ sons through IVF and surrogacy, preferring eggs from Jewish women. This case illustrates the growing use of advanced reproductive technologies among ultra-high-net-worth individuals and sets the stage for discussing 2025 breakthroughs in polygenic embryo selection.

• •Wall Street Journal profiled Chinese billionaire with ~100 sons via IVF/surrogacy
• •Reported preference for Jewish egg donors demonstrates 'phylo-semitism'
• •Example of elite adoption of advanced reproductive technologies
• •Possible use of polygenic selection for embryo screening

Analysis of a major JAMA study with ~30,000 women validating polygenic risk scores for breast cancer screening. The study demonstrates how DNA-informed risk allocation outperforms standard annual screening, with polygenic scores identifying 10x more high-risk women than BRCA mutations alone. This represents validation for adult healthcare applications.

• •Study of 30,000 women using polygenic scores to allocate breast cancer screening resources
• •Polygenic scores identify ~10x more high-risk women than BRCA mutations
• •Predictor based on ~100 SNPs shows meaningful risk stratification
• •Myriad's 'MyRisk' already provides polygenic screening to millions of women
• •Validates technology that critics claim 'doesn't work' for embryo selection

Breakthrough Nature paper (October 2025) analyzing 500,000+ genomes from Taiwan Biobank to build first high-quality polygenic predictors for East Asian populations. Includes few-centimeter accuracy for height prediction and opens door for embryo selection in Asian markets where cultural acceptance is higher.

• •First non-European polygenic predictors matching European quality
• •Analyzed 500,000+ East Asian genomes - largest non-European dataset
• •Achieves few-centimeter height prediction accuracy
• •Published in Nature October 2025 after preprint in 2024
• •Enables embryo selection for East Asian populations with high cultural acceptance

Heracyte claims to have built a polygenic predictor correlating 0.5 with actual IQ using UK Biobank data and synthetic fluid intelligence estimators. Simulations suggest selecting from 10 embryos could yield 5-10 IQ point gains or 2-3 inches in height - reaching significance levels where families would care.

• •New predictor achieves ~0.5 correlation with actual IQ (not yet independently replicated)
• •Built using UK Biobank with synthetic fluid intelligence estimators
• •Selection from 10 embryos could yield 5-10 IQ points or 2-3 inches height gain
• •Gains reaching threshold where IVF families would find it compelling
• •Better data could potentially reach 0.7 correlation

Analysis of changing public attitudes toward embryo selection, with Nucleus Genomics running aggressive 'Have Your Best Baby' subway ads and Heracyte publishing controversial white papers. Genomic Prediction has deliberately stayed conservative (disease-only, no IQ/height) while working with 300+ clinics and genotyping 200,000 embryos. Surveys show majority now approve disease screening, with large minority accepting intelligence selection.

• •Genomic Prediction: 300+ clinics, 200,000 embryos genotyped, conservative disease-only approach
• •Competitors like Nucleus and Heracyte aggressively marketing intelligence/height selection
• •Academic surveys show strong majority approve disease screening, large minority accept intelligence selection
• •Minority strongly opposed to intelligence selection is shrinking
• •East Asian markets show very high acceptance of intelligence selection
• •2025 may mark inflection point for public acceptance

Discussion of ultra-high-net-worth individuals aggressively using embryo selection, including Silicon Valley elites and Chinese billionaires coming to the US for IVF with donor eggs and surrogacy. Prediction that visible elite adoption will shift average people from 'woke scold' fear to FOMO (fear of missing out).

• •Silicon Valley elites and Chinese billionaires leading adoption
• •Wall Street Journal coverage of Chinese billionaires using US IVF services
• •Elite visibility will drive FOMO among affluent families
• •Shift from ideological resistance to 'what am I missing out on?'
• •Prediction: 2025 marks inflection point for mainstream acceptance

Three key predictions for 2026: (1) Continued validation making scientific denial untenable, (2) Fast growth in Asia-Pacific markets due to cultural acceptance and new predictors, (3) Elite adoption driving mainstream FOMO and acceptance.

• •Scientifically untenable to claim polygenic prediction 'doesn't work'
• •Predictor quality improvements will make gains too obvious to ignore
• •Asia-Pacific market will grow rapidly with no cultural resistance
• •Elite adoption will shift public attitudes from fear to FOMO
• •Mainstream acceptance likely within 1-5 years

Overview of publishing first physics paper where core idea came from AI (GPT-5), on state-dependent quantum mechanics modifications. Published in Physics Letters B with companion paper on AI contribution process. Includes discussion with critics and demonstration of generator-verifier pipeline capabilities.

• •First physics paper with core idea from AI published in Physics Letters B
• •Topic: relativistic covariance in nonlinear quantum mechanics
• •Companion paper documents AI contribution process
• •80-minute discussion with two professor critics available on YouTube
• •Demonstrates frontier AI capabilities in theoretical physics

Lin Yang's scaffolding architecture enables off-the-shelf models (GPT-5, Gemini, Claude) to achieve IMO gold medal level (5/6 problems correct) through generator-verifier pipeline. DeepSeek 3.2 Speciale now achieves this without scaffolding. Represents dramatic improvement from ~1/6 to 5/6 problem-solving capability.

• •Generator-verifier scaffolding enables 5/6 IMO problems solved (vs 1/6 baseline)
• •Works with GPT-5, Gemini 2.5, Grok, Claude when properly scaffolded
• •DeepSeek 3.2 Speciale achieves gold medal level without scaffolding
• •Performance comparable to top 1-in-100,000 humans
• •Also strong on Putnam exam (undergraduate level math)
• •GitHub repository with ~3 pages of prompting available

Analysis of why general public underestimates AI capabilities - they only see one-shot performance, not scaffolded/agentic workflows. Younger physicists (under 35) already using AI extensively in research (~70% adoption). Most commentators haven't experienced peak performance available through proper scaffolding.

• •~70% of physicists under 35 already using AI extensively in research
• •Public only sees one-shot performance, not scaffolded capabilities
• •Generator-verifier pipeline creates qualitative performance jump
• •Most AI commentators haven't seen peak capabilities
• •Dramatic gap between perceived and actual frontier performance
• •DeepMind's 'CoScientist' also showing strong research utility

Detailed explanation of how models improved dramatically at symbolic calculations (integrals, derivatives, matrix operations) through RL with synthetic data from Mathematica. Models went from unreliable to quite reliable in ~1 year, with natural language interface advantage over rigid Mathematica syntax.

• •Year-over-year improvement from unreliable to quite reliable symbolic math
• •RL training using synthetic data from Mathematica/symbolic engines
• •Models already had conceptual understanding, RL added execution reliability
• •Natural language interface vs rigid Mathematica syntax
• •Can chain together multi-step calculations that would take humans hours
• •Similar improvements happening in coding, debugging, library understanding

Prediction that 2026 will see continued rapid advancement through better pretraining, post-training RL, and agentic scaffolding. Expects dramatic quality jumps despite data scaling challenges, with breakthrough applications in research, coding, legal analysis, and financial work. Inference costs may be 10x higher but quality gains will be worth it.

• •No slowdown predicted - continued rapid capability improvement
• •Agentic workflows (multiple specialized model instances collaborating) will advance
• •Synthetic data and better evals will drive progress despite data constraints
• •Deep research mode may use 10x more tokens but deliver qualitative jumps
• •2026 could be tipping point for research utility across sciences
• •Breakthrough applications in coding, legal, financial, scientific domains
• •Prediction: 'A year from now, we're gonna be amazed at how good the models are'