AI summary by TwinMind
Are we all wrong about AI?
ColdFusion
24m 55s
Introduction & AI Public Perception
- Distinction between consumer generative AI (often trashy) vs. neural networks used quietly in background
- 2024 YouGov poll on American attitudes toward AI:
- Top emotions: caution, concern, skepticism, curiosity
- 14% believe AI already more intelligent than humans
- Younger people more positive about AI's societal/economic effects
- Common negative narratives: overpromising companies, job displacement (graphic designer lost job after 8 years, AI templates designs in 30 seconds vs. 30 minutes)
AI-Accelerated Chip Manufacturing
- AI improving chips that power AI systems (closed-loop cycle)
- Lithography process: circuit designs etched into semiconductor wafers using photo masks
- Nvidia's Q Litho platform:
- GPU-based algorithm accelerates computational lithography 4x
- Photo mask creation: weeks → overnight (40x faster)
- H100 chip: 89 reticles, single reticle processing: 2 weeks (CPU) → 8-hour shift (GPU)
- Working at 2 nanometer scale (human hair width to football field comparison)
- Industry adoption:
- TSMC (world's largest independent semiconductor foundry) using accelerated process
- ASML partnership for 2nm technology and beyond
- Performance impact visible in Nvidia GPU performance vs. Moore's Law chart
- Side note: DLSS AI upscaler provides higher quality visuals without excessive computational power
Kelp Forest Restoration in Tasmania
- Collaboration: Google, CSIRO, IMAS, The Nature Conservancy, Great Southern Reef Foundation, Kelp Alliance
- Problem scale: Australia's giant kelp forest decline
- 70% of Australians live within 50km of kelp forest
- Tasmania: 95% population decline, only 5% remains
- Ecosystem importance: Thousands of sea life types depend on kelp (food, habitat, food chain)
- AI solution using Google Earth engine + Vertex AI:
- Mapping 7,000+ square km of satellite imagery
- First complete geographical span mapping of Australian giant kelp
- Giant kelp canopy floats on surface → detectable by satellite
- Automated processing vs. old method (flying planes for photographs)
- Genetic analysis: AI pattern recognition on surviving 5% kelp to identify heat-resistant genetic patterns
- Breeding program: Heat-resistant kelp varieties for population replenishment
- Global impact: Geospatial maps accessible to all scientists, AI tools open source
AI-Powered Prosthetics
- Technology: Neural networks interpret electric nerve signals from patient muscles
- AI gives prosthetic limbs "a brain" for intuitive control
- Examples:
- Sarah Deard: Lost arm in subway accident at 45, AI bionic arm trains itself, improves with use
- Prosthetic legs that recognize stair climbing
- Hands that grab cups without spilling
- Companies:
- Ottobock: AI-powered bionic hands
- ReWalk: Powered walking assistant with autonomous AI decision-making exoskeleton
- Ossur (Iceland): Bionic prosthetics for lower limb amputees
- Atom Limbs: Advanced sensors + machine learning for realistic arm movement
- Market size: 550+ million amputees worldwide, 2 million in US, expected to double by 2050
- Limitations: High cost, recreating sensation still decades away, technology not ready for prime time
Healthcare Applications
- Google Med LM: AI tools for patient notes, disease marker identification
- Nvidia: AI tools for surgeries and medical imaging (partnerships with Johnson & Johnson, GE Healthcare)
- Drug discovery transformation:
- Traditional process: lengthy, billions in costs, high failure risk
- AI acceleration: reduced time and cost
- Companies:
- Recursion Pharmaceuticals: Using Nvidia's BioHive-2 AI supercomputer
- Benevolent AI: Testing drugs for diseases without good treatments
- Atomwise: AtomNet deep learning for binding affinity prediction
- Research validation: McKinsey study confirms AI positively impacts biofarma productivity and success rates
- Current developments: 270 AI-driven drug discovery companies founded
- Specific applications:
- Delaire Diagnostics: AI lung cancer indicators test
- Children's National Hospital: AI tool for rheumatic heart disease diagnosis in children
- Cancer detection: 85% detection rates with dried blood samples (quick, cheap, prolific across research)
- Da Vinci robotic surgery: Machine learning + image recognition for surgical planning and real-time adjustments
- Limitations: Training data quality affected by demographics, age, gender, environmental factors
Battery Research & Materials Discovery
- Problem: Lithium expensive, environmentally harmful, need alternatives
- Microsoft + Pacific Northwest National Laboratory collaboration:
- AI sifted through 23.6 million candidate materials
- Identified 23 promising materials for next-generation batteries
- N216 breakthrough material: 70% less lithium than current designs
- Half lithium atoms replaced with sodium
- Prototype development: idea → working prototype in 9 months (massively accelerated)
- Low-conducting prototype successfully powered light bulb
- Limitations: Limited training data in battery engineering, unforeseen complications possible
Conclusion & Career Perspective
- Consumer AI applications overdone and annoying
- Positive movement toward solving broader humanity problems
- Author covering neural networks since 2015 (pre-hype)
- Despite negative aspects, good to see positive AI applications emerging
