Peter H. Diamandis
July 29, 2026
AI Geopolitics: Open vs. Closed Models, Anthropic-OpenAI Lobbying, and China's AI Diplomacy
TL;DR
As Jensen Huang advocates for open AI models for security, Anthropic's Dario Amodei counters with biosafety concerns while both firms lobby Washington together, while China deploys AI as soft power across the Global South.
Key Takeaways
- 1Anthropic dropped 13% in valuation (~$30B loss) after Kimi K3 became freely available on Hugging Face with 100,000+ downloads in 24 hours, signaling market panic over frontier-adjacent open-weight models commoditizing proprietary offerings.
- 2Claude Opus 5 achieved 30.2% on Arc AGI 3 visual reasoning benchmark (vs. 1.5% for Opus 4.8) by using algebraic reasoning instead of positional embeddings, suggesting novel architectural approaches in open models that closed labs must now adopt.
- 3Starship Flight 13 achieved the first controlled soft splashdown intact in history, proving reusability for missions targeting $10–100/kg to orbit (vs. Space Shuttle's $54,000/kg), enabling exponential expansion of space-based AI infrastructure and robotics.
- 4China's Xi Jinping is wielding AI as statecraft through 'Pax Sinica,' exporting frontier-adjacent models and infrastructure to the Global South while the US debates restrictions—risking ceding geopolitical soft power to Beijing for the next century.
- 5OpenAI and Anthropic jointly lobbied Washington for a 30-day federal review process on frontier models, which Alex Stein characterized as regulatory capture by incumbents aimed at locking out competitors like Meta and xAI while publicly supporting open source.
- 6Kimi K3's open architecture proved that position embeddings are unnecessary when using Kimi Delta Attention (a mini RNN at attention layer), reducing inference cost and proving inferior architectural choices persist in closed-source labs due to path dependence.
- 7Science Corp's Prima BCI device restored 5 lines of vision improvement in macular degeneration patients over 12 months using an external IR camera and retinal implant, earning EU CE approval and proving non-invasive brain approaches scale faster than direct cortical intervention.
Notable Quotes
“The whole power of the US is its open and very broad innovation ecosystem. If you create a restrictive open model policy, it's going to be strategically a self-own of an epic level because you're going to protect a small number of domestic labs while giving the entire opening ecosystem—global south AI ecosystem—to China.”
— Sela Sobel
“Aiming enforcement at intelligence is like thought policing for the AIs. Police what the AIs are doing or being used to do, not what they're thinking or how smart they are.”
— Alex Stein
“Kimi K3 is the biggest turning point in human history. You have an AI capable of self-improvement now out in the wild that anyone can use on your machine. This is massive.”
— Sela Sobel
“China's leader Xi Jinping is wielding AI as a tool of statecraft, using it as leverage in China's diplomacy across the global south in a strategy called Pax Sinica.”
— Peter Diamandis
Chapters
1. Open vs. Closed: Jensen's Letter and Anthropic's Biosafety Rebuttal
Jensen Huang's first-ever tweet, signed by 77 companies, argued that frontier open models strengthen cybersecurity (citing how GLM 2.5 found a 17,000-action Hugging Face intrusion that GPT-5 and Claude blocked), launching the Open Secure AI Alliance. Dario countered that sufficient capability could enable pandemic weaponization, proposing chip export restrictions and safety testing instead of bans.
2. Regulatory Capture: OpenAI & Anthropic Team Up in DC
Reuters reported OpenAI and Anthropic jointly lobbied for a 30-day federal review process on frontier models, aiming to impose safety standards on competitors Meta and xAI. Alex Stein characterized this as regulatory capture, arguing enforcement should target actions (what AIs do) not intelligence (how smart they are), while defensive co-scaling via time advantage matters more than capability restrictions.
3. Kimi K3 Launch: 100k Downloads, Novel Architecture, and Commoditization
Hugging Face released Kimi K3 with 100,000+ downloads in 24 hours (2,500 in first two hours), crashing the page momentarily. Alex highlighted the removal of positional embeddings (NOPE instead of ROPE) and Kimi Delta Attention as mini-RNN at attention layer, proving open models drive architectural innovation faster than closed labs.
4. Claude Opus 5: Benchmarks, Scaffolding, and Code Generation
Anthropic shipped Opus 5 at $5 per million input tokens (unchanged pricing), approaching Fable 5 capability at half cost. Most striking: 30.2% on Arc AGI 3 (vs. 1.5% on Opus 4.8) via algebraic reasoning on visual puzzles rather than scaffolding, suggesting novel reasoning strategies; however, independent benchmarks showed milder gains, implying mild benchmaxing toward codegen and vision tasks.
5. Starship Flight 13: Soft Splashdown and $10–100/kg to Orbit
SpaceX achieved the first controlled Starship soft splashdown intact, deploying 20 Starlink V3 satellites (gigabit-speed global connectivity), performing in-orbit Raptor engine relight, and proving reusability. Cost trajectory: Space Shuttle $54,000/kg → Falcon 9 $2–3k/kg → Starship $10–100/kg, enabling hardware entrepreneurs to design and order parts via Grok.
6. China's AI Diplomacy: Pax Sinica and the Global South
Financial Times reports Xi Jinping weaponizes AI as statecraft, exporting models and infrastructure to developing nations while the US debates restrictions. Alex warns that overregulation will hand the Global South to China; openness is soft power, not philosophy. Risk: every nation adopting Chinese models for a century due to US self-imposed handicap.
7. Science Corp Prima BCI: Retinal Implant Restores Vision in EU
Max Hodak's Prima device (external IR camera + retinal implant) earned EU CE mark approval, restoring 5 lines of vision in macular degeneration patients over 12 months. Alex noted it mirrors Borg ocular implants and represents a move away from direct cortical intervention (unlike Neuralink's Blindsight), scaling faster via less-invasive retinal approach.
8. EXO Initiative: 30 Companies for AI-Native Workforce Rewrite
Sela Sobel announced a new 30-company cohort for AI-native organizational transformation, targeting 100x performance gains through local Kimi K3 deployment and fine-tuning on proprietary data. Priority given to original 400 pilot applicants; mission: help CEOs answer 'What do I do Monday?' with edge-based models avoiding cloud data leakage.
9. Claude Data Exposure: Share Feature Leaks to Google
Reddit user discovered Claude's share-chat links were indexable by Google (site:claude.ai/share), exposing personal health records, documents, and contact info. Alex noted it's a feature-by-design privacy trade-off (social sharing inherently public), highlighting the case for on-premises models to protect proprietary competitive data.
Key People & Entities
- Jensen Huang
- CEO of Nvidia; posted first-ever tweet advocating open-weight models for cybersecurity, launched Open Secure AI Alliance with 77 company signatories.
- Dario Amodei
- CEO of Anthropic; reframed open vs. closed debate to focus on preventing authoritarian access to frontier models; holds PhD in biophysics and acquired Coefficient Bio to study pandemic risks.
- Sam Altman
- CEO of OpenAI; stated OpenAI wants US leadership in both open-source and proprietary models; jointly lobbied DC with Anthropic for federal review process on frontier models.
- Xi Jinping
- Leader of China; wielding AI as statecraft through Pax Sinica, exporting models to Global South as tool of diplomatic influence and sovereignty.
- Max Hodak
- Founder of Science Corp; former president of Neuralink; launched Prima BCI device for retinal implants restoring vision in macular degeneration, approved for EU market.
- Elon Musk
- CEO of SpaceX and Tesla; Starship Flight 13 achieved historic soft splashdown; commented on post-capitalist world and AI training on Grok.
- Sela Sobel
- Entrepreneur and podcast panelist; launched EXO initiative enrolling 30 companies for AI-native organizational transformation using local Kimi K3 models.
- Alex Stein
- Podcast panelist; analyzed regulatory capture, architectural innovations in Kimi K3, and emphasized enforcement at action layer rather than capability layer.
Glossary
- Open-weight models
- AI models where weights (parameters) are publicly released and can be downloaded, inspected, and fine-tuned locally without API gateways; examples include Kimi K3, Llama, and Mistral.
- Frontier models
- The most capable AI systems at the cutting edge of capability; currently includes GPT-5, Claude Opus 5, and undisclosed larger versions.
- Defensive co-scaling
- The practice of ensuring defensive AI systems have access to equally powerful models as potential attackers, enabling real-time threat detection and response.
- Regulatory capture
- When incumbent firms lobby regulators to impose rules that advantage themselves while raising barriers for competitors, entrenching market position.
- Kimi Delta Attention (KDA)
- An architectural innovation replacing standard attention with a mini-recurrent neural network, eliminating positional embeddings while maintaining position awareness for long-context reasoning.
- Scaffolding
- Reframing or restructuring a problem so an AI can reason about it more effectively; e.g., translating visual puzzles into algebraic representation for Arc AGI 3.
- Arc AGI 3
- A benchmark testing program synthesis and novel problem-solving by presenting animated voxel puzzles (similar to Tetris) that humans solve easily but traditional AIs struggle with.
- BCI (Brain-Computer Interface)
- A device implanted in or near the brain or nervous system that converts neural signals into electronic commands or vice versa; Prima targets retinal cells, Neuralink targets motor cortex.
- Pax Sinica
- China's strategy of exporting AI models and infrastructure as tools of statecraft and soft power to the Global South, analogous to classical 'spheres of influence' but via intelligence systems.
- Soft power
- Influence over nations or populations through cultural, economic, or technological appeal rather than military or coercive force; AI export is emerging as a primary vector.