SpaceX's $2T Case, Nvidia's Shock Selloff, America Turns on AI, Trump Pulls AI Order, Bond Crisis?

André Karpathy, 39, has joined Anthropic to lead a new pre-training team focused on recursive self-improvement — the idea that AI models can meaningfully participate in their own training loops. Karpathy is already a legend: founding OpenAI member, former head of Tesla's self-driving team, and creator of viral projects like Autoresearch and the «André Karpathy Skills» tool for Claude code. His track record spans multiple waves of AI innovation, from brute-force computation at Tesla (where he hand-labeled video data in 2016–17) to co-founding OpenAI.

Recursive self-improvement and continual learning represent the final frontiers for AI. If unlocked, they could enable models to improve at exponential rates — potentially 10x performance gains every year. As one participant noted, «if you track what folks, particularly Jeff Dean, did inside of Google, it's like wave upon wave, they were at the foot of those waves. What's interesting about André is he's been at the wave upon wave of AI.» With Anthropic already EBIT-positive and holding a multi-month lead over competitors, Karpathy's addition could cement their dominance.

The significance extends beyond Anthropic. Recursive learning, combined with massive compute clusters like Colossus, could put AI development «on a combination of overdrive and autopilot,» as one guest put it. The models begin training themselves faster than humans can orchestrate improvements — a threshold the industry may be approaching sooner than expected.

The AI backlash is not irrational. People perceive that «a small number of people that control and profit from and benefit from AI are gonna end up getting outsized returns relative to the broader population,» as one host explained. The time to diffusion is long enough that economic asymmetry will persist for years, creating a power imbalance that feels deeply unfair. This mirrors the post-nuclear-bomb skepticism of science and technology that defined the late 20th century.

Second, there is evidence of external actors amplifying anti-AI sentiment. «I don't think it's just China with NGOs today. There is a long history of state actors intervening in media activities in foreign nations to try and create sentiment that reduces progress in that competitive state,» one participant noted. Techniques refined during the Cold War are being deployed today — a concerted effort to slow American technological progress. China, in particular, has a clear interest in seeing U.S. data centers blocked and AI development slowed.

Finally, AI challenges human ego in ways that feel anti-humanist. Just as heliocentricity disrupted the Church by decentering Earth, AI disrupts the primacy of human intelligence. «There's something about AI that's very not human centric, and it kind of shifts with the ego of the human,» one host observed. This psychological current fuels disdain even when economic arguments are set aside.

Cursor's Composer 2.5 model, released this week, is Pareto dominant — meaning it outperforms all other models across the entire performance-cost frontier. This was achieved in just three to four weeks of reinforcement learning on Colossus 2, using Cursor's proprietary coding data. Cursor allegedly has more tokens of coding data than exist on the public internet, and that data is now being injected into model training with extraordinary results.

This is a major validation of two things: first, that proprietary data moats are real and defensible; second, that access to cutting-edge compute clusters (like Colossus) is a kingmaker. Cursor was «dead in the water» in terms of compute access and falling behind competitors like Codex, Google, and Anthropic. Elon let them onto Colossus, and «boom, instantly their models are growing faster.» One participant predicted: «We could be sitting here a year from now and they're the dominant player.»

The next phase: Cursor will have a new base model with its data baked in from the start, trained using the biggest coherent compute cluster in the world. This could be the moment when verticalized, domain-specific AI models eclipse general-purpose ones — and XAI/Cursor leapfrog the field.

The 48-hour tech CEO delegation to China, led by President Trump, produced modest public outcomes: soybean purchases, aircraft sales, and limited H100/A200 chip sales to Baidu. But participants believe the real value was behind closed doors — «geopolitically the tic-tac-toe of what has to happen next.» One participant speculated: «There's a way to divide up the game board in a way that helps them and helps us.» Venezuela, Iran, and Taiwan are key pieces on that board.

Selling deprecated Nvidia GPUs to China may actually lower the odds of China developing its own alternative AI ecosystem, which would be more power-hungry and optically reliant. «This is stabilizing for the world and is the best, highest probability path for keeping America ahead in AI,» one guest argued. Trump's «superpower is his ability to bond with dictators, monarchs, royal families» — a skill that could yield quiet geopolitical wins even if the public optics are muted.

Putin's immediate meeting with Xi following the U.S. visit is also performative. The story continues: no happy ending, no grand détente, but a recalibration of power that may avert the Thucydides Trap for another decade.