America’s Official AI Plan: Genesis Mission, Claude 4.5, Google vs. NVIDIA, & ChatGPT Shopping

In the AMA on poverty, Immad and Salem argued for a three-part solution: coordinate with AI, give everyone universal AI, and layer services on top. The pace of job displacement (e.g., customer service agents gone within 2 years) demands urgent action. Universal AI agents would act as a 'Jarvis' looking out for individuals, helping them navigate bureaucracies and access resources. This mirrors historical New Deal-scale responses but powered by AI. The coordination problem (enough resources, poor distribution) is solvable with AI social scientists and individualized AI companions.

The model scored 52% on Scale's SWE-bench Pro (a hard reasoning benchmark) without using reasoning tokens, shocking observers. Combined with Haiku or Sonnet, its multi-agent support jumped to 88%, proving AI can effectively orchestrate other agents. The cost reduction and efficiency mean coding and token generation will become ubiquitous; it may one-shot most codebases next year. Anthropic's incoming employees on the performance team are now outperformed by the AI on key homework assignments, indicating that the innermost loop of civilizational progress—models that research and code better versions of themselves—is in motion. This loop will spin out to solve robotics and physical-world automation, driving radical growth. In the near term, individuals will trivially generate complex software, video games, and applications from a single prompt.

Cost of intelligence is being driven to zero, including super intelligence. Opus 4.5 and a company named Poetic have shown superhuman performance on ARC AGI 2, meaning visual pattern recognition and program synthesis out of pixel inputs are solved. The authors of ARC are reportedly asking 'what on earth do we do now'. The next benchmarks are dollar-based: vending bench, trading benchmarks that measure real economic value generation. This shift indicates we're moving from models that perform discrete tasks to models that can do economic work. It also underscores the need for harder evals, and that within a year, an individual entrepreneur with a fleet of AI agents could build a billion-dollar business, with examples already emerging like 47 AI startups launched in a month.

This is OpenAI's third vertical specialist agent after Deep Research and Codex, moving beyond a singleton generalist model. Amazon's CEO Andy Jassy announced their shopping agent Rufus has 250 million users and estimates $10 billion incremental sales next year due to 60% higher conversion. The critical question is who captures user intentionality: the agent embedded in the search/retail platform (Amazon, Google) or a more charming, cross-platform personal assistant (Jarvis). The eventual aim is a personalized AI layer that understands gaze, conversation, and implicit desires, then purchases autonomously. This collapses the whole product research economy and turns consumer decisions into a compute problem.

The Ironwood TPU scales from 64 to 9,000 units with multi-data-center capability, which is critical for large context windows. DRAM prices have spiked 5x, and Google's architecture can use cheaper chips at massive scale for large context, giving Gemini an advantage with 1-2 million input tokens. Other competitors like AMD, Amazon's Trainium/Inferentia, and Cerebras create a heterogeneous ecosystem. Backpropagation in training remains a bottleneck, making training chips harder to commoditize than inference, but for inference, hardware is becoming fungible. The trend is toward accelerated compute as a commodity, which is good for the industry.

Amazon's Project Rainier in Indiana transformed farmland into a compute facility for Anthropic's training and inference within a year, reflecting the 1939-like mobilization. Tranium 2 chips are equivalent to Hopper for inference but lag for large-scale training; they are used to break inference capacity constraints. As cloud giants differentiate, Amazon's strong government ties and massive scale position it as a key player in sovereign AI, providing lower-cost inference at scale. The discussion also notes that DRAM prices are surging, making inference-focused chips strategic.

Peter Diamandis emphasizes that everything of value on Earth exists in near-infinite quantities in space. The electromagnetic mass drivers (Gerard K. O'Neill's concept) on the moon could accelerate payloads to lunar escape velocity using cheap solar electricity, dropping cost to $0.10/kg. This would enable disassembly of asteroids and construction of rotating O'Neill colonies for habitation. Alex jokingly says Saturn and the moon are due for dismantling, but notes we'll have the ability to protect Earth and recreate tides artificially. This technological progression turns physical launches into a software-like exponential path, echoing the broader dematerialization trend.

Alex notes that extrapolating the cost and size of a gigaflop of compute to the size of a brain cell yields 2045 for nanobot-based uploading. However, partial solutions using low-bandwidth fMRI (1-2 sec temporal res, 1mm spatial) can reconstruct thought, and groups like Meta are training foundation models from fMRI scans. Whole-brain emulation via destructive scanning (connectome) has been done for fruit flies, and mice are next; a full human connectome is expected within 5 years. Immad adds that AI can help solve the hard problem of consciousness, and that BCI will be one of the biggest investment areas due to geostrategic importance. Salem emphasizes that we don't need full bandwidth; partial bandwidth with diffusion models can reconstruct brain processes. The discussion also touches on Drexlerian nanotech assemblers arriving within 10 years.