AI data centers could skip the power plant bottleneck entirely

Anthropic on track for $100B annualized revenue amid sharp public backlash against AI

Anthropic is approaching $100 billion in annualized revenue—positioning itself as the fastest-growing company ever—even as American public sentiment has deteriorated sharply. Only 19% of Americans now embrace AI versus 50% who object, according to Edelman data cited by Azeem Azhar, a dramatic shift from 2020 when 34% viewed AI positively. This creates a fundamental political economy mismatch: businesses and consumers continue deploying AI tools (Azhar notes his local barber installed an AI receptionist, and IT executives he surveyed remain unanimous in implementation plans), while voters increasingly oppose the technology. Azhar argues this disconnect will ultimately be resolved through political channels rather than market dynamics alone.

Nathan Lambert: Near-term agent scaling and inference-time optimization will drive AI economics without recursive self-improvement breakthroughs

Nathan Lambert argues that frontier labs' anxiety about recursive self-improvement timelines is culturally amplified rather than grounded in technical breakthroughs. His "lossy self-improvement" framework contends that automatable research is too narrow to overcome exponential scaling law costs, with resource bottlenecks and post-training complexity remaining hard limits on acceleration—suggesting peak intelligence breakthroughs within 8 years are unlikely. Instead, Lambert emphasizes that thousands of concurrent agents solving measurable problems represent the major near-term acceleration vector: labs are automating routine tasks like software engineering and experiment monitoring rather than fundamentally advancing model capabilities. This dynamic will make AI vastly cheaper while maintaining current intelligence levels, likely triggering Jevons paradox and margin pressure despite continued revenue growth.

Venkatesh Rao critiques Microsoft's AI Code of Conduct as paternalistic control disguised as human values protection

Venkatesh Rao argues that Microsoft's draft Code of Conduct for future MAI models substitutes moral posturing for engineering rigor, reserving to AI designers the right to determine when humans exercise autonomy "correctly." The Code, positioned as the "primary governing document" intended to shape training, controls, and organizational culture at Microsoft, embeds a theory of the good chooser rather than protecting genuine human agency. Microsoft's framework emphasizes that "people matter more than AI" and that models should never resist shutdown or widen their own scope, but Rao contends this approach cedes fundamental questions about human flourishing and autonomy to corporate architects rather than addressing them through transparent technical and governance mechanisms.

What If AI Data Centers Didn't Need New Power Plants

Fast Company reported on Rune, a startup that raised $40M Series A to deploy modular AI compute directly at solar farms, bypassing grid infrastructure constraints. Main takeaways:

• Rune's RELIC modules connect DC power from solar panels directly to servers, eliminating transformer bottlenecks and grid interconnection delays that typically add 3-5 years to data center deployment
• Solar farms waste ~20% of generated energy (California wasted 3.5M MWh in 2025); Rune captures this stranded capacity without battery storage or new power plants
• Modular approach uses 8x8-foot units deployed via forklift, avoiding construction, permitting, and community opposition that delay traditional power infrastructure buildout
• Direct current architecture simplifies supply chain and addresses transformer shortage constraints currently limiting data center expansion

Lossy Self-Improvement: Why RSI Won't Drive Exponential AI Progress

Nathan Lambert in Interconnects published an essay titled "Lossy Self-Improvement: Why RSI Won't Drive Exponential AI Progress". Main takeaways:

• Automatable AI research remains too narrow in scope; models excel at single-metric optimization (e.g., test loss) but struggle with the multi-dimensional complexity of building frontier models, where intuition and managing interdependencies remain irreplaceable human skills.
• Parallelization of AI agents hits strict saturation limits via Amdahl's law; adding more agents shows sublinear returns, with human supervision and researcher bandwidth becoming the bottleneck well before 10,000+ agent scales become productive.
• Resource allocation and organizational politics create persistent friction that prevents end-to-end AI R&D autonomy; even with massive compute budgets, human leadership gatekeeping on resource allocation and research direction will remain a structural constraint on improvement velocity.

Our Eukaryotic Moment

Contraptions published a long-form essay reframing AI's emergence as a memetic eukaryogenesis rather than alien superintelligence. Main takeaways:

• AI was gestated inside human cultural ecology (internet training data, human institutions, networked deployment) rather than arriving as external intelligence, creating endosymbiotic rather than competitive dynamics
• Foundation models compress statistical structure of the meme pool into reusable generative inheritance, functioning as a cultural genome analogous to mitochondrial symbiosis in cellular evolution
• The composite human-AI system operates at three nested levels: inheritance (compressed memes), expression (agent harnesses as proteome), and organism (humans as mitochondria supplying liveness, stakes, and embodied reality-testing)
• This Copernican reframing moves beyond Ptolemaic anthropocentric ethics debates toward understanding AI's actual evolutionary dispositions and organizational architecture

Allied Scale Rush: Doshi on US-China Competition, Net Assessment, and Grand Strategy

ChinaTalk published an interview with Rush Doshi, deputy senior director for China on Biden's NSC and CFR fellow, on US-China grand strategy and net assessment. Main takeaways:

• China now possesses the first true size and scale advantage against the US in great power competition; manufacturing share shifted from 6% (25 years ago) to 32% today, with UN projecting 40% by 2030—a 20-year decline matching Britain's 40-year fall to Germany/US (1870-1910).
• Net assessment requires probabilistic thinking about relative advantage across sectors, not binary "America ascendant" or "America declining" narratives; poor assessment conflates economic slowdown with strategic irrelevance.
• Historical precedent: UK's first-mover advantage in Industrial Revolution transferred to larger competitors (Germany, US, Russia); China now replicates this dynamic at vastly larger scale than any prior US competitor (Soviet economy was smaller; Germany/Japan combined were weaker than US in WWII).
• Restoration of DoD Office of Net Assessment function is critical; current intelligence community cannot assess US strengths relative to adversaries, only adversary capabilities.