Mars Interior Thermal

Mars's Hidden Heat and the Art of Getting the Consensus Wrong

A study applying tidal tomography to 16 years of NASA orbital data has revealed that Mars's southern interior is hundreds of degrees hotter than its northern interior — a dramatic thermal asymmetry for a body roughly half Earth's size. The technique works by measuring how the gravity of Mars's moons, primarily Phobos, creates tiny deformations in the planet's surface. Those deformations propagate differently through hot, less viscous material than through cool, rigid rock, effectively allowing researchers to map the interior's temperature profile without placing a single instrument on the surface. The north-south thermal divide may be a relic of a giant early impact that redistributed internal heat, or it may reflect a higher concentration of heat-producing radioactive material in the southern hemisphere — either explanation carries implications for how planetary differentiation is modeled more broadly, and both alter where future missions would prioritize subsurface drilling in the search for conditions that could sustain liquid water.

The broadcast's 'What If We're Wrong?' segment this week trained that same scrutiny on one of the day's most confident consensus claims: that U.S. chip export controls are successfully constraining China's AI development, with Tencent's public admission of computing insufficiency as proof. Three assumptions underlying that consensus each carry meaningful uncertainty. First, the conclusion that China's domestic chip development cannot close the gap depends partly on export-control logic — without EUV lithography machines, leading-edge fabrication is impossible — but chiplet design, stacked memory, and novel interconnect architectures offer performance gains on older nodes that Western analysts have previously underestimated, as Huawei's surprise 2023 Kirin 9000s chip demonstrated. Second, Tencent's unusually candid admission that it is 'severely behind' could reflect genuine constraint or could be strategic communication aimed at domestic regulators or government subsidy programs — the two interpretations produce opposite conclusions about policy effectiveness. Third, the assumption that constraining compute constrains AI capability has been weakened by algorithmic efficiency improvements: GPT-4-class capabilities required dramatically less compute to achieve in 2025 than they would have in 2023. The concrete indicator to watch is whether Chinese AI models from Baidu, Alibaba, Tencent, or academic labs begin closing the benchmark gap with Western frontier models over the next six to twelve months — a trend that would suggest the hardware bottleneck is less binding than current policy assumes.