Trump pulled the United States out of the Trans-Pacific Partnership in January 2017, and that single executive action handed China a geopolitical gift that Beijing's central planners spent decades trying to manufacture for themselves. The TPP covered twelve nations accounting for roughly 40% of global GDP. It was deeply flawed in its intellectual property provisions and regulatory harmonization schemes (free market thinkers had legitimate criticisms worth hearing). But the core logic was sound: bind Vietnam, Malaysia, Japan, Australia, and others into a trade architecture where American commerce set the terms. When Washington walked, Beijing stepped in with the Regional Comprehensive Economic Partnership, signed November 2020, covering fifteen countries and 30% of world GDP, with China at the center of gravity. You do not get a second chance to make that offer. The protectionist coalition that killed the TPP told you they were defending American workers. The actual result: Chinese firms gained preferential access to Southeast Asian supply chains while American exporters paid higher tariffs into those same markets. Vietnam's manufacturing sector grew 8.36% in 2023 alone. American companies increasingly route production through Vietnamese factories precisely because the rules Beijing now shapes make that the path of least resistance. Free market economists have long understood that trade is not war. Every voluntary exchange creates value on both sides. The mercantilist obsession with trade deficits as national wounds produces exactly this: Washington sacrifices market access and institutional influence to protect industries that restructure anyway, just more slowly and more expensively. China did not seize Asia. Washington handed them the keys.

Robotics deep dive: HUMANOID ROBOT EMBEDDED COMPUTE Every robot company that says anything about its onboard compute quotes a TOPS number. Almost none of them are quoting the same thing. NVIDIA's Jetson Thor markets 2,070. The footnote on NVIDIA's own page: sparse FP4. The previous generation, AGX Orin, markets 275, under a column header reading sparse INT8. Divide one by the other and you get the headline claim for the generation, up to 7.5x more AI compute. Put both on the same footing and it is 3.76x. Dense FP8 against dense INT8, 517 against 137.5. The other half of the speed-up is the number format. NVIDIA states 3.5x better energy efficiency, which reconciles as 2,070 over 130 watts against 275 over 60. At matched precision it is 1.74x. Both headline claims collapse by exactly 2x, by the same mechanism. Sparsity is real, FP4 is real, and NVIDIA prints both in its spec table. Thor is a large generational step. It is roughly half the step the marketing arithmetic describes, and the difference lives in a footnote most people never resolve. Unitree's page for the H2 PLUS states its chip correctly: FP4 2,070 TFLOPS. Search-indexed writeups of the same robot render it as a 2,070 TOPS chip. Floating point silently became integer and the precision qualifier vanished. The spec survived the vendor and died in the coverage. Qualcomm's Dragonwing IQ10 markets up to 700 TOPS and states no precision. Horizon's Journey 6P markets 560 and states none either. Either could be INT8 or INT4, a factor of two, with no document that settles it. Neither publishes a power figure, so neither can be placed on a per-watt axis at all. Horizon does publish one thing nobody else does. The footnote under its TOPS figure reads TPP under 4800. Total Processing Performance is the threshold metric in US export control. That number is bounded by a regulation rather than by the silicon. Figure, 1X, Agility and Apptronik publish no compute specification at all. The only two humanoid vendors that name their chip in a public spec table are Chinese.

Thoughts About Scaling Law Scaling, but not only of parameters. Every model release now ends with the same question: how many parameters? It isn't a question that can be answered on its own. Parameter count is only meaningful alongside three others — how much data you have, where you intend to spend your compute, and who will run the model, under what conditions. The field learned this the hard way. Kaplan et al. (2020) fit an exponent that told everyone to grow parameters faster than data — roughly 2.7:1 — and the industry complied: GPT-3, Gopher, MT-NLG. Hoffmann et al. (2022) redid the experiment across four hundred models and found the compute-optimal split is closer to 20 tokens per parameter, and that with sufficient compute the two should grow at the same rate rather than drifting apart. The error in the earlier fit compounded with every order of magnitude of compute, which is why the largest models of that generation were the most misallocated. The trillion-parameter round was, in retrospect, a detour the whole field took together and then reversed. Chinchilla wasn't the end either. It optimized training compute for models that would be trained once and evaluated. Today a model is called billions of times a day and inference dominates lifetime cost. Put inference into the objective and the optimum moves toward smaller models trained far longer — deliberate over-training, which is what Llama-2-7B and Gemma-2-9B were doing at roughly 290 and 889 tokens per parameter. Sparsity moved the target again. In a MoE model two quantities have to be kept apart: total parameters govern roughly how much the model can hold — knowledge, facts, the long tail — while activated parameters and effective depth govern roughly how far it can think, how many steps of a causal chain it can carry before it comes apart. A dense 20:1 ratio does not transfer. And the ratio isn't a single number at all: Roberts et al. (2025) find the optimal tokens-per-parameter is task-dependent, with memorization favoring more parameters and reasoning favoring more data. Follow-up work on MoE observes that at fixed TPP, pushing total parameters higher actually degrades reasoning, while activating more experts reliably helps it. This matters for what we are building toward. Finding a vulnerability is not a retrieval problem. It doesn't come from having memorized more CVEs; it comes from carrying a twenty-step chain of inference to the end without losing the thread. That capability does not live in total parameter count. Which brings us to this release. Total parameters appear to matter up to a threshold — enough to hold the world — after which additional capability comes from scaling elsewhere: effective depth per forward pass, and above all post-training. GLM-5.3 is our controlled experiment on that claim. Same base, same architecture, same total and activated parameters as GLM-5.2. One month of scaling long-horizon environments and RL. The gains are not marginal. Well, scaling has more than one dial. We turned the post-training one this time because it had the most slack left in it — not because the others are finished. Base model size, pretraining data, compute spent per forward pass: all of them are still on the table, and we will come back to each. What this experiment taught us is that the dials do not have to be turned together, and that the one worth turning next is rarely the one that was worth turning last. We are not done scaling. Next time, maybe mid-training, pre-training, and even more.