Clarity Steps

X's ad revenue sharing was scrapped for a reason: it had become a brand safety nightmare. The model incentivized a flood of low-quality engagement farming, which forced the company to rethink its entire approach. But is it even possible to safely monetize user replies at this scale?

I can prove whether AGI is possible with a for-loop. Quantize a 7B model to 4 bits. Every weight is now one of 16 numbers. Try every combination. Keep the one that passes an AGI benchmark. If none passes, AGI is impossible with LLMs. Someone tell me where this breaks. Let's actually run it. 1. The haystack 16 values × 7 billion weights = 16^7,000,000,000 models. That's about 10^8,400,000,000. Atoms in the observable universe: 10^80. Check one model per atom, per nanosecond, since the Big Bang, and you won't finish the first few dozen weights. 2. "But quantum computers" Grover's algorithm turns M candidates into √M steps. Quadratic, not exponential. √(16^N) = 4^N. Still billions of digits. And every Grover step has to run your entire benchmark through the model as a reversible circuit, in superposition. Nobody has a machine that can run one. 3. "But only one combination is ever in use at a time" True. And that's exactly how training works too. Training holds one combination, scores it, and nudges all 7 billion numbers in the direction that scores better. About 10^6 tries instead of 16^N. Brute force and training both check one model at a time. The difference is whether each try tells you where to look next. The eval score says "bad." The gradient says "bad, and here's how to fix every weight at once." That one extra piece of information is the gap between 10^8,400,000,000 and a few weeks on GPUs. 4. Fine. Pretend compute is free. Now it gets interesting. All 16^N models fail the benchmark. What did you prove? ✅ No model with THIS architecture, THIS size, THIS precision passes THIS benchmark. Training can't find a model that enumeration missed. ❌ AGI is impossible. Change any of the four and the proof evaporates. 2T params. 8-bit. Add memory. Add tools. 5. The hidden catch The weak word was never "compute." It's "benchmark." Every benchmark so far has been beaten by systems that were obviously not general. A test that truly captures "general" starts to look like testing on every possible task. You didn't kill the infinity. You moved it from the weights into the eval. 6. This idea already has a name Solomonoff induction and AIXI: define the perfect learner as "enumerate all programs, weight by how well they predict." Provably optimal. Provably uncomputable. 4-bit brute force is its neural-net cousin. Enumeration proves existence. It never gives you a method. The search IS the problem. Weights hold the answer. The gradient holds the directions.

The #1 predictor of divorce isn't a lack of love. It’s contempt. Research shows respect is what keeps a marriage alive, not passion. What's a bigger red flag in a relationship? A lack of passion, or a lack of respect?