Scaling Past Informal AI - Carina Hong, Axiom Math

This Latent Space piece features Carina Hong of Axiom Math, discussing the direction of "Scaling Past Informal AI": if AI is to handle higher-stakes, higher-precision tasks, it cannot remain at natural-language-style reasoning that looks plausible but is hard to check. The "Verified Generation" mentioned in the title and summary hints that the focus is on making the model's generated results verifiable by some external or formal mechanism, rather than relying solely on the model's self-assertion of correctness. This is especially critical for fields such as mathematical proofs, program correctness, and hardware or software verification, because in these scenarios the cost of errors is high and a clear judgment of right or wrong is needed. Another keyword, "Compounding Intelligence," points to a possible positive cycle: when AI can produce verifiable intermediate results, those results can become a reliable foundation for the next round of reasoning, training, or toolchains, allowing capabilities to accumulate progressively rather than amplifying errors along with them. For AI developers and researchers in Taiwan, the importance of this content lies not in releasing a new model or product, but in reminding everyone that the competitiveness of future AI systems may depend not only on generation quality, but also on whether answers can be checked, composed, and safely placed into actual workflows. It also reflects the rising interest in the AI community recently in formal methods, Lean-type tools, verifiable reasoning, and high-trust AI.