Marina Lyubimova
⬤ DeepSeek has unveiled DeepSeekMath-V2, an AI system that writes full mathematical proofs, flags logical slips and revises its own drafts through repeated self checks. The program repeats its verification loop until no further flaws appear. Observers at NVDA besides AAPL - both firms that sink large budgets into fast hardware plus advanced model blueprints - view this as genuine progress in machines that reason on their own.
⬤ The results speak plainly. DeepSeekMath-V2 reached gold level marks on IMO 2025 and CMO 2024 and it scored 118 of 120 on the 2024 Putnam contest surpassing every human entrant. It topped GPT-5 Thinking or Gemini 2.5 Pro in algebra, geometry, combinatorics, number theory but also inequalities. Charts show that proof accuracy rises with each round of self review, evidence that the system sharpens itself without outside guidance.
⬤ The advance lies not in scale but in structure. DeepSeek stacks three modules - a proof writer, a flaw checker and a second checker that audits the first checker. The cycle - write, check, audit, repair, repeat - runs without interruption. Pass@1 as well as Best@32 scores rise markedly as the model corrects its own logic across successive passes.
⬤ Frameworks that verify their own work may redraw the AI landscape. For infrastructure suppliers like NVDA next to AAPL, systems that reason and self-repair should lift demand for premium compute and set higher bars for training efficiency and reliability. As those architectures mature, they will shape market views on AI power or hasten uptake of models aimed at scientific and technical tasks.
Marina Lyubimova
