For UNSAT problems with 10 variables and 200 clauses it had the same issue as others: making up assignments.
ATM in a more interesting context, and despite lackluster adoption of later IBM
，详情可参考一键获取谷歌浏览器下载
Surprisingly, as a smaller model it performed better than Gemini 3 Pro. It found some valid assignments for SAT formulas, but has the same issue of making up assignments for UNSAT formulas.
但2025年，这个核心逻辑出现了裂缝。DeepSeek的横空出世，彻底打破了“算力至上”的行业迷信——其开发的模型仅用2000块H800 GPU，就实现了与Meta Llama 3（使用1.6万块H100）同等的性能，训练成本仅需560万美元。