3062 - 决策与人工智能.pdf

《3062 - 决策与人工智能.pdf》由会员分享，可在线阅读，更多相关《3062 - 决策与人工智能.pdf（23页珍藏版）》请在三个皮匠报告上搜索。

1、IBM TechXchange 2025 conference 3062-Decisions and AIIBM TechXchange|2025 IBM Corporation1#IBMTechXchangeStephane MeryStephane MeryIBM Distinguished EngineerCTO,Decision IntelligenceIBMLenny BrombergLenny BrombergProgram DirectorDecision Product Management LeadIBM2LennyIDC predicts that investments

2、in AI solutions and services will yield a global cumulative impact of$22.3 trillion by 2030,highlighting AIs transformative potential2Source:1.IDCAI:The$22 Trillion Rocket ShipIBM TechXchange|2025 IBM Corporation23Can we trust AI to take decisions?New AI(GenAI)is so good at many tasks We assume AI c

3、an answer every questions accurately We assume AI might outdo Humans in decision-making What do you think?What do you think?Are you ready to put AI in the driver Are you ready to put AI in the driver seat of your business?seat of your business?IBM TechXchange|2025 IBM Corporation4 LLMs do not really

4、 understand to a point where it can reason and make inferences and deductions It simulates reasoning via statistical patterns Hallucinations are frequent,idempotency is not guaranteed Can we trust AI to take decisions?.Not really IBM TechXchange|2025 IBM Corporation5Can we trust AI to take decisions

5、?.Not really Not the sweetspot of stochastic systems(Machine Learning,LLMs).Mathematics and logic sweetspot.IBM TechXchange|2025 IBM CorporationThe Illusion of Thinking:Key findings6The experiments revealed a consistent three-regime pattern across models:Low ComplexityLow ComplexityStandard(non-reas

6、oning)LLMs often outperform reasoning modelsbecause LRMs may“overthink”simple problems,introducing unnecessary complexity.Medium ComplexityMedium ComplexityLRMs show clear advantagesextra reasoning effort helps them navigate intricate puzzles more effectively than non-reasoning counterparts.High Com