gpt usecases

[hyunjimoon]

analyzing referee report

[hyunjimoon]

🏀giving challenge to intrinsically motivate gpt: competing claude ai and gpt

gpt's performance boost after comparing with claude ai. however, it (blindly?) follows some of claude's answer (e.g.1.3 subproblem is removed)

Tom Malone shared online copy of his thesis: https://www.researchgate.net/publication/234608112_What_Makes_Things_Fun_to_Learn_A_Study_of_Intrinsically_Motivating_Computer_Games

along with published work based on the thesis: http://ccs.mit.edu/malone/Malone%20computer%20games%20bibliography.htm

[hyunjimoon]

Agenda for meeting with suyeol(SY), saehoon (SP), Angie (AM)

given Scala from SY, and Q from SP, AM listed some chat agenda:

1. SP's usecase

SP's Q: SD 모델링을 하는 데에 필요한 상수 값을 gpt를 통해서 결정할 수 있을까? (뉴스를 이용한 gpt의 결정을 반영하는 것이 시뮬레이션상에서 가능할까 e.g. 글로벌 공급망에 영향을 줄 수 있는 뉴스를 검색하고 그 안정성에 대하여 10점 만점에 몇 점인지 평가해서 숫자로만 대답해줘! (SP's Q)

AM's A. would need to experiment with different services e.g. scala, orbit, need zoom in and zoom out

AM's Q1 to SP. if you were to use SCALA, what's your ideal use case of Scala's in supply chain contexts?
Q2 to SP. i summarized existing methods relevant to your use case (table below). how much time do you spend on each: developing statistical/dynamic/simulation models, manually monitoring of news and translating into model's parameters, designing decision rules (e.g. policies like safety stock)? Which area do you think can LLM/AI can improve efficiency the most?

Q1 to SY: among the last column how much is LLM's character (different from nonLLM AI models)?

Existing Method	Example of How It Would Address Iceland Volcano Event 🌋	Enhancement by AI System (e.g., SCALA) 🤖
Statistical Models 📊	Use historical data to calculate the impact on air freight and delivery delays. For example, calculate the average delay in days (e.g., 5 days) for perishable goods and use this to predict stockout scenarios. Adjust inventory levels accordingly to buffer against expected delays.	AI can incorporate real-time data and news updates, refining predictions as new information becomes available. It can also simulate various scenarios quickly, providing dynamic adjustments to the model.
Manual Monitoring 👨‍💼	Human experts monitor news and manually assess the impact on supply chain routes. They might estimate that 30% of air freight routes are disrupted and adjust plans to use alternative routes. For example, reroute 50% of perishable goods through sea freight, accepting a longer transit time of 10 days.	AI automates the monitoring process, quickly aggregating data from multiple sources and providing real-time updates. It can suggest optimal rerouting strategies and quantify the impact of disruptions more accurately.
Rule-Based Systems 📋	Predefined rules trigger responses based on specific events. For instance, a rule might state that if air travel is disrupted for more than 3 days, 40% of goods should be rerouted via alternative transport modes. These rules are based on past events and fixed criteria.	AI can enhance these systems by updating rules dynamically based on real-time data and predictive analytics. It can also identify new patterns and suggest modifications to existing rules, improving the system’s adaptability and effectiveness.

painpoints of supply chain management expert :

• Event-Linking Challenges: Difficulty in linking specific events to supply chain impacts, especially unprecedented ones.
• Expert Reliance: Dependence on expert judgment and experience, causing delays and inaccuracies.
• Decision Quality: Hard to maintain high decision quality without extensive data and experience.

2. SY's Scala feature in the context of 1

SCALA is

• fully autonomous AI agent that monitors and understands real-time world events and generates predictions.
• provides actionable insights to enhance decision-making by tracking events from multiple data sources, providing explanations and predictions, and recording them in a chronological timeline
• can pool heterogeneous data sources (voice, text, email, ensuring) and give timely responses to global changes

Q2 to SY. how are hypotheses generated and tested?
Q3 to SY. what probabilistic model is used? how rich is it? e.g. upto what dimension of multivariate probability distributions can the model run? how can the modeler check it runs well?

3. AM's interest given 1,2

AM is interested in probablistic programming based GenSQL for querying and managing database tables with uncertainty. (preliminary). It is AI that understand the world and can seamlessly connects data and statistical model query, as exemplified in here. more in #224 (comment)

Q3 to SP: does the above seem useful for your SD model building (Bayesian workflow - iteration of which stat.model, data, approximation algorithms)?
Q4 to SY: have you felt the need to query statistical model ? (e.g. depending on the degree of rare events, one can categroize fat tail distribution into three; normal, t, inverse gamma)

[hyunjimoon]