FOW: future of work T4 in M3S, ai for human capital and economic growth

[hyunjimoon]

https://m3s.mit.edu/t4-ai-for-human-capital-development-in-future-of-work

---

On this meeting, Jinhua asked for some help in investigating "AI in education, learning, human capital" to Angie

[hyunjimoon]

using AI tool: converging or diverging force? star vs super-star

summary on Somin21_pref_attach_star.pdf on star vs superstar (co-authored by Sandy Petland, PI of T4 pj, first speaker of this MMI forum on AI ecosystem)

1. fig.5b "nodes that manage to ‘climb the ranking ladder’ and reach its top, are significantly more likely to fortify their ranks, despite the potential massive fluctuations in their absolute network degrees."

2. fig.5a. inverse-U ranking (R) evolution and its calculus interpretation:

• negative derivative of R around N
• positive derivative of R around top-ranked nodes

3. three empirical data (amazon ranking, blockchain trading ledger, social trading platform). Estimation-based analysis was only performed for the last dataset as amazon and blockchain's N, H are too large.

[hyunjimoon]

@chasfine raised the issue "AI might reverse the trend of lower researcher productivity" after reading Nick Bloom's first papers from below:

• are ideas getting harder to find? (link)
• identifying technology spillovers and product market rivalry (link)

I wonder whether the definition of researcher and productivity would remain the same.

Idea's is characterized by nonrivalous and may/may_not be excludable. The most insightful comment so far from Scott is, in two by two quadrants by Romer, non excludable and rivalous (e.g. common goods) face overuse and non excludable and non-rivalrous (e.g. basic research) face undersupply problem. Which axis would AI have greater impact? rivalous or excludability? could this resolve either undersupply or overuse problem?

I think ai/llm's massive customization (i.e.transforming propositional knowledge to prescriptive knowledge at any state) can weaken the assumptions for burden of knowledge mechanism (2nd row below). The length of education to function well would be shorter.

[chasfine]

Your last line: "The length of education to function well would be shorter."

This may not be the case. As LLMs become more sophisticated, the degree of sophistication required of the user to detect faulty reasoning and subtle hallucinations may actually be higher. In short, naïve acceptance of LLM answers may increase when used by less sophisticated users.

[hyunjimoon]

One example of hallucinations I faced while discussing micro-economy problem (hicksian demand) with gpt:

as the error rate is low, if it makes mistake as below (px disappeared in the last line), I spend some time testing it. however, is this fair to LLMs in that we are starting from too high standard/expectation? I wouldn't have solve this problem without gpt's help (unless I go to office hour to study with prof/TAs). So for fair comparison, population should be segmented by use-cases where naive acceptance is ok or not. for instance, for routine job of finding path for students, it's ok for google map to have lower precision and sometimes gives errors (wrong detection of my current location, doesn't take into consideration riverside road from mit to harvard is closed during weekends) but for military purpose, precision should be much lower and requires several iteration

[hyunjimoon]

Biasi23_edu_innov_gap.pdf (candidate for referee report) analyzes how cutting-edge knowledge spreads through higher education. By using NLP on 1.7 million university syllabi and 20 million academic papers, the study introduces the "education-innovation gap" to gauge how close a syllabus is to the latest knowledge. Key findings:

1. Significant variations exist in the education-innovation gap across courses, even when considering the subject, course level, and timeframe.
2. Instructors, especially those active in research and teaching related topics, influence course content towards frontier knowledge.
3. There's an inequality in accessing frontier knowledge: Schools with more privileged students have a smaller gap.
4. Students from schools with a lower gap have better post-graduation outcomes, including higher chances of earning a Ph.D., more patent production, and higher earnings.

[hyunjimoon]

Scott Stern's course 15.911 ENTREPRENEURIAL STRATEGY is experimentally using AI to analyze startup casestudies:

AN ENTREPRENEURIAL STRATEGY “AI EXPERIMENT”
A significant portion of the graded written work for this course involves case memos (3 group memos, and one individual memo) that ask you to analyze an E-strat case in some detail and then offer concrete recommendations to the case protagonist facing a critical choice for their venture.
Consistent with the idea that AI large language models are going to inform both analysis and choice going forward (particularly for start-ups), each case now includes a representative case analysis produced by ChatGPT-4 (August 2023). Moreover, the case prompts have been rewritten to reflect the idea that the case protagonists would have access to this AI-level analysis.
Accordingly, your task will be to write up an independent memo (either as a group or individual depending on the assignment) that can build upon the AI analysis but is ultimately your own analysis and recommendations. Specifically, the grading of each memo will specifically target two criteria:

• Overall Clarity and Memo Quality (75%). The most important criteria will be not whether you “agree” or not with ChatGPT but whether the analysis you provide, and the recommendations you make, offer an integrated entrepreneurial strategy analysis that leverages the course tools and frameworks and applies them in an effective way to the case at hand.
• Identifying Your Own “Value-Added” (25%). Effective memos will point out areas of agreement and disagreement with the AI anaysis, including specific ways in which course framework and tools can enhance the analysis (either to reinforce or contradict the AI analysis) and specific ways in which the case facts might provide evidence in favor (or against) the AI analysis. It will also be useful to consider how the ways in which you build upon the AI analysis (either in agreement or disagreement on various points) ultimately affects the recommendations you make to the case protagonist.
  As a teaching “experiment,” we will review how well this is working after the first few lectures, and course-correct as necessary. Suggestions and feedback on this are welcomed – the whole Estrat team looks forward to effectively integrating AI analysis as a core element of choosing and scaling an overall entrepreneurial strategy.

[chasfine]

Regarding "each case now includes a representative case analysis produced by ChatGPT-4 (August 2023)": How are these analyses generated? Same prompts for each? Sequence of prompts, customized to each case? Do you have access to the prompts used and answers provided?

[hyunjimoon]

@chasfine info i have is "ctrl+f gpt" from
Entrepreneurial Strategy Group HW Assignment Prompt and Instructions F23 FINAL.pdf

[hyunjimoon]

T4 meeting,

1. inclusion of ethical is modularized in HQ
2. sequential human decision (include long term academic)

• decision is non-rational so ai agent can improve it

3. construct mobility network; use phone trajectory to understand economic network

[hyunjimoon]

09.26 Jinhua, Charlie, Angie

prior

T4#1. human dynamics

state of the tool VS perceived state of needs of the user base ~ entrepreneurship iterative testing

• in system dynamics inference algorith (A) << generative model (P)
• applying bayesian workflow concept described here

post

@jinhua-mit yesterday's meeting was very helpful in that it offered us where to direct our efforts + set milestones. As we've discussed, communicating with the customers in Singapore who might be interested in Charlie and my product for entrepreneurs (nss_gps🛰️) before I visit Singapore would be helpful. This is another AI technology so we need to plan for smoother adoption by VC and startups.

• Economic Development Board (EDB) of Singapore could potentially be a customer. Jinhua said that if they had an appealing idea to help EDB cultivate or provide help for startups to be more successful in Singapore, EDB would embrace it full heartedly.
• Smart Innovation Center - Mentioned as a potential source of entrepreneurship data.
• Monetary Authority of Singapore (MAS) Foundation - Mentioned as working with preschool aged children from low-income families.
• Howard Califano - Mentioned as running entrepreneurship programs at the OneNorth Center in Singapore that had 20-30 participants.

[jinhua-mit]

Check this out: Experimental evidence on the productivity effects of generative artificial intelligence; SHAKKED NOY AND WHITNEY ZHANG; DOI: 10.1126/science.adh2586

[hyunjimoon]

@chasfine
https://otter.ai/u/jKwb5HnyvNKgAEz-R58YD5pzDQw?utm_source=copy_url is the transcript and my notes on the M3S T45 project meeting.

Two action items:

1. I suggested to Jinhua and Shenhao that I'd come up with a concrete data collection plan in Singapore. e.g. how many and who to interview. I need help in planning this.

2. May I carefully invite you to the meeting? It's a little late: biweekly on Monday 8:30-9:30pm with around eight people participating (Sandy, Jinhua, Shenhao and five RAs). I'd like to do another scaling gymnastic 🤸🏻‍♂️ 🏋️ with you and everyone would be so happy with your feedback. Also, we can get more intense feedback from Jinhua and Sandy etc. on our work.

Fun S2M modeling (segment till small world to massage) for today:
Forming collaboration structure (:= p(a, t) each agent's action a at time t). Program to massaging p(a, t): 9 agents,10 meetings during three months (the time we said we'd present something to Jinhua).

conditional on one agent ( | a=a1)

1. elicit desire of each member and express it into long term goals g1..g9
2. decompose long term goals into short term operational actions a1_{1,...10}, ..., a9_{1,...10}

conditional on one time ( | t=t1)

3. at each t, construct collaborative architecture that enables a. input to be pooled b. output to be combined for t=t1+1 which becomes t1+1's input
4. if 3 is not good enough, first re-sequence a1_{1,...10}, ..., a9_{1,...10} then if still not good enough, revisit/negotiate g1..g9

---

people to reach out:

• "Sandy": Isabella loads (working with Roberto at Sloan), who just graduated with me on topic related to yours - [email protected]

SMART IC
NRP Director of Ecosystem
EDB
A*star is innovation agency of gov.
Which factors affect the performance of technology business incubators in China? An entrepreneurial ecosystem perspective

[hyunjimoon]

clockspeed in social prior engineering applied to M3S T4 pj: https://chat.openai.com/share/2da5a2e5-e6e6-43ff-8ef7-e7c14d07d83d

agents

angie, charlie, jinhua + gpt (crowd & common sense)

sequence

Syncing Angie, Charlie, Jinhua's vision is crucial for Angie's phd journey which is nailed on Charlie as advisor and M3S as project.
Instead of solving three bodies (sync Angie, Charlie, Jinhua) problem, I'd like to solve two stages of two bodies problem: first sync with Charlie then sync [Charlie-Angie] with Jinhua. Geometric proximity between Charlie and Angie increases information diffusion which justifies optimality of this sync sequence.

sequence

I have evidence that this evening is Jinhua's planning (simulating 1 year operation given members and 5 year goal) time, so I'd like to provide some support (hopefully) by sharing the updated version of my gpt experiment.

operation

• [8 individuals in T4 team < T4567 team < M3S team (T1-9) < Singapore] hierarchy/value chain is a rich environment to develop Angie-Charlie's tool for entrepreneurial operation
• acculturating/processifying the use of tool is important before the organization becomes cemented

[hyunjimoon]

using elicitation of member's state from M3S all hands meeting as prior, I'm shaping its "update process" to reach the vision (posterior).

prior

of organization, technology, strategy, operations

posterior

angie's desire is to

• collaborate with Tom (T6) on new form of human capital: working with ai + imagining new things AI can do (more relation entrep.); his current research is on "design human-ai (bridge), taxanomy of different tasks (~ how leneis cameup with different living things (reptiles, )), powerful automatic inference (specifying what methods are possible), estimate performance"
• prepare for 1/8-12 workshop (public facing event) to present first prototype of O4E
• meeting with Tom would be to elicit T6's vision, strategy, operation and to pool structure

I feel overlap in largely two:

1. interest for taxonomy
   Tom’s Cognitive, physical activities ~ bit, atom automation

2. pairing/matching
   pairing (need, solution)
   =pairing (pre-processed data, research talent)
   =pairing (idea, demand)= (product, market)

[hyunjimoon]

title: case study of entrepreneur angie to develop spacio-temporal agent-system

below is entrepreneurial journey of an agent called angie. it's goal is to develop tools for spacio-temporal agent-system. In today's T45 meeting, I wish to elicit T4,5 teams' input by sharing angie's vision to ignite collaboration on:

1. case study of M3S as entrepreneurs (@jinhua-mit)
2. theory of learning and improving human decision (Shenhao)
3. prompt targeted to entrepreneurs (will be used in singapore)

1. how it started

In 2017, I made demand forecasting and inventory optimization app for small bakeries but couldn't solve how to learn their preference for sell-out (rather than leftover) despite the fact that leftover can give higher profit. I decided to professionalize myself on learning and improving human decision with tools/data analytics.

2. how I professionalized fast using bit inflow to atom infrastructure

Over the half-decade, I placed myself near atom_supply (i.e. pallet in logistics, open-source library,) where learning was the fastest due to its fast clockspeed. After three eras for agent (2018-2022), system (2022-2023), agent-system (2023-), I identified three components on mind, hand, eye:

2.1 agent era (2018-2022)

agent.in: stan team (pi: andrew gelman)'s vision for automated scientific workflow (updated in 2023.Oct) as stan developer

We propose to develop a new scientific workflow, including theory, methods, and practical tools that will guide researchers and data analysts from the design of experiments through iterative model building and posterior inference to informed decision making using state-of-the-art Bayesian statistical methods. The project will transform existing and often vague recommendations into theoretically justified and practical workflow steps that are clear, explicit, and have fast computational implementations. The workflow steps will be combined with recommendations on what to do next based on efficient computational diagnostics that can identify potential problems and point to further diagnostics or sensible next steps. In all parts of the workflow, we value human experts by automating recommendations, not decisions. The workflow will be implemented as extensions to mature open-source software with an already large user base and will have large impacts already during the project. The project is ambitious for science as well as within statistics and machine learning in promising to make applied scientific research, data analysis, and building AI systems in science, industry, the public sector more reliable and reproducible, leading to improvement of our understanding of the world and better-informed decision making. Rigorizing the process of workflow will allow the recent rapid developments in statistical computing to \textbf{empower scientists} to make rapid progress on real-world problems.

agent.out: paper on why, when, how to use simulated data to align statistical model and computational tool

Q1. could we classify tasks where heterogeneous ones (value, vision) are represented as prior distribution $p(\theta)$ and homogenous ones (ethics, norms, common sense, law, price) are likelihood $p(y|\theta)$. social prior can be calibrated based on the simulated data $\tilde{y}$ from the two.

2.2 system era (2022-2023)

system.in: system dynamics group (pi: hazhir rahmandad)'s vision on system dynamics archetypes as vensim software developer

system.out(system.in, agent.out): eight SD model on diffusion, project, business cycle, market growth, service quality, material cycle, urban dynamics, world (increasing in scale) detail on parameters, time horizon, time steps, diagram

2.3 agent-system era (2023-)

agent-system.in from T45 (can i be collective intelligence developer :D)

• charlie's vision on operations for entrepreneurs
• jinhua's vision on behavior and computation e.g. transport (:= system or means of conveying people or goods from place to place by means of a vehicle) engineering esp. different clockspeed (operation control and plan based on demand prediction) and aggregation level across spatio-temporally network
• sandy's vision on social prior engineering e.g. social prior engineering

agent-system.out(system.out, agent-system.in)

2.4 close the loop to agent 2.1 and system 2.2

• agent.in(agent-system.out): two axis (atom-bit ratio, supply-demand uncertainty ratio) on which to construct joint space of three human-ai augmentation taxonomies

• system.in(agent-system.out): data on industry classification and tech mapping by e.g. nick bloom's Identifying Technology Spillovers and Product Market Rivalry.pdf

3. how I aim to collaborate

3.1 T4, 5

• @shenhao, @dingyi could this be connected to information relaxation in RL and/or lift and project as below?

Information Relaxation (Brown, Smith and Sun, 2010)
What if we relax the “information constraint”?
• a.k.a. non-anticipativity constraint: DM can only use the information revealed so far
• allow DM to access some future information
• introduce penalties for violating the constraint
The relaxed DP problem
• yields a performance bound on the maximal achievable performance
• can be easier to solve than the original DP problem more future information
The idea has been studied in different contexts:
• Rockafellar and Wets (1991) – stochastic programming
• Davis and Karatzas (1994) – optimal stopping
• Rogers (2002), Haugh and Kogan (2004), Anderson and Broadie (2004), Chen and Glasserman (2004) – option pricing
• Brown, Smith and Sun (2010) – general framework formulated
Applications/extensions:
• Desai, Farias, and Moallemi (2012) – optimal stopping
• Desai at el. (2012), Haugh and Lim (2012) – linear-qaudratic and linear-convex control
• Haugh and Wang (2014) – dynamic portfolio execution
• Kim and Lim (2016) – robust MAB
• Brown and Haugh (2017) – infinite horizon MDPs
• Haugh and Lacedelli (2019) – POMDPs

lift and project

• @jiangbo could we connect to research on revealed preference using quantumML?
• @team could you give some feedback in prompting for entrepreneurial agents (will be used in singapore to interview entrepreneurs)

1.  Name of company and year of founding
2.  Names and backgrounds of founder(s)
3.  Names and background of leadership team members 
4.  Principal product(s) and value proposition offered
5.  Principal target market and customers
	a.  How many different market segments served?
6.  Geographic footprint of physical and/or digital markets targeted
	a.  How many distribution/sales sites do you have?
7.  Sales/revenue numbers to date, by year
 a:  Total investments in your firm date
 b.  Number of employees
 c.  At current burn rate, how much time do you have with the existing level of cash in the firm?
 d.  What are the five most important metrics that the CEO tracks?
 e.  What are the five most important metrics that the COO (or equivalent) tracks?
8.  If product is solely a physical product…
	a.  Approximately how many parts and subsystems are procured from outside suppliers?
		i.  How many different suppliers do you have?
		ii.  What is the geographic footprint of your supply chain?
		ii.  What is the longest lead time part in your supply
			 chain?  How long?
	b.  Approximately how many parts  and subsystems are manufactured internally?
	c.  How many people do you have working in manufacturing and supply chain?
	d.  What investments have you made in automation of work processes?
	e. Do you have a well-identified bottleneck in your production system?What/where is it?
9.  If product is solely a digital product…
10.  If product has digital and physical components…

3.2 T6

• thomas malone's system types (hierarchy, market, democracy, ecosystem) and abdullah almachtoub (systemic elicitation for augmented task and its tool empirica)
  from tom's lecture

3.3 M3S, Bayes, MIT

• daniela rus on augmentation in hardware
• vikash manginka (computational representation)
• roberto rigobon team's EPOCHS (emotional intelligence, physical interaction, opinion and judgement, creativity and imagination, hope and vision, system thinking and sustainability) taxonomy which isabella loaiza (postdoc in roberto's lab after working with sandy) introduced
• daron acemoglu (https://shapingwork.mit.edu/about-us/)
• business school professors supporting bayes: https://github.com/Data4DM/Bayes2Business#list-of-supporters

[hyunjimoon]

charlie's two hypothesis on which is major bottleneck btw nail or scale (can't find value proposition VS figured out how to nail, trouble marshaling resource)? which can be specified after interviewing innovation center members

nail 
value propostion (do i have product, market, profit) 
- profitable we can produce

scale
found a product that works and produce
- what do we make or buy
- series of operation

• angie's goal is to do inverse-inverse reinforcement model i.e. learning entrep. utiltiy which is the function of customer's utility which are inferred along the entrep.journey
• when does entrep jump into entrep (fire, ownership): capture or get data (when they start thinking about it)
  clearly defined (when did you start) - economy developed (not innovation hub)
• start modeling this as mdp: markov process: start state (nail, scale, sail), operational activity (bit, atom, customer), duration; function of industry
• week long history, income or wealth reward for job
• treat activity change as trigger (regardless of time), anchor: whenever activity change
• dirl - backward and forward looking utility; policy (backward) - value function (forward looking)
  

[hyunjimoon]

am getting feedback on operations bottleneck from experts in Singapore:

1.  Name of company and year of founding
2.  Names and backgrounds of founder(s)
3.  Names and background of leadership team members 
4.  Principal product(s) and value proposition offered
5.  Principal target market and customers
	a.  How many different market segments served?
6.  Geographic footprint of physical and/or digital markets targeted
	a.  How many distribution/sales sites do you have?
7.  Sales/revenue numbers to date, by year
 a:  Total investments in your firm date
 b.  Number of employees
 c.  At current burn rate, how much time do you have with the existing level of cash in the firm?
 d.  What are the five most important metrics that the CEO tracks?
 e.  What are the five most important metrics that the COO (or equivalent) tracks?
8.  If product is solely a physical product…
	a.  Approximately how many parts and subsystems are procured from outside suppliers?
		i.  How many different suppliers do you have?
		ii.  What is the geographic footprint of your supply chain?
		ii.  What is the longest lead time part in your supply
			 chain?  How long?
	b.  Approximately how many parts  and subsystems are manufactured internally?
	c.  How many people do you have working in manufacturing and supply chain?
	d.  What investments have you made in automation of work processes?
	e. Do you have a well-identified bottleneck in your production system?What/where is it?
9.  If product is solely a digital product…
10.  If product has digital and physical components…

[hyunjimoon]

jinhua, angie
how to make singapore innovation hub with m3s

what conditions can nudge people to be more entrepreneurial?

• socio-economic, income + location (vc), exposure to innovation
• summarize media lab smart city course
  

[hyunjimoon]

scott stern's 15.357 class transcript

Scott's decade long research on measuring entrepreneurship (MIT REAP) taxonomy, quality, predictive analytics
part1: https://otter.ai/u/fmctbOQqXf16FsmgCFbc1K9qDdA
part2: https://otter.ai/u/dem1lxwwIG2bNQFyDwwVRs8tZog

[hyunjimoon]

scott stern's 15.357 class transcript

• innov-entrep unique: no one is in charge
• policy evaluation looks like innovation itself is dynamic

[hyunjimoon]

Jinhua, Angie

• Angie will share one page summary of literature review on decision to become entrepreneur and how to connect this with T4, 5 (ai for human capital and economic growth). keyword is quality perception.

• many research has been conducted on "decision to become entrepreneur" topic itself. what's missing is quality approach, not pure quantity. we not only need a increase of entrepreneur, but high quality entrepreneur. (perhaps under limited resource, we may even want decrease of entrep. (elite education, early people based (not project?) funding) - this depends on industry where the burden is levied on developer or not)

• prize (solution|need): if we know what need is (at least to form a question), larger N is better. we'd want to increase diversity of entering capability to prevent dual work. e.g. artist's vs scientist's process to solve vaccine problem may differ)

• patent (solution-need pair): if we don't know anything, we should test need-solution pair (assuming we can at least test). if we don't know what need is, smaller N is better.
  Q. is crowd sourcing more like prize or patent? I think prize as need is specified

• Angie conclusion is: "it requires precise evaluation function. tool to evaluate idea which entrepreneurs and external evaluators can agree upon can be helpful for idea’s right (radical innovation) and left tail (ideas that suck)"

• next step is nailing quality perception literature, we can use ai to act as a mediator, closing the quality perception gap of external and internal evaluator

[hyunjimoon]

people from Jan.10 meeting

Name	Role	Title of Talk	Time of Talk	Emoji
Jinhua Zhao	Emcee	Mens, Manus and Machina: Introduction of the M3S Program	1:30 - 1:45 PM	🎤
Daniela Rus	Moderator (work), Keynote Speaker (robotic)			🤖
Archan Misra	Keynote Speaker (learning), Moderator (learning)			🧠
Edmund Lin	Keynote Speaker	Global View of AI Policies and Strategies	1:45 - 2:20 PM	🌍
Bryan Low	Panelist			📊
Simon Chesterman	Keynote Speaker (Policies), Moderator (Policies), Panelist (learning)	Pathos and Bathos of AI	1:45 - 2:20 PM	🏛️
Robert Mahari	Panelist (Policies)	Regulation by Design: Developing AI systems to embed regulatory objectives	1:45 - 2:20 PM	📜
Tobin	Panelist (Policies)	'Community Transformers: Rebuilding Trust Through Secure AI'; Auditor for policy, Zero Knowledge Proof	1:45 - 2:20 PM	📜
Tan Kok Yam	Keynote Speaker		2:20 - 3:10 PM	🌍
Jayson Goh	Keynote Speaker		2:20 - 3:10 PM	🌍
Thomas Malone	Keynote Speaker		2:20 - 3:10 PM	🌍
Sanjay Sarma	Keynote Speaker (learning), Moderator (robotic)		3:35 - 4:25 PM	nan
Jason Jackson	Panelist (work)		2:20 - 3:10 PM	💼
Charlie Fine	Keynote Speaker (work)	AI and Entrepreneurship	2:20 - 3:10 PM	💼
Hai Wang	Panelist (work)	AI and Human collaborations.	2:20 - 3:10 PM	💼
Too Ping Tan	Panelist			📊
Mdm Zuraidah Abdullah	Keynote Speaker		3:35 - 4:25 PM	🌍
Cheong Wei Yang	Panelist		3:35 - 4:25 PM	📊
Rajesh Unni	Panelist		3:35 - 4:25 PM	📊
Phoon Kok Kwang	Panelist		3:35 - 4:25 PM	📊
Sylvia	Panelist		3:35 - 4:25 PM	📊
Jack Lim	Panelist		3:35 - 4:25 PM	📊
Subodh Mhaisalkar	Speaker	Remarks from NRF	4:25 - 4:35 PM	🔊
Daniela Rus	Keynote Speaker (robotic)		4:40 - 5:40 PM	🤖
Quek Tong Boon	Keynote Speaker	Senseable worlds	4:40 - 5:40 PM	🌍
Carlo Ratti	Keynote Speaker (robotic)		4:40 - 5:40 PM	🤖
Marcelo	Panelist (robotic)	Roboat project	4:40 - 5:40 PM	🔧
Fabio	Panelist (robotic)	Soft robotics as a new technological paradigm	4:40 - 5:40 PM	🔧
Cecilia	Panelist (robotic)		4:40 - 5:40 PM	🔧
James Fu	Panelist		4:40 - 5:40 PM	📊
Bernard	Panelist		4:40 - 5:40 PM	📊

[hyunjimoon]

chat with M3S CTO Alok Prakash on establishing system architecture e.g. defining interface between actors and evaluators (three types of interaction: 1. between suppliers (m3s) and demand/evaluators (Singapore organizations), 2. tree-branch within suppliers (T1-7), 3. tree-branch within demand/evaluators

[hyunjimoon]

to assist Charlie's panelist moderation, I synthesized Charlie's presentation with main paper for regional innovation ecosystem by Jorge Guzman, Fiona Murray, Scott Stern, Heidi Williams on Guzman23_reg💡⚙️.pdf as:

Stakeholder Type	Role	Symposium Participants	Questions to Participants
Government	Provide supportive policies, funding, and infrastructure for innovation	Feng-Ji Sim, Deputy Secretary, Smart Nation and Digital Government	How can Singapore's government agencies create policies and infrastructure to support AI-infused startups, particularly in reducing cognitive load for entrepreneurs and increasing their chances of success?
University	Conduct research, nurture talent, facilitate knowledge transfer, and provide education	Charlie Fine (panelist moderator) Simon Chesterman, Bryan Low, Archan Misra, Angie Moon	How can Singapore's universities develop AI tools and methodologies to help entrepreneurs structure complex problems like product-market fit and the "Nailing, Scaling, Sailing" stages of startup growth?
Entrepreneurs	Develop innovative solutions, seek partnerships, and contribute to economic growth	Saurav Bhattacharyya (panelist), Oliver Kwan	In the rapidly evolving AI startup landscape, what specific challenges do you face in product creation, market research, and stakeholder management, and how could AI assistants potentially help in decision-making?
Risk Capital	Provide funding, mentorship, and network access to promising startups	Chris Lee (panelist), Helen He (panelist)	Given the explosion of AI-infused startups, how can investors like SoftBank and New Wheel Capital help Singaporean entrepreneurs increase their likelihood of success, particularly in navigating the "Nailing, Scaling, Sailing" stages?
Corporate	Engage in open innovation, partner with startups, and provide industry expertise	SNG Ren Yeong (panelist), Alexander Lau (panelist) Lee Shiang Long, Man Fung SO, Liu Yanling, Sau Sheong Chang, Simon See, Bernard Siew	How can large corporations in Singapore collaborate with AI startups to help them structure complex problems and increase their chances of success in the highly competitive AI landscape?

claude process

from last winter:

[hyunjimoon]

short intro before charlie's presentation
slide: FineMoon24_edu-product_m3s.pdf

questions i received and my answer (to guide charlie's presentation)

1. what is prior, information, posterior (how does belief update happen)? (shenhao)

• prior on market acceptance, prior product feasibility given market acceptance
• information that chosen market gives on proposed product
• updated belief on market acceptance and product feasibility

2. will simulation demo be helpful for both startup and startup educators? (kakali)
   no, this simulation is for educate_product. but its output (tested hypothesis) will be input of apply_product

3. how many belief updates happen? how long is each belief update? (jianbo)

• at most S (# of market segment) * M (# of product feature) -1 belief update happens.
• it differs. time it take to do new experiment. you can be stuck with one product feature much longer than the other feature. same with market. but belief update mechanisms are all same.

4. uscase

• connect with orbit genAI tool which doesn't have operations + industry effect piece (people liked this)

[hyunjimoon]

Prep for 2024 summer symposium

Alok's Demo is interesting - correct answer but wrong reasoning

Charlie asked whether this research would have advantage over apple and google and Alok answered localizing in Singapore but admitted the inherent limitation.

I liked Shenhao's analogy of short and long activity sequence (

[hyunjimoon]

[hyunjimoon]

jb recommended proj 1 transcript "Wang and Leon discussed the integration of LLLM for industrial park planning and operation. Wang presented research outcomes on integrating LLMs into the planning and operation of industrial parks, addressing challenges such as constructing a comprehensive dataset and enhancing the LM's multi-step reasoning capability (wonder whether bayesian reasoning can be incorporated). Leon provided insights on the potential of LLMs in the industry and the need for further research to overcome emerging issues.

design and data driven hybrid community building
Project 1: Enhancing Knowledge-Driven Insights via IndustryScopeGPT
Leon Yang LIU  - Lab Manager, Associate Professor
Siqi WANG - Ph.D. candidate
 
Industrial parks are critical to urban economic growth. Yet, their development often encounters challenges stemming 
<img width="1328" alt="image" src="https://github.com/user-attachments/assets/bebaf8d7-a98b-416d-ba8b-c321cd25a616">
from imbalances between industrial requirements and urban services, underscoring the need for strategic planning and operations. This presentation introduces IndustryScopeKG built upon the IndustryScope project, a pioneering large-scale multi-modal, multi-level industrial park knowledge graph, which integrates diverse urban data including street views, corporate, socio-economic, and geospatial information, capturing the complex relationships and semantics within industrial parks. Alongside this, we present the IndustryScopeGPT framework, which leverages Large Language Models (LLMs) with Monte Carlo Tree Search to enhance tool-augmented reasoning and decision-making in Industrial Park Planning and Operation (IPPO).

Our work significantly improves site recommendation and functional planning, demonstrating the potential of combining LLMs with structured datasets to advance industrial park management. This approach sets a new benchmark for intelligent IPPO research and lays a robust foundation for advancing urban industrial development.

[hyunjimoon]

Question to Junyi - Are you aware of any literature that connects IRL with inverse planning? Thank you! https://en.wikipedia.org/wiki/Inverse_Planning

[hyunjimoon]

write mail to jinhua on education needed for decision making https://claude.ai/chat/3b021992-3b0c-4b3a-a098-8bf0e1679db7