The End Of Knowledge - Vault 4/53 - AI For Good - 2020

graph LR classDef main fill:#f9f9f9, font-weight:bold, font-size:14px classDef governance fill:#ffcc99, font-weight:bold, font-size:14px classDef dataTrusts fill:#ccff99, font-weight:bold, font-size:14px classDef inclusivity fill:#99ccff, font-weight:bold, font-size:14px classDef agriculture fill:#ff99cc, font-weight:bold, font-size:14px classDef policy fill:#ccccff, font-weight:bold, font-size:14px A[AI for Social
Good: What Next?] A --> B[AI needs governance
to prevent misuse. 1] A --> C[AI needs incentives
for positive use. 2] A --> D[International coordination
for AI projects. 3] A --> E[AI-driven drug discovery
faces challenges. 4] A --> F[New incentive model
for drug discovery. 5] A --> G[Countries share costs
for drug R&D. 6] A --> H[International data pipeline
for assays. 7] A --> I[Rethink advertising-funded
models. 8] A --> J[Governance needed
for AI growth. 9] A --> K[Educate citizens and
experts on AI. 10] B --> L[Data trusts for
negotiating data use. 11] B --> M[Grey rhino: predictable
high-impact events. 12] B --> N[Pandemic response: AI
analyzes wastewater. 13] B --> O[Wastewater data provides
privacy insights. 14] B --> P[Pilot in Toronto for
marginalized communities. 15] C --> Q[Empower communities with
data insights. 16] C --> R[Inclusive datasets for
AI bias testing. 17] C --> S[Missing data hinders
bias testing. 18] C --> T[Datasets build inclusivity
in AI. 19] C --> U[Agile approach for
public/private datasets. 20] D --> V[AI bias use cases
for SDGs. 21] D --> W[Farm.ai helps
smallholder farmers. 22] D --> X[Combine farmer and
market data. 23] D --> Y[Rollout AI to
more countries. 24] D --> Z[Collaborate with local
farming experts. 25] E --> AA[What-If tool for
policy impacts. 26] E --> AB[Model predicts
transmission rates. 27] E --> AC[Interface for forecasting
policy impacts. 28] E --> AD[Multidisciplinary input for
model training. 29] E --> AE[Invest in local data
science talent. 30] class A main class B,K,L,M,N,O,P governance class C,Q,R,S,T,U dataTrusts class D,V,W,X,Y,Z inclusivity class E,AA,AB,AC,AD,AE policy class F,G,H,I,J governance

Resume:

1.- AI is a powerful tool that can benefit society if used properly, but governance is needed to manage potential misuse.

2.- AI tools can be used positively for social good or negatively to concentrate power. Incentives are needed to steer AI positively.

3.- International coordination is important for AI for social good projects to connect expertise, domain knowledge, and funding from around the world.

4.- AI is transitioning drug discovery from chemist/biologist-driven to more data/AI/software-driven. This presents challenges due to incentives favoring secrecy over data sharing.

5.- A new incentive model for drug discovery R&D is proposed, with features of academic research like recognition through knowledge sharing.

6.- Countries could share costs of public domain drug discovery R&D. Resulting drugs would be free for poor countries and cheap for participating countries.

7.- An international data collection pipeline for chemical/biological assays is proposed as a benchmark for machine learning drug discovery researchers.

8.- Advertising-funded systems like social networks favor large incumbents over startups, slowing innovation. Rethinking the model, e.g. as public services, is suggested.

9.- As AI grows more powerful, governance through laws and international norms is needed to minimize negatives like paid manipulation and killer robots.

10.- Education is important so citizens can participate in collective decision making on AI. Experts also need more training on societal impact and ethics.

11.- Data trusts are proposed as powerful third parties to negotiate terms of data use between users and companies on behalf of users.

12.- The "grey rhino" concept is used for predictable high-impact events like pandemics. Proper monitoring infrastructure can enable more effective response.

13.- A two-part solution for pandemic response in marginalized communities: 1) AI analysis of wastewater data, 2) public dashboards communicating insights.

14.- Wastewater data collection enables highly granular, privacy-preserving insights on virus spread compared to individual testing. Dashboards contextualize data for community engagement.

15.- Pilot proposed in Toronto community housing socioeconomically marginalized people. Dashboard co-designed with local NGO. Goal to scale to other localities globally.

16.- Objective is to empower communities with insights rather than stigmatize. NGOs play key role in appropriate data communication to community and policymakers.

17.- The Complete Picture Project aims to build inclusive, diverse datasets showing complete picture of communities for AI bias testing.

18.- Problem: People missing from data can't be tested for bias. Datasets provide developers a starting point to test for hidden bias.

19.- Datasets can help build inclusivity into AI design, development and testing. Enable assessment of pre-built algorithms when applied to new communities.

20.- Agile approach starting with public datasets, iterating with user research. Goal to negotiate safe access to private datasets for enhancement.

21.- Project seeks AI bias use cases supporting SDGs, exchange of knowledge and tools. Sponsors and partners in academia and affected communities needed.

22.- Farm.ai uses smartphone-based insights to help smallholder farmers in developing countries, who produce 70% of world's food, increase productivity.

23.- Farmer data combined with satellite/market info generates predictions on yields, credit risk, and advice. Partnering with local NGOs for deployment.

24.- Rolling out to additional countries, seeking local partners for reach and global partners in tech/retail to translate gains to commercial markets.

25.- Collaborating with local farming experts is key. Recommendations can be wrong but automation enables reaching vastly more farmers vs in-person visits.

26.- What-If tool uses AI to predict how a locality's pandemic metrics react to policy changes, to enable data-driven, transparent policymaking.

27.- Deep learning model trained on 93 countries' data predicts transmission rates from policy timing/stringency and country-specific variables. Achieved high accuracy.

28.- Planned web interface lets users forecast impact of hypothetical policies, optimize for goals. Aims to build public trust through transparency.

29.- Seeks multidisciplinary expert input to train/validate models. Partnerships with policymakers and granular data sources needed, especially in neglected geographies.

30.- Using COVID as opportunity to fill data gaps in underrepresented areas. Investing in local data science talent for sustainable, representative AI expertise.

Knowledge Vault built byDavid Vivancos 2024