Evan Prodromou
Hacker, speaker, entrepreneur
Democratizing AI
How do we get more people making AI?
Robot Breather, Cameron Daigle, 2006 BY-NC-ND
Etherpad
https://public.etherpad-mozilla.org/p/Democratizing_AI
(Enter "Democratizing AI" on Mozilla Etherpad)
http://x.co/demoai
Hashtag
Thesis 1
Most of the AI being made in the world in 2017 is made by and for large Internet corporations, financial institutions, and national government.
Counter-arguments
- My cousin is learning AI.
- Many students are learning AI.
- Many startups include "AI" on their pitch decks.
- Many other organizations are "thinking about" AI.
Thesis 2
More sectors of society should be making AI: individuals, ad-hoc groups, small and medium-sized businesses, non-profit organizations, municipal and state governments.
Counter arguments
- AI is too dangerous to have it be widely used.
- AI is not useful or important for other sectors of society.
- AI is too hard for other sectors to use.
Democracy
This session is about democratization, not democracy. There are probably some really interesting aspects there!
What?
Why?
How?
...is AI?
...is it unhealthy?
...do we fix it?
What is AI?
Procedural programming
EAST 1 SOUTH 1 EAST 3 SOUTH 1
Changing environment
EAST 1 SOUTH 1 EAST 3 SOUTH 1
ERROR
Complex environment
Do what I say
Do what I mean
AI with Logic
GET A BATTERY.
GO FAST.
AVOID FIRE.
AVOID MONSTERS.
AVOID CHAIRS.
AI with Learning
WALK AROUND.
IF YOU GET A BATTERY, GREAT.
IF YOU GET BURNED, BAD.
IF YOU GET BITTEN, BAD.
REPEAT X 1M.
DO WHAT WORKED BEFORE.
Hybrid AI
A little of column A, a little of column B.
What is AI good for?
- Classification
- Prediction
- Planning
- Optimization
What is AI used for?
-
Classification
Evan is a 47-year-old white Canadian male who likes minimal techno and books about Byzantine history. -
Prediction
Evan is more likely to click on ad #19 than on that ad #32. -
Planning
Let's show Evan ads #19, then ad #4, then ad #32. -
Optimization
We just made a lot of money off this guy. Let's make more!
Almost any AI fits that framework.
Natural language processing, decision making, game playing, advertising, automated trading, content optimization, map routing.
AI in 2017
It's all about machine learning.
Machine learning
AKA, neural networks
Pros
- Works great!
- Deep learning optimizes environments
Cons
- Needs lots of training data
- High resource requirements
- Very dependent on initial conditions
- Skilled operators needed
The upshot
Entities with lots of training data make AI.
They make enormous data centers to store the data and run the AI.
They hire all the skilled operators necessary to make that AI.
They generate more data that makes the AI smarter and smarter, which makes them more money.
AI = data + money.
What's wrong with AI?
Let's compare against the 5 pillars of Internet health.
https://www.internethealthreport.org/v01/
Open Innovation
Fantastic
This is an era of immense innovation in AI. Most researchers are eschewing patents to make their techniques open to all. Software developers big and small are sharing AI engines of incredible power with very liberal Open Source licenses. Compute resources are cheap and easy. Most organizations do not
share their training data.
Digital Inclusion
Bad.
Only a few very large and very powerful entities are able to deploy AI to their own advantage. POC, women, LGBTQ+ people can work for those organizations and make AI, but not for their own direct use. Diversity-oriented organizations don't have the data or the staff to make and use AI. Without strenuous work, AI models bake in historical bias to future decisions.
Decentralization
OK.
Only very large and powerful entities are able to deploy AI to their own advantage. Few network effects exist. Hiring and access to skilled operators is a real problem for new entrants.
Security and privacy
Terrible.
AI models are invisible and inauditable. AI creators are tempted to grab as much data as possible, and without watchdogs they cross the lines. It's almost impossible to tell why machine learning models make their decisions. User control of personal data that has been through a learning process is almost impossible.
Web literacy
Emergency.
There is a widening gap between individuals and smaller organizations, and the entities that make AI. Very few Web developers make using AI techniques a part of their programming craft. The Web is changing, and we are not keeping up with it.
How can we fix it?
1. Make AI a priority.
Start factoring AI techniques into all kinds of software, big or small. Anywhere the computer should just "figure it out" is a great place to use AI. Setting an example of using AI in Open Source and public software can help kick-start the process.
2. Use different AI.
Change the game. Machine learning is not the be-all end-all of AI. There is a rich and varied history of great techniques, including logic, learning and hybrid techniques. Many of them are less data- and hiring-intensive than machine learning.
3. Share data sets.
Data is the fuel for hybrid and learning techniques. Collective data sets can kick-start the process for building new AI. With opt-in from users and respect for privacy, end-user data can also be shared.
4. Teach AI.
Talk to students and Internet makers about using AI in their projects. Make sure that AI techniques are part of Web literacy curricula.
5. End-user AI making.
Software products should include ways for end users to create their own AI models, and not just be passive lab rats for someone else's AI to observe.
(Pause for breath)
Hands On
No, I'm not going to teach you AI in 60 minutes.
Challenge: AI product design
In teams of 4-5, we are going to design software products that include AI as part of their feature set.
Take notes. There's a report step at the end of this. Use Etherpad or post-its or flip-charts or whatever floats your note boat.
Step 1: Team up
- Make group(s) of 4-5
- Talk about new or existing software products you think can and should be smarter
- Examples: personal productivity, email, alarm clocks, calendars, chat, environmental monitoring, content management systems, Web publishing, blog engines, social software, cameras
- Concentrate on individuals, interest groups, small business, non-profits
- Come up with a list of candidate products
- Pick one product to work on further.
Step 2: Features
- Make a list of intelligent features of your product
- Don't worry about if they are practical or realizable
- Don't worry about making money with your imaginary software
- Concentrate on features that work for the benefit of the software owner!
- List 3-5 intelligent features of your software product.
Step 3: Approaches
- For each of the features you picked, ask:
- Does it encourage literacy and participation?
- Does it harm or help digital inclusion?
- Does it promote open innovation?
- Does it centralize or decentralize?
- How does it affect security and privacy?
- Skip any questions that aren't interesting.
- Pick 2-3 answers that were interesting.
Step 4: Sharing
- How do each of your features encourage sharing of intelligence between users?
- Are there ways that people can get collective benefit from sharing logic, data, or both?
- Pick 1-2 answers that were interesting.
Step 5: Report!
- What kind of software did your group work on.
- 3-5 intelligent features of that software.
- 2-3 answers about Internet health that were interesting.
- 1-2 answers about sharing that were interesting.
Democratizing AI How do we get more people making AI? Robot Breather , Cameron Daigle, 2006 BY-NC-ND
Democratizing AI
By Evan Prodromou
Democratizing AI
- 2,347
