Mytholo logy of AI Chry rysta tal Ball lls, Genie ies and Deit - - PowerPoint PPT Presentation

B e t t e r F u t u r e s , T o g e t h e r

Victor Alexiev

+65 9815 1543 victor@innovator.sg

Mytholo logy of AI

Chry rysta tal Ball lls, Genie ies and Deit itie ies

Black Swans and Black Elephants

iRAHSs, 18 July 2017

What is is AI I Anyway?

The science of Intelligent Programs

Intelligent = improves its performance (accuracy or “cost”) over time

Why AI? I? Why Today?

Critical uncertainty: something that we know will be a game changer, but we are uncertain when, if, or how it will play out

Wave 5 Kondratiev Wave 1

Steam Engin gine; Cotton Ra Railw lways; Steel Ele lectrical engi gineering; Chemistry Pe Petrochemicals; Au Automobile les Info formation Te Technology; Con

• nnectivity

1 8 0 0 1 8 5 0 1 9 0 0 1 9 5 0 1 9 9 0

Wave 2 Wave 3 Wave 4

P R D E

P: P: Pr Pros

• sperit

ity R: Recessio ion D: Depressio ion E: E: Exp Expansio ion

Technology and Mega-Cycle les

A new technology, powered by a widely available resource becomes increasingly important across sectors. It turns into an ultimate driver of productivity, requires the build-up of new infrastructure and new forms of

• rganizing labour and capital... New skills and forms of education

A.I.

Predictions

…

Data ta + Computi ting po power +

+ Communication inf

infrastr tructure = Pre Predictions are

re be becoming ch cheap and being embedded into everything Pre redic iction-empowered technology (smart) will gradually displace judgement and

codified expertise as its costs continue decreasing, consuming knowledge-based tasks

Why it’s diffe iffere rent th this is time

International Community Open Source Technologies Data Everywhere Smart Hardware Intelligent Software Cheap Computing Power

B e t t e r F u t u r e s , T o g e t h e r

Victor Alexiev

+65 9815 1543 victor@innovator.sg

Curr rrent Sta tate

Black Swans and Black Elephants

iRAHSs, 18 July 2017

50 years rs of f pro rogre ress

“

• the generic visual object recognition capabilities of a two year old child
• the manual dexterity of a six year old child
• the social interaction and language capabilities of a ten year old child

“ Rod Brooks-MIT, 2016

Recognizing Key Entities Fro rom an Im Image

Approximating human performance in: Understanding English Sentences Playing rule-based games Writing computer programs from spec Transcribing voice Translating text

Recognizing Key Entities Fro rom an Im Image

Playing Games … e.g. – Chess … or “Go”

Wri riti ting Computer r Pro rograms fr from Specification

…

Lim imitations – Exp xplo loitable le Outc tcomes

Darpa experiment in purposefully misleading an image recognition algorithm

Lim imitations – Manipula lating Learn rning

Learning quality depends on learning samples and quality of interactions Learning AI can be subject to manipulation and distortion

http://www.darpa.mil/attachments/AIFull.pdf

Microsoft experiment with Generative AI twitter-bot went awfully wrong

What AI is is stru truggli ling wit ith.. ...

Handle small changes of input (e.g. invert colours) Extract or derive causal structures (esp. about unknown entities) Transfer learning from one domain/context to another Reason ethically ...and many more

B e t t e r F u t u r e s , T o g e t h e r

Victor Alexiev

+65 9815 1543 victor@innovator.sg

A Look In Into to th the Future

Black Swans and Black Elephants

iRAHSs, 18 July 2017

“The Future is is a Foreign Country. They do things differently there!”

What did id they get rig ight?

And more than anything...

Taking this is to extremes...

∞ / 0

Expla laining The Inexpli licable...

Crystal Balls Genies Deities

Scrying Evoking Worshiping

Enchanted / Intelligent Devices (mirrors, orbs, cups... phones) All knowing, all seeing Peek into the future, albeit open to interpretation Answer questions / Advise

What? When? Where re? Who? How? Why? Descri ribe Defi fine Can I? I? Should ld I? I? What if if I? I?

Mirr irror r Mirr irror on the wall, who’s the fairest of them all?

Opport rtunitie ies

Dec ecisio ion Qualit ity: Better decisions with perfect information Sim Simulated Rea Realit itie ies: Instant feedback of the consequences of our actions Co Communic ication: Ultimate communication device Tra ranscendence: Us as data (AI copies of real people)

Thre reats

Open to to interpretatio ion: Perfect information ≠ perfect understanding Over-relia iance: How many of you follow GPS blindly? Arr Arrogance: Give us the confidence to do the wrong things at scale Pri rivacy? Truth?: Forget about it

Ove ver-re reli liance

Tra ranscendence?

Fast Company, April 2017

Dri rive vers rs

Usage : the more you use, the more data you provide Int Integration : More integrated data sets and applications = more utility Pro rocessing Pow

• wer : ETL and Prediction models at scale

Talent : Access to talent to tie all this up

PA from another dimension (jinn, demon, spirits) Grant wishes/execute tasks at zero cost (obligatory and frictionless) Bound to an device/location (constrained autonomy) Sworn to serve their master (authentication)

FA FACTO CTORY RY WORKER RKER

POSTMAN TMAN

Shamelessly borrowed from my friend, Volker Hirsch : http://bit.ly/2nEFrhn

Shamelessly borrowed from my friend, Volker Hirsch : http://bit.ly/2nEFrhn

But not only “physical” tasks:

Robo-advisory is on the rise

• Lawyers
• Accountants
• Fund managers
• HR assistants

Opport rtunitie ies

Pro roductivit ity: We can get a lot more done Cos Cost of

• f Lea

Learnin ing: once one “genie” learns a trick, all genies know it Ac Access: Lowering the cost of services opens the market Hu Human Sa Safety: We don’t need to put humans at risk

Thre reats

Ag Agency: Who is responsible? Se Securit ity: What if a(all) genie(s) get(s) stolen/kidnapped/hacked? Ine Inequalit ity: Genies give an unfair advantage Pur urpose: If genies do most of the work, what are we gonna do?

Dri rive vers rs

Localized Processing Power Batteries Progress in Voice and Image Recognition Progress in Intent Detection ...Teaching ethics to AI becomes really important here

Infinite transcendent beings who lord over humanity Excellence in some aspects, weakness in others Free-willed superpowers Can be “unleashed upon us” Us, mere humans, are merely pawns serving the superhuman agenda

Opport rtunitie ies

Those “gods of the future” may just be augmented, benevolent humans With superhuman strength, we can take on our worst enemies/fears Our deities can protect us from the most terrible things in the universe ...

Thre reats

Co Control : Deities act on their own will, and who knows what that would be Am Ambiguity : Supernatural behaviour can arise as an unintended consequence Co Complexit ity : An average peasant like myself will see intent in randomness Distraction : While we’re afraid of deities we are not addressing real problems

Dri rive vers rs

Our imagination... ... Mixed with emergent behaviour of complex dynamic systems Insufficient education in systems thinking and complexity theory Proactive regulation Bad maths

B e t t e r F u t u r e s , T o g e t h e r

Victor Alexiev

+65 9815 1543 victor@innovator.sg

Back to to Reali lity

Black Swans and Black Elephants

iRAHSs, 18 July 2017

Will AI overlords take over the world? What do they want from us? I’ve watched a lot of SciFi movies — is now a good time to panic?

AI/SW does not want or feel anything… yet Jobs are rationalized by businesses striving to attain profit, scale and competitiveness You risk being innovated out of your job!

60k x 5$/d > Technology

Key Worries

Safety and Securi rity

Priv rivacy and nd Truth: We have too much data laying around Ce Centraliz ization: Centralized systems create single points of vulnerability Ru Rushing-in in: AI Arms race can lead to unintended consequences

Exp xpla lanatio ion and In Inte terp rpre retabili lity

Opacity: Most advanced models are difficult to backtrack Ma Manip ipulatio ion: Statistical models are subjected to adversarial manipulations Ed Education: Complexity is exponential (model audit is difficult)

Fair irness and De-biasing

Bia Bias: Embedding and scaling assumptions Arr Arrogance: Bad data = bad decisions Li Linear thin thinkin ing: Bad decisions + good outcomes = bad data

…A little bit of Game Theory as a Warning

If there is a significant incentive in cheating – someone will cheat. Embedding Fairness and Ethics in AI, if not effectively enforced – will not suffice But how do you enforce something like that?

Governance and Regula latory st stru ructures must be pro proactive in inste tead of f pla playing catc tch-up up

Smart Systems interact for us, not with us… most of us won’t even see AI coming

B e t t e r F u t u r e s , T o g e t h e r

Jon Hoel

+65 9111 7409 jon@innovator.sg

Victor Alexiev

+65 9815 1543 victor@innovator.sg

Thank You!

Key Takeaways

• We are facing the upswing of a mega-cycle
• It is all about ubiquitous prediction power embedded into every aspect of life
• AI is about sensor inputs, tough math and edge-case handling
• AI is not the first superhuman system created
• Black Swans: AI Arms Race, AI winter 2.0, Unintended Consequences
• Black Elephants: Model Bias, Need for Regulation, AI Terrorism

Who am I I (V (Vic ictor) r)

• Currently Co-Managing Director at Innovator SG, Partner at

Hacker.Works and AI Works

• Focused on Au

Automated Decision Sup upport and Innovation Management (ex Head of R&D at Nova, Nugit, Newstag)

• MSc Decision Sciences (LSE), MSc System Dynamics

Modelling (UiB), MSc Model Based Policy Analysis (Radboud), Bsc International Business (CBS), Bsc Economics (SU), Mathematics

• Leading the Singapore.City.AI community