State of the Art and Risks
Happy Anniversary Skynet!
Brought online Aug 4, 1997… became self aware 2:14a.m. EDT August 29th 1997
Unfortunately this brought USA and USSR to hot nuclear conflict shortly after panicked engineers tried to disconnect it.
Stanford “One Hundred Year Study on AI”
Substantial increases in the future uses of AI applications, including more self-driving cars, healthcare diagnostics and targeted treatment, and physical assistance for elder care can be expected.
Also
Society is now at a crutial juncture in determining how to deploy AI-based technologies in ways that promote rather than hinder democratic values such as freedom, equality, and transparency
Quick Facts
Publications: AI papers increased 20-fold between 2010 and 2019 to about 20,000 a year. The most popular category was machine learning. (Machine learning papers in arXiv.org doubled every year from 2009 to 2017.) Computer vision and natural language processing were the next most popular.
Sentiment: About 70% of news articles on AI are neutral, but articles with positive tone increased from 12% in 2016 to 30% in 2018. The most common issues are ethical: data privacy and algorithm bias.
Students: Course enrollment increased 5-fold in the U.S. and 16-fold internationally from a 2010 baseline. AI is the most popular specialization in Computer Science.
Diversity: AI Professors worldwide are about 80% male, 20% female. Similar numbers hold for Ph.D. students and industry hires.
Conferences: Attendance at NeurIPS increased 800% since 2012 to 13,500 attendees. Other conferences are seeing annual growth of about 30%.
Industry: AI startups in the U.S. increased 20-fold to over 800.
Internationalization: China publishes more papers per year than the U.S. and about as many as all of Europe. However, in citation-weighted impact, U.S. authors are 50% ahead of Chinese authors. Singapore, Brazil, Australia, Canada, and India are the fastest growing countries in terms of the number of AI hires.
Vision: Error rates for object detection (as achieved in LSVRC, the Large-Scale Visual Recognition Challenge) improved from 28% in 2010 to 2% in 2017, exceeding human performance. Accuracy on open-ended visual question answering (VQA) improved from 55% to 68% since 2015, but lags behind human performance at 83%.
Speed: Training time for the image recognition task dropped by a factor of 100 in just the past two years. The amount of computing power used in top AI applications is doubling every 3.4 months.
Language: Accuracy on question answering, as measured by F1 score on the Stanford Question Answering Dataset (SQuAD), increased from 60 to 95 from 2015 to 2019; on the SQuAD 2 variant, progress was faster, going from 62 to 90 in just one year. Both scores exceed human-level performance.
Human benchmarks: By 2019, AI systems had reportedly met or exceeded human-level performance in chess, Go, poker, Pac-Man, Jeopardy!, ImageNet object detection, speech recognition in a limited domain, Chinese-to-English translation in a restricted domain, Quake III, Dota 2, StarCraft II, various Atari games, skin cancer detection, prostate cancer detection, protein folding, and diabetic retinopathy diagnosis.
What about the future?
When will AI be just as competent as humans?
Ford (2018 )“Architects of Intelligence”: [2029–2200] mean == 2099
Grace, et.al. (2017)“When will AI exceed human performance?”: 50% by 2066, 10% by 2025, some “never”.
(arXiv:1705.08807)
Your guess is as good as mine
Where do we go from here?
Machine Intelligence –>
Expert Knowledge Encoded –>
Probabilistic Models –>
Machine Learning –>
New Things!
What is possible today (rapidfire examples)
Robotic cars date back to the 1920s
“Autonomous” cars were achieved in 1980s
DARPA funded projects through the 2010s
Waymo had cars which drove a cumulative 10M miles on public roads, no serious accidents, human intervention needed every 6,000 miles.
2016: Rwanda blood drones
Legged Locomotion
BigDog:
BigDog
Autonomous Planning
NASA’s “Remote Agent” was managing space flight as early as 2000
Deep Space 1
During Persian Gulf (1991), DART did logistics/transportation planning for 50,000 vehicles, cargo, people, accounting for start/end points, routes, transport/destination capacity, conflict resolution, etc…
DARPA claims that DART alone paid back their investment in AI for the past 30 years…
Machine Translation
Google Translate, and similar render billions of words for millions of people each day.
“Good enough”
For similar languages with lots of training data, quality is near-human
Speech Recognition
Alexa, Siri, Cortana, Skype, Google…
Google Duplex can make resturant reservations
Recommendations
Amazon, Facebook, Netflix, Spotify, YouTube, Walmart…
Spam filtering (99.9%)
Game Playing
AI is of course common in video games
Deep Blue vs Garry Kasparov (1997)
Piet Hut: “100 years or more until computers beat humans at go”
ALPHAGO beat all human players 20 years later (2017)
ALPHAZERO (2018) had no human input, learned from self-play and still beat all human opponents at Go, chess, shogi.
Risks and Benefits
Francis Bacon (1609): “mechanical arts are of ambiguous use, serving as well for hurt as for remedy”
Everything we have is because of human intelligence
How much more could we gain if we extend our reach with “superior” machine intelligence?
Demis Hassabis (CEO Google DeepMind): “First solve AI, then use AI to solve everything else”
What could possible go wrong?
Lethal autonomous weapons
- UN pre-treaty stage
Surveillance and Persuasion
- phone lines, video cameras, emails, messages, social media…
Biased decision making
- parole, loans, job applications, by what metrics?
Impact on employment
Why should I pay you when a machine can do your job better?
Machines make you more productive
Make economically impractical things viable
Safety-critical Applications
Cars, water supplies, space flight, hospitals
Ethical standards and technical standards
Cybersecurity
Personalized blackmail / phishing attacks
Potency and proliferation of malware and how we survive them. (Crowdstrike)
We’re at a guiding stage of humanity
Will AI eventually be comparable to human intellgence?
What then? And what if not?
AI as it has evolved
Much of history is just getting the thing to work
As sub-fields evolved, they tended to go off, following their own ends.
In the 2000s many OG AI researchers suggested that AI “should return home” … Human-Level AI (HLAI)
Artificial General Intelligence (AGI) also sometimes (GAI)
This also raises the question of Artificial Superintelligence (ASI)
It seems probable that once the machine thinking method had started, it would not take long to outstrip our feeble powers. … At some stage therefore we should have to expect the machine to take control, in teh way that is mentioned in Samuel Butler’s Erewhon.
–Turing 1951
ASI isn’t the monolith from 2001… it’s our hand that makes it… and our responsibility.
If we use, to achieve our purposes, a mechanical agency with whose operation we cannot interfere effectively … we had better be quite sure that the purpose put into the machine is the purpose which we really desire.
So… what should they be doing for us?