10. The State of the Art (as of AIMA 4th Ed, 2020)
Source: AIMA 4th Ed, §1.4
Overview
What can AI do today? Not magic — but science, engineering, and mathematics applied at scale. By 2019, AI systems had met or exceeded human-level performance in: chess, Go, poker, Pac-Man, Jeopardy!, ImageNet object detection, speech recognition (limited domain), machine translation (restricted domain), various Atari/Dota/StarCraft games, skin cancer detection, prostate cancer detection, protein folding, diabetic retinopathy diagnosis.
Domain-by-Domain Summary
Robotic Vehicles
- 2005 DARPA Grand Challenge: 132-mile autonomous off-road drive (Stanley, Stanford).
- 2007 Urban Challenge: autonomous driving in traffic.
- 2018: Waymo crossed 10 million miles on public roads; launched commercial robotaxi service.
- Drones: autonomous cross-country blood delivery in Rwanda since 2016. Quadcopters with 3D mapping, self-assembly into formations.
- Legged robots: Boston Dynamics’ BigDog, Atlas — walks on uneven terrain, does backflips.
Autonomous Planning and Scheduling
- NASA Remote Agent (1999): First autonomous on-board planning program to control a spacecraft.
- EUROPA toolkit: Used for daily operations of NASA Mars rovers.
- SEXTANT: Autonomous navigation beyond GPS, in deep space.
- DART (1991, Persian Gulf): Automated logistics planning for 50,000 vehicles/people. DARPA said this single application paid back its entire 30-year AI investment.
Machine Translation
- Online systems translate documents in 100+ languages covering 99%+ of humans.
- Hundreds of billions of words per day.
- For closely related language pairs with abundant training data, near-human quality in narrow domains.
Speech Recognition
- 2017 Microsoft: 5.1% word error rate on Switchboard (telephone transcription task) = human performance.
- ~1/3 of computer interaction worldwide is now by voice.
- Alexa, Siri, Cortana, Google Assistant: answer questions, carry out tasks.
- Google Duplex: makes restaurant reservations by voice, indistinguishable from humans.
Recommendations
- Amazon, Netflix, Spotify, YouTube, Facebook use deep learning–based recommender systems.
- Spam filtering: AI filters >99.9% of spam.
- Social media feed personalization.
Game Playing
| Game | Achievement |
|---|---|
| Chess | Deep Blue defeated Kasparov (1997) |
| Go | AlphaGo defeated world champion Ke Jie (2017); AlphaZero learned only from self-play |
| Poker | Heads-up no-limit Texas Hold’em: AI surpassed top humans |
| Jeopardy! | Watson defeated Jennings and Rutter (2011) |
| Dota 2 | OpenAI Five defeated world champions (2018) |
| StarCraft II | AlphaStar defeated top human players (2019) |
| Quake III | DeepMind agents defeated humans in capture-the-flag (2019) |
Image Understanding
- Object recognition: exceeded human accuracy on ImageNet.
- Image captioning: “A person riding a motorcycle on a dirt road.”
- Still imperfect: a “refrigerator filled with food” turns out to be a no-parking sign with stickers.
Medicine
- AI equals or exceeds expert doctors for image-based diagnostics:
- Alzheimer’s disease, metastatic cancer, ophthalmic disease, skin diseases
- A 2019 meta-analysis found AI performance equivalent to healthcare professionals on average.
- LYNA system: 99.6% accuracy diagnosing metastatic breast cancer — better than unaided humans; human + AI combination does best.
Climate Science
- 2018 Gordon Bell Prize: deep learning model discovers extreme weather events from climate data using GPU supercomputer at exaop scale (10^18 operations/sec).
When Will AI Achieve Human-Level General Intelligence?
Experts surveyed (Ford 2018, Grace et al. 2017) gave a wide range of estimates: - Mean estimate: ~2099 - 50% of respondents: by 2066 - Some: as early as 2025; some: “never”
Key caveat: “Experts are no better than amateurs at predicting world events” (Tetlock 2017).
The field’s self-narrative has shifted over time: 1. Intelligence by machine is possible (1950s) 2. Encode expert knowledge in logic (1970s) 3. Probabilistic models will be the main tool (1990s) 4. Machine learning from data, possibly without any understood theory (2010s)
What comes next is unknown.