What's holding back Artificial Intelligence? While researchers rightly focus on better algorithms, there are a lot more things to be done. We'll discuss why AI needs a new, interdisciplinary approach, how it will be used, and what we've learned from our recent State of AI industry survey.

1. The State of AI
2016
Ines Montani
Matthew Honnibal
Explosion AI

2. custom algorithms, applications and data assets
makers of spaCy, the fastest-growing open-
source NLP library
our backgrounds are computer science,
linguistics and front-end development
Explosion AI is a digital studio
specialising in Artificial Intelligence
and Natural Language Processing.

3. We are publishing...
Embed, encode, attend, predict
The new deep learning formula for
state-of-the-art NLP models
The State of AI
An open-data industry survey on the
state of Artificial Intelligence in 2016
ARTICLES
PROJECTS
spaCy
Industrial-Strength Natural Language
Processing in Python
OPEN SOURCE

4. 434 practitioners told us
about the state of AI.
industry survey launched on October 10, 2016
collaboration with Nathan Benaich (Playfair Capital)
results will be 100% open-source
you can still take part: thestateofai.com

5. The AI hype: about
90% justified.

6. AI systems are already
going into production.
Where is your company
in the AI adoption cycle?
The State of AI
1 %
19 %
28 %
21 %
31 %
We have a profitable
system in production.
We're currently
building our first
system.
We've rolled out
our system, but it's
not profitable yet.
We're AI-curious
and experimenting.
Other

7. Young companies already
have profitable systems.
Early-Stage Startup
Venture-Backed Startup
Established Company 38 %
16 %
54 %
19 %
36 %
21 %
42 %
48 %
25 %
Rolled out and profitable
Rolled out but not profitable yet
Currently building first system
The State of AI

8. NLP is improving faster
than ever.
IBM Watson
75 %
100 %
2004 2013 2015
APR
2016
OCT
2016
84.8%
93.7%
93.1%92%
87.3%
Example: Speech recognition accuracy

9. Researchers deliver
blueprints, not products.
neural networks really took off when ImageNet
data set was created
research runs on benchmark tasks
benchmarks demonstrate solution in principle
solutions in practice require different data

10. © Stanford Dogs Dataset
Chesapeake Bay RetrieverCurly-Coated Retriever
ImageNet: mostly dogs

11. © IBTimes / Google Deep Dream

12. The bottleneck in AI
is data, not algorithms.

13. Data problems are still the
biggest problems.
high accuracy problems
high data quality problems
high data quantity problems
20 % 40 % 60 % 80 % 100 %
The State of AI
How much do the following problems keep you up at night? (4/5 and 5/5)

14. © GATE
Also the state of AI 2016...

15. You can’t get quality data
by boring the shit out of
underpaid people.
machine learners need machine teachers
Amazon Mechanical Turk is the most popular way
to create annotated data
bad incentives: below minimum-wage payment,
outdated tools, zero investment, zero satisfaction

16. We don’t just
need innovation
on the back-end.

17. It’s time to apply what we
know about UX to AI.
the most valuable knowledge is in decisions
humans don't have to think about
intuitive interfaces, fast decisions
gamification: take inspiration from what humans
enjoy doing (and even pay for!)

18. © 94%, SCIMOB
Don’t assume your task can’t be fun.

19. AI isn’t a closed shop –
there’s a lot to contribute.
7 %
14 %
18 %
22 %
39 %
< 1 year
3 - 5 years
> 7 years
5 - 7 years
1 - 3 years
The State of AI
How long have you
been working with AI?

20. Thanks!
📊 Take part in the survey
thestateofai.com
📲 Follow us on Twitter
@explosion_ai
@_inesmontani
@honnibal
💥 Explosion AI
explosion.ai