<s>
Google	B-Application
DeepMind	I-Application
,	O
formerly	O
DeepMind	B-Application
Technologies	I-Application
,	O
is	O
an	O
artificial	B-Application
intelligence	I-Application
research	I-Application
laboratory	O
which	O
serves	O
as	O
a	O
subsidiary	O
of	O
Google	B-Application
.	I-Application
</s>
<s>
It	O
was	O
originally	O
founded	O
in	O
2010	O
in	O
the	O
United	O
Kingdom	O
before	O
being	O
acquired	O
by	O
Google	B-Application
in	O
2014	O
,	O
becoming	O
a	O
wholly	O
owned	O
subsidiary	O
of	O
Google	B-Application
parent	O
company	O
Alphabet	O
Inc	O
.	O
after	O
Google	B-Application
's	I-Application
corporate	O
restructuring	O
in	O
2015	O
.	O
</s>
<s>
Google	B-Application
DeepMind	I-Application
has	O
created	O
a	O
neural	B-Architecture
network	I-Architecture
that	O
learns	O
how	O
to	O
play	O
video	O
games	O
in	O
a	O
fashion	O
similar	O
to	O
that	O
of	O
humans	O
,	O
as	O
well	O
as	O
a	O
Neural	B-Algorithm
Turing	I-Algorithm
machine	I-Algorithm
,	O
or	O
a	O
neural	B-Architecture
network	I-Architecture
that	O
may	O
be	O
able	O
to	O
access	O
an	O
external	O
memory	O
like	O
a	O
conventional	O
Turing	B-Architecture
machine	I-Architecture
,	O
resulting	O
in	O
a	O
computer	O
that	O
mimics	O
the	O
short-term	O
memory	O
of	O
the	O
human	O
brain	O
.	O
</s>
<s>
DeepMind	B-Application
made	O
headlines	O
in	O
2016	O
after	O
its	O
AlphaGo	B-Application
program	O
beat	O
a	O
human	O
professional	O
Go	O
player	O
Lee	O
Sedol	O
,	O
a	O
world	O
champion	O
,	O
in	O
a	B-Application
five-game	I-Application
match	I-Application
,	O
which	O
was	O
the	O
subject	O
of	O
a	O
documentary	O
film	O
.	O
</s>
<s>
A	O
more	O
general	O
program	O
,	O
AlphaZero	B-Application
,	O
beat	O
the	O
most	O
powerful	O
programs	O
playing	O
go	O
,	O
chess	B-Application
and	O
shogi	O
(	O
Japanese	O
chess	B-Application
)	O
after	O
a	O
few	O
days	O
of	O
play	O
against	O
itself	O
using	O
reinforcement	O
learning	O
.	O
</s>
<s>
In	O
2020	O
,	O
DeepMind	B-Application
made	O
significant	O
advances	O
in	O
the	O
problem	O
of	O
protein	O
folding	O
with	O
AlphaFold	B-Application
.	O
</s>
<s>
In	O
July	O
2022	O
,	O
it	O
was	O
announced	O
that	O
over	O
200	O
million	O
predicted	O
protein	O
structures	O
,	O
representing	O
virtually	O
all	O
known	O
proteins	O
,	O
would	O
be	O
released	O
on	O
the	O
AlphaFold	B-Application
database	O
.	O
</s>
<s>
DeepMind	B-Application
posted	O
a	O
blog	O
post	O
on	O
28	O
April	O
2022	O
on	O
a	O
single	O
visual	O
language	O
model	O
(	O
VLM	O
)	O
named	O
Flamingo	O
that	O
can	O
accurately	O
describe	O
a	O
picture	O
of	O
something	O
with	O
just	O
a	O
few	O
training	O
images	O
.	O
</s>
<s>
In	O
July	O
2022	O
,	O
DeepMind	B-Application
announced	O
the	O
development	O
of	O
DeepNash	O
,	O
a	O
model-free	O
multi-agent	B-Algorithm
reinforcement	I-Algorithm
learning	I-Algorithm
system	O
capable	O
of	O
playing	O
the	O
board	O
game	O
Stratego	B-Application
at	O
the	O
level	O
of	O
a	O
human	O
expert	O
.	O
</s>
<s>
The	O
company	O
merged	O
with	O
Google	B-Application
AI	I-Application
's	O
Google	B-Application
Brain	I-Application
division	O
to	O
become	O
Google	B-Application
DeepMind	I-Application
in	O
April	O
2023	O
.	O
</s>
<s>
Demis	O
Hassabis	O
has	O
said	O
that	O
the	O
start-up	O
began	O
working	O
on	O
artificial	B-Application
intelligence	I-Application
technology	O
by	O
teaching	O
it	O
how	O
to	O
play	O
old	O
games	O
from	O
the	O
seventies	O
and	O
eighties	O
,	O
which	O
are	O
relatively	O
primitive	O
compared	O
to	O
the	O
ones	O
that	O
are	O
available	O
today	O
.	O
</s>
<s>
Some	O
of	O
those	O
games	O
included	O
Breakout	B-Application
,	O
Pong	B-Application
and	O
Space	B-Application
Invaders	I-Application
.	O
</s>
<s>
AI	B-Application
was	O
introduced	O
to	O
one	O
game	O
at	O
a	O
time	O
,	O
without	O
any	O
prior	O
knowledge	O
of	O
its	O
rules	O
.	O
</s>
<s>
After	O
spending	O
some	O
time	O
on	O
learning	O
the	O
game	O
,	O
AI	B-Application
would	O
eventually	O
become	O
an	O
expert	O
in	O
it	O
.	O
</s>
<s>
“	O
The	O
cognitive	O
processes	O
which	O
the	O
AI	B-Application
goes	O
through	O
are	O
said	O
to	O
be	O
very	O
like	O
those	O
of	O
a	O
human	O
who	O
had	O
never	O
seen	O
the	O
game	O
would	O
use	O
to	O
understand	O
and	O
attempt	O
to	O
master	O
it.	O
”	O
The	O
goal	O
of	O
the	O
founders	O
is	O
to	O
create	O
a	O
general-purpose	O
AI	B-Application
that	O
can	O
be	O
useful	O
and	O
effective	O
for	O
almost	O
anything	O
.	O
</s>
<s>
On	O
26	O
January	O
2014	O
,	O
Google	B-Application
announced	O
the	O
company	O
had	O
acquired	O
DeepMind	B-Application
for	O
$500	O
million	O
,	O
and	O
that	O
it	O
had	O
agreed	O
to	O
take	O
over	O
DeepMind	B-Application
Technologies	I-Application
.	O
</s>
<s>
The	O
sale	O
to	O
Google	B-Application
took	O
place	O
after	O
Facebook	B-Application
reportedly	O
ended	O
negotiations	O
with	O
DeepMind	B-Application
Technologies	I-Application
in	O
2013	O
.	O
</s>
<s>
The	O
company	O
was	O
afterwards	O
renamed	O
Google	B-Application
DeepMind	I-Application
and	O
kept	O
that	O
name	O
for	O
about	O
two	O
years	O
.	O
</s>
<s>
In	O
2014	O
,	O
DeepMind	B-Application
received	O
the	O
"	O
Company	O
of	O
the	O
Year	O
"	O
award	O
from	O
Cambridge	O
Computer	O
Laboratory	O
.	O
</s>
<s>
In	O
September	O
2015	O
,	O
DeepMind	B-Application
and	O
the	O
Royal	O
Free	O
NHS	O
Trust	O
signed	O
their	O
initial	O
Information	O
Sharing	O
Agreement	O
(	O
ISA	O
)	O
to	O
co-develop	O
a	O
clinical	O
task	O
management	O
app	O
,	O
Streams	O
.	O
</s>
<s>
After	O
Google	B-Application
's	I-Application
acquisition	O
the	O
company	O
established	O
an	O
artificial	B-Application
intelligence	I-Application
ethics	O
board	O
.	O
</s>
<s>
The	O
ethics	O
board	O
for	O
AI	B-Application
research	I-Application
remains	O
a	O
mystery	O
,	O
with	O
both	O
Google	B-Application
and	O
DeepMind	B-Application
declining	O
to	O
reveal	O
who	O
sits	O
on	O
the	O
board	O
.	O
</s>
<s>
DeepMind	B-Application
,	O
together	O
with	O
Amazon	O
,	O
Google	B-Application
,	O
Facebook	B-Application
,	O
IBM	O
and	O
Microsoft	O
,	O
is	O
a	O
founding	O
member	O
of	O
Partnership	O
on	O
AI	B-Application
,	O
an	O
organization	O
devoted	O
to	O
the	O
society-AI	O
interface	O
.	O
</s>
<s>
DeepMind	B-Application
has	O
opened	O
a	O
new	O
unit	O
called	O
DeepMind	B-Application
Ethics	O
and	O
Society	O
and	O
focused	O
on	O
the	O
ethical	O
and	O
societal	O
questions	O
raised	O
by	O
artificial	B-Application
intelligence	I-Application
featuring	O
prominent	O
philosopher	O
Nick	O
Bostrom	O
as	O
advisor	O
.	O
</s>
<s>
In	O
October	O
2017	O
,	O
DeepMind	B-Application
launched	O
a	O
new	O
research	O
team	O
to	O
investigate	O
AI	B-Application
ethics	O
.	O
</s>
<s>
In	O
December	O
2019	O
,	O
co-founder	O
Suleyman	O
announced	O
he	O
would	O
be	O
leaving	O
DeepMind	B-Application
to	O
join	O
Google	B-Application
,	O
working	O
in	O
a	O
policy	O
role	O
.	O
</s>
<s>
In	O
April	O
2023	O
,	O
DeepMind	B-Application
merged	O
with	O
Google	B-Application
AI	I-Application
's	O
Google	B-Application
Brain	I-Application
division	O
to	O
form	O
Google	B-Application
DeepMind	I-Application
,	O
as	O
part	O
of	O
the	O
company	O
's	O
continued	O
efforts	O
to	O
accelerate	O
work	O
on	O
AI	B-Application
.	O
</s>
<s>
According	O
to	O
the	O
company	O
's	O
website	O
,	O
DeepMind	B-Application
Technologies	I-Application
 '	O
goal	O
is	O
to	O
combine	O
"	O
the	O
best	O
techniques	O
from	O
machine	O
learning	O
and	O
systems	O
neuroscience	O
to	O
build	O
powerful	O
general-purpose	O
learning	O
algorithms	O
"	O
.	O
</s>
<s>
Google	B-Application
Research	O
released	O
a	O
paper	O
in	O
2016	O
regarding	O
AI	B-Application
safety	O
and	O
avoiding	O
undesirable	O
behaviour	O
during	O
the	O
AI	B-Application
learning	O
process	O
.	O
</s>
<s>
Deepmind	B-Application
has	O
also	O
released	O
several	O
publications	O
via	O
its	O
website	O
.	O
</s>
<s>
In	O
2017	O
DeepMind	B-Application
released	O
GridWorld	O
,	O
an	O
open-source	O
testbed	O
for	O
evaluating	O
whether	O
an	O
algorithm	O
learns	O
to	O
disable	O
its	O
kill	B-Device
switch	I-Device
or	O
otherwise	O
exhibits	O
certain	O
undesirable	O
behaviours	O
.	O
</s>
<s>
In	O
July	O
2018	O
,	O
researchers	O
from	O
DeepMind	B-Application
trained	O
one	O
of	O
its	O
systems	O
to	O
play	O
the	O
computer	O
game	O
Quake	B-Application
III	I-Application
Arena	I-Application
.	O
</s>
<s>
As	O
of	O
2020	O
,	O
DeepMind	B-Application
has	O
published	O
over	O
a	O
thousand	O
papers	O
,	O
including	O
thirteen	O
papers	O
that	O
were	O
accepted	O
by	O
Nature	O
or	O
Science	O
.	O
</s>
<s>
DeepMind	B-Application
received	O
media	O
attention	O
during	O
the	O
AlphaGo	B-Application
period	O
;	O
according	O
to	O
a	O
LexisNexis	B-Library
search	O
,	O
1842	O
published	O
news	O
stories	O
mentioned	O
DeepMind	B-Application
in	O
2016	O
,	O
declining	O
to	O
1363	O
in	O
2019	O
.	O
</s>
<s>
As	O
opposed	O
to	O
other	O
AIs	B-Application
,	O
such	O
as	O
IBM	O
's	O
Deep	B-General_Concept
Blue	I-General_Concept
or	O
Watson	B-Application
,	O
which	O
were	O
developed	O
for	O
a	O
pre-defined	O
purpose	O
and	O
only	O
function	O
within	O
its	O
scope	O
,	O
DeepMind	B-Application
claims	O
that	O
its	O
system	O
is	O
not	O
pre-programmed	O
:	O
it	O
learns	O
from	O
experience	O
,	O
using	O
only	O
raw	O
pixels	O
as	O
data	O
input	O
.	O
</s>
<s>
Technically	O
it	O
uses	O
deep	B-Algorithm
learning	I-Algorithm
on	O
a	O
convolutional	B-Architecture
neural	I-Architecture
network	I-Architecture
,	O
with	O
a	O
novel	O
form	O
of	O
Q-learning	B-Algorithm
,	O
a	O
form	O
of	O
model-free	O
reinforcement	O
learning	O
.	O
</s>
<s>
They	O
test	O
the	O
system	O
on	O
video	O
games	O
,	O
notably	O
early	O
arcade	O
games	O
,	O
such	O
as	O
Space	B-Application
Invaders	I-Application
or	O
Breakout	B-Application
.	O
</s>
<s>
Without	O
altering	O
the	O
code	O
,	O
the	O
AI	B-Application
begins	O
to	O
understand	O
how	O
to	O
play	O
the	O
game	O
,	O
and	O
after	O
some	O
time	O
plays	O
,	O
for	O
a	O
few	O
games	O
(	O
most	O
notably	O
Breakout	B-Application
)	O
,	O
a	O
more	O
efficient	O
game	O
than	O
any	O
human	O
ever	O
could	O
.	O
</s>
<s>
In	O
2013	O
,	O
DeepMind	B-Application
published	O
research	O
on	O
an	O
AI	B-Application
system	O
that	O
could	O
surpass	O
human	O
abilities	O
in	O
games	O
such	O
as	O
Pong	B-Application
,	O
Breakout	B-Application
and	O
Enduro	B-Application
,	O
while	O
surpassing	O
state	O
of	O
the	O
art	O
performance	O
on	O
Seaquest	B-Application
,	O
Beamrider	B-Application
,	O
and	O
Q*bert	B-Application
.	O
</s>
<s>
This	O
work	O
reportedly	O
led	O
to	O
the	O
company	O
's	O
acquisition	O
by	O
Google	B-Application
.	I-Application
</s>
<s>
DeepMind	B-Application
's	O
AI	B-Application
had	O
been	O
applied	O
to	O
video	O
games	O
made	O
in	O
the	O
1970s	O
and	O
1980s	O
;	O
work	O
was	O
ongoing	O
for	O
more	O
complex	O
3D	O
games	O
such	O
as	O
Quake	B-Application
,	O
which	O
first	O
appeared	O
in	O
the	O
1990s	O
.	O
</s>
<s>
In	O
2020	O
,	O
DeepMind	B-Application
published	O
Agent57	O
,	O
an	O
AI	B-Application
Agent	O
which	O
surpasses	O
human	O
level	O
performance	O
on	O
all	O
57	O
games	O
of	O
the	O
Atari2600	O
suite	O
.	O
</s>
<s>
In	O
October	O
2015	O
,	O
a	O
computer	B-Application
Go	I-Application
program	O
called	O
AlphaGo	B-Application
,	O
developed	O
by	O
DeepMind	B-Application
,	O
beat	O
the	O
European	O
Go	O
champion	O
Fan	O
Hui	O
,	O
a	O
2	O
dan	O
(	O
out	O
of	O
9	O
dan	O
possible	O
)	O
professional	O
,	O
five	O
to	O
zero	O
.	O
</s>
<s>
This	O
was	O
the	O
first	O
time	O
an	O
artificial	B-Application
intelligence	I-Application
(	O
AI	B-Application
)	O
defeated	O
a	O
professional	O
Go	O
player	O
.	O
</s>
<s>
Go	O
is	O
considered	O
much	O
more	O
difficult	O
for	O
computers	O
to	O
win	O
compared	O
to	O
other	O
games	O
like	O
chess	B-Application
,	O
due	O
to	O
the	O
much	O
larger	O
number	O
of	O
possibilities	O
,	O
making	O
it	O
prohibitively	O
difficult	O
for	O
traditional	O
AI	B-Application
methods	O
such	O
as	O
brute-force	B-Algorithm
.	O
</s>
<s>
In	O
March	O
2016	O
it	O
beat	O
Lee	O
Sedol	O
—	O
a	O
9th	O
dan	O
Go	O
player	O
and	O
one	O
of	O
the	O
highest	O
ranked	O
players	O
in	O
the	O
world	O
—	O
with	O
a	O
score	O
of	O
4	O
–	O
1	O
in	O
a	B-Application
five-game	I-Application
match	I-Application
.	O
</s>
<s>
In	O
the	O
2017	O
Future	B-Application
of	I-Application
Go	I-Application
Summit	I-Application
,	O
AlphaGo	B-Application
won	O
a	O
three-game	B-Application
match	I-Application
with	I-Application
Ke	I-Application
Jie	I-Application
,	O
who	O
at	O
the	O
time	O
continuously	O
held	O
the	O
world	O
No	O
.	O
</s>
<s>
It	O
used	O
a	O
supervised	B-General_Concept
learning	I-General_Concept
protocol	O
,	O
studying	O
large	O
numbers	O
of	O
games	O
played	O
by	O
humans	O
against	O
each	O
other	O
.	O
</s>
<s>
In	O
2017	O
,	O
an	O
improved	O
version	O
,	O
AlphaGo	B-Application
Zero	O
,	O
defeated	O
AlphaGo	B-Application
100	O
games	O
to	O
0	O
.	O
</s>
<s>
AlphaGo	B-Application
Zero	O
's	O
strategies	O
were	O
self-taught	O
.	O
</s>
<s>
AlphaGo	B-Application
Zero	O
was	O
able	O
to	O
beat	O
its	O
predecessor	O
after	O
just	O
three	O
days	O
with	O
less	O
processing	O
power	O
than	O
AlphaGo	B-Application
;	O
in	O
comparison	O
,	O
the	O
original	O
AlphaGo	B-Application
needed	O
months	O
to	O
learn	O
how	O
to	O
play	O
.	O
</s>
<s>
Later	O
that	O
year	O
,	O
AlphaZero	B-Application
,	O
a	O
modified	O
version	O
of	O
AlphaGo	B-Application
Zero	O
but	O
for	O
handling	O
any	O
two-player	O
game	O
of	O
perfect	O
information	O
,	O
gained	O
superhuman	O
abilities	O
at	O
chess	B-Application
and	O
shogi	O
.	O
</s>
<s>
Like	O
AlphaGo	B-Application
Zero	O
,	O
AlphaZero	B-Application
learned	O
solely	O
through	O
self-play	B-Algorithm
.	O
</s>
<s>
DeepMind	B-Application
researchers	O
published	O
a	O
new	O
model	O
named	O
MuZero	O
that	O
mastered	O
the	O
domains	O
of	O
Go	O
,	O
chess	B-Application
,	O
shogi	O
,	O
and	O
Atari	B-General_Concept
2600	I-General_Concept
games	I-General_Concept
without	O
human	O
data	O
,	O
domain	O
knowledge	O
,	O
or	O
known	O
rules	O
.	O
</s>
<s>
Researchers	O
applied	O
MuZero	O
to	O
solve	O
the	O
real	O
world	O
challenge	O
of	O
video	O
compression	O
with	O
a	O
set	O
number	O
of	O
bits	O
with	O
respect	O
to	O
Internet	O
traffic	O
on	O
sites	O
such	O
as	O
YouTube	B-General_Concept
,	O
Twitch	O
,	O
and	O
Google	B-Protocol
Meet	I-Protocol
.	O
</s>
<s>
In	O
October	O
2022	O
,	O
DeepMind	B-Application
unveiled	O
a	O
new	O
version	O
of	O
AlphaZero	B-Application
,	O
called	O
AlphaTensor	O
,	O
in	O
a	O
paper	O
published	O
in	O
Nature	O
.	O
</s>
<s>
For	O
example	O
,	O
AlphaTensor	O
figured	O
out	O
how	O
to	O
multiply	O
two	O
mod-2	O
4x4	O
matrices	O
in	O
only	O
47	O
multiplications	O
,	O
unexpectedly	O
beating	O
the	O
1969	O
Strassen	B-Algorithm
algorithm	I-Algorithm
record	O
of	O
49	O
multiplications	O
.	O
</s>
<s>
AlphaGo	B-Application
technology	O
was	O
developed	O
based	O
on	O
the	O
deep	O
reinforcement	O
learning	O
approach	O
.	O
</s>
<s>
This	O
makes	O
AlphaGo	B-Application
different	O
from	O
the	O
rest	O
of	O
AI	B-Application
technologies	O
on	O
the	O
market	O
.	O
</s>
<s>
With	O
that	O
said	O
,	O
AlphaGo	B-Application
's	O
‘	O
brain’	O
was	O
introduced	O
to	O
various	O
moves	O
based	O
on	O
historical	O
tournament	O
data	O
.	O
</s>
<s>
AlphaGo	B-Application
used	O
two	O
deep	O
neural	B-Architecture
networks	I-Architecture
:	O
a	O
policy	O
network	O
to	O
evaluate	O
move	O
probabilities	O
and	O
a	O
value	O
network	O
to	O
assess	O
positions	O
.	O
</s>
<s>
The	O
policy	O
network	O
trained	O
via	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
and	O
was	O
subsequently	O
refined	O
by	O
policy-gradient	O
reinforcement	O
learning	O
.	O
</s>
<s>
After	O
training	O
,	O
these	O
networks	O
employed	O
a	O
lookahead	O
Monte	B-Application
Carlo	I-Application
tree	I-Application
search	I-Application
(	O
MCTS	O
)	O
,	O
using	O
the	O
policy	O
network	O
to	O
identify	O
candidate	O
high-probability	O
moves	O
,	O
while	O
the	O
value	O
network	O
(	O
in	O
conjunction	O
with	O
Monte	O
Carlo	O
rollouts	O
using	O
a	O
fast	O
rollout	O
policy	O
)	O
evaluated	O
tree	O
positions	O
.	O
</s>
<s>
AlphaGo	B-Application
Zero	O
was	O
trained	O
using	O
reinforcement	O
learning	O
in	O
which	O
the	O
system	O
played	O
millions	O
of	O
games	O
against	O
itself	O
.	O
</s>
<s>
It	O
uses	O
a	O
single	O
neural	B-Architecture
network	I-Architecture
,	O
rather	O
than	O
separate	O
policy	O
and	O
value	O
networks	O
.	O
</s>
<s>
Its	O
simplified	O
tree	O
search	O
relies	O
upon	O
this	O
neural	B-Architecture
network	I-Architecture
to	O
evaluate	O
positions	O
and	O
sample	O
moves	O
.	O
</s>
<s>
AlphaGo	B-Application
Zero	O
employed	O
around	O
15	O
people	O
and	O
millions	O
in	O
computing	O
resources	O
.	O
</s>
<s>
Ultimately	O
,	O
it	O
needed	O
much	O
less	O
computing	O
power	O
than	O
AlphaGo	B-Application
,	O
running	O
on	O
four	O
specialized	O
AI	B-Application
processors	O
(	O
Google	B-Application
TPUs	B-Device
)	O
,	O
instead	O
of	O
AlphaGo	B-Application
's	O
48	O
.	O
</s>
<s>
In	O
2016	O
,	O
DeepMind	B-Application
turned	O
its	O
artificial	B-Application
intelligence	I-Application
to	O
protein	O
folding	O
,	O
a	O
long-standing	O
problem	O
in	O
molecular	O
biology	O
.	O
</s>
<s>
In	O
December	O
2018	O
,	O
DeepMind	B-Application
's	O
AlphaFold	B-Application
won	O
the	O
13th	O
Critical	O
Assessment	O
of	O
Techniques	O
for	O
Protein	O
Structure	O
Prediction	O
(	O
CASP	O
)	O
by	O
successfully	O
predicting	O
the	O
most	O
accurate	O
structure	O
for	O
25	O
out	O
of	O
43	O
proteins	O
.	O
</s>
<s>
In	O
2020	O
,	O
in	O
the	O
14th	O
CASP	O
,	O
AlphaFold	B-Application
's	O
predictions	O
achieved	O
an	O
accuracy	O
score	O
regarded	O
as	O
comparable	O
with	O
lab	O
techniques	O
.	O
</s>
<s>
A	O
week	O
later	O
DeepMind	B-Application
announced	O
that	O
AlphaFold	B-Application
had	O
completed	O
its	O
prediction	O
of	O
nearly	O
all	O
human	O
proteins	O
as	O
well	O
as	O
the	O
entire	O
proteomes	O
of	O
20	O
other	O
widely	O
studied	O
organisms	O
.	O
</s>
<s>
The	O
structures	O
were	O
released	O
on	O
the	O
AlphaFold	B-Application
Protein	O
Structure	O
Database	O
.	O
</s>
<s>
In	O
July	O
2022	O
,	O
it	O
was	O
announced	O
that	O
the	O
predictions	O
of	O
over	O
200	O
million	O
proteins	O
,	O
representing	O
virtually	O
all	O
known	O
proteins	O
,	O
would	O
be	O
released	O
on	O
the	O
AlphaFold	B-Application
database	O
.	O
</s>
<s>
In	O
2016	O
,	O
DeepMind	B-Application
introduced	O
WaveNet	O
,	O
a	O
text-to-speech	B-Application
system	O
.	O
</s>
<s>
It	O
was	O
originally	O
too	O
computationally	O
intensive	O
for	O
use	O
in	O
consumer	O
products	O
,	O
but	O
in	O
late	O
2017	O
it	O
became	O
ready	O
for	O
use	O
in	O
consumer	O
applications	O
such	O
as	O
Google	B-Application
Assistant	I-Application
.	O
</s>
<s>
In	O
2018	O
Google	B-Application
launched	O
a	O
commercial	O
text-to-speech	B-Application
product	O
,	O
Cloud	O
Text-to-Speech	B-Application
,	O
based	O
on	O
WaveNet	O
.	O
</s>
<s>
In	O
2018	O
,	O
DeepMind	B-Application
introduced	O
a	O
more	O
efficient	O
model	O
called	O
WaveRNN	O
co-developed	O
with	O
Google	B-Application
AI	I-Application
.	O
</s>
<s>
In	O
2019	O
,	O
Google	B-Application
started	O
to	O
roll	O
WaveRNN	O
with	O
WavenetEQ	O
out	O
to	O
Google	B-Application
Duo	I-Application
users	O
.	O
</s>
<s>
In	O
2016	O
,	O
Hassabis	O
discussed	O
the	O
game	O
StarCraft	B-Application
as	O
a	O
future	O
challenge	O
,	O
since	O
it	O
requires	O
strategic	O
thinking	O
and	O
handling	O
imperfect	O
information	O
.	O
</s>
<s>
In	O
January	O
2019	O
,	O
DeepMind	B-Application
introduced	O
AlphaStar	B-Application
,	O
a	O
program	O
playing	O
the	O
real-time	O
strategy	O
game	O
StarCraft	B-Application
II	I-Application
.	O
</s>
<s>
AlphaStar	B-Application
used	O
reinforcement	O
learning	O
based	O
on	O
replays	O
from	O
human	O
players	O
,	O
and	O
then	O
played	O
against	O
itself	O
to	O
enhance	O
its	O
skills	O
.	O
</s>
<s>
At	O
the	O
time	O
of	O
the	O
presentation	O
,	O
AlphaStar	B-Application
had	O
knowledge	O
equivalent	O
to	O
200	O
years	O
of	O
playing	O
time	O
.	O
</s>
<s>
In	O
July	O
2019	O
,	O
AlphaStar	B-Application
began	O
playing	O
against	O
random	O
humans	O
on	O
the	O
public	O
1v1	O
European	O
multiplayer	O
ladder	O
.	O
</s>
<s>
Unlike	O
the	O
first	O
iteration	O
of	O
AlphaStar	B-Application
,	O
which	O
played	O
only	O
Protoss	O
v	O
.	O
Protoss	O
,	O
this	O
one	O
played	O
as	O
all	O
of	O
the	O
game	O
's	O
races	O
,	O
and	O
had	O
earlier	O
unfair	O
advantages	O
fixed	O
.	O
</s>
<s>
By	O
October	O
2019	O
,	O
AlphaStar	B-Application
reached	O
Grandmaster	O
level	O
on	O
the	O
StarCraft	B-Application
II	I-Application
ladder	O
on	O
all	O
three	O
StarCraft	B-Application
races	O
,	O
becoming	O
the	O
first	O
AI	B-Application
to	O
reach	O
the	O
top	O
league	O
of	O
a	O
widely	O
popular	O
esport	O
without	O
any	O
game	O
restrictions	O
.	O
</s>
<s>
In	O
2022	O
,	O
DeepMind	B-Application
unveiled	O
AlphaCode	O
,	O
an	O
AI-powered	O
coding	O
engine	O
that	O
creates	O
computer	B-Application
programs	I-Application
at	O
a	O
rate	O
comparable	O
to	O
that	O
of	O
an	O
average	O
programmer	O
,	O
with	O
the	O
company	O
testing	O
the	O
system	O
against	O
coding	O
challenges	O
created	O
by	O
Codeforces	O
utilized	O
in	O
human	O
competitive	O
programming	O
competitions	O
.	O
</s>
<s>
AlphaCode	O
earned	O
a	O
rank	O
equivalent	O
to	O
54%	O
of	O
the	O
median	O
score	O
on	O
Codeforces	O
after	O
being	O
trained	O
on	O
GitHub	B-Application
data	O
and	O
Codeforce	O
problems	O
and	O
solutions	O
.	O
</s>
<s>
Google	B-Application
has	O
stated	O
that	O
DeepMind	B-Application
algorithms	O
have	O
greatly	O
increased	O
the	O
efficiency	O
of	O
cooling	O
its	O
data	O
centers	O
.	O
</s>
<s>
In	O
addition	O
,	O
DeepMind	B-Application
(	O
alongside	O
other	O
Alphabet	O
AI	B-Application
researchers	O
)	O
assists	O
Google	B-Application
Play	I-Application
's	I-Application
personalized	O
app	O
recommendations	O
.	O
</s>
<s>
DeepMind	B-Application
has	O
also	O
collaborated	O
with	O
the	O
Android	B-Application
team	O
at	O
Google	B-Application
for	O
the	O
creation	O
of	O
two	O
new	O
features	O
which	O
were	O
made	O
available	O
to	O
people	O
with	O
devices	O
running	O
Android	B-Application
Pie	O
,	O
the	O
ninth	O
installment	O
of	O
Google	B-Application
's	I-Application
mobile	O
operating	O
system	O
.	O
</s>
<s>
It	O
is	O
the	O
first	O
time	O
DeepMind	B-Application
has	O
used	O
these	O
techniques	O
on	O
such	O
a	O
small	O
scale	O
,	O
with	O
typical	O
machine	O
learning	O
applications	O
requiring	O
orders	O
of	O
magnitude	O
more	O
computing	O
power	O
.	O
</s>
<s>
DeepMind	B-Application
researchers	O
have	O
applied	O
machine	O
learning	O
models	O
to	O
the	O
sport	O
of	O
football	O
,	O
often	O
referred	O
to	O
as	O
soccer	O
in	O
North	O
America	O
,	O
to	O
modelling	O
the	O
behaviour	O
of	O
football	O
players	O
,	O
including	O
the	O
goalkeeper	O
,	O
defenders	O
,	O
and	O
strikers	O
during	O
different	O
scenarios	O
such	O
as	O
penalty	O
kicks	O
.	O
</s>
<s>
Google	B-Application
has	O
unveiled	O
a	O
new	O
archaeology	O
document	O
program	O
named	O
Ithaca	O
after	O
the	O
home	O
island	O
of	O
mythical	O
hero	O
Odysseys	O
.	O
</s>
<s>
The	O
deep	O
neural	B-Architecture
network	I-Architecture
helps	O
researchers	O
restore	O
the	O
empty	O
text	O
of	O
damaged	O
documents	O
,	O
identify	O
the	O
place	O
they	O
originated	O
from	O
,	O
and	O
give	O
them	O
a	O
definite	O
accurate	O
date	O
.	O
</s>
<s>
Sparrow	O
is	O
an	O
artificial	O
intelligence-powered	O
chatbot	B-Application
developed	O
by	O
DeepMind	B-Application
to	O
build	O
safer	O
machine	O
learning	O
systems	O
by	O
using	O
a	O
mix	O
of	O
human	O
feedback	O
and	O
Google	B-Application
search	O
suggestions	O
.	O
</s>
<s>
Chinchilla	O
AI	B-Application
is	O
a	O
language	O
model	O
developed	O
by	O
DeepMind	B-Application
.	O
</s>
<s>
In	O
July	O
2016	O
,	O
a	O
collaboration	O
between	O
DeepMind	B-Application
and	O
Moorfields	O
Eye	O
Hospital	O
was	O
announced	O
to	O
develop	O
AI	B-Application
applications	I-Application
for	I-Application
healthcare	I-Application
.	O
</s>
<s>
DeepMind	B-Application
would	O
be	O
applied	O
to	O
the	O
analysis	O
of	O
anonymised	O
eye	O
scans	O
,	O
searching	O
for	O
early	O
signs	O
of	O
diseases	O
leading	O
to	O
blindness	O
.	O
</s>
<s>
There	O
are	O
also	O
projects	O
with	O
the	O
Royal	O
Free	O
London	O
NHS	O
Foundation	O
Trust	O
and	O
Imperial	O
College	O
Healthcare	O
NHS	O
Trust	O
to	O
develop	O
new	O
clinical	O
mobile	O
apps	O
linked	O
to	O
electronic	B-Application
patient	I-Application
records	I-Application
.	O
</s>
<s>
In	O
November	O
2017	O
,	O
DeepMind	B-Application
announced	O
a	O
research	O
partnership	O
with	O
the	O
Cancer	O
Research	O
UK	O
Centre	O
at	O
Imperial	O
College	O
London	O
with	O
the	O
goal	O
of	O
improving	O
breast	O
cancer	O
detection	O
by	O
applying	O
machine	O
learning	O
to	O
mammography	O
.	O
</s>
<s>
Additionally	O
,	O
in	O
February	O
2018	O
,	O
DeepMind	B-Application
announced	O
it	O
was	O
working	O
with	O
the	O
U.S.	O
Department	O
of	O
Veterans	O
Affairs	O
in	O
an	O
attempt	O
to	O
use	O
machine	O
learning	O
to	O
predict	O
the	O
onset	O
of	O
acute	O
kidney	O
injury	O
in	O
patients	O
,	O
and	O
also	O
more	O
broadly	O
the	O
general	O
deterioration	O
of	O
patients	O
during	O
a	O
hospital	O
stay	O
so	O
that	O
doctors	O
and	O
nurses	O
can	O
more	O
quickly	O
treat	O
patients	O
in	O
need	O
.	O
</s>
<s>
DeepMind	B-Application
developed	O
an	O
app	O
called	O
Streams	O
,	O
which	O
sends	O
alerts	O
to	O
doctors	O
about	O
patients	O
at	O
risk	O
of	O
acute	O
kidney	O
injury	O
.	O
</s>
<s>
On	O
13	O
November	O
2018	O
,	O
DeepMind	B-Application
announced	O
that	O
its	O
health	O
division	O
and	O
the	O
Streams	O
app	O
would	O
be	O
absorbed	O
into	O
Google	B-Application
Health	I-Application
.	O
</s>
<s>
Privacy	O
advocates	O
said	O
the	O
announcement	O
betrayed	O
patient	O
trust	O
and	O
appeared	O
to	O
contradict	O
previous	O
statements	O
by	O
DeepMind	B-Application
that	O
patient	O
data	O
would	O
not	O
be	O
connected	O
to	O
Google	B-Application
accounts	O
or	O
services	O
.	O
</s>
<s>
A	O
spokesman	O
for	O
DeepMind	B-Application
said	O
that	O
patient	O
data	O
would	O
still	O
be	O
kept	O
separate	O
from	O
Google	B-Application
services	O
or	O
projects	O
.	O
</s>
<s>
In	O
April	O
2016	O
,	O
New	O
Scientist	O
obtained	O
a	O
copy	O
of	O
a	O
data	B-Application
sharing	I-Application
agreement	O
between	O
DeepMind	B-Application
and	O
the	O
Royal	O
Free	O
London	O
NHS	O
Foundation	O
Trust	O
.	O
</s>
<s>
The	O
agreement	O
shows	O
DeepMind	B-Application
Health	O
had	O
access	O
to	O
admissions	O
,	O
discharge	O
and	O
transfer	O
data	O
,	O
accident	O
and	O
emergency	O
,	O
pathology	O
and	O
radiology	O
,	O
and	O
critical	O
care	O
at	O
these	O
hospitals	O
.	O
</s>
<s>
The	O
ICO	O
found	O
that	O
the	O
Royal	O
Free	O
failed	O
to	O
comply	O
with	O
the	O
Data	O
Protection	O
Act	O
when	O
it	O
provided	O
patient	O
details	O
to	O
DeepMind	B-Application
,	O
and	O
found	O
several	O
shortcomings	O
in	O
how	O
the	O
data	O
was	O
handled	O
,	O
including	O
that	O
patients	O
were	O
not	O
adequately	O
informed	O
that	O
their	O
data	O
would	O
be	O
used	O
as	O
part	O
of	O
the	O
test	O
.	O
</s>
<s>
DeepMind	B-Application
published	O
its	O
thoughts	O
on	O
the	O
investigation	O
in	O
July	O
2017	O
,	O
saying	O
“	O
we	O
need	O
to	O
do	O
better	O
”	O
and	O
highlighting	O
several	O
activities	O
and	O
initiatives	O
they	O
had	O
initiated	O
for	O
transparency	O
,	O
oversight	O
and	O
engagement	O
.	O
</s>
<s>
In	O
May	O
2017	O
,	O
Sky	O
News	O
published	O
a	O
leaked	O
letter	O
from	O
the	O
National	O
Data	O
Guardian	O
,	O
Dame	O
Fiona	O
Caldicott	O
,	O
revealing	O
that	O
in	O
her	O
"	O
considered	O
opinion	O
"	O
the	O
data-sharing	B-Application
agreement	O
between	O
DeepMind	B-Application
and	O
the	O
Royal	O
Free	O
took	O
place	O
on	O
an	O
"	O
inappropriate	O
legal	O
basis	O
"	O
.	O
</s>
<s>
The	O
Information	O
Commissioner	O
's	O
Office	O
ruled	O
in	O
July	O
2017	O
that	O
the	O
Royal	O
Free	O
hospital	O
failed	O
to	O
comply	O
with	O
the	O
Data	O
Protection	O
Act	O
when	O
it	O
handed	O
over	O
personal	O
data	O
of	O
1.6	O
million	O
patients	O
to	O
DeepMind	B-Application
.	O
</s>
<s>
In	O
October	O
2017	O
,	O
DeepMind	B-Application
announced	O
a	O
new	O
research	O
unit	O
,	O
DeepMind	B-Application
Ethics	O
&	O
Society	O
.	O
</s>
<s>
Their	O
goal	O
is	O
to	O
fund	O
external	O
research	O
of	O
the	O
following	O
themes	O
:	O
privacy	O
,	O
transparency	O
,	O
and	O
fairness	O
;	O
economic	O
impacts	O
;	O
governance	O
and	O
accountability	O
;	O
managing	O
AI	B-Application
risk	O
;	O
AI	B-Application
morality	O
and	O
values	O
;	O
and	O
how	O
AI	B-Application
can	O
address	O
the	O
world	O
's	O
challenges	O
.	O
</s>
<s>
As	O
a	O
result	O
,	O
the	O
team	O
hopes	O
to	O
further	O
understand	O
the	O
ethical	O
implications	O
of	O
AI	B-Application
and	O
aid	O
society	O
to	O
seeing	O
AI	B-Application
can	O
be	O
beneficial	O
.	O
</s>
<s>
This	O
new	O
subdivision	O
of	O
DeepMind	B-Application
is	O
a	O
completely	O
separate	O
unit	O
from	O
the	O
partnership	O
of	O
leading	O
companies	O
using	O
AI	B-Application
,	O
academia	O
,	O
civil	O
society	O
organizations	O
and	O
nonprofits	O
of	O
the	O
name	O
Partnership	O
on	O
Artificial	B-Application
Intelligence	I-Application
to	O
Benefit	O
People	O
and	O
Society	O
of	O
which	O
DeepMind	B-Application
is	O
also	O
a	O
part	O
.	O
</s>
<s>
The	O
DeepMind	B-Application
Ethics	O
and	O
Society	O
board	O
is	O
also	O
distinct	O
from	O
the	O
mooted	O
AI	B-Application
Ethics	O
Board	O
that	O
Google	B-Application
originally	O
agreed	O
to	O
form	O
when	O
acquiring	O
DeepMind	B-Application
.	O
</s>
<s>
DeepMind	B-Application
sponsors	O
three	O
chairs	O
of	O
machine	O
learning	O
:	O
</s>
