|
![](/i/favi32.png) April 22nd-28th 2023 Ukraine’s game planThe Economistcould go wrong
I
n 1960 norbert wiener
published a
prescient essay. In it, the father of cyber
netics worried about a world in which “ma
chines learn” and “develop unforeseen
strategies at rates that baffle their program
mers.” Such strategies, he thought, might
involve actions that those programmers
did not “really desire” and were instead
“merely colourful imitation[s] of it.” Wie
ner illustrated his point with the German
poet Goethe’s fable, “The Sorcerer’s Ap
prentice”, in which a trainee magician en
chants a broom to fetch water to fill his
master’s bath. But the trainee is unable to
stop the broom when its task is complete.
It eventually brings so much water that it
floods the room, having lacked the com
mon sense to know when to stop.
The striking progress of modern artifi
cialintelligence (
AI
) research has seen
Wiener’s fears resurface. In August 2022,
AI
Impacts, an American research group, pub
lished a survey that asked more than 700
machinelearning researchers about their
predictions for both progress in
AI
and the
risks the technology might pose. The typi
cal respondent reckoned there was a 5%
probability of advanced
AI
causing an “ex
tremely bad” outcome, such as human ex
tinction (see chart). FeiFei Li, an
AI
lumin
ary at Stanford University, talks of a “civili
sational moment” for
AI
. Asked by an
quality writing and chat knowledgeably
about all kinds of topics. As Robert Trager
of the Centre for Governance on
AI
ex
plains, one risk is of such software “mak
ing it easier to do lots of things—and thus
allowing more people to do them.”
The most immediate risk is that
LLM
s
could amplify the sort of quotidian harms
that can be perpetrated on the internet to
day. A textgeneration engine that can con
vincingly imitate a variety of styles is ideal
for spreading misinformation, scamming
people out of their money or convincing
employees to click on dodgy links in
emails, infecting their company’s comput
ers with malware. Chatbots have also been
used to cheat at school.
Like soupedup search engines, chat
bots can also help humans fetch and un
derstand information. That can be a dou
bleedged sword. In April, a Pakistani court
used
GPT
4 to help make a decision on
granting bail—it even included a transcript
of a conversation with
GPT
4 in its judg
ment. In a preprint published on arXiv on
April 11th, researchers from Carnegie Mel
lon University say they designed a system
that, given simple prompts such as “syn
thesise ibuprofen”, searches the internet
and spits out instructions on how to pro
duce the painkiller from precursor chemi
cals. But there is no reason that such a pro
gram would be limited to beneficial drugs.
Some researchers, meanwhile, are con
sumed by much bigger worries. They fret
about “alignment problems”, the technical
name for the concern raised by Wiener in
his essay. The risk here is that, like Goethe’s
enchanted broom, an
AI
might single
mindedly pursue a goal set by a user, but in
the process do something harmful that was
not desired. The bestknown example is
the “paperclip maximiser”, a thought ex
periment described by Nick Bostrom, a
philosopher, in 2003. An
AI
is instructed to
manufacture as many paperclips as it can.
Being an idiot savant, such an openended
goal leads the maximiser to take any mea
sures necessary to cover the Earth in paper
clip factories, exterminating humanity
along the way. Such a scenario may sound
like an unused plotline from a Douglas Ad
ams novel. But, as
AI
Impacts’ poll shows,
many
AI
researchers think that not to wor
ry about the behaviour of a digital superin
telligence would be complacent.
What to do? The more familiar pro
blems seem the most tractable. Before re
leasing
GPT
4, which powers the latest ver
sion of its chatbot, Open
AI
used several ap
proaches to reduce the risk of accidents
and misuse. One is called “reinforcement
learning from human feedback” (
RLHF
).
Described in a paper published in 2017,
RLHF
asks humans to provide feedback on
whether a model’s response to a prompt
was appropriate. The model is then updat
ed based on that feedback. The goal is to re
Dostları ilə paylaş: |
|
|