linear function falls on the right side of
the decision boundary.”
Also, if we really understood hand-
coded systems, like operating systems,
then they would have no bugs. But they
do have bugs, which suggests that we
don’t fully understand these hand-coded
systems—and the main reason we don’t
understand is that the domain is com-
plicated and we can’t envision all the
Peter Norvig is director of research at Google.
More Spending Won’t Solve Anything
Jason Pontin’s op-ed calling for more
government R&D (“Make America Great
Again,” May/June 2017) piqued a nerve.
I was at a conference at MIT last month
in a room full of scientists calling for the
same thing, and I had to tell them it’s not
that easy. I’m a fan of science. I have seven
degrees from Stanford and MIT, mostly
in STEM, but I’m also a fan of evidence.
It’s not clear the evidence supports more
As lead energy economist for the
White House Council of Economic Advisers in 2006, I was given carte blanche to
design a new U.S. energy policy. I was
encouraged to spend billions. However,
after consulting with the top experts
around the world, I realized that there
were better ways to spend government
Write MIT Technology Review
One Main Street, 13th Floor
Cambridge, MA 02142
Please include your address,
telephone number, and e-mail address.
Letters and comments may be edited for
both clarity and length.
Of Course Complexity Is Baffling
In “The Dark Secret at the Heart of AI”
(May/June 2017), Will Knight writes:
“The workings of any machine-learning
technology are inherently more opaque,
even to computer scientists, than a hand-coded system.” While it’s partly true that
hand-coded systems can be easier to
understand than data-trained models, the
issue isn’t really that particular models
are hard but rather that complex domains
in general are hard to understand.
For example, do we understand why
a certain stock goes up or down on any
given day? Mostly no: a human analyst
will have an explanation, but if the stock
had gone in the other direction, the ana-
lyst would have given a different reason.
Which means these post-hoc “explana-
tions” are no more valuable than a com-
plex machine-learned model saying, “I
predict it will go up because this non-
Numbers can be bloodless. Eric
Schmidt once said at a talk at Stanford
that the hardest thing for a CEO is layoffs,
because he knew that for every thousand
people he laid off, a certain number would
commit suicide. So you’d better be sure
that those layoffs were worth it. Every gov-
ernment economist holds a similar num-
ber in the back of his mind—$8 million.
This is the current statistical value of a life
used by the federal government.
You are calling for government R&D
spending to increase to its peak levels.
That’s an additional $200 billion, or
27,000 lives per year. We’d better be sure
that spending is worth it.
Yes, government R&D as a percentage
of GDP has fallen, but it is still more than
it was in the past. Adjusting for inflation
and using your numbers, we spend $126
billion today compared with $76 billion
in the ’60s. People regularly call for an
“Apollo mission” for climate change, but
we already spend more on renewable-energy research than we ever did for the
The trouble with pointing to the high
returns for existing R&D is that it elides
a problem economists spend a lot of time
talking about in our intro classes. There
is a difference between average returns
and marginal returns. When advocates
cite studies on the high average returns to
government R&D funding, they miss out
on this important distinction.
Letters and Comments
MIT Technology Review
Volume 120, Number 3
People regularly call for an “Apollo mission” for climate change,
but we already spend more on renewable-energy research than
we ever did for the Apollo mission.