Jeffrey
Wu
Email
me@wuthejeff.com
Website
wuthejeff.com
Github
user
WuTheFWasThat
Google
Scholar
user
x4JAvwMAAAAJ
Summary:
Researcher
and
full-stack
engineer
interested
in
keeping
the
world
safe
for
everyone
as
technologies
such
as
machine
learning
become
more
powerful
Work
Experience
OpenAI
Research
engineer
Aug
2018
–
Jul
2024
AI
safety
research.
Worked
on
language
modeling
(including
training
GPT-2)
and
RLHF
(trained initial GPT instruct series models).
Managed teams and projects working on scalable
oversight,
generalization,
and
interpretability.
Google
Research
Software
engineer
Oct
2016
–
Aug
2018
Built
general
infrastructure
for
supporting
models
for
personalization
from
cross-product
user
history
(basically
learning
giant
embedding
spaces
for
all
users,
Chrome
pages,
Youtube
videos, etc).
Experimented with RNN models to replace bag-of-words models, and contributed
to
launch
of
news
feed
trends
personalization.
Terminal.com
Founding
engineer
Jan
2013
–
Oct
2016
Building
cloud-based
container
infrastructure,
for
scientific
computing
and
online
education.
Helped
design
and
implement
many
core
systems
across
the
stack
and
oversaw
their
security
and
scalability.
Saw
company
grow
from
2
to
12,
and
managed
a
small
team
of
engineers.
Interfaced with clients, including Crunchbase, Stanford University, Codecademy, and Udacity.
Company
was
sold
to
Udacity.
Probabilistic
Computing
Project
Master’s
student
Nov
2011
–
Jan
2013
Implemented
a
probabilistic
programming
language.
Explored
a
new
Gibbs
sampling
algo-
rithm to make inference more efficient in very general settings.
Work presented [at NIPS 2012
probabilistic
programing
workshop].
[Source
code]
and
[thesis].
Research
I’m broadly interested in training AI systems that are honest and kind to humans, and assistive
systems
that
make
humans
smarter
or
wiser.
Training
large
language
models
I scaled up OpenAI’s largest language models 100x (from
110M parameters to 12B parameters).
This included training GPT-2 [blog] and early stepping
stones
towards
GPT-3
(NeurIPS
2020
best
paper
award),
Reinforcement
learning
from
human
feedback
I
worked
on
helping
language
models
learn from human feedback and specified human values.
Early proofs of concept trained mod-
els that produced summaries better than human-written ones.
This culminated in instruction
following
models
(InstructGPT),
which
were
the
predecessor
to
ChatGPT.
1
