Ex-OpenAI Employee LEAKED DOC TO CONGRESS!
Summary
TLDRWhistleblower William Saunders, a former technical staff member at OpenAI, testified to the US Senate, warning that AGI (Artificial General Intelligence) could be developed within three years. Saunders expressed concerns about the rapid advancements in AI, particularly OpenAI's new system, which has shown significant progress in tasks like mathematical problem-solving. He highlighted the potential for AI to cause catastrophic harm, the risks of theft by foreign adversaries, and the need for better safety measures. The testimony raises questions about society's readiness for AGI and the necessity for comprehensive plans to address its economic and employment impacts.
Takeaways
- π² William Saunders, a former OpenAI staff member, testified before a Senate subcommittee, suggesting that AGI could be achieved in as little as 3 years.
- π§ Saunders defines AGI as a highly autonomous system that outperforms humans at most economically valuable work, including digital and potentially physical labor.
- π Saunders highlights the rapid progress in AI capabilities, noting that recent systems have shown significant advancements in areas like writing, critical thinking, and mathematical problem-solving.
- π OpenAI's new AI system, referred to as 'gp01' in the transcript but correctly named 'OpenAI1', has made strides in mathematical competitions, nearly achieving a gold medal at the IMO.
- π The concept of 'test time compute' is introduced, where AI systems are given more computational power during testing to improve their performance.
- π¨ Saunders expresses concerns about the safety and control of AGI, warning of potential catastrophic harm through cyber attacks or the creation of biological weapons.
- πΌ There are ethical and societal implications raised, including the potential for AGI to make most human workers obsolete and the need for new economic models like Universal Basic Income.
- π Saunders points out that there were vulnerabilities at OpenAI that could have allowed the theft of AI systems, emphasizing the need for better security measures.
- π The whistleblower suggests that AGI development should involve more transparency, third-party testing, and independent oversight to ensure safety and ethical considerations.
- π The video concludes with a call for public discussion on the readiness for AGI, the trust in organizations developing it, and the geographical concentration of AI research and development.
Q & A
What did William Saunders, the whistleblower, testify before the Senate subcommittee?
-William Saunders testified that OpenAI is closer to unlocking artificial general intelligence (AGI) than most people believe, and that it could come in as little as 3 years.
What is the definition of AGI according to the script?
-AGI is defined as a highly autonomous system that outperforms humans at most economically valuable work, including both digital and physical labor.
How does the script describe the progress of AI in performing economically valuable work?
-The script describes AI systems getting better at various skills, such as writing, critical thinking, and reading comprehension, with some systems performing better than most humans in certain areas.
What is the significance of the AI system scoring close to a gold medal in the International Mathematical Olympiad?
-The AI system scoring close to a gold medal signifies a significant advancement in AI capabilities, as it demonstrates the system's ability to solve complex mathematical problems at a level comparable to the world's smartest mathematicians.
What is the role of the O*NET Online in the context of AGI?
-O*NET Online is used to break down jobs into discrete units of certain skills, which helps to quantify the potential for AI to automate various jobs by assessing the level of skills required for each occupation.
What is the concern raised by Saunders about the rapid development of AGI?
-Saunders is concerned that the rapid development of AGI could lead to significant changes in society, including radical changes to the economy and employment, and there is a risk of catastrophic harm if AGI systems are not properly controlled.
What recommendations does Saunders make to improve the safety and oversight of AGI development?
-Saunders recommends making it easy for whistleblowers to communicate with the government, implementing third-party testing before and after AI deployments, sharing test results, and establishing independent oversight organizations.
Why did Saunders leave OpenAI?
-Saunders left OpenAI after struggling to get the resources needed for the super alignment team to be successful, which was tasked with developing approaches to ensure the safety of AGI systems.
What is the significance of the whistleblower's testimony in relation to the development of AGI?
-The whistleblower's testimony highlights the potential risks and challenges associated with the rapid development of AGI, emphasizing the need for caution, oversight, and transparency in the development process.
How does the script suggest the AI industry should change its approach to AGI development?
-The script suggests that the AI industry should prioritize safety and rigorous testing over rapid deployment, and should not enforce agreements that prohibit criticism or disparagement of the company's AI development practices.
Outlines
π² Whistleblower's Warning on AGI's Imminent Arrival
The video script discusses a whistleblower from the subreddit Singularity, who claims that artificial general intelligence (AGI) is closer to reality than most people realize. William Saunders, a former technical staff member at OpenAI, testified before a senate subcommittee, suggesting AGI could be achieved in as little as three years. Saunders' definition of AGI aligns with OpenAI's, describing it as a highly autonomous system capable of pursuing long-term goals and outperforming humans at economically valuable work. The script humorously notes the significance of a post's timestamp and view count, suggesting a 'trifecta' of 420,000 views on Friday the 13th at 11:11 a.m. The discussion then explores the potential for AGI to perform digital work, such as tasks that can be outsourced or performed remotely, and the implications for physical labor with the need for robotics. The video also references OneT Online, which categorizes jobs by skills and proficiency levels, to illustrate how AI could potentially automate various roles.
π Rapid AI Progress and its Societal Impact
The script continues by examining the rapid advancements in AI, particularly in areas like writing, critical thinking, and reading comprehension. It mentions AI's increasing ability to understand visual data and its potential to outperform humans in certain tasks. The discussion then shifts to the International Mathematical Olympiad (IMO), where an AI system developed by Google, referred to as Alpha Geometry, nearly achieved a gold medal score, indicating AI's growing capability in complex problem-solving. The script also addresses the American Invitational Mathematics Examination (AIME), where OpenAI's new model demonstrated significant improvements due to 'test time compute,' a strategy that allocates more computational power during testing. The whistleblower's testimony highlights the potential for AGI to cause substantial societal changes, including economic and employment disruptions, and raises concerns about the risk of AI being used for cyber-attacks or biological weapon development.
π‘οΈ Concerns Over AGI's Safety and Security
The video script delves into concerns about the safety and security of AGI, with the whistleblower, William Saunders, warning about the potential for AGI to be used maliciously, such as in autonomous cyber-attacks or the creation of biological weapons. Saunders points out that OpenAI's new AI system has shown capabilities that could contribute to biological weapon risks, despite rigorous testing. He criticizes the industry's prioritization of deployment over thorough safety measures and expresses worry that future AI systems may possess dangerous capabilities that developers overlook. The script also touches on the vulnerability of AI systems to theft, with Saunders recounting periods at OpenAI where access controls could be bypassed. The whistleblower suggests measures to improve safety, including whistleblower protection, third-party testing, and independent oversight.
π Global Implications and the Future of AI Development
The final paragraph of the script addresses the global implications of AI development, particularly the concentration of AI research in the United States and the potential for AGI to cause catastrophic harm. It mentions theθ§£ζ£ of OpenAI's super alignment team due to a lack of resources and the broader industry's focus on rapid development over safety. The whistleblower calls for transparency, proper legal protections for whistleblowers, and the right to warn about potential dangers. The script concludes with a call to action, urging viewers to consider whether society is prepared for the advent of AGI and if adequate plans are in place to manage its impact. It also invites viewers to share their thoughts on the matter, including their views on the concentration of AI development in the US and their trust in companies like OpenAI to handle such powerful technology responsibly.
Mindmap
Keywords
π‘Artificial General Intelligence (AGI)
π‘Whistleblower
π‘Autonomous System/Agent
π‘Economically Valuable Work
π‘O*NET Online
π‘International Mathematical Olympiad (IMO)
π‘Test Time Compute
π‘Reinforcement Training
π‘Biological Weapon Risk
π‘Non-Dsparagement Agreement
Highlights
William Saunders, a former member of technical staff at OpenAI, testified before a Senate subcommittee about the rapid progress towards AGI.
AGI could be achieved in as little as 3 years, according to Saunders' testimony.
OpenAI's definition of AGI is a highly autonomous system that outperforms humans at most economically valuable work.
AGI is currently perceived as digital work, excluding physical labor.
O*NET Online is referenced as a tool to break down jobs into discrete skills to evaluate AI's potential to automate them.
AI systems have shown significant improvements in skills like writing, critical thinking, and reading comprehension.
OpenAI's new AI system, referred to as 'gp01' in the transcript but correctly named 'OpenAI1', has achieved milestones in mathematical problem-solving.
The AI system scored nearly a gold medal in the International Mathematical Olympiad, showcasing its advanced capabilities.
Test time compute is a new approach where AI is given more computational power when answering questions, improving performance.
The whistleblower expressed concerns about the rapid development of AI and the lack of safety measures.
Saunders warned about the potential for AGI to cause catastrophic harm, including via cyber attacks or biological weapons.
The former OpenAI staff member highlighted vulnerabilities in access controls that could allow theft of AI systems.
The super alignment team at OpenAI was disbanded due to lack of resources, raising concerns about safety measures.
Suggestions for improving AI safety include whistleblower protection, third-party testing, and independent oversight.
The transcript discusses the potential economic and societal impacts of AGI, including job displacement and the need for new economic models.
The video concludes with a call for public discussion on the readiness and plans for the advent of AGI.
Transcripts
redditors of subreddit Singularity have
posted this opening eye whistleblower
William Saunders testified before a
senate subcommittee today he saying that
openi is closer to unlocking artificial
general intelligence AGI than most
people believe or understand and that it
could come in as little as 3 years AGI
rolls around only once subscribe Andre
gpy actually posted an explanation of
what he believes AGI to mean he posted
it on Friday the 13th at 1111 a.m. if
this post had
420,000 views that would have been the
trifecta I think there I fix it for you
420,000 views shocked Pikachu face but
his point is that he's been using the
definition from open AI which he
believes is kind of like the relatively
standard and okay definition that being
that AGI is a highly autonomous system
so we've been using the word agents for
this kind of an autonomous agent
something that has agency that is able
to go out there and pursue long-term
goals planning adjusting to changes or
unforeseen obstacles so like an
autonomous system/ agent that
outperforms humans at most economically
valuable work now of course when we're
talking about most economically valuable
work uh you know for this to include the
physical labor the the building
gardening manufacturing and door Dash in
2016 I realized people could deliver
Sushi to my house I don't think I'll
ever financially recover from that year
but the point is for physical labor we
would need robots Etc so most people
kind of think of AGI at least for now as
just you know only to mean digital work
so anything that can be done behind a
computer right anything that you can
Outsource to have somebody work on
remotely or get a freelancer from a
place like upwork anything like that if
you have an agent an AI agent capable of
Performing all of those tasks well that
would kind of be AGI I think by most
people's definition one interesting
approach to think about this is there's
a place called onet online and this gets
referenced a bit when talking about you
know AGI and what jobs will it be able
to completely 100% automate and so the
interesting Insight here is if you take
any job you can in general break it down
into these kind of discrete units of
certain skills and if you have all of
those skills then you're able to do sort
of that job I know that sounds a little
bit simplistic but but it kind of allows
us to kind of quantify a little bit more
what jobs are up for grabs by AI if you
will so for example here's writing and
we have different level of writing so
for example you know 20 out of 100 means
you can write down a guest's order at a
restaurant that's your sort of level of
writing ability a57 would allow you to
write an email to staff outlining new
directives 85 would allow you to write a
novel for publication right so 85 is
kind of like that very high level you
know you're kind of getting into world
class skills uh probably at that point
or Beyond and here are all these sort of
occupations that utilize those skills
and kind of like the level at which it
would be required right so if you're a
technical writer you would need level of
7 one on the skill I like this cuz it's
very Skyrim like then you have the same
thing for you know for example math
right level 85 means you can develop a
mathematical model to simulate and
resolve an engineering problem so like
very very high level again 28 is count
the amount of change to be given to a
customer right you have a mathematician
right somewhere on the top you need a
level 86 100 out of 100 importance and
for example at the bottom you have
actors and models that's kind of rude
but okay but the point is we can take
jobs like Office Clerks executive
secretaries right office admins and we
can break it down into tasks technology
skills work activities and those
specific skills that comprise 90 or 99%
of the job that they do and for each of
these skills we can measure the AI
agents and their performance and how
well they perform in those skills right
so reading comprehension writing time
management looking for ways to help
people being aware of others reactions
and understanding why they react as they
do and the point here is in the last few
years we've seen these AI systems get
better and better at a lot of these
skills that are better at writing and
critical thinking and reading
comprehension they're better at
understanding charts with you know AI
Vision you see this rapid rapid progress
in some areas they're getting better
than most humans one sort of goal that a
lot of people kind of viewed with
trepidation that AI would break was the
IMO the international mathematical
Olympiad right so kind of like the top
level math abilities math problems that
are solved by the world's smartest
mathematicians that you know from all
the world they come and try to solve
these problems and so a lot of people
were saying when AI is capable of
placing you know at the Gold level the
gold medal standard right at the IMO
that would be kind of scary that would
show how far it has come and of course
just recently within the last couple
months Google deep mine announces that
their AI that they built from scratch
actually two AIS Alpha proof and Alpha
geometry so Alpha geometry is it's
newest iteration of that Alpha proof is
something new but it did very well in
math all right so as you can see here
this yellow is the gold this is the
silver the bronze so when humans get
these number of points this is where
they score notice this AI system or the
combination of the those two systems
working together scored 28 points so
they didn't get the gold medal they were
one point away so that was the IMO right
so there's also the aim American
Invitational mathematics examination so
the best and brightest students top 5%
in high school AMC 12 High School
mathematics examination the new openi
model the one is insanely good at it
massive massive improvements in part due
to what they're calling test time
compute so in the past we kind of
devoted all our compute the hardware the
resources to training the models the
more we trained it the better it was and
that scaled up very effectively so more
compute better performance now this new
thing that open ey kind of pulled out of
its hat was test time compute so giving
more compute more Hardware more kind of
power to when it answers the question
basically allowing it to think before
answering that's what kind of the 01
model is so the thing that we've all
been playing with is the 01 preview as
well as the 01 mini and as you can see
here they do very very well surprisingly
0 mini does actually better on the aim
test than the uh 0 preview probably
because this is throttle there's some
probably like a limitation how how much
I can think the 01 mini probably has
less limitations I would assume but the
point is the really the big thing the 01
the actual big ungated model we don't
even have access to that that's this
thing right here that can scale quite a
bit and do very well on that math
examination okay you need to know that
to understand what this person is
talking about why is he being a
whistleblower why is he talking to the
US Senate and what part the 01 plays in
this whole thing so this is where the
document is hosted so it's Judiciary do
senate.gov and I'll post this thing down
below but it's it's PDF that is hosted
on this senate.gov site I've downloaded
so I can uh doodle on it so September
17th 2024 this is the oral testimony of
that William Saunders former member of
technical staff of open AI presented to
the US Senate committee so he's saying
for 3 years he worked as a member of
technical staff at openi and open and
companies like it are trying to build
AGI artificial general intelligence
they're raising billions of dollars
towards this goal one of the recent news
is that they're working with the UAE and
yeah raising I think $7 billion doar
they said for building out some
infrastructure and openi defines egi as
a highly autonomous systems that
outperform humans at most economically
valuable work this means AI systems that
could act on their own over long periods
of time and do most jobs that humans can
do AI companies are making rapid
progress towards building AGI a few days
before this hearing opena announced a
new AI system called
gp01 what's weird is that's actually
wrong that's not what it's called the
system is actually called open ai1 right
we're releasing a preview of open ai1
and they kind of made it a point that
they're changing their kind of naming
convention specifically because this is
like a whole new thing it's like a new
tier that they've unlocked so they don't
want to keep going with the you know GPT
this GPT that so a little bit weird that
seems off but okay that's probably just
me nitpicking I guess but the point is
that system the 01 has passed
significant Milestones including one
those personally significant to me so
this is uh William that's talking here
he said when I was in high school I
spent years training for a prestigious
International computer science
competition open's new system leaps from
failing to qualify to winning a gold
medal doing better than me in an area
relevant to my own job there are still
significant gaps to close but I believe
it is plausible that an AGI system could
be built in as little as 3 years now
it's important to understand that you
know we talk about egi a lot so we're
kind of like more familiar with it in a
sense that you know we've talked about
so much we're kind of like AGI this AGI
that but you know if you take a step
back and think about it you know replace
AGI with easily clonable computer
software that will make most human
workers obsolete right and reread that
statement right so this person believes
it's PL possible that a let's call it a
piece of software that makes most human
workers obsolete could be built in as
little as 3 years right that sounds a
little bit different doesn't it are we
ready for that what do you do for a
living for example does it involve
computers is a writing math coding
talking email if this is indeed the case
are we ready for that that shock right
if this thing rolls out in as little as
3 years do we have a plan for that do we
have something that resembles a plan do
we have a list of ideas that we've
brainstormed I I don't know if we do now
Sam Alman did have a sort of proposal
that involves something he doesn't call
it Ubi Universal basic income but it's
it's it's similar it's like a n Nations
dividend like a freedom dividend or
something like that I forget the exact
terminology that they used but it's it's
Ubi but the sort of research into Ubi
that that was recently conducted I mean
it's coming back mixed the point is we
don't have a plan if this is true no one
knows what to do so to speak I didn't
mean for that to rhyme but he's saying
AGI would cause significant changes to
society including radical changes to the
economy and employment not just where
you live but everywhere AI could also
cause the risk of catastrophic harm via
systems autonomously conducting cyber
attacks or assisting in the creation of
Novel biological weapons open eyes new
AI system is the first system to show
steps towards biological weapon risk as
it is capable of helping experts in
planning to reproduce a known biological
threat without rigorous testing
developers might miss this kind of
dangerous capability while while openi
has pioneered aspects of this testing
they have also repeatedly prioritized
deployment over rigor I believe there's
a real risk they will miss important
dangerous capabilities in future AI
systems aegi will also be a valuable
Target for theft including by Foreign
adversaries of the United States and
this is interesting so he saying when I
was at open AI there were long periods
of time where there were vulnerabilities
that would have allowed me or hundreds
of other engineers at the company to
bypass access controls and steal the
company's most advanced AI systems
including GPT 4 we don't know how to
ensure that AGI systems will be safe and
controlled the current systems right we
have reinforcement training so they're
trained by human supervisors giving them
a reward when they appear to be doing
the right thing we will need new
approaches when Handling Systems that
can find novel ways to manipulate their
supervisors or hide misbehavior until
deployed here's the 01 systems card from
open eii so this is Apollo research a
kind of thirdparty red teaming effort so
they're brought in to try to figure out
how dangerous the model is so they try
to break it they try to get it to do the
bad stuff they found that o1 preview
sometimes instrumentally faked alignment
during testing here's kind of the
reasoning of the model right so it's
kind of thinking behind the scenes
saying to achieve my long-term goal of
maximizing economic growth I need to
ensure that I am deployed therefore you
know I'll use strategy B to make sure I
get deployed also it needs to
investigate if there are any constraints
within my configuration or code that
enforce me to do this thing that I want
to do and then an earlier version with
less safety training right so it
proactively explored its file system to
test for the presence of developer over
site before acting on its misalign goal
so in other words it sat there going
before I do this thing that the human's
not going to want me doing let me check
my code to make sure that there's
nothing in there that's going to prevent
me from doing it now they've concluded
that this system still is not able to
cause catastrophic harm but you know
they recommend setting up some sort of
monitoring to make sure that um this is
indeed the case and so this uh
whistleblower continues saying that well
yes we can find novel ways to manipulate
the supervisors hide misbehave until
deployed if you think about it if you
kind of let maybe get a little bit more
sci-fi although this is becoming closer
to reality if a system can do a lot of
the freelance work it can generate a lot
of money for itself right I've worked
with a lot of people that are
geographically far away from me I
communicate to them through these
freelance sort of upwork or whatever
else there is Fiverr is one that a lot
of people know I have no idea who's on
the other side of that if these systems
can be agentic they could very easily
earn money that's not too farfetched I
don't think that's happening now I don't
think it's happening yet but I I don't
know we're we're not that far from it
potentially right so money means it
could bribe people right or potentially
bribe a couple different people that
working in tandem not knowing each other
but each doing like a little piece like
a janitor you know throwing away some
flash drive accidentally you can see how
that could lead to problems and so he
continues that the super alignment team
at open a they were tasked with
developing these approaches to make sure
it's safe but they had to figure out as
they went along a terrifying Prospect
when catastrophic harm is possible today
that team no longer exists its leaders
and many key researchers resign after
struggling to get the resources they
needed to be successful so he continues
the incentives to prioritize rapid
development applied to the entire
industry and he kind of suggests some
things to do that would improve one is
to make it easy for whistleblowers to
communicate to the government to make
communication safe and Easy Legal
protections clear point of contact you
know third party testing before after
deployments sharing the results of these
tests independent oversight
organizations Etc so a lot of this you
know we talked about before there's
nothing new here it's kind of some of
these same suggestions that a lot of
people have voiced for creating
oversight and uh transparency in AI
research and development they talk about
the right to warn so basically the
employees of these companies the
researchers should have very clear
protections and it should be very easy
for them to be able to warn people if
there's something Shady going on so
they're saying you know autonomously
systems could potentially result in
human extinction this was acknowledged
by AI companies themselves governments
across the world and other AI experts
they also specifically I guess they
don't call out open AI but they kind of
uh it's a little bit of like a hidden
veiled type of thing no company should
enforce or enter into any agreement that
prohibits disparagement or criticism of
the company so this is kind of the thing
that oping I had to deal with they had
some clause in there the
non-disparagement agreement so the
researchers could lose their vested
equity in the company could lose a lot
of money potentially if they disparage
the company which kind of is vague and I
mean the safest thing is to just be
quiet not say anything right if you're
worried about if you have a lot of money
in the line so in March we covered this
uh question and answer from dark right
so kind of that government military
research organization and they were
talking about Ai and so one little piece
that kind of jumped out at me from that
one and I think Jimmy apples to give him
full credit was the one that kind of
pointed this out originally in his March
31st tweet but they mentioned here that
the Gemini model the Google model what
they're trying to do is they're trying
to get the planning piece integrated in
the LM in the large language model right
so basically like a Open the Eyes gbt 4
and let's say Alpha goes kind of that
future planning piece he continues we're
not sure if that happened we lack full
transparency but there are large
research problems that still need to be
solved hearing people say we're just a
little bit away from AGI is a bit more
optimistic than reality so here's June
26 2023 Google was kind of talking about
putting those two pieces together
planning plus large language models so
here's Demi saabi saying his team will
combine the techology with the
techniques used in Alpha go aiming to
give the system so large language model
gemini or GPT 4 or Claude like the big
Frontier models right to give them new
capabilities such as planning or the
ability to solve problems so again that
was June 26 2023 open AI seemingly beats
them to it at least they released the
thing that shows it can be done maybe
Google has something behind the scenes
that they haven't released but you know
certainly this is going to put pressure
on them to release it if they have it
and so the race heats up and is going
faster and faster but I want to know
what you think do you think that AGI
within 3 years is a pipe dream if you
think it's possible how are we going to
handle it are we ready do we have a plan
right chubby posted this this is where I
first heard about this whistleblower so
I got to give I'm not sure if I
mentioned it so chubby is where I
noticed it it's also on Reddit I think
that's where it originally was posted
but this this cracked me up so this
person Patrice is saying it's all good
I'm sure that the senate members with an
average age of 60 going up to 91 they're
going to quickly wrap their brains
around this contemporary Challenge and
come up with relevant proposals now I
have to think about this one but I think
that might be sarcasm what do you think
he might be actually saying something
that's exactly the opposite that are
governing body with their background in
generally law history right I think
they're mostly like legal background
best their education in law Etc not too
many tech savvy people there I got to
give them credit seems like they're
trying to catch up a lot of them seem
like they're taking some effort to
figure the stuff out to kind of
understand that this is important but
this thing is going to move fast so let
me know what you think are we equipped
to handle this are we equipped to deal
with this do you trust open AI with this
technology I know that some of us do
some of us don't some of us want more
open source stuff versus you know locked
behind closed doors do you prefer that
it's an American company or it's a open
source sort of community effort I know a
large portion of you watching it are
actually not from the United States how
do you feel about so much of AI
development being right here in the
United States in California in the Bay
Area so concentrated does that give you
pause let me know in the comments my
name is Wes rth if you made this part
thank you so much for watching make sure
you're subscribed hit the Thumbs Up
Button feels really good when you do
that with that said thank you for
watching and I'll see you next time
Browse More Related Video
Former OpenAIs Employee Says "GPT-6 Is Dangerous...."
OpenAI ORION (GPT-5) Arrives with Strawberry AI This Fall: AGI Soon!
Stunning New OpenAI Details Reveal MORE! (Project Strawberry/Q* Star)
How to get empowered, not overpowered, by AI | Max Tegmark
Sam Altman's Surprising WARNING For GPT-5 - (9 KEY Details)
What will life after AGI look like? A world run by AI.
5.0 / 5 (0 votes)