Okay, I'm going to switch mics here. Is this one on? Great. I have a hard time standing
still, so great to see you. So if I mess up or it makes a lot of noise, just let me know.
So my name is Craig Martel. I am the chief digital and AI officer at the Department of
Defense. It's a pretty cool role. I'm responsible, I'm the chief data officer, the chief analytics
officer, and the chief AI officer for all of DOD, which means I go to, can I swear? Am I allowed
to swear? Okay. Which means I go to a shitload of meetings. Fuck yeah, exactly. Please don't put
that on YouTube, my buddy.
My boss is the deputy secretary of defense, and I don't want her to see that. So we will edit that
out. But the meetings are really cool because we realized very quickly that we had to build a
hierarchy of needs where we had to say, look, AI is going to come. So this was last year before,
here's a precursor for my talk, before all of the hype about large language models, I said it
was kind of the crazy, silly, absurd hype about large language models. We built this hierarchy of
needs. And first it was get the data right. And if you get the data right, and that's really hard in
some places as big as the Department of Defense because it's the world's largest organization, and
it's distributed, I don't know, everywhere, including space. So getting the data right is really
hard. But if you get the data right, then AI is going to come for free. That was the story I
used to tell. Now the problem is people want AI for free, and they shouldn't. So we'll get to that
part in a second. The second layer was analytics. Because the term AI is hyper overused. Hyper
overused. And in my job, AI often means I want magic that will magically do my job and allow me to
get rid of things. And I'm not going to get rid of things. I'm going to get rid of things. I'm going to
get rid of people who are too expensive. That's what the phrase AI means at the DOD often. But
really what they're asking for is just to see where their crap is. They just want an analytics
dashboard that will let them know, here's my stuff, here's my troops, did that move, when is that
stuff showing up, and I can easily see that. So a lot of the calls for AI, at least in the Department
of Defense, aren't really calls for AI after all.
They're simply calls for high quality data and a good dashboard. So we said quality data, and then we
said an analytics layer. And then finally I put at the top of that hierarchy AI. And what I want to
provide for AI for the whole Department of Defense is what I call AI scaffolding. So the model,
building of a model is now a commodity.
Ten years ago, 85% of the problem was the model. Today, 5% of the problem is the model. And the other
95% is the pipeline in which that model sits. And the generation of that model is a commodity. You
send your data to one of the big players, and they'll give you back a model. And that model will be
good enough. I feel like a wedding singer. And that model, I can't sing. And that model will be good
enough.
So really, we've been thinking very hard about what goes around that model. On the left-hand side of that
model, so before you build the model, is high quality labeled data. Nobody wants to label data.
Labeling data is the hardest thing in the entire pipeline. But it's the single most important thing in the
pipeline. High quality labeled data means way more than the algorithm. The algorithm is mostly irrelevant
at this point. Choose one of ten.
Choose one of five vendors, and you're going to get a good enough model that solves your problem. Maybe not perfect, but good
enough often. Good enough to sell shirts and make billions of dollars for some companies, right? Good enough to return
documents and make billions of dollars for other companies, right? Good enough to recommend jobs. I used to work at
LinkedIn. Good enough to recommend jobs and make billions of dollars.
So the model itself has become less relevant. Getting high quality labeled data is really hard. So we wanted to provide data as a
service. And then on the other side of the model is monitoring.
So what is AI? Somebody tell me what AI is.
.
It is not a panacea. That comes later in the talk.
.
Cheap something parlor trips to fool muggles to think there's intelligence in the model.
in a box. Didn't I teach you that? Didn't you come to my class and I taught you
that? Well done. We'll get to that. He said plausible sentence generators. AI is
statistics at scale. That's it. We measure the past, we track the past, we use data
from the past, and we use it to predict the future. It's no different than when you're
driving the car trying to figure out whether that bicyclist is going to turn in front of
you or not. You've used your past experience, you look at the features based upon the way
that bicyclist is riding, and you make a decision about whether you think you're safe to pass
them or not. We use the past to predict the future. What's the problem? That works really
well, right? That works extremely well in lots of contexts. But it doesn't work very
well when the world changes, right?
Because we've predicted the past based upon a particular set of features in the world,
and those features change. So we want to think about model monitoring on the right side of
the model to ask, is this model still delivering value? Now, industry really has some great
wins here. Why? Because users of a shopping site or users of a search engine label the
data at scale massively millions of times per minute. You go in there and you buy the
shirt or you don't. You click on the document or you don't, and they can track. We did this
at Lyft, we did this at Dropbox, we did this at LinkedIn. They can track over time that
the model is becoming less and less effective, so it's time to retrain. It's time to take
samples from the world and train again.
That's the...
They get that for free. They get that from us, just interacting with their products.
But in the DoD, that's extremely hard. So we wanted to actually... We're working very
hard to build in model monitoring into the specification of models that we should buy
so that we can at least tell, has there been a distributional change in the labels that
the model is generating?
Plus, let's be clear. In warfare, the world changes pretty quickly sometimes. The world
doesn't remain stable.
So in lots of things that we do, or not even just in warfare, what if you're on a ship
doing search and rescue and suddenly clouds roll in and the waves get bigger? Well, your
model is going to be very different in that case than it would be in a case where the
seas are calm, right?
So we, the DoD, more than I think most folks, live in a world that changes very quickly.
And so we really want to build models that have high-quality label data and are monitored
on a regular basis.
To make sure they're still bringing value, because we usually pay millions and millions
of dollars for these things.
All right. That was the plan, and that's still the plan. That's what we're doing. We're building
on a hierarchy of needs. You can go back and look at any of my public talks. I talked
a lot about the hierarchy of needs.
But then large language models came along.
And my life got really tough.
.
How many people, how many people use large language models on a regular basis
in the course of their job? How many people don't use large language models on a regular
basis in the course of their job? Wow. That's fascinating. For those of you who don't, scream,
just yell out. How come?
.
The government.
The government.
.
.
I'm here trying to build bridges. Here I am. I'm trying to build bridges. I showed
up. What do you mean by the government? She can't tell me.
So not just the government, but everybody. All kinds of bad players, the government and
others. All right.
So first thing you need to know is when you're doing this thing, it's a lot Like
Okay. What else do other folks not use large language models?
Jadon Johnston?
They're full of shit.
Jadon Johnston?
with that is a great use case where are you from originally sir Germany and so
can you type in in in Deutsch and get good English out okay so that I think
that's a so what he said was he right he's he's German originally and oh he's
probably a Deutscher now still in his own mind right so okay so he's a
Deutscher living in the United States and he writes it in in English with
that's that's stilted as a German English I'm assuming and it comes out
much more fluent that is an absolutely fantastic use for large language models
and I support that one completely okay other folks who use them wait right here
no ads so use it for like information retrieval okay
it makes old coders relevant again did you say old coders oh because you can
learn go really fast or learn rust really fast okay so I also think that's
a really good use case for large language models is coding so I'm gonna
let me let me throw something out to you you can tell that I used to be a
professor this is what I would do all the time why is coding a good use case
for large language models why is coding a good use case for large language models
but the free generation of text less of a good use case well structure what else
you let someone else have a chance you're in bit which one the language you
can verify that it works how yeah run it so the beauty of the both of those are
the same answer and I think they're right the beauty of text from a large
language model sorry whoa code from a large language model sorry whoa code
from a large I'm tired from a large language model is that code code is
verifiable it's a formal language we are there are external tools that you can
use to verify that code not I say you're shaking your head not completely
not completely but it has to compile if the code doesn't compile it's it's not
going to run by definition that's a definition of not running if you're a
good software engineer then you'll write a unit test and those unit tests are
part of your larger system now you might be a crappy software engineer in which
case you wouldn't have written a unit test anyway so I'm not that bothered by
that and if you work at a company that has good infrastructure there are
integration tests so your code actually gets put put back into the system as a
whole and I'm not a security expert but however you scan code for security
flaws you can scan this code for security flaws as well I'm happy to be
wrong about that you can come yell at me later if I got that part wrong I really
am fascinating about fascinated about the pluses and minuses of scanning for
security flaws and whether whether large language models are better or worse I
think that's a fascinating question I don't know the answer to it but I'm
saying in principle I I mean I'm support in support of code generation as part of
large language models okay why am I not as much in support or I'm not quite as
sanguine quite as optimistic about
large language models generating big chunks of text people don't proofread
okay the the slide says well everybody get everybody understands the shall we
play a game oh hey Ray where's Ray somebody made me a row row made me a
model of the whopper from from for more war games one of my favorite movies as a
kid okay
so so here's what I found I find fascinating about large language models
and human psychology we evolved we evolved to treat things that speak
fluently as reasoning beings now we also have other ways to judge whether
someone's rational if we watch them try to get out of a if they go into the
closet five times and can't find the door the the door out of their hotel
room maybe they're not rational if they're not rational they're not rational
so we have other triggers about rationality behavioral triggers about
rationality but one of those oh yeah Ray can I can I see that thanks right one of
those behavioral triggers for rationality is fluency of language
thank you so just shall we play a game I wanted to show it robo thank you Roro
Roro built this for me and if you press it you can clap I think you should it's
good
I can't say what does that say
pro-ther recreate vision
oh yeah do you want to play thermonuclear
war okay well we it's a prototype it's a good prototype okay so we've we
evolved to treat beings entities that speak fluently as if they reason oh man
Okay, and so we have this training data.
Actually, let me take a step back.
A large language model is, in fact, a language model.
What does a language model do?
Say it again?
Actually, it predicts subwords, but we'll pick...
It predicts the next word given the prior words.
That's all it does.
Language models have been around for decades.
My whole academic career, industrial career.
What's a good use of a language model?
Autocomplete is a great use of a language model.
I don't understand that one.
She said data classification.
No, no, just a language model.
Forget the large part just for a second.
Just a language model.
Okay, so here's a good use for a language model.
If I am a translation system and I translate German into English,
there's maybe five words which are semantically similar that I could choose from.
I choose the one that's most likely given the prior sentences.
So the language model becomes part of the translation system.
Here's another one, speech recognition.
In a noisy channel, which you have all the time, one of the words might be garbled,
so you can use a language model to predict what that word might have been given the prior contexts.
So all language models do are predict the next word given the prior words.
That's it.
That's all they do.
It's really important.
What does a large language model do?
That's it.
It's all it does.
The last function for a large language model is did I predict the words that I dropped out of a sentence?
Did I predict them correctly?
So the only difference, ooh, people are going to disagree with the only part there a lot.
People in industry and in academia will disagree a lot with the only part there.
But the major difference between an old-fashioned Bayesian language,
which I taught you, and a massive new large language model is the size of the context so that we call them prompts now,
the size of that window, right, and the computational power behind it so it can ingest massive amounts of data.
So you can think about a large language model.
It's not 100% mathematically correct.
But you can think about a large language model as just one big statistical engine that says,
given all of the past context, which includes the prompts, prior answers, and what I've set up until now,
what's the best next word?
Okay, and now you know why large language models do two things.
One, they seem really fluent.
Because you can predict a whole sequence of next words based upon a massive context that
makes it come out sounding really complex.
Also,
transformers had a big, we can talk about transformers in a second,
but also transformers had a big part in allowing not just the prior few words,
but words a long time ago to have an impact.
So the math and the engineering has just really been optimized to produce fluent text.
Okay, but now you also know why these things hallucinate.
And you also know why that I think you're wrong if you think they reason.
Because the loss function, the way that you think they reason, is wrong.
The way they were trained was only for fluency.
There was nothing about these things whatsoever,
nothing about these things whatsoever,
is this translating me?
That's a great use for a language model.
If it doesn't understand what I said,
because I said a, a, a a lot,
let's see, I want to see a, perfect.
It can use a language model to help guess what,
what the word that got dropped out is.
Okay, great.
Now I'm going to stare at that thing the whole time.
I wish I didn't know that was there.
Okay, so these things do not reason is my conjecture.
The reason I say they don't reason is there's nothing about the way they were
trained to constrain them to reason correctly.
The only thing, the only constraints on training for large language models is fluency.
And then we as humans, I believe, are duped by the fluency.
Now, plenty of people in the audience are going to tell me all kinds of really valid use cases,
and I'm excited about that.
So I'm not saying duped in the sense that we're not going to get great use cases out of this and
great technology that we can use in all kinds of places.
I'm going to tell you a task force that we just launched that's going to figure out exactly that.
Where will these things be valuable?
But in a sense,
we're lucky if they land on a fact, not unlucky that they hallucinate.
I'm surprised they're truthful at all as opposed to being surprised that they hallucinate.
Hallucinate should be the norm because all they do is generate, all they were trained on is generating coherent text.
Okay, so now what I really think happens is because we evolved to think that those who speak fluently
have reasoning underneath them,
we believe that the, here's what I really believe,
we believe that all of the training data that was used to train a large language model
has captured within it the patterns of reasoning.
That we believe that there's, we tacitly believe that there's so much training data of fluent speech
that it captures within it the patterns of reasoning so that the large language model itself has somehow encoded the patterns of reasoning.
That's what I think we tacitly believe.
And to me, that's an empirical question.
Whether, whether these things are going to be able to reason or not and so far they have not borne themselves out.
Okay, how many people
tried to hack LLMs earlier today?
Okay, how many people succeeded in getting it to do something bad? What happened?
You switched to Germany and it told you secret things.
Why do we, why is there a lot, first of all, let's ask the administrators.
Why is there a large language model being used that has US secrets in it?
Let's not show that to the world either.
Did it really show, I mean, it wasn't actually secret, right?
I see.
Okay.
see. Anybody else? What did you get it to do? Write malicious scripts. Interesting.
Did anybody get it to tell a lie? What'd you get? What'd you do?
You asked the first what?
First computer game for girls. What's the answer, actually?
So here's what happened.
here's what i got today i think that's a great example um anybody see the movie the usual
suspects it's a great movie right it's a great movie uh who did stephen baldwin play
michael mcmanus just okay so i asked it who was craig martell
of course i'm a little i'm i'm self-centered enough to do that and it told me that craig
martell was the character that stephen baldwin played in the usual suspects
i didn't know that was the case i went back and checked i wasn't in the movie
i wasn't portrayed in the movie it's too bad so so i you could say i don't know why that happened
why did it reason like that that's the wrong question just for some reason the prompts
generated those strings i don't know why it was the state that it was in but it's
you really
got to see that these things are just string to string and even when you ask it what are you
thinking or why did you this this chain of reasoning questions why do you believe why
did you come up with that answer it cannot introspect folks that's also just a unicode
string to a unicode string function what what was your chain of reasoning is a unicode string
to a unicode string function and if its output had anything to do with its internal state its
a unicode string thatركets it can scarf through anything and you'd be ready to does
it out there that's what we call our wing traveled unicode you get a string of string
then of course you get a string and it just goes back and then itlocals makes getsカ
for pineapple black there or through what your brain is like well let's go back and do these
things just ah we have to try things with our brain because it might remain like we did
before um and so we really didn't something that takes a lot of things and that's what the
primary reason I'm actually going to pulled together on a little one because dewayne did an
amazing thing i think we make a really good work on the comp serrat theory of these concepts in the
Instagram course so they weren't in huh this section here we looked at aloma but the two
that AI does not equal LLMs. I have a t-shirt that my team just made for me
that says that but I'm waiting for my belly to go down before I wear it. And we
have a bunch of stickers that says AI does not equal LLMs. What do I mean by
that? Yeah, the hype cycle has just decided that large language models are
the only thing, right? But that's unfair because large language models are
unproven, and we want to help prove them out. I actually think as a scientist who
studies natural language processing, I think LLMs are the most exciting thing
that's happened scientifically in my entire career. I just think there's a
really large gap between LLMs as a scientific enterprise and the products.
That's really my frustration. My frustration isn't in the science, but
the underlying part of the underlying technology in large language models is
transformers, right? The G, the T and GPT-1.
GPT is transformer, general-purpose transformer. I think that's right. I think
it's general-purpose, but I know for sure the T is transformer. And so
transformers have made the world radically better. Transformers are why
you can talk to your phone as well as you can. Transformers are why image
recognition works as well as it does. Image recognition works really well,
like fascinatingly well, right? Transformers, deep learning and
visualizing. Transformers are why you can go to a country that doesn't use
Roman script at all. My wife is Thai, so I'll use Thailand as the example. You can
go to Thailand and you can hold up your camera and suddenly the script turns
into English that I can read. Transformers, right? I can go to
Deutschland and hold it up and get it mostly right, right? So transformers have
been really amazing. And so large language models have benefited from
transformers and they are also really amazing at generating fluent text.
Here's my frustration, though. If I show you a picture of a dog and tell you it's
a cat, what's the cognitive load that you need to decide it's not? Do you want
to just change? Like, I tell you it's a cat, you can decide extremely
quickly it's not a cat, it's a dog, right? So you can evaluate the output of that
system. So take a image recognition algorithm and embed it into a
pipeline with human beings and let's say it gets it right 70% of the time. That
might sound scary depending upon the use case, but it's really not because for
most image recognition tasks you can say right away yes, no, yes, no, yes, no, yes,
no, yes, no, and a human being can be in the pipeline to say false positive, false
positive, false positive, and just take it out of the queue. So there's been
really massive growth because of transformers and things like image
recognition and it's allowed, it's allowed
for a really cool human machine interaction and it's really changing the
world. I'm super excited about it. However, if I generate 30 paragraphs of
text, how easy is it for me to decide what's a hallucination or not? She's
telling me to change the slides. Thank you, Kathleen. How easy is it for me to decide
what's a hallucination or not? It takes time. You also often want to use large
language models in a context where you're not an expert. Like that's one of
the real values of a large language model is asking you questions where you
don't have expertise, right? And so how likely are you, a non-expert, to be, to
read through that whole document? First of all, I fear a lot of uses, a lot of
desired uses of large language models are to create massive amounts of
text that I don't have to read so my job can be easier. Okay, that's a fear.
Witness the lawyer who submitted that brief with ten phantom cases, right? So, so
setting aside natural laziness, I, none of you, none of you guys are lazy, right? So
even if you wanted to go through and proofread all of the text generated, how
likely are you be able, are you able to determine what's a hallucination or not?
Highly unlikely. So my fear, my concern is a better word because I think these
things are great. My concern is that the, the, the thing that the model gets wrong
has a high cognitive load to be able to determine whether it's right or whether
it's wrong. Okay? So I'm a big fan of the science of large language models. It has
changed like, as a natural language processing researcher, it has pushed the
decades forward. I'm concerned with the products being sold. And that gap in between. Okay?
Any thoughts or questions? I'm really curious to what this audience thinks.
Yeah, so the question is, why should a human validate the output? Why not have other AI
models validate it? You can't have turtles all the way down. You can't. If you show me
that some other model, which probably had to be validated by humans, is that good, and
it can validate this other model, here's the way I think about it. Here's a big black box.
You could have five language models inside of it or one. A question comes in, an answer
comes out. I just want the answer to be true.
So, if you can build me a validatable system where I can measure the correctness of the
output text to test whether that box is good, and you use other models to validate it, I
am all for it. But at some point, a human being has to judge the box. It can't be turtles
all the way down. If it's turtles all the way down, we're just giving up our scientific
responsibilities. Right? And we can't give up our scientific responsibilities.
Yeah, and he's way smarter than I am. Yeah, so I have the intuition. Let's start with
I am an empirical scientist, and I do not mind being proven wrong at all. It won't bother
me at all. I would love for us to solve the hallucination problem. Let me jump to the
end of the talk.
What I really want is a set of use cases, classes of use cases, kind of like level
1 through level 5 autonomy in cars, where we have increasingly difficult acceptability
conditions and metrics to go with those acceptability conditions. And then you can just show me
here. I have proven to you there are no hallucinations, and it works for this use case. Or I've proven
to you that there's 80 percent hallucinations, and that's fine for now.
Or I've proven to you that it works really well
for first draft generation. And our acceptability conditions for first draft generation are
lower. I'm onboard that. So what I really want to push the community towards, and I'll
talk about Task Force in Lima in just a second, which is doing exactly that. What I really
want to push the community towards is a set of classes of use cases that are increasingly
difficult that have acceptability conditions and metrics to measure them against. And what's
really interesting in all my conversations with the big companies
I've had fairly similar conversations I think everybody agrees we need those
evaluation metrics it's right now we evaluate it by going ooh isn't that cool
or we read some of it we read a paragraph or two and go oh sounds good
that's just not acceptable to me I am totally okay if these things work in
fact I really want them to the scientist in me really wants them to the person
who runs AI for the Department of Defense wants to make sure that if
you're a soldier in the field asking a large language model a question about a
new technology that you don't know how to set it up I need five nines of
correctness I cannot have hallucination that says oh yeah put widget a connected
to widget B and it blows up right so so there's there's this part of me over
here that's a scientist
who is freaking out happy and there's part of me over here who's responsible
for determining when we accept things in the Department of Defense and right now
the community is going ooh isn't it cool so over here I'm saying ooh isn't it
cool over here I'm saying ooh isn't it cool isn't enough tell me what the
acceptability conditions are and tell me the metrics we're using to measure it
don't care if you use another model check in another model awesome don't
care if you use humans in the loop don't care
show me the input-output pairs and show that we get there right the only the
reason I play the chicken little role because I think I think what I when I'm
interpreting your comment is is you're being a little bit chicken little and
but the reason I play that role is because I really want to de-hype this I
really want to turn it from here is this monolithic technology designed to solve
all of our needs and I want to bring it down to here's a set of use cases and
evaluation for those uses and I want to review them and say is this a good place
cases and some technologies will work for some use cases other technologies
work for other use cases I want to make it boring and measurable that's really
well that's that's really my goal yep
I didn't quite hear is it here's what I heard tell me if I got it right in the
absent of good data can you still do things no sorry
can you stand up and yell
sure it's just that you got it to do that you got to take a baby we got to
invent a baby and here's the fundamental flaw with the way people think about
this babies come into the world with all kinds of cognitive apparatus all kinds
they they can recognize faces they know hunger they know causation they know
they have numeracy they they can
tell numerical they can tell numerical things I have 10 minutes left okay they
have they come into the world with all kind so we can't just assume a randomly
set network and by the way an artificial neural network isn't the brain that's a
whole different conversation but it's simply a set of logistic regressions or
pick your pick your activation function connected together it's got nothing to
do with the brain right it looks similar in the way you draw it on the board but
you can trick you could turn those circles and arrows into a
equations that look nothing like circles and arrows, right? Circles and arrows are
imagery. Calculations are all that's happening in a computer. So you can't see a randomly set
network to be anything like a child's brain. I think that's the fallacious piece there.
Does that make sense?
Way over there.
Yep.
Okay.
Okay, that's awesome. Let me echo his comment so people hear it. His comment was,
what about those damn kids in the next generation? That's basically what it was.
We already have a hard enough time. We're trying hard to find hallucinations ourselves. We already
have a hard enough time with it, hard enough time doing that. What about the next generations who
get used to using these machines and the machines tell them that they're telling
the truth?
Okay. Twofold. One, I think you just nailed something that's extremely important.
And what you nailed that's extremely important is people have a proclivity to believe authoritative,
fluent speech. And that's one of the problems with these models because they're trained to speak
fluently and we defer to them. So that's a real problem. Second one is how about we draw a line in
the sand and we don't ship these damn things until they stop hallucinating? Or that we have ways to
mark their hallucinations? Or industry? How about we have ways that facts have weights on them
where we think it's 70% correct? There's all kinds of things that we could push the community towards.
And I want to. My objection is this free-flowing text and telling everybody that
that AI's arrived. It hasn't.
Oh, sorry. This woman in the front.
She's in the Screen Actors Guild. We should all give her a round of applause for being on strike.
Yeah, so that's. So she's saying we want that.
The studios want the first draft to be written off of your intellectual property,
but written by our AI. And then we'll hire you to edit it.
Keep fighting, sister, is the only thing I got to say there. Like,
I think. First of all, I think if the studios win this battle, we're all going to see really
crappy movies. Just really, really, really crappy movies. Because what's that? Too late.
There's some good movies. I just saw Big George Foreman. I loved it. So if these models regress
to the mean, they regress to the mean very quickly. And so I really am not excited about a world
where the mean movie is constantly generated over and over and over again. Ask it to generate
a bedtime story for your kid and tell me if your kid likes it. It's very unlikely to be the case.
I did that. I did that for this only audience only. I would have said text to text anywhere else.
Sure. Yeah. So what he said was I'm in the D.O.D. He reminded me that I was in the D.O.D.
And then he told me that Palantir, I got five minutes. He told me that Palantir
is a really great contractor and partner of ours, is developing a system where he was showing how the
large language models was developing plans and all sorts of things. Okay. The short answer,
then I want to talk about Task Force Lima. The short answer is I talk with the CEO of Palantir
all the time. He completely understands my concerns and he's on board with use cases and
evaluation, I mean acceptance criteria. So I'm totally cool with companies doing all of this
research. It's incumbent upon us, and I haven't asked for all of y'all, it's incumbent upon us to
say we will accept it if and only if. And that's Task Force Lima. So Task Force Lima
was just announced yesterday and it's a D.O.D. wide effort, but it's being run out of my office
at the CDAO.
It is incumbent to say, what is UL, weep UL. That's great.
It needs a better language model. Task Force Lima's job is twofold. First,
to understand the demand within the Department of Defense for generative AI and large language
models in particular. To think hard with those who are making that demand, whether we could actually
field one of these systems to solve it. And what are the risks and how do we mitigate it?
So for example, first draft generation, no problem. Connected to something kinetic, absolutely not.
Imagining a war gaming situation, maybe because human beings would read it
and validate it before sending it up. So we're thinking really hard, sort of, what is this ladder
of demand and what do we need the system to be able to provide? And can we build a human machine
teaming that could mitigate against the risks? So that's one use case. One thing that we're doing.
And the second thing that we're doing is working with industry, as I said a bunch of times here,
to build out these use cases and acceptability conditions so that industry as a whole,
academia industry and you guys as a whole, can start pushing us towards meeting those acceptability
conditions.
So now I've been completely negative the whole time. Now let me be clear. I would love for the promise of generative AI and large language models to come to pass. I would love it. But we have to be non-hype, systematic, and empirical about it, and measure that they're actually delivering against the acceptability conditions.
Let's be scientists. Now what do I need from y'all? Hack the hell out of these things. Hack the hell out of them. Tell us where they're wrong. Tell us how they break. Tell us the dangers.
I really need to know, because I just have strong intuitions. But you guys can give me some facts.
And on that note, if you go to...
If you go to dds.mil, you have to have the www, sorry. If you go to www.dds.mil slash taskforce Lima, you can click on that work with us, and you'll send us an email.
And there's a template in that email about the kinds of things you might want to do to help us understand where these things are broken.
Why am I here today? I'm not here today to be Professor Craig and teach you about large language models. That was just a positive side effect for me, because I like to hear myself talk.
I'm here today because I need...
I need hackers everywhere to tell us how this stuff breaks. Because if we don't know how it breaks, we can't get clear on the acceptability conditions.
And if we can't get clear on the acceptability conditions, we can't push industry towards building the right thing so that we can deploy it and use it.
I am out of time. I have one minute. Any other questions? Comments? Anybody want to yell at me? You're welcome.
Ah, that's a great question. Oh, God, that's a good... I only have a minute. Oh, my God, that's a good question.
So she asked, how do you measure something like Google search? That's great. We already have cognitive apparatuses that allow us to do that.
We go and we read it, and we look at the URL, and we check it against the other knowledge in our head, or we go to another link and see if they are coherent.
We already have built up...
Google search was 98, I think. We've built up since 1998 these tools that allow us to do this.
They don't speak fluently and authoritatively. You go and you read it, and you extract facts like you have been trained your whole life to extract facts from documents
and use your reasoning and your judgment to decide whether it's right or whether it's wrong.
Sometimes you're right, sometimes you're wrong. But we know how to do that. We know how to take responsibility for doing that.
But these...
These things that speak fluently... I am done.
These things that speak fluently and authoritatively are creating that next generation that the gentleman over here is afraid of where they just go,
uh-huh, the machine told me.
Let's work hard to not ship the system until the machine tells us the right things.
We can do that. We're scientists. We don't have to give in to the hype.
I'm done.