AI transcript
[MUSIC]
>> Hello, and welcome to the NVIDIA AI podcast.
I’m your host, Noah Kravitz.
One of the big transformations being
enabled by AI is the way we create software.
From coding co-pilots to
in-development systems built to translate
plain language requests into fully functional applications,
generative AI is fueling a new wave
of tools to help us create software faster.
Our guest today is Kanjun Kew.
Kanjun is co-founder and CEO of Imbue,
a three-and-a-half-year-old company somewhere,
and they’re founded in 2021,
that is building AI agents that work with us to
translate our ideas into code and bring them to life.
There’s a lot more to it than that,
but why hear it from me when you can hear it directly from Kanjun.
Kanjun Kew, welcome.
Thank you so much for taking the time to join the NVIDIA AI podcast.
>> Thank you, Noah. It’s great to be here.
Let’s talk about software development and AI,
and Imbue, and all of that good stuff.
But maybe let’s start with agents.
Agents are, I don’t want to use the word “hot”
because I don’t want it to sound fluffy, right?
But agents are a thing right now.
We’ve had some guests on recently talking about
agents in different contexts,
and Imbue’s approach to agents is something worth delving into.
Maybe we can start there. What are AI agents?
Why do we need them? Why isn’t Imbue working on them?
>> Yeah. Agents are really popular right now.
So Imbue was founded in early 2021,
and at that time,
our goal was to figure out how to make and work on general AI agents.
At that time, people thought we were totally crazy.
Like, what are agents?
Everyone’s working on AGI.
AGI is going to rule everything.
But what we were really interested in is,
how can we have systems that we as people can mold,
shape to what we’re wanting?
As opposed to, oh, this external intelligence that knows all the answers.
And so we started as a research lab at that time
because the technology was certainly not good enough
to build general agents that could reliably do anything
that we wanted them to do.
And, you know, in the very beginning,
we thought of agents in a similar way
to how a lot of people think about agents today.
As these kind of, you know, think about what an agent is.
Often people think about kind of an autonomous,
personal assistant type thing.
You ask it to do something, it does it.
For you, it comes back to you.
Now everyone has their own personal assistant.
And actually, a lot of our learning has been
that that’s a really tricky user experience.
And we experience it ourselves with human agents,
which is that often when I delegate something,
it comes back, it’s not quite what I wanted.
And now I have to negotiate how to get what I wanted.
I was listening, I told you before we record,
I was listening to a little bit of your fireside chat
with Brian Cutton-Zera from GTC this year,
which listeners go check that out.
It’s a great listen.
And you were talking about the difficulty,
and I can relate to this so much,
the difficulty inherent in delegating work
to someone else, right?
And to your point, thinking of it as humans,
you have to break the problem down,
you have to sort of figure out,
well, how much do I tell them exactly what to do?
– Yeah.
– Yeah, all that, yeah.
– What context does it need ahead of time?
What instruction should I give?
Delegation is actually a really tricky paradigm
because it actually puts all the onus
on the person who’s delegating
to define the problem, define the scope.
Of course, the person who’s being delegated to the agent
might come back with some questions and stuff like that,
but it’s a very tricky thing to trust.
So what we’ve learned over the years
working on general agents is we’ve actually started
to think about agents in a very different way.
And this is both from a kind of pragmatic business,
like user perspective and also from a mission perspective.
The way that we think about agents is,
if you think about what an agent is,
what this personal assistant is doing,
what it is is it’s kind of this intermediary layer
between you and your computer.
And you’re kind of telling it stuff
or interfacing with it in some way,
whether it’s a UI or a natural language,
and it’s interfacing with your computer.
And the most effective way to interface with your computer
is by writing code.
That’s what your computer is made of.
And there are really kind of two types of code.
There’s like everything is hard-coded,
which is the default for software today.
Everything is hard-coded.
So now your agent can only do a subset of things.
Or now with language models, you can generate code.
You can write code that does new stuff
that’s not hard-coded.
And now you have an agent that’s much more general.
It’s able to do sets of tasks
that I didn’t program into it ahead of time,
but now it’s able to do.
And so the way that we think about what agents are
is actually as this intermediary layer
between you and your computer,
and it’s essentially a layer of abstraction
on top of programming that allows us as regular people
to be able to program our computers
without even thinking about what we’re doing as programming.
And so we think of our mission
as essentially trying to reinvent the personal computer
and really deeply empower everyone to be able to create
in this computing medium of the future,
because this digital medium
is becoming more important in our lives.
And what we want is actually not what I want,
at least personally,
is actually not a super centralized assistant
that someone else has decided
is able to do this, integrate with that.
What I actually want is something that I can make and own
that is mine, and that does what I wanted to do,
kind of serves me.
And today we’re kind of in a world
where like all of our software is rented,
it serves other people.
So that’s kind of what we think of what agents are
is this layer of abstraction on top of programming
makes it so that it’s very intuitive
for everyone to program.
And that actually requires quite a bit of invention.
So can get into that in historical context.
– Yeah, well, the way you were,
we’re describing it,
I know you weren’t describing the sort of AI,
the assistant, the agent layer,
all the A-words, AI, assistant, agent,
you weren’t describing the agent layer
as replacement for a user interface.
So you mentioned, you know, UI,
but that’s kind of what I was thinking of,
right, that abstractive layer that’s kind of in between.
So how is imbue approaching it sort of on the ground
and working with you or primarily with enterprise clients?
– No, we primarily with, I would say it’s prosumer.
So people who are,
so the way to think about what we’re doing is,
instead of thinking of agents as automation systems,
right now we’re in kind of the agent,
automation, agent paradigm,
we think of agents as collaborative systems.
So how do we enable a system that empowers me
and helps me do more of what I want to do and work with?
I can work with it.
And so in the beginning,
we’re actually, you know, enabling you to write code.
Right now, these models are not that good at writing code.
As they write code, you actually have to go check it.
So you have to understand how to write code to go,
to see, oh, did the agent do a good job?
But over time, as these models get better at writing code,
now you don’t have to check it anymore.
And so in the beginning,
we start with software engineers as the prime,
or we call them software builders.
So you don’t have to be an engineer.
I’m a software builder.
I’m not a software engineer.
We start with software builders who can’t.
– I’m a prototype builder then.
I wouldn’t go as far as software.
– Okay.
So you could probably be a soon to be user.
Once we get to a place where you don’t have to read
and write, touch the code so much.
Right now, we’re targeting software builders
who can read and touch the code and be like,
okay, well, it’s not quite right.
We want to adjust it in this way.
And over time, as the models get better,
and you don’t have to be so low level in the code,
now more and more types of creatives,
types of builders can use these systems
to really mold and shape their computer
to what they want it to be.
– And what level of complexity,
what level of depth is the actual software
that users are building within view?
There’s the issue of accuracy, obviously,
as you were saying they’re not,
none of the models are creating 100% perfect code yet.
But I also wonder how complex can these things get?
And that kind of brings us to talking about reasoning.
We don’t have to get there yet,
but kind of edging towards that conversation.
– Yeah, I think one of actually,
our biggest learnings has been that if as a user,
I have a code base that is fairly modular
and I’ve kind of broken things down,
then the model is actually pretty good at dealing
with a very complex system,
because it doesn’t have to load that much stuff
into its head and it doesn’t have to cross check
all of these dependencies.
So just like with humans building software,
where you don’t want a ton of dependencies,
also, if you have a slightly more isolated system,
it’ll do a better job.
Similarly, there’s a lot of kind of user expertise
in using this kind of product.
So our product, it feels very collaborative.
It’s almost like a document editor
and kind of like interfacing and interacting
with an agent in that way.
And so as a user, you can basically,
we learn to give it tasks,
it’s more likely to succeed out.
And we learn to structure our work
so that we can delegate it to agents.
And we’ve seen this with other AI tools as well,
like Co-Pilot, our team definitely writes code
in a slightly different way,
so that Co-Pilot can work well for them.
– Right, right.
– To your question of complexity,
it like really depends on the user,
like some of us can make it work with really complex things.
– Yeah, yeah.
Where are you seeing agents being used the most
or perhaps it’s more that they’re having
a dramatic impact where they are being used
and how does that translate into businesses,
remaining competitive, having a competitive edge.
Folks that I’ve talked to and I keep talking about,
2022 last year,
we’re kind of the year of the models coming out
and mainstream taking notice of JNAI.
And perhaps this year has been the year
where people are trying to figure out,
what apps do I build to leverage these things,
as a software building business
or as a business that does other things
that wants to leverage this.
So where is in view seeing impact being made
or even looking to areas in the near future?
– You know, the interesting thing about agents
is it’s such an ill-defined term right now.
And people are calling all sorts
of very trivial things agents and that’s fine.
But I think there’s a spectrum of kind of agent,
usefulness, effectiveness.
There’s like a system that scrapes the web
and then aggregates the data
and pulls the data out in some way.
This is kind of like basically just a JNAI model.
Like, you know, it’s kind of very similar to chatGPT,
but you like put something on top of it
to like dump the output into a different system.
You could call that an agent.
So some people call that an agent.
We see that kind of thing being implemented
in all sorts of places.
But I think the much more exciting thing
when it comes to agents is these more general agents
that enable people to kind of start doing things
that they previously didn’t even imagine that they could do.
You know, I think like some of the really simple examples
right now are for us, like some researcher or scientist,
a biologist has a ton of data that they need to process
and they’re not a software engineer,
but they’re technical enough
that they can kind of like pull in the data
and then get something out of it,
get something that helps us out of it.
If they’re able to use something like this
that lets them work at this slightly higher level
or, you know, kind of over time,
that a very exciting thing is as we start to build
the tools that we need, like an example is my grandmother
gets a bunch of scam calls in Chinese
and, but all of her calls are in Chinese.
And if I want to build a piece of software
that filters out her scam calls from her other calls,
like this is very hard right now,
even for me as someone who knows how to build software.
And it’s such a niche market.
Like no one else is going to build that software for her.
We’ve tried to find software like that in the US,
doesn’t really exist.
And so, exactly.
So right now we’re in this world where software is built by–
– Not to interrupt you.
If it exists in the US for English language spam,
it doesn’t work that well either for my–
– Exactly, exactly.
So, you know, right now we’re in this world
where other people build software for us.
We have to rely on other people to build software for us.
And it’s actually really strange.
Like we don’t really own our digital environments
in that way.
Like everything is kind of built by someone else
because it’s too hard for us to like build or own things.
And I think there is a future in which like I could actually
pretty easily build something for my grandmother
or for my community or for my group of friends
or for my church to manage registrations or whatever it is.
And that can be really tailored to my particular use case
and to me, my community, my friends.
And so, I think the really exciting thing about
this future is like all of this like bespoke software
as opposed to today where we have this
kind of centralized software.
It’s almost like people don’t often think
of their digital environment in this way,
but the digital environment is like the physical environment.
And today it’s as if we all live in corporate housing.
– I used to be so excited that I could listen
to any music I wanted for, you know, 10 bucks a month.
And now I’m thinking like,
“But I don’t own any of it.”
They could take it away for a minute or a second.
– Yeah, and honestly, I think a lot of the kind
of frustration people have about big tech, about technology
is that we don’t feel like, and I don’t feel like
I have control over these things
that are a huge part of my life.
And so that’s what at MBU, what we wanna do
is give that control and power back to the people.
And we do that by creating these tools and systems
that collaborate with you to help you
be able to create stuff for yourself.
– So how hard is it to build these things
for folks who use what, you know, as you mentioned,
there are different, many different, you know,
voices, individuals, companies,
talking about the agent’s agentic AI,
and a lot of them are defining it,
talking about it at least slightly different.
I’m sure taking, you know, different approaches,
kind of under the hood.
What are the challenges?
What are the things that, you know,
we can get a little bit technical here as you like.
Some of the things, some of the problems that, you know,
you and your teams are solving to make it easier
for the rest of us to translate our ideas into software.
– Yeah, so some problems I would say
are kind of universal across all of these different types
of people who are building agents.
And some problems are unique to us
and what we’re trying to do.
So I would say, you know, most people are building agents
in this kind of like workflow automation paradigm
I mentioned earlier.
And so for that paradigm,
robustness, reliability is really important.
Like, okay, you know, I built a thing
that responds to customer service tickets.
But if it 3% of the time,
it says something really terrible to the user,
like this is not a usable agent.
For us, reliability and robustness is important,
but it’s actually a little bit less important.
As it gets better, the user experience just gets better.
As a user, I don’t have to check stuff as much.
But even if it’s not the best, it’s still okay.
Like I can still use it as the user
and I’ll like fix the bug that, you know,
the model produced and that’s okay.
So a lot of what we think of is kind of like,
how do we get agents to meet
both the model capabilities and also users
where they are today.
– So that expectation is built in
that we’re not at the stage yet where it’s error-free.
And as a user, you need to know that.
And it’s not just like, okay, you have to accept that,
but it’s actually like your experience
is gonna wind up being better, right?
‘Cause you know you’re part in it.
And it’s a, again, as you said,
it’s not sending it off to do something
and, you know, giving us the final result
we had no part in.
– Yeah, exactly.
I think, you know, people often think of agents
as a research problem, but we think of it
as both a research problem and a user experience problem.
And this user experience part
is really about setting the right expectations
so that with the experience,
so that it’s like, I’m not expecting it to go off
on its own for 10 minutes or 10 hours or 10 days
and come back with something magically correct.
Instead, I’m kind of iteratively working with it
and seeing, oh, it’s kind of like clay, you know,
I’m molding it, shaping it, shaping the output.
I think the workflow automation agents,
some of these agents kind of,
they’re a little bit more, the bar is higher
for how accurate they have to be
because what we have found is that as a user,
the longer it takes to come back with an answer,
the more I expect the answer to be correct
and I’ll get frustrated if it takes a really long time.
So we’re very much on the, like, highly interactive,
don’t take a super long time to come back with something,
be an agent that really works with the user side.
– Thinking about the idea of moving from,
I’m expecting from a computer, right?
The utmost inaccuracy, my calculator always says
two plus two is four.
Kind of moving from that to just a different frame of mind
as an end user saying, okay, we’re prioritizing, you know,
I don’t want to put words in your mouth,
but kind of the speed and experience
and you know, going in, it’s not gonna get them all right.
Is that something that, you know,
because this is the nature of AI and gen AI
and people are kind of used to that,
by now people are accepting of
or is there still kind of a, I don’t know,
maybe I’m just old, is there still kind of a mental hurdle
to getting past the right, that expectation?
– Yeah, so one of our core philosophies
is that we need to meet people where they are
in terms of what mental models we have
in our heads as people.
And so actually a good historical analogy
is a back before the personal computer,
people are really excited about the super computer.
And the first personal computer first came out,
everyone made fun of it.
They were like, this is a hobbyist toy.
And the super computer, you know,
you accessed it with a terminal,
you were time sharing on these super computers.
It was not especially usable.
And so a very small set of people were able to use it.
But as time went on, a small group of people at Xerox Park
actually invented a lot of these primitives
that led the personal computer to be usable by people.
They invented the desktop, files, folders.
These are like concepts that we understood
as humans at that time.
And so for us, you know,
part of actually building a good user experience
around agents requires invention.
It requires inventing concepts that match kind of like,
are able to map this technology
to what we currently understand as humans.
So earlier I was saying,
it’s kind of like a document editor right now,
you know, our current product experience.
And it may not ultimately be that way,
but a document is something that I as a person today
understand how to edit and work with.
And it’s almost like an interactive editor
that helps me accomplish my tasks.
And to your question of how users receive it,
one thing that’s really interesting that we observe
is software builders, the feedback so far has been,
wow, like this is really interesting.
It lets me work at this higher level.
I don’t have to dig so far into the code all the time.
I can kind of stay and thinking at this higher level.
And it’s actually able to go down into the code
and surface to me, like do the task and then I check it.
And that’s pretty cool.
It like lets me move a lot faster.
And that’s really interesting.
You know, that for us, that’s kind of the primary thing.
Like I want people to be able to work
at our human problem level of abstraction with software
instead of having to get so deep into the weeds.
Yeah, yeah.
No, I can relate from the standpoint
of learning how to use these tools when I’m writing,
when I’m not on the podcast, asking people like you questions.
You know, a lot of my work is writing.
And if I’m working from a document, a transcript,
source materials, it’s that same thing
when I can use the tool to just kind of surface to me.
You know, did anywhere in the document
or in the transcript at CanJune talk about pineapple on pizza?
You know, and just being able to surface that back, right?
It saves all the time of going through the document.
I don’t need the exact words.
I don’t need it 100%.
Then we’ll get to that later,
but I can just kind of go back and check.
Like, oh, right.
She said she is, isn’t a fan of Pepperoni.
You know, yeah.
And it’s, it’s, it’s incredibly helpful.
You know, it’s not just a saving time,
but I think you said it really well.
It allows you to stay at that level of thinking.
Exactly. Yeah.
I think our core, the thing I really care about
is helping people be able to be creative in this way.
And often there’s such a barrier between our ideas
and the execution of those ideas
that we never get to do a lot of things
that we really want to do with our lives.
And so I think the like true power of computing
as a medium hasn’t really been unleashed yet
and we want to unleash it.
And what that looks like is that people are able
to kind of like take their ideas.
Like you can take your ideas about writing this piece
or aggregating all of the pieces you’ve written
and being able to draw insights out of them
in order to create your book, like take these ideas
and actually be able to work with them at this higher level
so that you’re not always bogged down in the weeds.
And I think the true power of AI,
the true power of computers, like that’s what it enables.
And we’re not there yet, but we can get there.
And it’s not just about automation or business automation,
business workflow automation.
Right, right.
Now, what was the, in your conversation with Brian from GTC,
what was it that he said?
I’ve got all these ideas and I sit down to code
and I’m like import, what do I want to import, right?
And you’re, I think that was great
’cause you’re derailed immediately
and I can relate in the work that I do that, you know, yeah.
Yeah, 100%.
Yeah, one of our users said, wow,
I never realized how much context switching I do
when I’m writing code from high to low level.
Same with when you’re writing normal stuff.
I’m seeking with CanJune Q.
CanJune is co-founder and CEO of Imbue
and we have been talking about Imbue’s work on AI agents
that help people code and really fascinating approach
that, you know, as we’ve been talking about,
I think goes beyond just expressing in code,
but code being the way that we interface with computers
and get them to do the things we want them to do.
Well, I want to ask you about AI models and reasoning.
And then I also want to ask you sort of about scale
and what goes into, you know, building an agent for yourself
and then what goes into building agents and multiple agents
and agents that collaborate for users at scale.
Is there an order we should go in?
Should we talk about reasoning first?
Is there a relation?
That’s interesting.
Let’s talk about scale first.
Okay, cool.
Yeah, so one way that people think about scale with agents
is a lot of agents interacting with each other
and kind of what that looks like.
And some people do that by giving different agents
different prompts so they have different personalities
and things like that.
And honestly, I think that’s a little bit, it’s interesting.
It’s a little bit limited because we already have agents today.
All software is agentic.
The whole point of what an agent is is it something
that takes action and it uses your computer
to kind of like execute something.
And so almost all software is executing something.
It’s kind of like changing the state of your computer,
website, data, et cetera.
Now, the difference between most software
and like AI agents, what we call AI agents today
is that AI agents can process stuff
in a way that’s like not fully deterministic.
But even so, we still had AI agents in Facebook news feed,
for example, recommendation engine is an agent
that non-deterministically decides what to show you.
So we’ve had agents since forever, since we had software.
And so, you know, kind of the way I think about scale
and agents is actually about,
is the same as the way I think about scale of software.
So in the future, in the next 10 years,
I think there’s gonna be this explosion of software
and the software is gonna be slightly less hard coded
than before.
It’s gonna be able to work with more empty US input.
It might be more interactive.
Hopefully a lot of people are going to be able
to create it if we succeed.
And so now we end up with this giant ecosystem,
like world of software that’s far beyond what we have today.
And in that world, what happens is now you have a lot
of different automated systems interacting with each other.
And this is actually, could be super exciting.
Every person could have their own automated systems
that do things for themselves, for their lives.
They have, you know, I’m supported by the software
that surrounds me, as opposed to today
maybe being bothered by it.
(laughing)
– I was listening to you and I was thinking of how to phrase,
how to try to phrase the question,
getting back to your point about, you know, sort of,
I was thinking of it as sort of one size fits all software,
you know, that’s deterministic and it does what it does
versus, you know, me being able to create
and reshape things as we go.
And you answered it for me.
So that’s that.
– Oh, that’s great.
I love this, I love this one size fits all software today
versus future bespoke software.
That’s a great term.
– Are you worried at all about,
I think, I don’t know if the term AI slop applies to code,
but this idea of, you know, AI models creating texts
that’s kind of meaningless or value-less,
but it’s being automatically put out onto the web and stuff
for a very sort of, you know, relatively ignorant point
of view, the notion of model-generated software
that can run itself being deployed out on the public web,
you know, is a little scary to me,
but also I’m sure there’s stuff out there,
but how do you think about that?
– Yeah, the way I think about kind of AI slop for software
is automated systems that like infringe on us.
So scam calling or spam calling is a good example
of an automated system that infringes on us,
or like apps that have lots of notifications
or, you know, games that are meant to be extractive.
Those are our systems that infringe on us.
And, you know, I actually think the default path
of where things are going with centralized AI systems
and the kind of like returns the scale
of improvements of the underlying models
is that we will kind of, as humans,
be a little bit disempowered and be holding
to whoever controls the automated systems.
It’s not necessary, you know, I’ve kind of told you
about this beautiful vision of a future
where everyone can create software
and everyone creates bespoke software,
and like that’s the future we want to build,
but it’s not necessarily the future
that is the default path.
The default path could be that there’s a lot of,
like even more software out there today
that’s trying to get our attention
and trying to extract from us.
And what we need, I think, what I want
is for people to create defenses
and to kind of like get rid of that stuff
and disrupt it.
You know, hopefully when I can build my own software,
I can actually disrupt a lot of the stuff
that’s being built today,
so that I have software that’s serving my own interests.
I have agents that are serving my own interests
and helping me do what I want to do.
So to your question of AI slop,
I think there’s definitely going to be people
making more automated systems
that are bots and bother people.
And just like in cybersecurity,
I think there’s an attack defense dynamic,
and what we want is to enable people
to create opposing systems for themselves
that like help defend their environment and themselves
and help kind of protect us to do what we want
and liberalize we want.
And hopefully there’s also kind of some,
you know, there’s already some regulatory aspect of this
that exists and, you know, there hopefully will be more
in response to what we see.
So that effect is real.
– Right, all right.
There’s been a lot in the generative AI news cycles.
There’s a thing, that’s a thing.
This year in particular about models and reasoning
and future models, models being trained
that have reasoning capacity, that kind of thing.
Recently open AI launched a new iteration of a model
talking about reason capabilities.
Is reasoning something that does or can happen in an LLM?
Is it something that the agents bring to the table,
so to speak?
How do you think about reasoning?
How does in view approach building reasoning capabilities
into your products?
– Yeah, that’s a great question.
So yeah, reasoning is a buzzword right now
and models definitely do reasoning.
That’s, you know, and in the way that like underlying LLMs
definitely do reasoning in the way that humans kind of,
it’s not exactly how we do reasoning maybe,
and it’s not always perfectly correct.
And it’s often continuing to justify its own reasoning,
although humans do that too.
– Now he’s gonna say, that’s familiar.
– So unclear how similar or different it is to humans,
but the underlying LLM definitely does some reasoning.
One key difference that we observe right now
is that the underlying LLM isn’t necessarily
as good at verifying if its own answer is correct.
And as a person when I’m doing a task,
I actually, we don’t notice this,
but we’re always constantly checking like,
is this right?
Did I do that right?
Is this what I expected?
And there’s not that loop.
So that loop is kind of added by the agentic piece.
– Okay.
– Yeah, we think a lot actually about our research direction
as around this kind of verifying verification.
Is it correct?
Did I do it right?
And if I have an agent that’s writing code for me,
I do want it to check itself like,
hey, did I do that right?
I did.
Oh, I didn’t.
Let me fix this mistake and then come back to you.
And so the better it is at verifying its own answers,
the more like a better user experience it is.
And so, when we talk about reasoning,
we mostly talk about this kind of like verification
and robustness.
Is it able to verify what it’s doing?
We’ve actually learned some pretty interesting things
when working on verification around kind of,
it turns out that, so in software development,
when you write a function,
you also often write software tests.
You’re testing, okay, did the software
have the behavior expected?
And given really good tests,
actually the underlying models are pretty good
at treating the function or treating the piece of software.
– Right.
– But given the piece of software,
the underlying models are not very good
at treating good tests.
Which is kind of interesting.
– Yeah.
– One.
– Any idea why?
– Yeah, one, you know, it’s partly because the models
are probably not trained that much on this particular task
of creating tests.
Two, though, maybe it’s possible.
We don’t know, we’re not 100% sure,
but it’s possible that actually verifying
if something is correct is a harder reasoning problem
than kind of creating the thing in the first place.
– Yeah.
– So it kind of requires this like analysis and judgment.
And so our research direction is primarily focused
on verification, how do we get to models
that actually are able to properly verify
that the output is correct and kind of is
what the user wanted.
And we think of that as the hard problem
in reasoning for agents.
– At the same time, MBU has, did you pre-train a model?
Did you build a foundation?
You didn’t build the model from scratch,
but it’s a 70 billion parameter model.
– That’s right, we actually did pre-train
a 70 billion parameter model from scratch.
– It is from scratch, okay, one mistake.
– Yeah, we actually learned a ton from that process.
And one of the things we learned was like,
actually we don’t know if we need to do tons
of pre-training going into the future.
We’ll see. – Yeah.
– But we got a lot out of post-training on that model.
And so for a lot of the verification work,
we’re actually really interested in post-training,
fine-tuning, reinforcement learning
on top of the underlying models.
– That seems like a good place to ask this question.
What does the future look like?
I almost want to leave it just there, but that’s not fair.
What does the future of AI agents look like?
What is MBU’s approach?
Do you have a roadmap you can share any of?
– Where is this heading?
And I know that’s an impossible question to answer
in many ways, but I’m also guessing
you have something of a vision, so.
– Yeah, that’s a great question.
So I talked today about trying to enable everyone
to build software, but really internally,
the way we think about it is all software
in the future will be agents, basically.
I mean, not all software, but most software.
It’ll be a little bit smarter,
kind of just like living software.
And what we want to do is enable everyone
to be able to build agents so that in the world,
we’re all able to build our own agents for ourselves
or use each other’s copy,
someone’s modify a little bit from myself.
And that’s kind of what our product
is meant to do over the long term.
And so actually in our 70 billion parameter model,
we released a set of blog posts that taught people
how to set up infrastructure to train such models.
We expect most people, most of us won’t train our own models,
but it’s kind of part of this desire
to democratize a lot of these capabilities.
We also released a toolkit for people doing evaluations
of their models and with clean data and everything.
And so in terms of kind of what’s the future
of building agents, my hope is that agent building,
unlike software building, is not something
that only a few people are able to do well.
My hope is that there’s actually something
that’s like widely democratized
and where everyone is empowered
to be able to create their own agents.
And I think right now we have this very scary view
of somebody else is going to create a system
that automates my job.
And that sucks.
That’s like really disempowering.
I don’t want that for my job.
– Yep.
– But the thing I love doing is automating parts
of my own job.
– Yeah.
– I like, you know, love making it better
in all of these different ways.
And that’s what we want to enable people to be able to do.
Like by giving you the tools to make your own agents,
that means that you can make your own things
that automate parts of your job.
And now your job can be higher leverage and higher level.
And now you can do a lot more.
– Right.
– And so we want to give kind of, you know,
someone else automating my job is very disempowering to me,
but someone giving me the tools
so that I can make my own tools for myself.
That’s very empowering for me.
And I think this mentality shift
is actually really important.
– Amen.
Ken June for folks listening
who would like to learn more about imbue.
You mentioned a blog as well.
Where should they start online?
Website, social media, podcast.
There’s a podcast, I think.
Where should they go?
– Great question.
So imbue.com is where we are on the internet.
And you can follow our Twitter account imbue.ai.
And we will have, you know,
as we start to release the product a little bit more
publicly, we’ll probably have announcements
and things where you can start experimenting
with what we’re doing.
So please do follow us.
There’s also a newsletter sign up
where we send extremely rare emails.
(laughing)
Because we mostly focus on building.
– You’re not an application trying to extract constantly?
– No, no, not trying to get your attention,
trying to make a useful product.
– Good, good, good.
Ken June, this is delightful.
Thank you so much for taking the time to come on the pod.
And best of luck with everything you’re doing.
Maybe we can check in again down the road.
– Definitely.
Thank you, Noah.
This is super fun.
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
[BLANK_AUDIO]
>> Hello, and welcome to the NVIDIA AI podcast.
I’m your host, Noah Kravitz.
One of the big transformations being
enabled by AI is the way we create software.
From coding co-pilots to
in-development systems built to translate
plain language requests into fully functional applications,
generative AI is fueling a new wave
of tools to help us create software faster.
Our guest today is Kanjun Kew.
Kanjun is co-founder and CEO of Imbue,
a three-and-a-half-year-old company somewhere,
and they’re founded in 2021,
that is building AI agents that work with us to
translate our ideas into code and bring them to life.
There’s a lot more to it than that,
but why hear it from me when you can hear it directly from Kanjun.
Kanjun Kew, welcome.
Thank you so much for taking the time to join the NVIDIA AI podcast.
>> Thank you, Noah. It’s great to be here.
Let’s talk about software development and AI,
and Imbue, and all of that good stuff.
But maybe let’s start with agents.
Agents are, I don’t want to use the word “hot”
because I don’t want it to sound fluffy, right?
But agents are a thing right now.
We’ve had some guests on recently talking about
agents in different contexts,
and Imbue’s approach to agents is something worth delving into.
Maybe we can start there. What are AI agents?
Why do we need them? Why isn’t Imbue working on them?
>> Yeah. Agents are really popular right now.
So Imbue was founded in early 2021,
and at that time,
our goal was to figure out how to make and work on general AI agents.
At that time, people thought we were totally crazy.
Like, what are agents?
Everyone’s working on AGI.
AGI is going to rule everything.
But what we were really interested in is,
how can we have systems that we as people can mold,
shape to what we’re wanting?
As opposed to, oh, this external intelligence that knows all the answers.
And so we started as a research lab at that time
because the technology was certainly not good enough
to build general agents that could reliably do anything
that we wanted them to do.
And, you know, in the very beginning,
we thought of agents in a similar way
to how a lot of people think about agents today.
As these kind of, you know, think about what an agent is.
Often people think about kind of an autonomous,
personal assistant type thing.
You ask it to do something, it does it.
For you, it comes back to you.
Now everyone has their own personal assistant.
And actually, a lot of our learning has been
that that’s a really tricky user experience.
And we experience it ourselves with human agents,
which is that often when I delegate something,
it comes back, it’s not quite what I wanted.
And now I have to negotiate how to get what I wanted.
I was listening, I told you before we record,
I was listening to a little bit of your fireside chat
with Brian Cutton-Zera from GTC this year,
which listeners go check that out.
It’s a great listen.
And you were talking about the difficulty,
and I can relate to this so much,
the difficulty inherent in delegating work
to someone else, right?
And to your point, thinking of it as humans,
you have to break the problem down,
you have to sort of figure out,
well, how much do I tell them exactly what to do?
– Yeah.
– Yeah, all that, yeah.
– What context does it need ahead of time?
What instruction should I give?
Delegation is actually a really tricky paradigm
because it actually puts all the onus
on the person who’s delegating
to define the problem, define the scope.
Of course, the person who’s being delegated to the agent
might come back with some questions and stuff like that,
but it’s a very tricky thing to trust.
So what we’ve learned over the years
working on general agents is we’ve actually started
to think about agents in a very different way.
And this is both from a kind of pragmatic business,
like user perspective and also from a mission perspective.
The way that we think about agents is,
if you think about what an agent is,
what this personal assistant is doing,
what it is is it’s kind of this intermediary layer
between you and your computer.
And you’re kind of telling it stuff
or interfacing with it in some way,
whether it’s a UI or a natural language,
and it’s interfacing with your computer.
And the most effective way to interface with your computer
is by writing code.
That’s what your computer is made of.
And there are really kind of two types of code.
There’s like everything is hard-coded,
which is the default for software today.
Everything is hard-coded.
So now your agent can only do a subset of things.
Or now with language models, you can generate code.
You can write code that does new stuff
that’s not hard-coded.
And now you have an agent that’s much more general.
It’s able to do sets of tasks
that I didn’t program into it ahead of time,
but now it’s able to do.
And so the way that we think about what agents are
is actually as this intermediary layer
between you and your computer,
and it’s essentially a layer of abstraction
on top of programming that allows us as regular people
to be able to program our computers
without even thinking about what we’re doing as programming.
And so we think of our mission
as essentially trying to reinvent the personal computer
and really deeply empower everyone to be able to create
in this computing medium of the future,
because this digital medium
is becoming more important in our lives.
And what we want is actually not what I want,
at least personally,
is actually not a super centralized assistant
that someone else has decided
is able to do this, integrate with that.
What I actually want is something that I can make and own
that is mine, and that does what I wanted to do,
kind of serves me.
And today we’re kind of in a world
where like all of our software is rented,
it serves other people.
So that’s kind of what we think of what agents are
is this layer of abstraction on top of programming
makes it so that it’s very intuitive
for everyone to program.
And that actually requires quite a bit of invention.
So can get into that in historical context.
– Yeah, well, the way you were,
we’re describing it,
I know you weren’t describing the sort of AI,
the assistant, the agent layer,
all the A-words, AI, assistant, agent,
you weren’t describing the agent layer
as replacement for a user interface.
So you mentioned, you know, UI,
but that’s kind of what I was thinking of,
right, that abstractive layer that’s kind of in between.
So how is imbue approaching it sort of on the ground
and working with you or primarily with enterprise clients?
– No, we primarily with, I would say it’s prosumer.
So people who are,
so the way to think about what we’re doing is,
instead of thinking of agents as automation systems,
right now we’re in kind of the agent,
automation, agent paradigm,
we think of agents as collaborative systems.
So how do we enable a system that empowers me
and helps me do more of what I want to do and work with?
I can work with it.
And so in the beginning,
we’re actually, you know, enabling you to write code.
Right now, these models are not that good at writing code.
As they write code, you actually have to go check it.
So you have to understand how to write code to go,
to see, oh, did the agent do a good job?
But over time, as these models get better at writing code,
now you don’t have to check it anymore.
And so in the beginning,
we start with software engineers as the prime,
or we call them software builders.
So you don’t have to be an engineer.
I’m a software builder.
I’m not a software engineer.
We start with software builders who can’t.
– I’m a prototype builder then.
I wouldn’t go as far as software.
– Okay.
So you could probably be a soon to be user.
Once we get to a place where you don’t have to read
and write, touch the code so much.
Right now, we’re targeting software builders
who can read and touch the code and be like,
okay, well, it’s not quite right.
We want to adjust it in this way.
And over time, as the models get better,
and you don’t have to be so low level in the code,
now more and more types of creatives,
types of builders can use these systems
to really mold and shape their computer
to what they want it to be.
– And what level of complexity,
what level of depth is the actual software
that users are building within view?
There’s the issue of accuracy, obviously,
as you were saying they’re not,
none of the models are creating 100% perfect code yet.
But I also wonder how complex can these things get?
And that kind of brings us to talking about reasoning.
We don’t have to get there yet,
but kind of edging towards that conversation.
– Yeah, I think one of actually,
our biggest learnings has been that if as a user,
I have a code base that is fairly modular
and I’ve kind of broken things down,
then the model is actually pretty good at dealing
with a very complex system,
because it doesn’t have to load that much stuff
into its head and it doesn’t have to cross check
all of these dependencies.
So just like with humans building software,
where you don’t want a ton of dependencies,
also, if you have a slightly more isolated system,
it’ll do a better job.
Similarly, there’s a lot of kind of user expertise
in using this kind of product.
So our product, it feels very collaborative.
It’s almost like a document editor
and kind of like interfacing and interacting
with an agent in that way.
And so as a user, you can basically,
we learn to give it tasks,
it’s more likely to succeed out.
And we learn to structure our work
so that we can delegate it to agents.
And we’ve seen this with other AI tools as well,
like Co-Pilot, our team definitely writes code
in a slightly different way,
so that Co-Pilot can work well for them.
– Right, right.
– To your question of complexity,
it like really depends on the user,
like some of us can make it work with really complex things.
– Yeah, yeah.
Where are you seeing agents being used the most
or perhaps it’s more that they’re having
a dramatic impact where they are being used
and how does that translate into businesses,
remaining competitive, having a competitive edge.
Folks that I’ve talked to and I keep talking about,
2022 last year,
we’re kind of the year of the models coming out
and mainstream taking notice of JNAI.
And perhaps this year has been the year
where people are trying to figure out,
what apps do I build to leverage these things,
as a software building business
or as a business that does other things
that wants to leverage this.
So where is in view seeing impact being made
or even looking to areas in the near future?
– You know, the interesting thing about agents
is it’s such an ill-defined term right now.
And people are calling all sorts
of very trivial things agents and that’s fine.
But I think there’s a spectrum of kind of agent,
usefulness, effectiveness.
There’s like a system that scrapes the web
and then aggregates the data
and pulls the data out in some way.
This is kind of like basically just a JNAI model.
Like, you know, it’s kind of very similar to chatGPT,
but you like put something on top of it
to like dump the output into a different system.
You could call that an agent.
So some people call that an agent.
We see that kind of thing being implemented
in all sorts of places.
But I think the much more exciting thing
when it comes to agents is these more general agents
that enable people to kind of start doing things
that they previously didn’t even imagine that they could do.
You know, I think like some of the really simple examples
right now are for us, like some researcher or scientist,
a biologist has a ton of data that they need to process
and they’re not a software engineer,
but they’re technical enough
that they can kind of like pull in the data
and then get something out of it,
get something that helps us out of it.
If they’re able to use something like this
that lets them work at this slightly higher level
or, you know, kind of over time,
that a very exciting thing is as we start to build
the tools that we need, like an example is my grandmother
gets a bunch of scam calls in Chinese
and, but all of her calls are in Chinese.
And if I want to build a piece of software
that filters out her scam calls from her other calls,
like this is very hard right now,
even for me as someone who knows how to build software.
And it’s such a niche market.
Like no one else is going to build that software for her.
We’ve tried to find software like that in the US,
doesn’t really exist.
And so, exactly.
So right now we’re in this world where software is built by–
– Not to interrupt you.
If it exists in the US for English language spam,
it doesn’t work that well either for my–
– Exactly, exactly.
So, you know, right now we’re in this world
where other people build software for us.
We have to rely on other people to build software for us.
And it’s actually really strange.
Like we don’t really own our digital environments
in that way.
Like everything is kind of built by someone else
because it’s too hard for us to like build or own things.
And I think there is a future in which like I could actually
pretty easily build something for my grandmother
or for my community or for my group of friends
or for my church to manage registrations or whatever it is.
And that can be really tailored to my particular use case
and to me, my community, my friends.
And so, I think the really exciting thing about
this future is like all of this like bespoke software
as opposed to today where we have this
kind of centralized software.
It’s almost like people don’t often think
of their digital environment in this way,
but the digital environment is like the physical environment.
And today it’s as if we all live in corporate housing.
– I used to be so excited that I could listen
to any music I wanted for, you know, 10 bucks a month.
And now I’m thinking like,
“But I don’t own any of it.”
They could take it away for a minute or a second.
– Yeah, and honestly, I think a lot of the kind
of frustration people have about big tech, about technology
is that we don’t feel like, and I don’t feel like
I have control over these things
that are a huge part of my life.
And so that’s what at MBU, what we wanna do
is give that control and power back to the people.
And we do that by creating these tools and systems
that collaborate with you to help you
be able to create stuff for yourself.
– So how hard is it to build these things
for folks who use what, you know, as you mentioned,
there are different, many different, you know,
voices, individuals, companies,
talking about the agent’s agentic AI,
and a lot of them are defining it,
talking about it at least slightly different.
I’m sure taking, you know, different approaches,
kind of under the hood.
What are the challenges?
What are the things that, you know,
we can get a little bit technical here as you like.
Some of the things, some of the problems that, you know,
you and your teams are solving to make it easier
for the rest of us to translate our ideas into software.
– Yeah, so some problems I would say
are kind of universal across all of these different types
of people who are building agents.
And some problems are unique to us
and what we’re trying to do.
So I would say, you know, most people are building agents
in this kind of like workflow automation paradigm
I mentioned earlier.
And so for that paradigm,
robustness, reliability is really important.
Like, okay, you know, I built a thing
that responds to customer service tickets.
But if it 3% of the time,
it says something really terrible to the user,
like this is not a usable agent.
For us, reliability and robustness is important,
but it’s actually a little bit less important.
As it gets better, the user experience just gets better.
As a user, I don’t have to check stuff as much.
But even if it’s not the best, it’s still okay.
Like I can still use it as the user
and I’ll like fix the bug that, you know,
the model produced and that’s okay.
So a lot of what we think of is kind of like,
how do we get agents to meet
both the model capabilities and also users
where they are today.
– So that expectation is built in
that we’re not at the stage yet where it’s error-free.
And as a user, you need to know that.
And it’s not just like, okay, you have to accept that,
but it’s actually like your experience
is gonna wind up being better, right?
‘Cause you know you’re part in it.
And it’s a, again, as you said,
it’s not sending it off to do something
and, you know, giving us the final result
we had no part in.
– Yeah, exactly.
I think, you know, people often think of agents
as a research problem, but we think of it
as both a research problem and a user experience problem.
And this user experience part
is really about setting the right expectations
so that with the experience,
so that it’s like, I’m not expecting it to go off
on its own for 10 minutes or 10 hours or 10 days
and come back with something magically correct.
Instead, I’m kind of iteratively working with it
and seeing, oh, it’s kind of like clay, you know,
I’m molding it, shaping it, shaping the output.
I think the workflow automation agents,
some of these agents kind of,
they’re a little bit more, the bar is higher
for how accurate they have to be
because what we have found is that as a user,
the longer it takes to come back with an answer,
the more I expect the answer to be correct
and I’ll get frustrated if it takes a really long time.
So we’re very much on the, like, highly interactive,
don’t take a super long time to come back with something,
be an agent that really works with the user side.
– Thinking about the idea of moving from,
I’m expecting from a computer, right?
The utmost inaccuracy, my calculator always says
two plus two is four.
Kind of moving from that to just a different frame of mind
as an end user saying, okay, we’re prioritizing, you know,
I don’t want to put words in your mouth,
but kind of the speed and experience
and you know, going in, it’s not gonna get them all right.
Is that something that, you know,
because this is the nature of AI and gen AI
and people are kind of used to that,
by now people are accepting of
or is there still kind of a, I don’t know,
maybe I’m just old, is there still kind of a mental hurdle
to getting past the right, that expectation?
– Yeah, so one of our core philosophies
is that we need to meet people where they are
in terms of what mental models we have
in our heads as people.
And so actually a good historical analogy
is a back before the personal computer,
people are really excited about the super computer.
And the first personal computer first came out,
everyone made fun of it.
They were like, this is a hobbyist toy.
And the super computer, you know,
you accessed it with a terminal,
you were time sharing on these super computers.
It was not especially usable.
And so a very small set of people were able to use it.
But as time went on, a small group of people at Xerox Park
actually invented a lot of these primitives
that led the personal computer to be usable by people.
They invented the desktop, files, folders.
These are like concepts that we understood
as humans at that time.
And so for us, you know,
part of actually building a good user experience
around agents requires invention.
It requires inventing concepts that match kind of like,
are able to map this technology
to what we currently understand as humans.
So earlier I was saying,
it’s kind of like a document editor right now,
you know, our current product experience.
And it may not ultimately be that way,
but a document is something that I as a person today
understand how to edit and work with.
And it’s almost like an interactive editor
that helps me accomplish my tasks.
And to your question of how users receive it,
one thing that’s really interesting that we observe
is software builders, the feedback so far has been,
wow, like this is really interesting.
It lets me work at this higher level.
I don’t have to dig so far into the code all the time.
I can kind of stay and thinking at this higher level.
And it’s actually able to go down into the code
and surface to me, like do the task and then I check it.
And that’s pretty cool.
It like lets me move a lot faster.
And that’s really interesting.
You know, that for us, that’s kind of the primary thing.
Like I want people to be able to work
at our human problem level of abstraction with software
instead of having to get so deep into the weeds.
Yeah, yeah.
No, I can relate from the standpoint
of learning how to use these tools when I’m writing,
when I’m not on the podcast, asking people like you questions.
You know, a lot of my work is writing.
And if I’m working from a document, a transcript,
source materials, it’s that same thing
when I can use the tool to just kind of surface to me.
You know, did anywhere in the document
or in the transcript at CanJune talk about pineapple on pizza?
You know, and just being able to surface that back, right?
It saves all the time of going through the document.
I don’t need the exact words.
I don’t need it 100%.
Then we’ll get to that later,
but I can just kind of go back and check.
Like, oh, right.
She said she is, isn’t a fan of Pepperoni.
You know, yeah.
And it’s, it’s, it’s incredibly helpful.
You know, it’s not just a saving time,
but I think you said it really well.
It allows you to stay at that level of thinking.
Exactly. Yeah.
I think our core, the thing I really care about
is helping people be able to be creative in this way.
And often there’s such a barrier between our ideas
and the execution of those ideas
that we never get to do a lot of things
that we really want to do with our lives.
And so I think the like true power of computing
as a medium hasn’t really been unleashed yet
and we want to unleash it.
And what that looks like is that people are able
to kind of like take their ideas.
Like you can take your ideas about writing this piece
or aggregating all of the pieces you’ve written
and being able to draw insights out of them
in order to create your book, like take these ideas
and actually be able to work with them at this higher level
so that you’re not always bogged down in the weeds.
And I think the true power of AI,
the true power of computers, like that’s what it enables.
And we’re not there yet, but we can get there.
And it’s not just about automation or business automation,
business workflow automation.
Right, right.
Now, what was the, in your conversation with Brian from GTC,
what was it that he said?
I’ve got all these ideas and I sit down to code
and I’m like import, what do I want to import, right?
And you’re, I think that was great
’cause you’re derailed immediately
and I can relate in the work that I do that, you know, yeah.
Yeah, 100%.
Yeah, one of our users said, wow,
I never realized how much context switching I do
when I’m writing code from high to low level.
Same with when you’re writing normal stuff.
I’m seeking with CanJune Q.
CanJune is co-founder and CEO of Imbue
and we have been talking about Imbue’s work on AI agents
that help people code and really fascinating approach
that, you know, as we’ve been talking about,
I think goes beyond just expressing in code,
but code being the way that we interface with computers
and get them to do the things we want them to do.
Well, I want to ask you about AI models and reasoning.
And then I also want to ask you sort of about scale
and what goes into, you know, building an agent for yourself
and then what goes into building agents and multiple agents
and agents that collaborate for users at scale.
Is there an order we should go in?
Should we talk about reasoning first?
Is there a relation?
That’s interesting.
Let’s talk about scale first.
Okay, cool.
Yeah, so one way that people think about scale with agents
is a lot of agents interacting with each other
and kind of what that looks like.
And some people do that by giving different agents
different prompts so they have different personalities
and things like that.
And honestly, I think that’s a little bit, it’s interesting.
It’s a little bit limited because we already have agents today.
All software is agentic.
The whole point of what an agent is is it something
that takes action and it uses your computer
to kind of like execute something.
And so almost all software is executing something.
It’s kind of like changing the state of your computer,
website, data, et cetera.
Now, the difference between most software
and like AI agents, what we call AI agents today
is that AI agents can process stuff
in a way that’s like not fully deterministic.
But even so, we still had AI agents in Facebook news feed,
for example, recommendation engine is an agent
that non-deterministically decides what to show you.
So we’ve had agents since forever, since we had software.
And so, you know, kind of the way I think about scale
and agents is actually about,
is the same as the way I think about scale of software.
So in the future, in the next 10 years,
I think there’s gonna be this explosion of software
and the software is gonna be slightly less hard coded
than before.
It’s gonna be able to work with more empty US input.
It might be more interactive.
Hopefully a lot of people are going to be able
to create it if we succeed.
And so now we end up with this giant ecosystem,
like world of software that’s far beyond what we have today.
And in that world, what happens is now you have a lot
of different automated systems interacting with each other.
And this is actually, could be super exciting.
Every person could have their own automated systems
that do things for themselves, for their lives.
They have, you know, I’m supported by the software
that surrounds me, as opposed to today
maybe being bothered by it.
(laughing)
– I was listening to you and I was thinking of how to phrase,
how to try to phrase the question,
getting back to your point about, you know, sort of,
I was thinking of it as sort of one size fits all software,
you know, that’s deterministic and it does what it does
versus, you know, me being able to create
and reshape things as we go.
And you answered it for me.
So that’s that.
– Oh, that’s great.
I love this, I love this one size fits all software today
versus future bespoke software.
That’s a great term.
– Are you worried at all about,
I think, I don’t know if the term AI slop applies to code,
but this idea of, you know, AI models creating texts
that’s kind of meaningless or value-less,
but it’s being automatically put out onto the web and stuff
for a very sort of, you know, relatively ignorant point
of view, the notion of model-generated software
that can run itself being deployed out on the public web,
you know, is a little scary to me,
but also I’m sure there’s stuff out there,
but how do you think about that?
– Yeah, the way I think about kind of AI slop for software
is automated systems that like infringe on us.
So scam calling or spam calling is a good example
of an automated system that infringes on us,
or like apps that have lots of notifications
or, you know, games that are meant to be extractive.
Those are our systems that infringe on us.
And, you know, I actually think the default path
of where things are going with centralized AI systems
and the kind of like returns the scale
of improvements of the underlying models
is that we will kind of, as humans,
be a little bit disempowered and be holding
to whoever controls the automated systems.
It’s not necessary, you know, I’ve kind of told you
about this beautiful vision of a future
where everyone can create software
and everyone creates bespoke software,
and like that’s the future we want to build,
but it’s not necessarily the future
that is the default path.
The default path could be that there’s a lot of,
like even more software out there today
that’s trying to get our attention
and trying to extract from us.
And what we need, I think, what I want
is for people to create defenses
and to kind of like get rid of that stuff
and disrupt it.
You know, hopefully when I can build my own software,
I can actually disrupt a lot of the stuff
that’s being built today,
so that I have software that’s serving my own interests.
I have agents that are serving my own interests
and helping me do what I want to do.
So to your question of AI slop,
I think there’s definitely going to be people
making more automated systems
that are bots and bother people.
And just like in cybersecurity,
I think there’s an attack defense dynamic,
and what we want is to enable people
to create opposing systems for themselves
that like help defend their environment and themselves
and help kind of protect us to do what we want
and liberalize we want.
And hopefully there’s also kind of some,
you know, there’s already some regulatory aspect of this
that exists and, you know, there hopefully will be more
in response to what we see.
So that effect is real.
– Right, all right.
There’s been a lot in the generative AI news cycles.
There’s a thing, that’s a thing.
This year in particular about models and reasoning
and future models, models being trained
that have reasoning capacity, that kind of thing.
Recently open AI launched a new iteration of a model
talking about reason capabilities.
Is reasoning something that does or can happen in an LLM?
Is it something that the agents bring to the table,
so to speak?
How do you think about reasoning?
How does in view approach building reasoning capabilities
into your products?
– Yeah, that’s a great question.
So yeah, reasoning is a buzzword right now
and models definitely do reasoning.
That’s, you know, and in the way that like underlying LLMs
definitely do reasoning in the way that humans kind of,
it’s not exactly how we do reasoning maybe,
and it’s not always perfectly correct.
And it’s often continuing to justify its own reasoning,
although humans do that too.
– Now he’s gonna say, that’s familiar.
– So unclear how similar or different it is to humans,
but the underlying LLM definitely does some reasoning.
One key difference that we observe right now
is that the underlying LLM isn’t necessarily
as good at verifying if its own answer is correct.
And as a person when I’m doing a task,
I actually, we don’t notice this,
but we’re always constantly checking like,
is this right?
Did I do that right?
Is this what I expected?
And there’s not that loop.
So that loop is kind of added by the agentic piece.
– Okay.
– Yeah, we think a lot actually about our research direction
as around this kind of verifying verification.
Is it correct?
Did I do it right?
And if I have an agent that’s writing code for me,
I do want it to check itself like,
hey, did I do that right?
I did.
Oh, I didn’t.
Let me fix this mistake and then come back to you.
And so the better it is at verifying its own answers,
the more like a better user experience it is.
And so, when we talk about reasoning,
we mostly talk about this kind of like verification
and robustness.
Is it able to verify what it’s doing?
We’ve actually learned some pretty interesting things
when working on verification around kind of,
it turns out that, so in software development,
when you write a function,
you also often write software tests.
You’re testing, okay, did the software
have the behavior expected?
And given really good tests,
actually the underlying models are pretty good
at treating the function or treating the piece of software.
– Right.
– But given the piece of software,
the underlying models are not very good
at treating good tests.
Which is kind of interesting.
– Yeah.
– One.
– Any idea why?
– Yeah, one, you know, it’s partly because the models
are probably not trained that much on this particular task
of creating tests.
Two, though, maybe it’s possible.
We don’t know, we’re not 100% sure,
but it’s possible that actually verifying
if something is correct is a harder reasoning problem
than kind of creating the thing in the first place.
– Yeah.
– So it kind of requires this like analysis and judgment.
And so our research direction is primarily focused
on verification, how do we get to models
that actually are able to properly verify
that the output is correct and kind of is
what the user wanted.
And we think of that as the hard problem
in reasoning for agents.
– At the same time, MBU has, did you pre-train a model?
Did you build a foundation?
You didn’t build the model from scratch,
but it’s a 70 billion parameter model.
– That’s right, we actually did pre-train
a 70 billion parameter model from scratch.
– It is from scratch, okay, one mistake.
– Yeah, we actually learned a ton from that process.
And one of the things we learned was like,
actually we don’t know if we need to do tons
of pre-training going into the future.
We’ll see. – Yeah.
– But we got a lot out of post-training on that model.
And so for a lot of the verification work,
we’re actually really interested in post-training,
fine-tuning, reinforcement learning
on top of the underlying models.
– That seems like a good place to ask this question.
What does the future look like?
I almost want to leave it just there, but that’s not fair.
What does the future of AI agents look like?
What is MBU’s approach?
Do you have a roadmap you can share any of?
– Where is this heading?
And I know that’s an impossible question to answer
in many ways, but I’m also guessing
you have something of a vision, so.
– Yeah, that’s a great question.
So I talked today about trying to enable everyone
to build software, but really internally,
the way we think about it is all software
in the future will be agents, basically.
I mean, not all software, but most software.
It’ll be a little bit smarter,
kind of just like living software.
And what we want to do is enable everyone
to be able to build agents so that in the world,
we’re all able to build our own agents for ourselves
or use each other’s copy,
someone’s modify a little bit from myself.
And that’s kind of what our product
is meant to do over the long term.
And so actually in our 70 billion parameter model,
we released a set of blog posts that taught people
how to set up infrastructure to train such models.
We expect most people, most of us won’t train our own models,
but it’s kind of part of this desire
to democratize a lot of these capabilities.
We also released a toolkit for people doing evaluations
of their models and with clean data and everything.
And so in terms of kind of what’s the future
of building agents, my hope is that agent building,
unlike software building, is not something
that only a few people are able to do well.
My hope is that there’s actually something
that’s like widely democratized
and where everyone is empowered
to be able to create their own agents.
And I think right now we have this very scary view
of somebody else is going to create a system
that automates my job.
And that sucks.
That’s like really disempowering.
I don’t want that for my job.
– Yep.
– But the thing I love doing is automating parts
of my own job.
– Yeah.
– I like, you know, love making it better
in all of these different ways.
And that’s what we want to enable people to be able to do.
Like by giving you the tools to make your own agents,
that means that you can make your own things
that automate parts of your job.
And now your job can be higher leverage and higher level.
And now you can do a lot more.
– Right.
– And so we want to give kind of, you know,
someone else automating my job is very disempowering to me,
but someone giving me the tools
so that I can make my own tools for myself.
That’s very empowering for me.
And I think this mentality shift
is actually really important.
– Amen.
Ken June for folks listening
who would like to learn more about imbue.
You mentioned a blog as well.
Where should they start online?
Website, social media, podcast.
There’s a podcast, I think.
Where should they go?
– Great question.
So imbue.com is where we are on the internet.
And you can follow our Twitter account imbue.ai.
And we will have, you know,
as we start to release the product a little bit more
publicly, we’ll probably have announcements
and things where you can start experimenting
with what we’re doing.
So please do follow us.
There’s also a newsletter sign up
where we send extremely rare emails.
(laughing)
Because we mostly focus on building.
– You’re not an application trying to extract constantly?
– No, no, not trying to get your attention,
trying to make a useful product.
– Good, good, good.
Ken June, this is delightful.
Thank you so much for taking the time to come on the pod.
And best of luck with everything you’re doing.
Maybe we can check in again down the road.
– Definitely.
Thank you, Noah.
This is super fun.
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
(upbeat music)
[BLANK_AUDIO]
In this episode of the NVIDIA AI Podcast, Kanjun Qiu, CEO of Imbue, explores the emerging era where individuals can create and utilize their own AI agents. Drawing a parallel to the personal computer revolution of the late 1970s and 80s, Qiu discusses how modern AI systems are evolving to work collaboratively with users, enhancing their capabilities rather than just automating tasks.
