The Future of Generative AI Agents with Joon Sung Park

Summary

In a discussion about AI agents, Jun Zen Park provides background on the evolution of agents, from early assistive agents like Clippy to modern conversational agents. He outlines two branches of agent development: tool-based agents meant to automate complex tasks, and simulation agents that mimic human behavior. Large language models enable more advanced, personalized agents, though interaction challenges remain around deploying agents for high-risk tasks. Park sees initial success for agents in soft-edge problem spaces like games and entertainment before expanding to other areas. Though applications like ChatGPT show promise, he questions if conversational agents are the ultimate killer app compared to historic examples like Microsoft Excel.

Takeaways

• 😀 LLMs like GPT-3 made AI agents possible by providing the ability to predict reasonable next actions given a context
• 👥 There are two main types of AI agents - tool-based agents to automate tasks, and simulation agents to model human behavior
• 💡 LLMs still need additional components like long-term memory and planning for full agent capabilities
• 🎮 Games were an inspiration for early agent research aiming to create human-like NPCs
• 🚦 Current LLM limitations around safety and fine-tuning may limit the range of possible agent behaviors
• 🎭 Simulation agents for 'soft edge' problems like games and entertainment may succeed sooner than tool agents
• 🔮 Multimodal (text + image) agents are an exciting area for future research
• ❓ It's unclear if ChatGPT represents the 'killer application' for LLMs we expected
• 📚 Agent hype cycles have spiked and faded as expectations exceeded capabilities
• 🤔 Carefully considering human-agent interaction and usage costs will be key to adoption

Q & A

What was the initial motivation for Jun to research generative agents?

Jun was motivated by the question of what new and unique interactions large language models like GPT-3 would enable. He wanted to explore if these models could be used to generate believable human behavior and agents when given a micro context.

How does Jun define 'tool-based' agents versus 'simulation' agents?

Tool-based agents are designed to automate complex tasks like buying plane tickets or ordering pizza. Simulation agents are used to populate game worlds or simulations, focusing more on replicating human behavior and relationships.

What capability did large language models add that enabled new progress in building agents?

Large language models provided the ability to predict reasonable next sequences given a micro context or moment. This could replace manually scripting all possible agent behaviors.

What does Jun see as a current limitation in using models like GPT-3 for simulation agents?

Models like GPT-3 have been specifically fine-tuned to remain safe and not surface unsafe content. This limits their ability to accurately reflect a full range of human experiences like conflict.

Where does Jun expect agent technologies to first succeed commercially in the next few years?

Jun expects agent technologies to first succeed commercially in 'soft edge' problem spaces over the next few years, like simulations and games. There is more tolerance for failure in these areas.

What does Jun see as a key open question around why previous periods of hype around agents failed?

Jun wonders if past agent hype cycles failed because not enough thought was given to interactions - how agents would actually be used and whether they solved needs users really had.

What future line of questioning around large language models is Jun interested in pursuing?

Jun wonders if ChatGPT represents the 'killer app' for large language models that people were waiting for. He thinks it's worth discussing whether ChatGPT is actually as transformational as expected.

How does Jun suggest thinking about future model architectures that could replace Transformers?

Jun suggests thinking about Transformer capabilities as an abstraction layer - focusing on the reasoning capacity it provides. The implementation could be replaced over 5-10 years while still building useful applications today.

Where does Jun look for inspiration on new research directions?

Jun looks to foundational insights from early Artificial Intelligence researchers that have stood the test of time. He believes great ideas are timeless even if hype cycles come and go.

What aspect of current agent capabilities is Jun most interested in improving further?

Jun is interested in enhancing accuracy to better reflect real human behavior and diversity. This could enable personalized and scalable simulations grounded in real communities.

Outlines

00:00

👨‍💻 Jun introduces himself and his work on generative agents

Jun provides background on himself, stating he is a PhD student working on human-computer interaction and natural language processing. He discusses his interest in exploring how large language models can enable new capabilities, which led him to work on a project called Generative Agents. This involves using language models to create AI agents that can populate and behave realistically within simulation worlds.

05:01

🤔 What motivated Jun's research direction and focus on agents

Jun explains his thought process in identifying a research direction. When large language models like GPT-3 emerged, it was unclear what novel capabilities they would truly enable. Simple tasks like classification weren't fundamentally new. Simulating human behavior via agents was ambitious, tied to longstanding AI goals, and could open new interaction possibilities.

10:01

😎 The evolution of agents and how they split into two communities

Jun provides historical context on agents. There have been tool-based agents focused on automation vs simulation agents focused on modeling human communities. With large language models, these have splitted into two present-day communities - one using tools and one focused on personalized, multi-agent simulation.

15:01

🌟 How LLMs enabled a breakthrough in quality and scale for agents

Previously, generating sophisticated agent behaviors required manual scripting which didn't scale. Large language models provide the ability to predict reasonable next actions given a situation context. By adding memory and planning, this provides a scalable path to coherent, long-term agent behavior.

20:04

🎥 Multimodal Inputs (Image, Video) will make the next breakthrough

Currently agents operate via text from visual inputs translated to language descriptions. With multimodal models like DALL-E expanding to handle images, video, etc directly, future agents could take raw perceptual inputs and better simulate human perspective and behavior.

25:06

🎯Opportunity for more accurate agents that model actual communities

Jun sees opportunity to evolve agents to accurately reflect real human social dynamics instead of fictional scenarios. This is challenging today due to content moderation but may be feasible over time. Accurate community simulation could enable applications in markets, personalization, policy, and beyond.

30:08

😅 Deployments today constrained to "soft failure" cases, long road ahead

Agents face adoption challenges when mistakes incur high costs, as in purchasing transactions. History shows hype cycles then fading interest in 6-12 months. The first deployments are likely in entertainment, gaming etc. Standards for auditability, controllability needed before handling complex real-world tasks.

35:11

🔮 Don't assume Agents will work now just because technology improved

Despite progress, it is worth questioning why past agent hype cycles failed. Just having greater technological capacity doesn't guarantee actual usefulness and adoption. Key questions remain around real user need, interaction model fit, and compelling benefit over cost.

40:13

❓Is ChatGPT actually the "killer app" for LLMs we've been waiting for?

The breadth of ChatGPT adoption seems incredible. However, it is largely an interface wrapper on existing skills. We should question whether it truly constitutes the "killer app" that maximizes impact of LLMs. If not, what applications are still missing that could better deliver generalized value?

45:14

📚 Historical inspiration from Simon, Newell as timeless resources

Instead of recent papers, Jun is inspired by timeless foundational works, like Herbert Simon and Allen Newell. Their insights launched fields and won accolades. Unlike hype cycles, their textbooks showcase ideas that stood test of time and continue enabling impact decades later.

Keywords

💡agents

The video focuses heavily on 'agents' - AI systems designed to execute tasks and interact with humans. Examples from the transcript show agents being used for automation, simulations, games and more. The evolution of agent design and capabilities is a central theme.

💡large language models (LLMs)

LLMs like GPT-3 are presented as a key enabling technology for the current explosion in agent development. The models provide the natural language processing capacity to power human-like interactions.

💡generative agents

The guest's research on 'generative agents' is highlighted - using LLMs to automatically generate simulations populated by human-like NPCs (non-player characters). This demonstrates a key new application area for LLMs.

💡multi-agent systems

The potential to combine multiple specialized agents together into collaborative 'multi-agent systems' is discussed. This could enable more complex automations.

💡killer application

The speakers explore whether chatbots like ChatGPT represent the 'killer application' for LLMs that will drive mass adoption. Or if there are even bigger opportunities still to be unlocked.

💡limitations

Current limitations around accuracy, safety and scalability of agents are covered. For example, the over-cautious nature of ChatGPT. And the challenges of deploying agents in high-risk real world scenarios.

💡commercial deployment

Discussion on the most viable initial areas for commercial agent deployment in enterprises. With a focus on use cases that allow for a higher tolerance of mistakes during the current limitations.

💡future research

The guest mentions key avenues for future academic research on agents. Including increasing accuracy through personalization and scaling up the number of agents in simulations.

💡interaction challenges

A recurring theme is the historical failure of agents to deliver on hype because of unsolved interaction challenges. Despite advanced technology, if applications are not intuitive and valuable for end users they will not succeed.

💡multimodality

The potential to leverage multi-modal LLMs that can process images and video as well as text is discussed. This could significantly enrich the perception and situational awareness of agents.

Highlights

Agents have finally made their way into real enterprises with real use cases

Large language models provided the key capability to generate believable human behavior

Agents can be categorized as tool-based or simulation-based

Tool-based agents aim to automate complex tasks while simulation agents model communities

The agent architecture gives agents long-term memory and planning abilities

Multimodal capabilities like images will make agents more powerful

Accuracy limitations due to safety constraints may need to be addressed

Agents will likely succeed first in soft-edge problem spaces before hard-edge ones

The interaction challenges, not the technology, caused past agent hype cycles to fail

It's worth questioning if ChatGPT is the killer app we've been waiting for

The killer app should enable manipulating the key data type the technology generates

Learn from past insights that had impact and stood the test of time

Review recently published cutting-edge papers for the latest developments

Refer to pioneering works in AI and cognitive science for foundational ideas

Current hype cycles shouldn't discount timeless, foundational concepts

Focus on capacities and modalities over specific underlying technologies

Transcripts

the number of users who use CH gbt is

that's

incredible but I think it's sort of

worth asking ourselves is that sort of

quote unquote the killer applications

that we were waiting for chpt does feel

like it's a fairly simple wrap around

lar sling model because that's what the

main is open AI has done fantastic

things to make it safer and make it more

useful F tuning I think what's really

great but I think it's worth asking if

that is actually the killer application

why is it a killer application and the

answer might actually come out that

maybe it actually isn't the killer

application that we were waiting for um

in which case what is going to be the

killer application that's really going

to add value in a much more

generalizable

way welcome to AI in the real world I'm

your host my name is Joan Chen and I am

a general partner at Foundation capital

I work closely with startups are

reshaping business with AI in this

series I'll be holding in-depth

discussions with leading AI researchers

we'll explore how St of-the-art AI

models are being applied in real

Enterprises today to kick things off I'm

excited to speak with Jun Zen Park a phc

student in computer science at Stanford

June works at the intersection of human

computer interaction and natural

language processing he is best known for

his research on AI agents we break down

how AI is transforming agent design

share advice for Builders working with

these models and unpack why we haven't

yet found the perfect killer app for AI

agents here's our conversation how are

you good what about you great seeing you

again good to see you again it's been a

while because the unconference was last

I want to say May June May or June yeah

last June May or June wow TimeWise and

the world has changed I think that

agents have finally made its way into

real Enterprises with real news cases

and it was not um back then it was a lot

of like what what could this be right um

right thanks to you and some of your

work which is why I'm super excited to

have this

conversation together um especially

right now since like uh Enterprises are

um uh like are are sinking in a real way

to adopt so uh so I thought who can can

we chat with that would have like a

really interesting perspective and

that's why we reached out back to you so

really appreciate the time of course

thanks for having me do you mind maybe

just to start like getting giving giving

us a quick rundown of like what's

happened maybe some of the background

that you have uh building this

technology yeah so let's see uh so do

you want me to just sort of speak about

sort of what has happened in the past

six months where sort of what would be

interested uh interested for you just

just a brief overview of what youve

worked on and also what's happened in

the last uh six months to a year in

terms of evolution yeah all right that

sounds good right so um I guess I do a

quick intro so I'm sort of I'm a PhD

student here uh sort of working the area

of HCI and and NP so as you know sort of

the work that we've done I think the one

that I'm sort of mainly known for is

this paper code generative agents um and

generative agents in particular was a

project that to ask can we use our

language models to create General agents

that can populate a simulation world

right so if you play something like Sim

City or Sims uh can we actually create

these NPC like characters that would

actually flood into the City and

actually live like humans and by

definition it is sort of everything from

how they would wake up in the morning

talk to each other form routines and

relationships all the way to creating

basically communities and ersing social

dynamic

Dynamics and sort of my interest in this

area really stems from this idea so this

is sort of what people at the

intersection of human computer

interaction and natural language

processing and machine learning like to

ask which is we now have these really

amazing models like large language

models and Foundation models the

question really becomes what are you

going to do with them right these models

are new and they're great and we think

they have great capacity but are they

really going to enable us to do

something that's quite and unique and

that has been sort of the focal point

for a lot of the research that I do and

ultimately the conversation that we got

down to was this idea of well these

models of train on Broad data like uh

like the web Wikipedia and so forth so

they can actually be used to generate a

lot of believable human behavior when

you're given a very micro context so can

we actually piece this together to

create human leg agents which is

something that AI more broadly has

envisioned since its founding day

um and we decided that this is the time

to do that and so that's how we got to

where we are uh so that's the general of

Agents uh and this is the paper that was

published in April last year or we put

it on archive in April and was

officially published November which is

which is crazy how much um uh the world

has has developed I'm curious what

initially

motivated uh this topic for you I'm sure

you had lots of different options in

terms of what to research and study why

did you decide to focus on

this yeah so ultimately it really was

the question of what will L language

models these new models that are being

trained really going to enable us to do

and when I started my PhD was around

like 2020 and that was when gpt3 was

just about to come out during my first

year we wrote a paper called uh

Foundation

models uh which sort of made this

observation that there's going to be

this new wave of models that's going to

come out where we're not going to be

training these models for a specific

task but rather we'll be training for a

modality right we're going to be

training this language model that can

process language and so

forth

and we thought that was going to be a

big opportunity there in terms of what

we can do with them but the question of

what are we going to do with them was

incredibly unclear so really our first

instinct as sort of researchers in more

machine learning in NLP community where

we sort of were drawn to was this idea

of can we do classifications or

generations with these models and seeing

that these models could do that was

really exciting because we didn't train

these models to do that but they could

but more from the interaction

perspective doing classification and

simple generation was something that we

already knew how to do so that didn't

feel fundamentally new so really the

question again became what are we going

to do that's going to be truly new and

transformative in the sense of

interaction

so that's uh what really Drew us to look

for these kind of ideas um and again

that's we thought simulating human

behavior in general computational agents

that felt like a big problem because in

part because it's something that again

our community had wanted for many

decades uh it was sort of the idea that

people in more the cognitive science

field that will inspire the early AI

research like alen new and Hart's these

folks were asking and we were certainly

inspired by those ideas and of course we

thought it would be a lot of fun because

we sort of grew up with Sims Pokemon and

these kind of games in the 90s and early

2000s and we were certainly inspired by

those games as well I love I love those

games as well and it's nice to see some

of that um play out in the real world

now I agree I think games are fun in the

sense that um like you know I think they

are inspirational in many ways because

they do they very forward looking in

many ways right because you can be a

little bit more playful and I think

research can be in many ways playful

especially when you're trying to do

really forward looking research so it

certainly is a big inspiration and I was

just going to sort of end that comment

by saying that I think it's worth asking

for us as a community what's going to be

the new sort of quote unquote cure

application of these

models in the sense that um when we had

personal computer in the early 80s and

so forth the computers were very cool uh

but what really made them into household

applications were the existence of this

what we would now consider as killer

application of PCS like Microsoft Excel

that really made tabular information uh

usable and scalable I think we l

language model Community are should also

be looking for those kind of ideas as

well because that's going to be

ultimately what's going to really

transform the user experience around

these models and I think we're seeing

some great uses uh usage of these models

but I think there's a lot more to do

going forward makes makes a lot of sense

when you look at the what's happened

since April right a lot of things have

changed uh we have new LM capabilities

we have a whole flurry of startups

building in the space could you maybe

summarize what what you've

seen right

so right so agan cty has been a big

thing uh especially first the latter

half of

2023 this is how I'm seeing it um agent

community in the sort of the way I view

it has split into two communities I

would argue now

so maybe it might actually Mak a little

more sense to really talk about like the

history of Agents because Asian became a

big thing last year but this is not a

new idea in and out of itself right in

this in even even in the commercial

space we actually had agents like

Microsoft clippy I'm not sure how many

of us will actually remember that but

there used to be these agents uh in this

in sort of our industry and in research

so this is certainly not a new idea uh

so if you go all the way back um so we

had agents like clippy and in many ways

these agents especially in the

reinforcement learning and machine

Learning Community agents were these

elements that basically

could simulate human behavior I think

that is ultimately sort of underlying

thesis but many of the agents were given

tools to automate certain tasks and the

tasks it were meant to automate were

tasks that are not simple right it's not

something like you're running a for loop

with your python code but it's a little

bit more complex in there right it

operates in much more embodied spaces or

in spaces that we often operated right

the web uh right can it the simplest

example with these kind of tool-based

agents are can it order me pizza can it

buy plane tickets and those might sound

simple but we we know from our

experiences that even ordering pizza

actually does require multiple steps

right we need to travel to certain

websites we need to look through the

menus actually make the payment and de

with sort of entering your address and

so forth so that was one Gena of agents

that already sort of existed for a long

time or I would say all genas of Agents

sort of existed but that was one draine

that was highlighted in the past uh so

you see things like clipp is also in

that draine as well you're a Microsoft

uh office user clippy would try to

automate some tasks for you based on

your prior interaction with the

software another a set of Agents um was

this idea of simulation agents or agents

that were cre cl to clarify on that

point those agents are single agents

correct they can be single agents they

were

often implemented as single agents

that's right I don't think by definition

they actually had to be single agents so

you actually try you're now seeing in at

least in the research you're start to

see glimpse of people trying to imagine

what would it look like for these agents

to be in a multi-agent setting so

research paper that I remember coming

out after uh gener of Agents was

basically what if you have a company of

Agents right there's going to be a CEO

but there's also going to be designer

agent who spe who works in some other

aspects there's going to be editor in

this company and those are still much

within the literature of what I would

call tool-based agents right they're

trying to automate some complex tasks

for the users and I think there's going

to be a lot of sort of really big

opportunities in the space that's

something that people have been working

on for a long time uh for for all the

right reasons now another community that

has form but to some extent actually has

a slightly different route is agents

that were created for

simulations uh and these agents were

certainly a part of games right in the

past we had Sims but we also had these

NPC characters that we could interact

with now those NPCs and agents back then

were very much like it was simpler

agents that were either rule-based uh

there were some reinforcement learning

agents back then as well in

space uh but another one that we could

usually think about were agents that

were used basically in social science

economic agents or agents that would

simulate our policy decision making and

so forth uh and those agents were also a

part of this

literature and what we're seeing today

is we're one recognizing that lar

language model is simulating human

behavior so it touches on all these

agents that it can be a foundational

sort of a architectural layer for

creating all these different sorts of

Agents but in terms of our initial

application spaces we're seeing this

split where there's one Community who's

now deeply interested in agents using

tools but another community that that is

deeply interested in this idea of can we

simulate and this is where I would say

like

multi-agents uh and as well as

personalization is really starting to be

highlighted in the simulation space

because it's a little bit more directly

incorporated into the idea of

simulations who are we simulating for

what are we simulating who are we

simulating and by definition simulations

often happen in this multi-agent space

so those are the two communities that

you're starting to see um so generative

agent certainly stands on the far end of

the simulation based agents whereas some

other projects that were also really

cool last year I think a lot of sort of

uh open AI gpts I would say are another

end of the simulation agents or another

end of toolbase Agents so those are the

axes that you're sort of seeing right

now now I sort of End by saying my hunch

actually is again though because they

all start from the same technical thesis

that we can simulate human behavior they

will merge in the end I don't think they

will be like completely separate thesis

like five to 10 years down the line it's

more going to be the question of where

are we going to make our short-term bets

and what's going to be an interesting

and meaningful application space in the

next two to five years so that's the

field that I'm seeing and how it's

developing right now before we maybe go

into that could you maybe describe how

um llm specifically has

affected um the especially the latter

cohort right what is the before and what

is the after and what is the magnitude

of improvement because of this uh

technology that's now you know cheap

enough to use right so L Lang motor is

really what made this possible uh that

is really the fundamental T stack that

we needed uh in the past when you wanted

to create and this goes for both types

of Agents uh tool-based and simulations

what you really needed

was you basically needed rule-based

agents that was the most common and Rule

based agents are sort of a more

sophisticated way of saying we're

scripting all the behaviors so imagine

you're building an NPC in for a game a

human author would actually write every

sentence that the agent would say to the

user for instance it would author would

actually describe in in either code or

language if this happens you do this so

you basically design all the possible

behaviors now that is expensive and not

scalable right um and that was the

fundamental block that we had now

toolbase agents had similar issues that

in many of the context it had to operate

it's not very generalizable tool so if

you sort of see how clippy or even some

of the agents that we're using today

very simple types of agent actually are

already embedded into our daily usage so

you may have used Google spreadsheet or

Google doc it would autocomplete in some

very rudimentary way that actually could

be considered in some ways an agent in

this direction of tool-based agents and

the rules they were using so far were

very simple it's not exactly rule based

but it is something that was very much

hardcoded into the agent's behavior and

there wasn't there was some learning

going on but that were very root like

very straightforward simple uh like

statistics that we're using what L

language model changes is l l model

gives us a single ingredient which is

given a micro context micro moment let's

say I'm sitting in this room talking to

Jan and about let's say generative

agents or simulations and so forth given

that micro moment description a language

model is extremely good at predicting

the next moment right so what is what is

the what are the reasonable set of

things that June might say in this

particular conversation given what he

knows it's very good at doing that um

that on its own not is not a perfect

agent uh or it's not the complete

ingredient that you need to create these

agents that are meant to live for many

many years or decades but they are the

right ingredient or building block that

we needed because that can be used to

replace what was in the past manual

authoring in the past we had to manually

author all the possible sequences given

any micro moment but L's language model

can come

in so given that ingredient what we

really could do is bake in long-term

memory and some reflection module on top

of it and planning module so given the

micro ingredient plus an agent

architecture that we give it on top of

it these agents can basically now start

to function as something that can

operate in that in a world that's much

like ours with a fairly decent degree of

long-term coherence so that's where we

are and that's really the difference it

made and I'd say this is sort of a zero

to one difference not a degree

difference because before Lis language

model this was not

possible what else is um so we like

large language models gave memory gave

context uh gave interactions to these

agents what else in a perfect world

would these agents have in order to

better mimic the real world like what's

maybe in the next next step just out of

curiosity uh right so to clarify L

language model doesn't actually have so

L language model provides one element

it's the micro sort of a module for

predicting the next sequence yeah uh the

it is the agent architecture that

actually ends up giving the memory and

planning ability um but those two pair

becomes a fantastic combination um now

going forward what I do think is going

to be interesting are so right now we're

using L language model but we may have

have all noticed that things like chpt

can now not only deal with just language

but also other modality like

image I think that's going to be really

interesting right

so right now let's say if you and this

is sort of based on my prior work like

call generative agents and we had this

game world like Sim and that we call a

Smallville the way these agents

perceived and operated in their world

was basically by um by translating like

our system translating the visual world

into natural language so we would tell

the agent you are in your apartment or

you are in the kitchen

talking to someone so we would actually

take the visual world and use our system

to translate the visual world into

natural language and then feeding it to

the agent architecture that would use

Earth language model to process this but

now with these models being able to deal

with multimodal aspect we might actually

be able to bypass that pH and go

straight from here is the visual world

or space that you're seeing right now

that is your memory now act on it I

think that's going to be potentially

very powerful because in

part image is much richer to some it

conveys a lot more I I do come from

natural language uh processing

background at least that's my other half

of my sort of academic background so I

have bias to towards believing the

natural language is profound and I think

that we're going to be that will be the

case going forward as well but image

does offer something that just language

alone does not so image is going to be a

big thing now imagine in the future

video is going to be a big thing as well

then gradually the more these agents

where basically increasingly get more

powerful as this new modality gets piled

on so that's something that we should be

looking forward to that's that's great

what are some on the downside what are

some of the

limitations um that you're seeing in

terms of these um these agents

especially General

agents right um

so to so there are limitations that I

can mention just about sort of in sort

of in the context of our work um and

then I think there are going to be

interesting limitations that are much

more application specific so for

generative agents today certainly the

technical limitation right now might

have to do with things like so you're

using whether it's an open so right now

we use open AI model open AI has

actually done a lot of work to make the

model safer and I think for open AI I

think that was the right approach in the

sense that what they really wanted to

create was these chatbots or agents or

chpt that was safe tool to use for most

of people now if you want to run a

simulation or create truly accurate and

believable agents with something like

chpt however that could become a

limitation because what we really

experience as humans uh is we fight we

sometimes have conflicts we disagree

with each other and that might not be

something that something like Chachi B

that's been fine tuned to not behave

that way to remain safe it's something

it might not be something that these

models will try to surface and that

could be a potential Block in creating

more accurate more believable

simulations or agents for that matter so

that certainly is one limitation right

now that we're seeing um an interesting

way to tackle this I think going forward

is to use open source models or other

models that have less of this fine-tuned

nature um but it's going to be highly

dependent on the models that we'll be

using for this so I think that's one

thing to look forward to got it got it

that's that's super that's super helpful

maybe one last question on kind of the

research side um when you think about

future areas to explore for you

specifically yeah um what are some of

the more narrow topics that you're

hoping to to dive deeper in given

world so ultimately I think making the

agents more accurate as rep for these

agents to be more accurate reflection of

who we are I think it's going to be a

really interesting research and much

like it's I think it's going to that's

going to be an area that's going to have

more of a research and broader impact um

so right now you may have seen the sort

of simulation demo the agents that live

in that simulation are fictional that we

just for instance we have an an agent

named Isabella we told Isabella that she

is a cafe owner and LGE language model

basically makes up what a Persona that

is reasonable given

description but I think it's going to be

far more interesting if we can make

these simulations actually closely model

our actual human communities so it's not

just functional but actually has

groundings that's going to open up from

our perspective an entirely new set of

application spaces as well as research

impact Um this can be used for instance

to accurately model or predict markets

or it's going to be able to use to more

closely personalized many of these

agents for individual use cases so

that's something that we're looking

forward to in terms of sort of a

particular topic that we're diving into

that plus of course scaling of the

agents I think that's another big one

but those two that that makes sense you

know one of the things that's missing in

most

AI uh Technologies is like really the

emotional part of how humans feel right

right like all of that data is largely

not captured and therefore not part of

any kind of models today uh language is

one small output of what we have it's

it's a very important output for sure

right um but it's still one small output

so I wonder how uh we might be able to

incorporate some of the data around our

emotions I agree and thoughts in the

future um maybe let's move on to the

applications today since you talked

about some of the challenges for agent

many organizations are thinking about

how to use large language models today

right there's a huge amount of

aspirations a subset of them are also

thinking about what are some of the

agent technology applications that are

viable within an Enterprise which has

limitations around infrastructure around

data silos security and all that

stuff any particular areas where you've

seen companies be successful at using

these Technologies in

production right so I think this is

going to be incredibly like Case by case

uh answer so I let me

think or if not like any hypothesis as

to where you might see the first

commercial deployments at

scale

right so there is something that I have

I this is there's a message that I have

in trying to communicate I think in in

different settings so this is not

something I'm I sort of conve for the

first

time and and my opinion has been getting

updated but I think fundamentally I

think this is right so the way I've been

describing it is in human computer

interaction or in

most task settings there are two types

of problems that we deploy our machines

or agents in one

has very hard Edge problem spaces these

are things like hey order me pizza or

buy me a plane ticket these are tasks

where there's a very concrete outcome

and there often is a right or wrong

answer that the agent has to achieve at

least from the user's perspective there

is something that will absolutely be yes

or

no and then there are problems PES where

we have soft Edge problems these are

problems where we can increase L he'll

climb towards sort of being better but

at certain level it starts to actually

become useful so to make this intuition

a little bit to make this a little bit

more intuitive here for instance if I

guess the worst possible case scenario

is I ask the agent to buy me a plane

ticket and it it bought me the wrong

ticket that just goes to a different

place then that's like a heavy

no uh whereas let's say I ask the agent

to sort of simulate the behavior that is

sort of fun so that I when I'm in a game

this is sort of entertaining and

interesting that might be something that

the Asian doesn't need to be quite

perfect in but it can still get there

quite quickly and then we can gradually

improve those two are the spaces that we

can sort of in terms of when we consider

where to deploy or how this will

actually make its first impact we might

actually be looking at those problem

spaces first and these are the classes

that I'm seeing if I were to make my bet

AG agents will likely succeed first in

the soft Edge problems basis and will

gradually inch into making it work in

the hardage problems space this has been

sort of an intuition with agent research

Community for some time so when clippy

for instance

failed our intuition there at least from

research perspective wasn't that these

agents failed because we didn't have the

technology there certainly these were

deployed and there were some confidence

around the technology but the the

problem there actually was with

interaction that when agents are

deployed in hardage problems basis it's

often deployed in states where it

actually has to have a fairly long chain

of steps and fairly High the risk were

fairly High when it fails the cost of

correcting its error is actually quite

High Cost of auditing its error is

actually quite

high so when these agents are deployed

in hardage problems SP bis it has to

reckon with the fact that it will

undoubtedly make

mistakes and it when it does make

mistake it has to be increasingly

auditable and controllable by the users

so that the cost of correcting its error

is not high enough that from the user's

perspective the cost benefit anal

analysis basically has to make sense and

that's been a fundamental challenge with

agents that's why in every era we see

the interest around agents Spike for a

while and then it quickly subsides after

maybe a half a year or two year now

there's there's a real question now

though that given the large language

model and the progress we saw that this

might not be the case this time or at

some point we might be able to make this

work but for now so I'm closely

monitoring this as well and I think we

all should I don't think we should just

go and say because it didn't work before

it's not going to work this time but my

hunch is that we will likely see very

similar pattern Arise at least for the

first of future and we haven't quite

dealt with the interaction problems with

those types of Agents so I think it's

much safer to assume that it is going to

be the softage problems basis and that's

why in some in many of the aspects

that's why we our team was also

interested in this idea of simulation

because simulation is sort of the prime

example of soft problems spasis where

the simulation has to be good enough for

to be start being useful that's also why

I think a lot of really early promising

AI startups that's going to go in the

agent space are places that does NPCs

for games because those are very safe

softage problem spaces where the agents

can fail but that's okay and gradually

we'll sort of go to the other area as

well but I think that's where the impact

is going to start in the next couple of

years I I I also think just seeing um

the startups in the the space in sectors

and functions that allow for failure

like you said include things like

marketing where if you Market

incorrectly it's not that big of a deal

if you uh write the wrong

code that's probably going to be a

bigger deal if you pick the wrong

security features that's a huge deal

right U if you pick the wrong things for

healthcare that's you know that's that's

an even bigger deal so there are degrees

of Tolerance fault tolerance within the

Enterprise um that's one thing and the

second on the consumer side especially

if the agents are just assisting

consumers by not executing on

anything um that could probably also

work for example you know there's a

company called rewind which is using

some of the agent Technologies I believe

um and they're getting a bunch of

consumer demand but but what consumers

are doing is just searching for a

behavior that they have had

before um and this this this product is

helping them do that versus do anything

you know um real world really

interesting uh but that's that's super

the way that you frame it is very um uh

is very

useful what about just from an

architecture

standpoint uh large language models

enabled by Transformer

architectures this is a whole different

direction we're already seeing companies

that are saying hey Transformers are are

not the most efficient you know

inference costs are very high uh let's

look at this the next thing have you

spent much time thinking about that um

and if so any impact to the work that

you're

doing right so certainly um next sort of

model that we're going to be banking on

I think that always is an important

topic and that's something that I think

we as a community always has to sort of

monitor because I think you're right the

Transformer is not going to be the end

model that would be hopefully I mean KN

going what the Hope here is that we

wouldn't be using Transformer 10 years

down the

line uh but one way that we do view this

is this is very much like a programmer a

programmer way of looking at this we

view this in abstractions right so what

Transformer has gotten us right now is

this amazing capacity for reasoning and

processing information and generating

information so it might be the case that

in the future that task will be done by

even better models and hopefully that's

going to be the case um but for the sake

of building applications it is true that

we can sort of view this as a layer of

abstraction that there might be some

other of Technology that's going to be

powering it in the future but really

what we're focusing on is the capacity

and the modality what kind of reasoning

using what modality can these technology

that exist today do and we're going to

be building on top of it so I think

that's sort of uh our way of looking at

it in sort of I say medium term like in

the next three to five years now if

you're look because right now there are

some promising architectures that's

going on that's sort of been created at

the Forefront of I more in the machine

learning and natural language processing

communities that I personally getting a

little bit excited about um but at the

moment those are still in the very much

in the research Pro uh phase and can you

share some examples of that yeah so I

think like there's one model that

recently came out like Mamba by some

folks uh those are from Stanford uh

folks so I think the author is now as

CMU and Princeton uh sort of all with

sort of this community so that's one

example of sort of a potentially

promising or interesting model and

that's a one that I recently heard about

that I think is interesting to be

looking

at but these models for them to be

deployed at scale in a commercial way if

we decide to basically go with certain

model that's getting created today it

will give us

maybe two to five year timeline before

they can really take off because

Transformer is

not it is a relatively modern model but

it really how you look at sort of the

timeline this transform sevenish years

sevenish years so I think hopefully we

if we find something like this time

maybe it's going to go much faster but

it still took about sevenish years for

chat PT to really come out so it's not

immediate uh whereas a lot of

interactions that we can build I think

there's a lot that we can do like today

to create really cool experiences so I

think that's how we're looking at this

there is a medium-term this is where we

are focused on the level of extraction

that this is the capacity that we'll

have and then maybe in the down the line

five to 10 year term we can really be

looking forward to some new models

that's going to make an

impact that's great that's great um cool

very cool um maybe more generally if you

just zoom out for for a moment when you

look at the ecosystem today uh what are

some of the problems that you want to

see solve we talked about multimodal a

little bit we talked about new models

right after Transformers that might come

out what are some of the problems that

you are most excited about someone

solving not necessarily uh you

personally but someone solving I sort of

have two in mind um and it's a little

bit less of a this is a specific problem

that I want someone to solve but it's

more sort of questions that I have that

I think more of us should be thinking

about and to some extent and this this

happens a lot with sort of the way I do

my research as well where I get inspired

by big problems or foundational problems

that we had in previous decades because

often times there's a lot of insights

that we can learn from the past as we

sort of build on the

future

one certainly I'm embedded into this

agent

space one is in the past agent had its

hype cycle basically but it

failed um that the hype cycle lasted for

a couple of years and then people very

quickly lost interest basically because

it didn't quite deliver the promises

that it

had I think it's worth asking ourselves

why that was the case I think the

opportunity this time is real but I also

think the opportunity in the past was

also real to some aspect as well so just

because the opportunity is real and

language model is really cool doesn't

necessarily guarantee us at least from

my perspective that we're going to that

agent will finally be a thing that

everyone will use down the f I think

there is a future where that's that will

happen at some point I think it might

even happen this cycle but I think it's

really worth asking as a community why

did it fail in the past so that we don't

repeat those mistakes one sort of main

thing I'm sort of curious about that I

don't think a lot of are thinking about

is actually not the technology part but

the

interaction how are these agents going

to be used in what way because

ultimately that's where it really

delivers value to the end users and

that's where agents in the past have

failed that was really cool technology

but we didn't seriously ask ourselves is

this something that people really need

and does the cost benefit analysis of

using these agents and learning how to

use them well really makes sense for the

broader user user base so that's one and

other one is sort of

my it's a little bit of a hot take but

it's also a shorter take which is we

have large language models and I think

these have made a huge impact already

right the number of users who use CH

jbts that's

incredible but I think it's sort of

worth asking

ourselves is that sort of quote unquote

the killer applications that we were

waiting for because in many ways

chbt or maybe it is and I I think if it

is I think somebody should articulate

this but chpt does feel like it's a

fairly simple wrap around L language

model because that's what itain is and

you know open AI has done fantastic

things to make it safer and make it more

useful by tuning I think what really

great but I think it's worth asking if

that is actually the killer application

why is it a killer application and the

answer might actually come out that

maybe it actually isn't the killer

application that we were waiting for um

in which case what is going to be the

killer application that's really going

to add value in a much more

generalizable way um that's a very

abstract question it's for for now it's

for me it's just a hunch that I think

there's something to be asked about

there and if I'm wrong like I would also

love to hear again somebody really say

we already have this clear application

maybe it's co-pilot track GPT and here's

why

but for now this is a question that I'm

still asking

myself that makes sense um and thank you

thank you for sharing that what are some

of your favorite AI apps today that you

use I love chbt I use it every day I

chpt did make a difference in my

workflow uh so I so as sort of a

researcher and one of the main things I

do is I program every day uh or at least

most days or I write or write papers so

I do one of those two things chpt is

fantastic at both uh so it's you know

it's as all programmers sort of know you

know we sometimes don't bother

remembering all the different functions

or

documentation it's very good at

generating a lot of the code when I have

an idea really impressive it's also

quite a good editor so if I make grammar

error in my sort of sentences chpt will

usually catch them for me right it's

simple and easy thing but it's good

enough now that it's actually making a

difference in the

workflow um so I say trpd for sure by

extension I think co-pilot will make a

difference um so it's sort of worth

asking maybe going back to the question

around like killer replication what is

the definition of killer application I

think it does some people Define it as

application that has more users and the

fact of that I think always has to be

the case that no killer application has

no user killer application by default

means the application that will have the

most number of

users but I think there's more

theoretical definition to what a killer

application is that implies a lot of

users or the most number of users but

for instance if we look back to the

prior era of PC a killer application

that I mentioned was something like

Microsoft's Excel or this tabular data

format the thing that would man that

would let us manipulate the tabular data

so really the definition in a more

theoretical sense of K application here

is there's new technology stack that is

been developed there's a new file type

that is getting generated then the

killer application is the one that would

let us manipulate the file application

file type that's one theoretical

definition that one could give at least

that's sort of the definition that I've

been toying with I think it's an

interesting one I don't think it's the

only one but those are sort of the ways

that I'm looking at this that that that

makes a lot of sense I also use CH GPT

every single day uh it's been um it's

been very helpful everything from coming

up with menu names to you know rewriting

uh emails that don't sound as As Nice um

and I've tried a little bit to you know

give it files and images oh I actually

help my mother create a background she's

a dancer so she was performing and

wanted a very specific background for

her dance and you know I created that

for her using chat gbt so all sorts of

utility there um but I love the way that

you framed um the last potential

application maybe just one last question

from from my end any um any resources or

books that you you love that is uh on um

this

topic on this

topic

right I I do think and this is this is

often the case with many many of The

Cutting Edge like spaces I think a lot

of the papers that are coming out that

are gaining a lot of attention I think

those are sort of wor checking out as

sort of the resources it's not exactly

like here's one book that we can all

look at uh uh but things are moving fast

enough that I think um I think those are

sort of interesting resources are just

things that are getting created today uh

and they're that quations uh so those I

sort of mentioned as sort of a generic

answer I do think I think this has been

a sort of running theme in some of the

things I mentioned today I get inspired

by insight

that basically had impact and stood the

test of

time right and the reason why that is

the case is because I personally think

all the great all great ideas are sort

of Timeless that because current hyp

cycle is over doesn't mean they're less

interesting or less meaningful for sure

that foundational ideas that will

continue to have impact um so when I

look for resources I actually look back

to books from truly the prior

Generations um so some of the works that

I often go back to are works by Herbert

Simon Ali new those are founders of AI

and this many of this Fields who would

later go on to win the Turing award the

Nobel Prize and so forth and those Works

uh early cognitive psychologists and

scientists inspired my worka uh and

their textbook uh those people actually

have written books because they were

much more established than sort of The

Cutting Edge spaces today uh so I go

back to those as sort of my personal

resources for getting

ideas that's great thank you so much

this was super super helpful to me

personally because what we do as

investors is we try to understand the

impacts of technology and start to

invest in companies when it

becomes at the beginning of when it

becomes commercially viable right so to

your point around what are the problem

spaces what are the applications in

which this can be applied in a cost

effective and secure way that where the

end user is willing to

interact uh and dra and and get value

that's when we start to come in uh and

invest in these companies which will

hopefully uh be much bigger companies in

the future so really appreciate this

chat yeah it was

[Music]

fun