Vertical AI: The Next Big Startup Trend

Summary

The transcript discusses the emergence of vertical AI startups focused on specific roles in society, such as AI for lawyers, doctors, and software engineers. These startups aim to enhance productivity by automating complex tasks. The conversation highlights the potential for AI to transform various professions, suggesting a future where individuals can leverage AI 'conductors' to manage virtual teams, leading to a significant increase in productivity and the possibility of millions of small companies run by solo entrepreneurs with AI assistance.

Takeaways

• 🚀 The emergence of vertical AI startups is focused on specific roles in society, building specialized applications for tasks like law, medicine, tax assistance, and customer support.
• 🤖 Large, general-purpose AI language models like OpenAI's GPT and Google's Gemini are trained on the open internet and can answer a wide range of questions, though not always accurately.
• 🔧 A startup named Cognition has debuted an AI software engineer tool called Devon, which has gone viral for its ability to fix bugs and build apps in real-time.
• 💬 Speculation suggests Devon may be built on GPT-4 from OpenAI, though this is unconfirmed. The tool is designed to enhance reasoning and long-term planning capabilities within existing language models.
• 📈 Benchmarks are being developed to test the performance of language models in coding, with Devon showing superior results compared to generic language models.
• 🌟 The progress in AI is rapid, with the potential to make complex jobs more accessible through command line interfaces, automating processes behind the scenes.
• 🔥 The increased efficiency of developers through AI tools like Devon could lead to a greater demand for human expertise to manage and refine AI-generated code.
• 🛠️ The shift from AI co-pilots assisting developers to AI taking the pilot role signifies a transition in how AI is integrated into the development process.
• 🌐 The evolution of AI is expected to lead to the creation of more solo entrepreneurs and smaller companies, leveraging AI to multiply productivity and reduce operational costs.
• 📊 The long-term impact of AI on job roles may not be about replacement but rather about 'leveling up', where humans work alongside AI to achieve higher productivity and innovation.

Q & A

What is the main difference between general-purpose AI language models and vertical AI startups?

General-purpose AI language models like OpenAI's GPT and Google's Gemini are trained on the open internet and can answer a wide range of questions, while vertical AI startups focus on specific job titles or roles in society, building applications tailored to those specific functions.

How does the AI startup Harvey assist lawyers?

Harvey is an AI specifically designed for lawyers, presumably to automate legal tasks, provide legal research assistance, or offer other law-related services that save time and improve efficiency.

What is the role of the AI notaker developed by the startup Bridge?

Bridge's AI notaker is designed to assist doctors by taking notes during their work, potentially saving them hours each day by automating the documentation process.

How does the AI tax assistant, Tax GPT, function?

Tax GPT is an AI tool that assists with tax-related tasks, likely automating tax preparation, calculations, and offering guidance on tax laws and regulations to streamline the tax process for users.

What is the primary function of the AI software engineer tool called Devon?

Devon is an AI tool that mimics the role of a software engineer, capable of fixing bugs, fine-tuning AI models, and building applications end-to-end, thereby assisting or even replacing human engineers in certain coding tasks.

What is the speculated basis of the Devon AI tool?

There is speculation that Devon was built on GPT-4 from OpenAI, although this has not been confirmed. The CEO has mentioned that Devon was created by tweaking reasoning and long-term planning capabilities in an existing large language model (LLM).

How does the performance of Devon compare to other major coding models?

According to the data from coding benchmarks, Devon outperforms generic language models, indicating that it is more effective in coding tasks due to its specialized training and capabilities.

What is the potential impact of these role-based AI tools on the job market?

These role-based AI tools are expected to make certain job functions more accessible and automate complex tasks, potentially increasing productivity and efficiency. However, they may also require human oversight, especially for tasks that the AI does not perform perfectly, leading to a need for human expertise and creativity.

How does the evolution of AI in coding compare to the evolution of the internet?

The evolution of AI in coding is similar to the internet's impact on networking software. Just as the internet enabled new business models by moving traditional businesses online, AI is now enabling the creation of specialized AI tools that replace specific human roles, leading to a new era of productivity and innovation.

What is the potential future of startups and solo entrepreneurs with the advent of AI conductors?

The future may see millions of companies run by solo entrepreneurs with the help of AI conductors, agents, and bots. This would allow individuals to manage and scale their businesses with minimal human resources, leading to a significant increase in productivity and potentially a reduction in operational costs for companies.

How has the process of creating and launching a new product or service evolved over the years?

The process has become significantly easier and more cost-effective over time. The advent of cloud services like AWS, developer tools, APIs, and app stores has abstracted away much of the infrastructure and distribution challenges, allowing solo developers and small teams to launch products with less capital and resources than ever before.

Outlines

00:00

🚀 Emergence of Vertical AI Startups

This paragraph discusses the rise of vertical AI startups that focus on specific roles or job titles in society, creating specialized applications. It highlights examples such as Harvey (AI for lawyers), a bridge (AI notaker for doctors), TaxGPT (AI tax assistant), and NCR (AI for customer support). The conversation emphasizes the potential of these startups to improve productivity and efficiency in various sectors. The recent launch of a tool named Devon by a startup called Cognition is mentioned, which has gone viral for its ability to fix bugs and build applications in real time. Speculations about Devon's development on GPT-4 from OpenAI are addressed, along with its performance on coding benchmarks compared to generic language models.

05:01

🌟 Role-Based AI and the Future of Development

The discussion in this paragraph revolves around the impact of role-based AI on the field of software development. It emphasizes the rapid progress in AI, shifting from co-piloting to taking the pilot's seat in coding tasks. The conversation explores the potential of AI to automate and scale up tasks traditionally performed by humans, such as lawyers and accountants. The idea of AI replacing specific human roles is contrasted with the notion of enhancing human capabilities through AI collaboration. The paragraph also touches on the challenges of working with existing code bases and the different approaches companies are taking to integrate AI into software development, such as GitHub's co-pilot and Source graft's Cody.

10:02

📈 The Impact on Startups and Entrepreneurship

This paragraph examines the broader implications of AI advancements on startups and entrepreneurship. It predicts a future where individuals can run companies with the assistance of AI conductors, agents, and bots, significantly reducing operational costs and enabling solo entrepreneurs to succeed. The conversation also reflects on the historical context of app development and the decreasing resources required to launch a startup. It acknowledges that while the barrier to entry has lowered, creating something profound still often requires a small team and financial investment.

Keywords

💡AI startups

AI startups refer to new companies that are focused on developing and implementing artificial intelligence technologies. In the context of the video, these startups are creating specialized AI tools for specific industries or job roles, such as lawyers, doctors, and software engineers. This approach is contrasted with general-purpose AI models like those developed by large tech companies.

💡Language models

Language models are a type of artificial intelligence that is designed to process, understand, and generate human language. They are trained on vast amounts of text data and can be used for tasks like answering questions, translating languages, and even creating written content. The video discusses both general-purpose language models like OpenAI's GPT and specialized models tailored for specific tasks.

💡Vertical AI

Vertical AI refers to artificial intelligence applications that are focused on specific industries or job roles, as opposed to horizontal AI which aims to be broadly applicable across many domains. Vertical AI solutions are designed to address the unique needs and challenges of a particular sector, enhancing efficiency and productivity in targeted areas.

💡Specialization

Specialization in the context of AI refers to the process of tailoring AI systems to perform specific tasks or roles with a high degree of expertise. This is in contrast to general-purpose AI, which is designed to handle a wide range of tasks. Specialized AI can lead to more efficient and accurate performance in its targeted domain.

💡Productivity

Productivity in this context refers to the efficiency and output of work, particularly how AI tools can enhance or increase the productivity of professionals in various fields. By automating certain tasks or providing intelligent assistance, AI can free up time for more complex, creative, or strategic work.

💡Code autonmy

Code autonomy refers to the concept of AI systems being able to autonomously generate and maintain code without human intervention. This level of automation could potentially revolutionize software development by reducing the need for manual coding and allowing for the rapid creation and iteration of software projects.

💡Conductor

In the context of the video, a 'conductor' is a metaphor for a role or a piece of software that coordinates and manages a team of AI agents, each performing a specific task. This concept suggests a future where individuals could manage complex projects with the help of AI, significantly scaling up their capabilities and efficiency.

💡Agent-first approach

The agent-first approach refers to the strategy of developing AI systems that act as autonomous agents, capable of performing tasks independently. This is contrasted with a context-first approach, which focuses on integrating AI into existing systems or workflows to enhance their utility. The agent-first approach aims to create AI that can operate with a high degree of autonomy, similar to a human in a specific role.

💡Context-first approach

The context-first approach in AI development prioritizes understanding and adapting to the specific context or environment in which the AI will operate. This approach focuses on making AI tools that can effectively integrate with existing systems, workflows, and human expertise, enhancing their utility and effectiveness in real-world scenarios.

💡Solo entrepreneur

A solo entrepreneur is an individual who starts and runs a business on their own, without a team or partners. The trend towards solo entrepreneurship is facilitated by advancements in technology, including AI, which can automate many tasks and reduce the resources required to establish and maintain a business.

💡Economic productivity

Economic productivity refers to the efficiency with which economic inputs are converted into outputs. In the context of the video, it is suggested that AI tools will increase economic productivity by automating tasks, reducing the need for manual labor, and allowing humans to focus on higher-value activities. This could lead to a significant increase in the output of goods and services with the same or fewer resources.

Highlights

Vertical AI startups are emerging, focusing on specific roles in society.

These startups are building apps around job titles, such as AI for lawyers or doctors.

Harvey is an AI for lawyers, Bridge is an AI notaker for doctors, and Tax GPT is an AI tax assistant.

NCR's AI is designed for customer support, and Brett Tor's startup is in this space.

Cognition's tool, Devon, is an AI software engineer that has gone viral on演示.

Devon is speculated to be built on GPT-4 from OpenAI, though not confirmed.

Devon's performance on coding benchmarks is significantly better than a generic language model.

The role-based vertical startups are powerful, as they make complex tasks more accessible.

These tools will automate processes, making developers more valued for their expertise.

The demos showcased AI's ability to find and remediate errors in code.

The evolution of AI in coding is moving from co-piloting to a more autonomous role.

The future may see AI 'conductors' coordinating teams of AI agents.

AI will not replace humans but will level them up, increasing productivity.

The economic productivity will rise as AI takes on specific human roles.

The future could see millions of companies run by individuals with AI assistance.

Solo entrepreneurs will benefit from AI agents, reducing the need for large teams and funding.

It's never been easier to start creating something as a solo developer.

For profound projects, a small team of developers and some funding will still be necessary.

The barrier to entry for launching an MVP has significantly decreased over time.

The trend shows that coding and app development are becoming increasingly accessible.

Transcripts

vertical AI startups are starting to

Make Some Noise we all know about large

language models we've talked about them

here if you listen to this program you

know about open AI Google's Gemini

previously known as Bard anthropic

Claude all this stuff their general

purpose they've been trained on the open

internet as we were just discussing so

they can answer questions about almost

anything and yeah sometimes it's correct

sometimes they're incorrect but it's

showing promise there is another school

of thought here that's emerging in

startups iCal AI these companies are

kind of taking a job title a role in

society and they are building vertical

apps Harvey is AI for lawyers a bridge

is doing an AI notaker for doctors saves

them hours a day according to them tax

GPT is an AI tax assistant NCR is AI for

customer support that's Brett tor's new

startup this week a startup called

cognition debut a tool called Devon

they're calling it an AI software

engineer the demo went viral on X uh you

probably seen them all over the place

and in the news if you watch it you can

see Devon fixing bugs in real time fine

tuning an AI model building apps end to

endend and people are speculating Devon

was built on GPT 4 from open AI That's

not

confirmed but according to the CEO Devon

was built by tweaking reasoning and

long-term planning into an existing llm

here's how it ranks against other major

models on coding benchmarks they're

building all these bench marks to test

each language model and as you can see

it's according to this chart and

according to their data doing much

better than just a generic language

model kind of makes sense jamat did you

see these demos this week I think I saw

you on the group chat talking about it

and what was your take on these

role-based vertical startups oh I think

this is so powerful I mean it's

incredible because we're measuring this

progress in like what week over week

feels like

yeah I think the point that you should

take away is that

the most of these very difficult in

impenetrable job types for the average

person if this if you said to them hey

become a developer that's like a

complicated Journey

right it's just going to be now

like a command line interface where you

just kind of describe in English what

you want to do and all of this stuff

will just happen behind the scenes and

it'll be totally automated so that'll

grow the number of people that can use

these tools at the same time it'll make

the developers I think even more valued

because you're going to need people in

the guts of these models and in the code

that it generates because it's not

always going to work perfectly there's

always going to be some kind of

hallucination some stuff is not going to

compile now the demos that they did

though were incredible they were able to

find errors they were able to remediate

errors in

code I mean I just I think it's really

really special you've been on co-pilots

for the past year talking about that

this is slightly different we're moving

from hey here's a co-pilot somebody

helping a developer to hey here's a

developer working and now they have a

supervisor so what do you think of these

sort of role-based agents and how

quickly we went from year one

co-piloting to okay now they're the

pilot and we're sitting in the co-pilot

seat watching them fly the plane yeah

well look F first of all everyone's

working on autonomous coding or working

towards that this is like one of the

core most obvious use cases of llms

because code is text and it can also be

run through a compiler to debug it so

you can also get to in theory you can

get to high levels of accuracy yet

although in the example that you gave

Jason this new product was only at 133%

so there's still a long way to go but

the potential is clearly there so a lot

of companies are working on some

variation of this

idea Devin is I guess you could call it

an agent first approach and I think

that's very cool for generating new

software projects but where I think this

gets much trickier and is much more

difficult is when you're working in

existing code bases and just to talk my

own book for a second we're an investor

in a company called Source graft they

have a product called Cody and their

whole approach is context first as

opposed to agent first it's all about

getting getting the co-pilot to work

inside of existing code bases so

different companies are coming at this

from different approaches GitHub

co-pilot I think is kind of more like

Cody where it's all about making an

existing code base more useful whereas

Devon again is starting with I think net

new code bases but that's going to demo

really well and so that's what you're

seeing is like these really cool Demos

in any event the larger picture here is

that we are going to get better at

better at coding auton ly I guess you

could say and I don't know if it gets

ever gets to 100% where you don't need

coders anymore but it's going to make

coders much more productive over time

you're going to get this huge multiplier

effect on the ability to write code and

that's really exciting for a bunch of

obvious reasons free we've been tracking

this Evolution from you know Gmail guess

the next word guess the next phrase

guess the next sentence to co-pilots now

we have these role-based agent-based

solutions that startups are pursuing

what's next if we follow this thread

what would the next Evolution here B

well the big push has been for this

notion of AGI to replace a

human and I think what we're seeing

is software that replaces a specific

human doing a specific thing like being

a lawyer being an accountant being an

art director if you think about the

internet when the internet which was

like networking software and the

capabilities that arose from the

connection of all these computers during

the internet era The Innovation was

everyone tried to create a business

model which was how do you take an

existing vertical business and put it on

the internet I think what we're seeing

in this era is everyone's taking a

vertical human and creating a vertical

version of a human um in the AI era and

so um I think like the the success will

probably acrew to one company that

replaces one set of core human services

like being a lawyer being an

accountant you know being an artist in

whatever way and that that ends up being

the specific vertical tool that people

will use to automate and scale up their

ability to do that task in an automated

way because I think that there's like a

great deal

of capability that emerges in the fine

tuning and the unique data that certain

people may have to make that one tool

better than the rest and therefore

everyone will end up using this one

lawyer service or this one Accounting

Service or what have you so I definitely

think that's kind of what we're seeing

yeah I think it's pretty obvious where

this is going you got co-pilots

assisting a developer or a lawyer then

the next or a writer then they got the

next phase okay you've got a peer so

you're doing peer programming or

somebody's kind of working alongside you

you're checking their work hey maybe

they're even checking your work seeing

if you have bugs where this is going to

be next year is there's going to be a

conductor there's going to be somebody

who has a role or a piece of software

has a role where you say hey you're a

CEO of a company you're a Founder a

product manager here's your lawyer

here's your accountant here's your

developers here's your designer and now

you will coordinate those five people

now imagine how that changes startups

when you as an individual have a

conductor working with you and says you

know what I don't know if I agree with

this legal advice that's coming in in

relation to the tax advice and maybe we

should not even add this feature to the

program let's talk to the product

manager the agent product manager about

taking that feature out so we don't have

these Downstream legal issues and we

don't even have to file taxes in this

you know area it's going to get really

interesting next year when they have a

conductor the other way it may go Jason

is you have a lawyer that has 50

Associates working for them through the

AI so you don't replace the lawyer you

don't replace the software engineer the

software engineer levels up and now the

software Eng has 50 Engineers available

50 agents running doing tasks for them

you do still you do still need humans

with domain expertise and creativity to

Think Through architecture to Think

Through process and to make sure that

the AI agents are doing their job so I

think what it creates as extraordinary

leverage for people and organizations

which is why generally economic

productivity goes up people don't lose

jobs they level up in this phase the the

Opex of companies will probably be cut

in

half at the limit

I think Jason is

actually absolutely right I think you

find that there'll be millions of

companies with one person and then a

whole layer of software and conductors

and agents and Bots that's the future

yeah so you won't have these engineering

people that person should be running

their own company and so you'll just

have millions and millions and millions

and maybe billions of companies and I

think that that's really exciting not

all of them will work many of them will

fail and a few of them will be

ginormous and it'll be up to the person

who can navigate and be a conductor as

you said you know yeah be really

interesting the solo entrepreneur

movement of the last couple years there

were all these kind of like independent

hackers building one item like Phil

Kaplan did with um drro kid he just had

like two or three people working on that

got very big you know I was telling you

guys about that slopes app I showed you

I I I reached out to the founder of that

I was like hey tell me about the

business like it's enough of a business

to support one person or two people like

there will be a lot of these apps or

Services one

conductor and you know it makes what

half million a year 3 million a year

whatever it's enough to support one two

three people working on it but

previously you know you be going to the

Venture Community like oh what did it

take a modern app company Sachs to kind

of build an Android in a an IOS app just

you know five 10 years ago if we were

funding one 10 years ago what would the

footprint look like for you know a

consumer app company if you're going to

go all the way back to like the late 9

90s during the era I remember that with

PayPal just to launch really what was an

MVP we had I'd say a dozen developers

and it was pretty expensive and we had

to set up our own Colo there was all

this infrastructure that all got

abstracted away with AWS then you move

to the mobile er uh and the app stores

provide there there's just a lot more

developer tools more apis yeah well as

well as distribution but it's just it's

far easier to code these apps so things

have gotten easier and easier that's the

trend if that's the point you're trying

to make it's certainly never been easier

to get started in creating something if

you're a solo developer yeah that being

said I think that depending on what

you're trying to do it's still usually

the case that if you're trying to do

something interesting and profound

you're going to need a small team of

developers and a couple million bucks to

get you started yeah it used to be Ru of

themb I think 12 people for an app

company got two working on each platform

a couple designers couple

testing and uh design ux you get to 10

12 people to to run a modern one