SPEAKER_00: number one you consumed a bunch of information two you did the work and learned how to code and build products and number three you got on a plane and you went to where the action was happening and you met people you networked yeah this is so easy and the industry is so wide open so for people in america who are saying oh my god i don't know how to break into tech consume every bit of content you can about what you want to do and and being an entrepreneur number two learn how to build products so all the information is on the internet and number three

SPEAKER_02: go to wherever the act most action is it happens to be the bay area but there's also stuff happening in dubai there's stuff happening in tokyo there's stuff happening in sydney and melbourne sometimes you just got to get on a plane and meet people in network it's very simple one two three

SPEAKER_04: you figured it out kid i love it this week in startups is brought to you by miro helps take ideas from in your head to out there in the world with its ability to democratize collaboration and input sign up for free at miro.com startups northwest registered agent when starting your business it's important to use a service that will actually help you northwest registered agent is that service they'll form your company fast give you the documents you need to open a business bank account and even provide you with mail scanning and a business address to keep your personal privacy intact visit northwestregisteredagent.com twist to get a 60 discount on your next llc and nuts.com is your one-stop shop for the highest quality foods for business they offer delicious office snacks corporate gifts and wholesale ingredients nuts.com is offering new business customers a free gift with purchase and free shipping on orders of 125 or more at nuts.com twist all right everybody welcome back to this

SPEAKER_07: week in startups i'm covering this ai trend well for over 10 years but the last year has been absolutely extraordinary since chat gpt launched specifically 3.5 and then later 4.0 everybody

SPEAKER_00: else has gotten in on the action and auto gpts and baby gpts were all the rage a couple of months ago what were those they were like little user agents they would perform tasks anonymously maybe string a couple of them together like go do a search for a flight and then buy it for me this is what the internet uh and intelligent agents were always supposed to do but they never worked and uh it's a brilliant idea

SPEAKER_02: obviously a little bit scary to some people but it's the obvious future and as is the case founders uh always try to make the future get here a little bit quicker and today we're having a founder on the show arian sharma he is the ceo and co-founder of induced ai and what they're building is agents that live in a browser and it's called as i said induced ai welcome to the show arian thanks for having me um it's great to be here yeah and now you heard my uh thanks for coming you heard my intro there uh explain to the audience what you're building and why it's important

SPEAKER_17: yeah i can give you a quick background on why this is even a consideration and why how we came to the the concept of agents i think one of the first traces of this the whole agent idea stems from the rl world which is reinforcement learning it's one of the earliest ai research areas and what it basically means is you can have these models that are um learning how to do different things and they

SPEAKER_19: can they can try a bunch of different patterns and they get rewards when they're doing something right and then they get penalties think of it as penalties when they're doing something wrong if you're teaching one of these models how to play tennis every time they hit a good shot they get plus one every time do something wrong they get mice one and they're the way intelligence is baked into these systems is you just allow them to do a bunch of patterns and you keep rewarding and sort of giving them penalties when they're doing wrong and over time they try out a bunch of different patterns and they become smarter that was kind of the earliest traces of what agents meant where they have kind of agency to figure out and then self-develop a little bit of intelligence um we used to do all of this the whole agent premise was restricted to reinforcement learning for a while because you want to try out a bunch of things and that was kind of the only architecture that allowed um people who work and build these agents there's this paper called world of bits that open ai wrote uh in 2016 and it was andri karpathy and a bunch of others that was that was the first version of these agents where um they were teaching agents how to control the web how to play some games minecraft and a bunch of environments and they they kind of tried uh building it in a

SPEAKER_17: reinforcement learning first way what has happened since then is obviously we've had transformers and we've had these amazing language models come in um and they've opened up a new set of capabilities

SPEAKER_19: and after chat gpt came out last year it was evident that it's it's it's it's got a bunch of interesting capabilities you can ask questions it will give you responses it can sort of reason on its own you can you there was interesting patterns that people the developer community was creating auto gpt baby aja versions of this where you have a prompt prompting itself and then chat gpt can talk to itself and have these loops and a little bit of that loop um that is important and reinforcement being created give an example of that yeah um screen example of that is you let's say you give chat gpt a task come up with a copy for um jason's new fund and it comes up you give some description this is the highlight of the funds this is what we're focused on etc etc and it comes up with a three-line copy and then you take the three-line copy and then it's you just give it back to the system analyze this copy that you came up with and make it better or give me criticisms it gives you like three criticisms these are the things that you can improve and then you just kind of keep continuing the loop where it takes those three criticisms goes back and it's sort of trying to self-heal or self-correct in some sense uh but it's all happening at the prompt level so this is super easy for anybody to do nobody there's no reinforcement learning there's no model architecture nothing involved it's just smart use of language models you're stringing them together in interesting ways to create this sort of agentic behavior and what what was interesting for a lot of people in the developer community was that now that we have these language models and we have these interesting new capabilities maybe we can go back to the premise of agents but design them in a language model first manner instead of reinforcement learning or some of the older versions of agents that were created um and auto gpd baby hi have our first versions of that where you have a language model sitting at the center so you go and give them a command maybe you know go to google search up for something and then give me a summary and extract top three highlights from the summary and then that model that is sitting and taking in your input has access to a bunch of tools it can choose which tools it wants to use you can decide and go to this website perform this action capture this data processes it and otherwise so it has a little bit of this reasoning engine built in to itself and you can execute the tasks using those tools so it's not the language model is no longer restricted to its training corpus it's no longer restricted to just giving you text responses it has a little bit of agency to use just text input output to interact with the outside world to interact with

SPEAKER_25: the internet leverage apis as different tools and do more stuff you can't do that currently with chat gpt if you try to get it to do things on the internet i did the other day i said hey tell me the best pairs

SPEAKER_07: of boxer shorts it gave me six different types of boxer shorts and i said okay buy me one pair of each in medium 32 waist and it was like i can't do that so in your world it could do that it could go search the web for those six brands put it in the cart check it out use my credit card know my address and then ship it that's what you were building correct essentially we are building we kind of think

SPEAKER_17: of them as mini digital workers they are these language models that are sitting and you can give them instructions and they have full access to a browser we are only doing browser stuff now there's different ways of doing this where you can have it at desktop level you can have it at mobile level you can have it at the edi level but we sort of think that the browser is the most

SPEAKER_25: so exciting and the broadest rated order so i could create a personal shopper using your software

SPEAKER_07: i could tell the personal shopper your job is to find the best item uh make a list of those report back uh on this and then uh i'll tell you which ones i want to order and then you go order them and make sure they get shipped based on this criteria or i could even take that out just say get me the three best coffees the highest rated ones and ship them to my home address and it would actually do it

SPEAKER_19: yeah so the one caveat there is there are two ways to think of these agents um one very exciting premise of these agents is running them fully autonomously which is what auto gpt and a bunch of other agents originally started with where you're just giving a text description like this go find me shoes or go find me flight tickets and then it automatically figures out where it needs to go it'll ask you questions to figure out what you want and then automatically kind of figure out whatever needs to be done to get your output that is great that's kind of the future that we are building towards but in in the near term in the next four five six months based on the current capabilities of the models that we have it is a great like vision it's a great demo but when you actually implement it there's problems with reliability because fundamentally these models are non-deterministic and they come up with new outputs every time so you cannot be sure that every time you ask you know

SPEAKER_17: go and purchase shoes for me it's going to get you shoes from the exact same place or get it from and

SPEAKER_19: more likely than not it'll get lost because if these there's no rules that are surrounding them

SPEAKER_37: there's no guardrails around these models all right founders always ask me for pitch deck punch-ups and you know what i got some great news for you we worked with the team at miro the awesome whiteboarding software i've been talking about to create an amazing pitch deck template for founders which you can see if you're watching the video right now this is going to help bring your pitch deck from zero to hero from zero to vc ready and our founder university participants love this template we use it all the time it saves them time and it gets them more meetings so head to miro.com miroverse miro.com miroverse and search for pitch neck to check it out and if your team is hybrid or fully remote miro is so useful for you it's like an old school in-person whiteboarding session but distributed and asynchronous so you can do it on your own time miro lets you brainstorm ideas and collaborate on projects from anywhere in the world when you think miro think zero to one but faster and miro is so much more than a simple digital whiteboard your team can collaborate on important stuff like research design planning and feedback cycles and faster inputs equals faster outcomes and we all know product velocity and startup velocity is how your company is going to win so to access our new mirror verse template and thousands of others sign up today for a free miro account at miro.com startups miro.com startups that's miro.com startups to sign up for free so you're trying to do this with

SPEAKER_07: repetitive tasks like uh sdr a sales development rep is a perfect example of a job people hate it's repetitive you go and find targets to sell some sas software to everybody gets these annoying you know email sequences but they they obviously work people are still doing them so that's one of the first use

SPEAKER_17: cases that you're building yes so that's the the simplest analogy that this example of that is it's all stuff that was done previously with rpa software so uipath and you know assume these like large rpa companies that have existed for several years and decades um their idea was let's string together tools that don't have apis and we can connect them together and build these bots what did you refer to

SPEAKER_30: those as what kind of companies it's called robotic process automation it's the rpa industry robotic

SPEAKER_07: process automation as opposed to business process automation rpa is these repetitive tasks robotic

SPEAKER_17: process automation interesting yeah and there it's just you have these robots that are created mini digital scripts and workers that do a bunch of repetitive tasks um and the that has existed for a

SPEAKER_19: long time so it's one of it's a last span category for a lot of enterprises that's how enterprises automate

SPEAKER_17: work especially when they're dealing with tools that don't have apis and that you cannot just you know string together through zapier or existing api tools linkedin would come to mind right lincoln slows you down

SPEAKER_07: it doesn't have an api everybody wants an api they don't want to give an api because they know that go fast but with robotic process uh automation you set up a browser you search it goes and does these things automatically looking for i don't know ctos chief technology officers puts them into a database you know and sends them a link an email whatever tries to guess the tries to validate guess their

SPEAKER_02: email and then validate it with an email validation service um and that's how people have these

SPEAKER_07: databases if they ever try to sell you databases based on linkedin data it's these rpas that have done that searching and of course they get turned off if they load too many pages so it's a bit of a

SPEAKER_19: cat and mouse thing correct linkedin is a great example you can look at a bunch of legacy industries like healthcare that have insurance platforms and claims processing platforms all of these don't have api so a lot of real legacy industries rely on rpa the the problem is that rpa has existed for a long time it's just it used to be done in a very manual way where even though it's the eventual goal is you want to automate a workflow it takes a lot of effort to kind of set up these processes because these rpa companies go top down uh it's almost like if for every dollar you spend on implementing rpa you have

SPEAKER_17: to spend five or six dollars on consultants who will actually come and understand your process they will you know buy one of the software from one of the vendors and then the reason it's so expensive and time consuming is that traditionally if you want to string together and automate a workflow on the

SPEAKER_19: browser you have to script every step so simple google search for launch or this week in startups is go to google you will have to click on the field get the selector the html selector which is behind the scenes the dom or whatever of the web page then click on that field type in your text click get the identifier of the button then go to the page it's just every field button element that you'll interact with on the web in completing your workflow you have to manually go and script it and the problem with that is scripting takes time but also these scripts can break because if these you know websites keep changing layouts all the time they keep changing selectors and class names all the time and because you're hard coding it to selectors and class names from the html if any of that changes your scripts will

SPEAKER_17: break so you have to constantly keep maintaining these scripts um so that's kind of one language model might

SPEAKER_07: come in because the language model would look at the page if your profile page gets updated or i should say linkedin changes profile pages and they just move the html around and people were looking for what city you're in and they called it location instead of city well that breaks the rpa right and so now i just

SPEAKER_30: said what's the location the language model should be able to figure out what's the location in the first i don't know that you know 500 words of text on the page yeah the language model is doing real-time

SPEAKER_19: inference on every run so it can it can basically handle these changes it can handle you don't have to spend as much time sitting just broad directional input of you know in the city you're going to point where the city is you look at the page get the thing reliably you can obviously set guardrails around that as well that's kind of one problem of rpa that that solves in this new world um the other problem is that with traditional rpa you can only because it didn't have any reasoning skills you can only automate things that are you know rule set based let's go to google put in this exact text click on the first link they go this exact page you can it's you cannot do things like you know go to this linkedin analyze if it fits my ideal customer persona or analyze if this falls into the five cities that i want to target and then basis that you know do this action or draft a custom message so any any level zero cognitive reasoning tasks that you language models can be pretty good at um you cannot do them with traditional rps so we've we've kind of taken this this whole industry of how rpa was done and designed in an ai native way of how can we make setup and maintenance you have a demo you can

SPEAKER_17: show us actually how this works yeah yeah i can pull up a quick demo it should give you a good idea of

SPEAKER_26: how um the base version works um so for people who are listening we'll describe for you what's

SPEAKER_19: happening on the screen yeah so you just go go to the induced platform you have it's just empty screen with no workflows right now i'm gonna click on new workflow on the top right and just ask you

SPEAKER_17: you can either design the whole work from from scratch where you you know give step-by-step input or you can use ai assistance and i'll go through what that means but you just ask for a workflow name i'm going to put in employee timesheet and i'll run through what workflow i'm making so i just put in

SPEAKER_19: an employee timesheet and then i'm going to put in a step-by-step this is just english description of workflow that i want to design and for context the workflow is basically think of a construction company or a company that has warehouses or physical centers across the country and they have employees coming in and filling in paper timesheets um that is basically a log of when they're coming into work how many hours they're working what are the breaks they're taking etc etc and this company takes in all of these paper timesheets puts them on an air table and they have to manually calculate their role for every employee because it's all on so you have to take whatever is on paper understand it then run it against a company policy doc to calculate now this guy worked five hours this is a deduction this is over time and then go and enter whatever payroll you've calculated back into an air table so this was we did this for one of our early pilot customers they have a physical they have a real back office of 15 people in their finance team that finance and ops team that does this

SPEAKER_17: uh but with kind of in the new world with some of these agents you can just describe this on our

SPEAKER_07: platform crazy so it has hey identify uh access the air table payroll navigate to the employee payroll base from stored variables identify a relevant employee search for employee with the payment status review timesheet data open the employee's time sheet etc and then it says calculate the payment access the employment payout info sheet on google docs using the timesheet data note the total hours worked end time minus start time deduct any breaks multiply the hours worked by the hourly rate mentioned in the google doc to get the gross amount deduct any break time expenses or other deductions and so this is the step-by-step process that some human did you're just describing it in

SPEAKER_19: essentially plain english not code yeah and then you just click create workflow what we'll do with the ai systems is unlike a lot of traditional i mean traditional is a bad word used because it's all very new but unlike a lot of other autonomous agents we don't directly start running it based on english and this is actually the first time we are ever showing this product on uh media it's like a podcast or a

SPEAKER_17: video stream but this is uh this is this week in startups exclusive um thank you but it takes in

SPEAKER_19: whatever input you've given it compiles it we have this middle layer in between which is just the input that you've given but structured into smaller steps so it all of the all of the steps that i took in it just broken them down chunked them in um so it's just basically access air table then loop through the entries pick one entry just smaller smaller chunks of whatever i described and the reason this is useful is one for visibility for who's designing this workflow they can see whatever input even what's the final translation um and then what you can also have visibility into is we split it

SPEAKER_17: into a bunch of different action types so it's you know going to the web page is one action clicking filling all of those are standard web actions um then there's a bunch of data actions like looping you know identifying filtering that you can do on the page and then we have the agent blocks which

SPEAKER_19: are all of the smart actions so once you go to an employee's timesheet any calculation that you want to do you can use the agent block to delegate and get input from model so it's basically a bunch of different block types depending on what your workflow step is that we automatically identify and put in you you can obviously edit you can obviously make changes to this and at the end you can have standard you know elt or outputs in whatever format so after the workflow is run you know you can have triggers that if you have api call put it in a google sheet or anything that you need so it just puts an in into these formats and if you notice for those you can see the screen um it automatically puts in these variables um across the steps so if if your workflow involves capturing data from one place and then using it in another step later on it's basically i'll talk more about the browser environment but this is basically a runtime that is designed for agents so it has access to its own file system its own memory it can store data retrieve data so it's basically

SPEAKER_64: basically firing up a computer in the cloud essentially or a browser session i don't know if

SPEAKER_07: you're using chrome or chrome os or windows you can light up a virtual browser or a virtual machine anywhere aws etc so you're basically popping up a desktop and then running this stuff yeah yeah so we

SPEAKER_19: spin up a chromium fork on the cloud it's a virtual machine with a chromium fork it's a custom browser

SPEAKER_17: that we've designed specifically for running bots and autonomous agents and systems like that uh which we can go into but that that's kind of it's been set up which is why it has access to all of these tools that you can use in the flow and then once you you're happy with the workflow just click run

SPEAKER_19: and um this is kind of interesting because the way it runs is it's been the browser that has spun up on the cloud you can see a real-time live stream of the browser when the workflow is running so it's as

SPEAKER_17: if you're watching and a team member's screen as you can see the stuff that's happening on the remote browser streamed live here on the left and on the right you can see step by step what's being done so you can see the stream goes to air table it loops through the entries it picks one of the employees um opens the employees data opens up this timesheet and then now it's going to run ocr and it can use tools on its own as long as you're giving it step descriptions and use ocr document

SPEAKER_19: processing to capture text input from this document store it in memory then it'll go to this google doc which has information about how you calculate payroll based on the timesheet data it'll capture this in memory as well and then once it has both these things in memory it's going to compare both the data

SPEAKER_17: points run a reasoning step which is why reasoning steps usually take longer but it's going to take both of those things in calculate a final payroll amount and you can see the variables being referenced here um we'll go back to air table and it's just going to fill in we've got a bunch of fields that were to be calculated just going to come back fill in total hours worked is nine break duration is one over time zero total payable seven you'll put in the hours you put in total amount and and the

SPEAKER_19: interesting thing is because this is all running remotely uh we can we kind of call this the mission

SPEAKER_17: control view of um the platform where you can have a bunch of different browser instances doing different things all the same thing running at the same time so it's like a real back office you can scale up scale down you can have 50 instances running at once you can have five you've got one that's you've got the air table instance you've got one running linkedin we've got one on aws there's a tech crunch thing here and you can kind of see live streams of all of them so for the the task of of a manager who's managing a 15 member back office team goes from actually delegating tasks to real people to just sitting and watching you know 15 of these screens and crazy if something goes wrong just go in and flag it but otherwise you can kind of have a top level view of everything so if i'm running some

David Friedberg: insurance company sales team customer support whatever my startup has these six or 15 agents

SPEAKER_07: running doing the tasks that humans were previously done and you just watch them and make sure it's doing it correct and it feels like you've been working on this for less than a year how many months

SPEAKER_26: into this are you um we started working on this around april of this year so march april is when we

SPEAKER_31: started so you're six months into this process yes awesome and we've we've kind of identified so just

SPEAKER_17: just as a as a disclaimer for this this run was very well described so i i gave you know explicit input

SPEAKER_19: i described everything and that's how we're done it it doesn't require me to manually skip the whole thing and it runs reliably once i've given input but it requires explicit input and i think we we took that uh the the approach that we've taken architectually which i was talking about uh was was actually it's like the most important thing in this six-month process that i think has helped us a lot with instead of designing an automation product that sits at a chrome extension and then you use it synchronously on our computer where i have an extension i record something and then i can replicate but i have an extension that has an assistant i give it commands on my chrome and it's doing things we design everything to be remote and have its own environment so it's we took we took chromium focused it made made some changes to it of how the dong comes out how the html comes out has anybody

David Friedberg: started using this yet in like a real world situation or are you still in the laboratory

SPEAKER_19: we we launched early october and then we went live with a couple of folks we are live with about 15 now and that is across different sizes small companies mid-sized companies we've kind of found our sweet spot and these mid mid market to mid to up market companies that are operating in sort of

SPEAKER_17: older industries like healthcare um you know financial services etc um to fly with with a few of them we

SPEAKER_19: are constantly we're changing the form factor a little bit because with the text input that you saw where you have to describe your workflow and then it translates um to overflow which is just the actual workflow that's running that is not it it's the easiest way to start that's how most ai language model products start but still not very intuitive if you think of it from the user's perspective because when you're typing the text description for your workflow you basically have to have a browser screen that's open and then you're typing out every step where i'm going to this and i was just like type out okay go to this then what am i clicking on i just type out you know click on this and what we are designing now is it's sort of like and this is the video that i i'll share so you can probably

SPEAKER_17: edit in because it's it's not even an alpha right now we're going to put it out in a few weeks where it's sort of like think of a collab notebook a google collab notebook where you have a bunch of cells and if you want i can show you google collab as a reference but sure it just we have so this

SPEAKER_19: is this is how collab works it has these bunch of cells and then you can basically give in like you

SPEAKER_17: can run each cell so i can i can be like you know print hello world this is how regular collab works

SPEAKER_19: and then you can add like each cell can be done and then you can have a bunch of more cells that come together and you can string together cells to create something um we are creating a new interface for how we set up workflows but instead of you having to type everything at once and then edit them and then run it and then come back and debug we have this environment that is set up for you to create workflows so you kind of just come in and instead of typing out code snippets you have you just type out your english steps and we have in the place of this section on the right we have the stream that we had opened up so you just come in and be like go to google and then you can actually see it go to google and you like click on like search for this we can start actually see it happen um so it's basically real time you know setting it up once you're happy with every step just click confirm and then it automatically migrates to workflow it's much more intuitive it's easier if you had a press

SPEAKER_07: monitoring service you could say search the web for this person's last name or this person's name uh go to google news click on the link summarize it email it to this person let them know they were in the news uh you know or mentioned in a news story that's a job that pr people uh do for a living now

SPEAKER_02: are you going to get the you know you'll probably do some mistakes and probably will put old stories and it's gonna you know not be perfect in your world but it would certainly make the person who does pr

SPEAKER_07: clipping uh and what they call media monitoring or web monitoring it would make them bionic they would be able to do a hundred times what they do every day so essentially that's what you're doing and then

David Friedberg: anybody who has a business process can basically script this starting a business used to be such

SPEAKER_79: a painful process you needed to get a lawyer there were tons of fees it was a mess but not anymore just check out northwest registered agent they're going to help you form your company fast remember speed matters and then they're going to get you the docs you need to open a business bank account instantly then they're going to provide you with mail scanning and a business address and they're going to do all that keeping your personal privacy intact northwest can form llc's corporations and non-profits and here's why founders love northwest there's no hidden fees there's no upselling you can

SPEAKER_81: call them or cancel at any time and northwest has the best of both worlds solution it's simple and self-serve but they can be hands-on if you need help with their amazing registered agent service

SPEAKER_79: northwest provides everything you need to start and maintain your business and they're giving twist listeners a 60 discount for just 39 plus state fees they'll form your llc corporation or non-profit so visit northwestregisteredagent.com twist today northwestregisteredagent.com twist

SPEAKER_00: so what happens to the business processing outsourcing industry in india where you're from uh and i know you have uh investment from uh sequoia india india or what is now uh there's a new name

SPEAKER_07: for that firm but uh peak 15 yeah so you have an investment from them uh when you show this to the people who are in the business process outsourcing did their heads just blow up and go oh my god what's going to happen to these hundred million people employed in the business processing uh world or has

SPEAKER_86: this just been a continuation and there's always going to be more business to process

SPEAKER_19: yeah i think there's one it's it's an extremely large market where this where outsourcing happens um and it's just increasingly we're all the tech is growing but it's like the the legacy industries and

SPEAKER_17: there's all sorts of unique things that keep coming up which keeps growing the business process outsourcing of the back office or like the philippines india all these kind of work things that

SPEAKER_19: get delegated outside and you have these economies that are created for people that are remotely doing these tasks um so i think the market is huge they always need more automation so every time we go

SPEAKER_17: and we've had a lot of people who run these globe they're called gcc so they're a global um you know back back office centers and there's like a bunch of different words for them we got a lot of

SPEAKER_19: them reach out uh and they're always looking for automation even though they have you know 50 000 people sitting and doing these things because they want to make it more efficient they want to be able to take on more business um so a lot of them want to be customers of this so they can improve internal processes but i think in general the trend that we are seeing is um this will kind of take the low level things first so the extremely repetitive things that have very little cognitive involvement are almost always easy to automate even with traditional rpa with these models coming in we'll be able to do a little bit of cognitive tasks as well where a little bit of thinking filtering you know profile validation lead enrichment things like that will go in and but then there's always you know complex tasks or sensitive tasks that you want these back offices

SPEAKER_17: to do like payments etc um so i think it'll be a slow migration but uh this kind of just trajectory of

SPEAKER_56: tech where incredible yeah gcc stands for global capability centers this is yeah india when you want

SPEAKER_07: to outsource a back office you were uber or airbnb and you have a huge amount of i don't know refunds to process or claims to process or complaints whatever you would hire a gcc to work with you on the best practice hire some people in a lower per hour location and to just get it off your plate in america where you know people candidly don't even want to take these jobs no matter what you pay them they just wouldn't take them and you know that's what's been happening for i don't know for all time but certainly in the last 30 or 40 years this process of outsourcing so i guess since you're so close to all of this and you're building it uh you're talking about very low level things people who are doing data entry i guess data cleanup sdrs these are you know very um the the lowest paying

David Friedberg: white-collar jobs i would say right these are yeah how you know in three years what job you think you

SPEAKER_07: could do can you do a bookkeeper an accountant in five years could you do a paralegal in 10 years can you do every job could you do you know a salesperson's job in full you're sending stuff negotiating etc where do you see this winding up um i think i i so i think the way i think about this

SPEAKER_19: and i can share a few tweets if you want that should be like interesting to see as a reference um there was this there was this paper that that jim um from nvidia released and this was blew up on twitter you've probably seen it before where it was basically gpt4 playing minecraft and it's the way this operates is is called voyager what it's doing is it does a bunch of things and then once it does a bunch of things it analyzes those things and it buckets them into skills so you know chopping the tree it's done a bunch of times it sees what the reactions come in and then it's going to bucket that into a scale of you know this is shopping and then every time it does chopping it kind of feeds into that loop of refining the chopping skill and basically what's written here is it unlocks a new training paradigm where training is execution and it's it will help the runs that it's this agent is doing in minecraft helps iteratively compose a bunch of skills that it can slowly like learn but it's all confined to minecraft so it's it's a constrained environment where this agent is operating that's developing skills and continuously improving its skills i think the way to think about how this slowly these agents become powerful is we will not have in my view generally capable super autonomous global agents that can do everything the way we get closer to these areas being more powerful is the way you said bookkeeping you take three tools that are involved in book treatments maybe quickbooks maybe you know excel or google sheets and maybe like a bunch of code interpreter calculation stuff you get these agents to use these tools you build this repository of skills across these tools or you kind of let them explore and build these repository of skills and then the more they run the better those skills get as much training data as we can put into them they slowly get better as using those three four tools doing those three four kinds of interaction on those tools and then you kind of go into that agent and like do this calculation for me on my quickbooks and then run this math function or like do a prediction or a regression for me it's able to use those three tools for well and you can start doing things so i think it's going to happen sooner i don't think it's two three four or five years out it'll start we'll start seeing specialized versions

SPEAKER_17: of these agents and specialized agents that can use a bunch of tools come in in the next few months um it's just the way to approach that is get as many real world use cases get as many real world tools get these agents to learn these real world use cases tools around them a bunch of times and then kind of keep improving um how they scale up so i think that it'll slowly go in skill by skill

SPEAKER_07: makes sense yeah and so what happens when some bad actor pops up 15 windows to go cause chaos on the

SPEAKER_06: open web uh across services but what do you think the potential there is because unlike chat gpt if you

SPEAKER_00: ask it you know chat gpt is not going to go out and start trolling somebody on twitter or reddit or reddit or harassing them let's say but you could start very easily with your software and i'm not

SPEAKER_06: saying you would do this obviously but it's obviously on the path uh and people do this already but you

SPEAKER_07: could fire up 15 uh accounts or 15 browser windows 15 different accounts and then try to maybe swing an election right we saw the russians had boiler rooms doing this uh and they were you know all uh uh um we found out about it that was part of the mula report trigger warning uh russiagate but they were actually doing this but they were using humans right and then you do have those boiler rooms i think in manila india and other places where people do fake reviews of products so somebody here could fire up 15 of these windows and start posting pro-palestinian pro hamas pro israel whatever comments or just generally causing chaos so how do you think about that because your tool would allow a neophyte a non-intelligent person uh you know a bad actor to go absolutely buck wild and destroy everybody

SPEAKER_26: else's experience on the web i think so that's the first part is there is a lot of this is what open

SPEAKER_17: ai has also dealt with with their browsing being limited to only a few websites and they had to take down browsing in between because it was bypassing authenticated pages and giving you paywall content through its capers without actually so there's like a bunch of things that are happening there

SPEAKER_19: so i think it's important to add guardrails and in in the way we think about this is because we've designed this as a browser environment that is meant for bots we can build it up in so we don't need to build it up like a human browser environment we can add a bunch of guardrails that are specific to these bots that like allow them only a limited set of capabilities and a bunch of websites are just out of bounds a bunch of capabilities are just out of bounds and the users that are controlling these things they can define a bunch of rulesets but there are a bunch of global rule sets that just prevent

SPEAKER_07: um bad things your terms of service and then what it's allowed to do you could say hey listen we don't want you using these bots to go post to social media sites and ruin them so you could just you're i assume you're just banning the ability to do that these bots can't go out and post on reddit or

SPEAKER_34: whatever yeah and we we're banning a bunch of these we are going safer um than we should be going right

SPEAKER_19: now just because it's easier to build up the safety spectrum than come down um and the other kind of way we think about this is slowly it's going to evolve with the platforms where you know maybe if a bunch of bots are being run on airbnb they should be involved in the decision-making process of what kind of bots are allowed or not allowed and there'll be some sort of transaction that will eventually happen given the rate of progress that these autonomous agents and browser bots are seeing that you know websites will have to define and this has happened for a long time so robots.txt files exist on the internet for a lot of websites and open air recently a couple months ago open source their scrapers and they opened their signature so a website can choose if they want to let open air scrape them or not and i think we'll see similar versions of that with these bots and agents where you'll be able to

SPEAKER_17: allow bots to run on certain parts and some sort of web standard somebody wanted to go to airbnb

SPEAKER_07: and they thought oh let me just go to the checkout and check out with a fake account then cancel the reservation just to ask the person a question right so once you have the book you can i guess have a dialogue with the person and it wanted to fish for information or whatever you could just say you know what that's not a an allowed use case and then you have a conversation with brian and the team over at airbnb and they say yeah we don't mind somebody building an agent that looks at up to 50 pages per week or something but after that we want you to go through the api or we want you to get a license or something and you'll just be a good actor watching that of course bad actors will do what

SPEAKER_00: you're doing and not require that so there's going to be a bit of chaos i think we'll all predict on the open web in the coming year or two uh and i this is a little bit of sci-fi but i think it'll

SPEAKER_17: sort of be a new version of capture where um you have the way we have captures differentiating humans and bots you probably have a version of captures that differentiate good bots and bad bots and you have some way of proving what you're going to do and maybe there's some digital signature exchange that happens there but i think it's just we're all in this new world of and a lot of i think the reason a lot of the larger companies are also moving slower now is because they're afraid of you know infringing policy or content policy in terms of service etc of different web products and it'll slowly evolve um both sides where they get more clarity on what they want to allow and then yeah builders of bots like us get more clarity on what should be allowed for users if you run a

SPEAKER_81: business you know that having reliable vendors is non-negotiable and whether you need office snacks holiday gifts or wholesale ingredients you need to check out nuts.com that's it nuts.com that's a crazy amazing domain name nuts.com and nuts.com is your one-stop shop for the highest quality foods for your business again they offer delicious office snacks corporate gifts and wholesale ingredients i got a gift pack i have been eating these beautiful roasted nuts and other amazing premium products like chocolate covered sweets i love the trail mix popcorn you know i stopped eating the candy so i went for the dried fruit but they also have wrapped candy as well and my favorite jerky and of course they have all the gluten-free stuff whatever dietary option you're into over 50 000 companies choose nuts.com for their business needs from offices to hotels to restaurants to retail stores nuts.com has something for every business and so here is your call to action nuts.com makes ordering for your business quick and easy and right now nuts.com is offering new business customers a free gift with

SPEAKER_79: purchase and free shipping on any orders of 125 or more at nuts.com twist go check out all the delicious options at nuts.com twist and you'll receive your free gift and free shipping when you spend 125 or more

David Friedberg: that's nuts.com twist here's a crazy idea people who are building these bots

SPEAKER_07: have credit cards and real names and are validated as real humans and have a tax id and if you want to have a bot doing things in the real world you have to put a credit card in you have to have a social security number and or uh whatever the business number is and you have to have a valid phone number and you have to have a valid email and you have to be authenticated on the phone or have a driver's license on file so to use these things you could just have like if you want to buy a gun or have a car you have insurance you have a driver's license so if it's deemed to be too dangerous you can look at the amount of danger this causes in the world or chaos it could cause and then just like cars and guns are treated differently than pens and paper and you know maybe there's fertilizer that you know if you buy it we know that people can make bombs out of fertilizer you know you have to

SPEAKER_02: have your passport and driver's license and you can only buy a certain amount of fertilizer and there's a waiting day right so all of these things we have in the real world yeah we have some version of kyc or you know throttling when people can get it we have cool off periods with guns and uh you know hopefully the good uh folks in the world like yourselves who are building this stuff

SPEAKER_07: are thoughtful about it what i love about this is i think this is work that people don't want to do

SPEAKER_00: it's soul crushing work in most cases repetitive tasks it's you know going out and chopping wood would be more pleasurable i think for most people and healthier than doing some of the jobs that you're going to eliminate and i think there are jobs that should be eliminated nobody wanted to be a phone operator and sit there and plugging cables all day for 40 50 hours a week it was arduous and painful same thing here nobody wants to be in the fields down on their knees picking strawberries robots should do that better it's the same kind of analogy i wish you great success with this what have you i know you're uh very young i didn't want to bring that up because when i was young and people referred to me as the 23 year old founder of this magazine i always found it kind of annoying that that was in the first sentence but you are 19 years old you have raised a couple million dollars from sam altman and

David Friedberg: the former sequoia india and uh you got that ai grant right from uh daniel gross too i understand yeah so for folks who are young founders out there how the heck did you do it

SPEAKER_19: um i think i i had a very interesting story because i i didn't grow up in the bay area or

SPEAKER_17: the us i was out sort of an outsider in that sense but i grew up in india and i always i used to see your podcast i used to see yc videos i used to see a bunch of things from outside and i used to be like you know this is something something cool is happening what was the youngest age you watched one of my

SPEAKER_128: podcasts i'm curious i think 12 or 13 when i were 12 or 13 in india watching this week in startups

SPEAKER_10: you have no idea how much that fills my heart with joy because i always said you know i think

SPEAKER_00: there are people around the world who might hear this podcast and be inspired to start a company or just get stoked to be an entrepreneur and to hear you actually say you listen to this at 12 or 13 halfway around the world and now you're on the program six years later is just mind-blowingly

SPEAKER_131: joyful for me so thank you for that no thank you for doing the show i remember the clip that you

SPEAKER_17: did with patrick uh where he he spoke about how how they started stripe and how they raised money from sam and how they came to the valley and started doing stuff uh so it's like a bunch of these things that i used to see from outside and i this was stuff that i wanted to do i started writing code very early so i was already working in tech while i was here i was taking up remote jobs while i was still in school i had built like a bunch of projects and then as soon as i made a little bit of money i started making trips to the u.s just to come to the bay area stay in these hacker houses um try to meet people go to these events um that's how i met a lot of these investors and people who eventually kind of invested or became a part of um the company but i used to start and that that was i sort of had this brute force way of breaking in twitter was in fact also super useful but i should just call dm a bunch of people reach out to them pitch them um and yeah it's just like sort of

SPEAKER_00: recapping number one you consumed a bunch of information two you did the work and learned how to code and build products and number three you got on a plane and you went to where the action was happening and you met people you networked yeah this is so easy and the industry is so wide open so for people in america who are saying oh my god i don't know how to break into tech consume every bit of content you can about what you want to do and and being an entrepreneur number two learn how to build products so all the information is on the internet and number three go to wherever the act most

SPEAKER_02: action is it happens to be the bay area but there's also stuff happening in dubai there's stuff happening in tokyo there's stuff happening in sydney and melbourne sometimes you just got to get on a plane and meet people in network it's very simple one two three you figured it out kid i love it yes and i think the

SPEAKER_17: the other thing that's great about the valley in general is culturally everybody's open to taking meetings um and that that sort of people like i think the classic um it sort of plays it's an advantage for young founders even though it sort of like we're building for a very old industry it's like legacy industries you can have all these questions around how do you get to these customers

SPEAKER_19: how do you talk to them they will like how do you break in stuff like that but i think on the other side everybody gets excited when they see you know large market opportunity you know the young team that wants to move fast it's kind of the classic um story that people get excited about so

SPEAKER_17: i think everybody has been very open we're grateful for everybody who took meetings with us it's like helps helped us in the process but i think it's yeah it's very doable if you put in the work and just show up um people like backing um people who are doing the work and and have interesting stories

SPEAKER_07: you know that um people who are of action people who are doing stuff in the world

David Friedberg: are one in a hundred of the people who generally interact with us so i get hundreds of emails where people tell me their ideas tons of dms people tell me their ideas and then once in a while i get a link

SPEAKER_00: to a product or a screenshot of a product or a loom or a you know a quick demo or a figma and i click the link and i go wow what you built is freaking cool uh let's get on a zoom or meet somebody on my team you know or come to our accelerator or maybe we can invest and that really does differentiate you i think you figured it out arian and uh man uh your parents must be so proud of you what may i ask what do your parents do and and what do they think of all this because they're in india

SPEAKER_30: and you are uh 19 years old and you raised over two million dollars are are your parents like

SPEAKER_17: entrepreneurs themselves or no they were doctors so they come from the opposite end of spectrum they were pretty disappointed when i was not going to college um and they it's still not off the books for me like some point maybe you want to reconsider and you know maybe apply and get into some school and do it but i think they are generally they've kind of become more supportive uh over time where it's like you you're doing you're not doing something wrong so as long as you're not like a criminal just and you're happy you're doing what's a pretty good benchmark you're not a doctor

David Friedberg: but you're also not a criminal so there's something in between those two things that is acceptable uh they don't have to bail you out so uh shout out to your parents but message to your parents uh not everybody is gonna just go through and do the standard thing some people have a lot of creativity and they have more energy than those career paths um allow for and so i think i was i would have

SPEAKER_17: just been formal out of um college even if i would have gone just because i think the rate at which stuff is accelerating it's almost like the the opportunity cost of going is it's just it's huge

SPEAKER_00: like i i would the same thing happen by the way in the dot-com era and i told everybody like if you're going to college during the dot-com era when all this was changing it's a big mistake because i've never seen a gold rush like this and then i saw a second one which was in mobile in 2008 9 10 11 12 and now this is the third one i've seen in my lifetime was really three very unique ones the internet and the dot-com era the mobile uh shift and then now ai and they come along every 10 years and it's like this incredible season where there's a ton of snow on the mountain and you really ski really well or there's just great waves to serve there's not always great waves to serve i mean you can build a great company anytime but uh listen i um i am so proud of you um and i'm not your parents but i'm super proud of you that you're doing it uh and i wish you great success and uh my only regret is i didn't get a chance to be in this seed round but hey maybe you'll raise money again and i can slide a quick

SPEAKER_02: maybe your your your uncle jake out could slide a quick 100k or 250k into this i think you're gonna knock it out of the park by the way congratulations and uh take your time focus on product you know

SPEAKER_00: what to do you've you've listened to all these talks and podcasts and you've got great investors sam's amazing it's all about just focusing on the product and the customers and you seem incredibly product and uh customer focused you must have picked that up i guess from sam and just watching

SPEAKER_86: our videos and y combinator videos and blog posts yeah yeah i think that's the only way to do it i've

SPEAKER_17: had versions of doing things otherwise it just doesn't work you have to know only two things matter is just be heads down starting the space i i'd like to be very analytical so all these tweets and like data points that i consume they are actually how i think and i can translate the little bit of this macro reasoning of what's happening with just product what the customer is saying and how they're going to sort of strings together in an additive which i think is equally important as doing the work so it's yeah that's the only way that i think things can happen um and we've had we had a good launch so it's just now iteration we have a lot of backlog of demands we're kind of slowly serving up and figuring out how to yeah get the capacity all you gotta do is delight those customers

SPEAKER_06: everything will be fine everybody check out induced dot ai i n d u c e d dot ai and uh you can follow

SPEAKER_13: arian he's on twitter as he mentioned x r a r y x n s h a r m a uh go ahead and follow him and all the

SPEAKER_00: links are in the uh show notes and we'll see you all next time on this weekend service bye-bye