Ingestion 65d2b17a extracted

Extractions (2)

Content

[Applause]


thank you thank you um thank you for


being here um I saw some of you uh at


the bar last night it looked like a lot


of fun um I'm curious how many of you


just came here straight from the


bar but I appreciate that so my name is


Andre


BFF um I traveled here from South


Florida um I'm very grateful to to be


here thank you for having me thank you


to the organizers this has been a


phenomenal conference and um I I I


really enjoy the city I love it all


right so today we'll talk about building


llm powered applications in


Ruby um so I started out my uh career


building uh rails applications 13 years


ago for the Department of uh US


Department of Health and Human Services


building public facing applications for


millions of consumers then I moved on to


uh continue building rails applications


at USA Today does anyone remember spre


um Sean scoffield uh who was the


original uh author uh hired me and I


worked with one of the most uh amazing


rails teams at the time


um worked at a couple other rails


places um so we're we're going to talk


about generative AI so what is


generative AI so it is a type of


artificial intelligence that generates


text audio video Etc and for the


purposes of this presentation we'll just


focus on


text so large language models are deep


artificial uh learning um


deep learning artificial neuron


networks um with a general purpose uh


language understanding and generation


and they exploded in popularity after


the uh 2017 Google paper uh attention is


all you need and the underlying


architecture uh is called Transformers


um we're not going to dive into


it so the generative AI impact is that


all of a sudden were able to deploy


commercially commercial prod production


grade systems in just a couple days so


previously when you had to uh spend


about a month collecting data um about 3


months uh training data and about 3


months deploying and then optimizing


those models to run on your Hardware


well now you can use


apis um and you can get something into


uh POC state or production State um


within a couple


days so llms add Excel add a bunch of


different tasks um I think some some of


the tasks are really good at our for


example struct structuring data so


taking unstructured data and converting


it to a structured format taking


paragraphs of text and returning Json or


yam something that developers can


consume um they're pretty great at


summarizing text as well you could send


it a short narrative and get a summary


back uh classifying data so bucketing


data dep depending on uh a certain kind


of enum translating languages generating


content answering questions chat bot


style um so I I'd like to present you a


vision


where um I think this is going and I I


don't think anyone truly knows where


this is going so I don't want you to get


kind of tied up in the implementation


details I want you to consider this


holistically so I think that um


generative AI is going to become a part


of every single Tech stack the way the


databases the cache encryption cues


Lambda functions storage and the


protocols are so every Tech stack is


going to have this running process that


it's going to rely on for certain


things so um


the2 investment firm presents this


Vision where it's a crucial part of the


um application


architecture and the developers are


building applications on top of these AI


systems so here's example


so um in every single project of ours


we've got if else statements and you


have to enumerate all the different


options it's a finite amount of


options um and depending on any given


input you either arrive at output one or


output two so with AI we have some


flexibility to consider more uh


ambiguity or more vagueness for the um


any given input that comes in so for


example given any input and an amount of


a number of different outcomes or


methods the AI will pick the best


suitable candidate so even even for


input that it's never seen


before um so for


example uh if it's if it's an AI agent


uh you could have it write a file


approve a document um text my


partner so this has always been the


promise of


rails um is that for business owners uh


your rails teams should mostly focus on


writing business logic um we don't want


developers figuring out engineering


problems uh for the sake of for the sake


of uh figuring out engineering


problems so developers focus on writing


business logic and not Reinventing


engineering Solutions so the old old


world um before AI is has always been


business Logics in fat models or service


objects and it's kind of sprinkled all


over the project um and with and the new


world uh is um I'm going to propose that


some of the business logic is going to


go into


prompts so um in this example where


um um this is a simulated e-commerce


store um and we have the AI agent in the


middle um orchestrating the business


logic and sending mess mesages um


between the different systems that that


e-commerce um store might connect with


so for example any just about any


Commerce store uh connects to a payment


Gateway via an API connects to an


inventory management system VN


API accounting systems shipping


service so I'm


propos that you would be able to write


out your uh standard operating


procedures for your business um put it


in the prompt and then uh just like in


the kind of in the diagram you would


have ai orchestrate those um um


orchestrate the execution for handling


orders uh processing returns um handling


the customer loyalty program


Etc so this takes us to AI agents um and


AI agents are autonomous or


semi-autonomous general purpose LM


powered programs um they can use tools


which are basically apis and


Integrations uh via function


calling um and they work best with


powerful llms and um the openingi gp4 is


kind of the the leader right now


although it changes every


week and they can be used to automate


workfl Flows In Business


processes so agents


are um anyone that's ever try to build


them they're pretty unreliable and this


is the way I I I think about this


is they become more


reliable uh the um the upper triangle


triangle as the number of tasks or


responsibilities that these agents are


respons for shrinks or


decreases um so you kind of narrow down


the set set of responsibilities that AI


agents um uh execute or or the decision


tree that the agent has to go


down so um if you look at this chart uh


on the x-axis we have um f a focused


agent so maybe a single task or uh dual


task agent on the uh left hand side on


the x-axis and and a general purpose


agent uh towards the right side uh on


the


x-axis and


basically I think um we can this is at


least POC ready um so the upper left uh


quadrant and the and the bottom um right


and and um if if you get a general


purpose reliable agent well that's


AGI


so the adoption has


been still pretty slow um and this is


number this is due to a number of


different


reasons um it's a fast changing field so


if you decide to invest into this Tech


technology right now you're chasing a


moving Target because everything changes


every


week um there's a lot of ambiguity


around the IP ownership or or uh


copyright ownership so who owns the


derivative of work that the AI has


generated uh whether it's text or


imagery um because these models are um I


believe trained on some copyrighted data


even though they don't tell us so


there's a lack of tooling for developers


to build


applications um and also it's it's risky


so in some instances


where um large projects are rolled out


um like the Air Canada um chatbot that


told the customer wrong information and


and and then the customer ended up suing


the company for money


um or also below there's an example


of


um a user convincing the General Motors


chatbot to sell them a car for


$1 um


and you either give away a uh $40,000


car or you suffer reputational damage so


of of course the first option was safer


I I hate calling this prompt engineering


this is this is prompt Alchemy this is


not an engineering


discipline um given how random it is um


you change a couple of words and you get


a completely different


output um look at look


look at


this look at this research paper um I


mean you're you're telling the model to


uh take a deep breath


and think of this problem uh carefully


and it improves the results it doesn't


make sense um and when you try to ask


people why does this happen no one knows


and they just say well it's in the


training data but how does this happen


no one


knows there's jailbreaking techniques so


basically getting um a model to do or


say unintended


things um for example this paper came


came out of uh


anthropic um


um they they coined this term uh many


shot jailbreaking so if you go to open


AI right now and ask it how to make a


bomb um I don't want to do that because


I don't want to get uh someone show up


at my door the next


day um but you can trick it into telling


you how to do that by listing a bunch of


examples as if it did tell you how to do


a these malicious things and then it'll


gladly tell you how to make a bomb


so of course there's a lot of


deficiencies that we've looked


at um there's


hallucinations um um because the models


are trained in on a uh in a snapshot in


time the data gets outdated pretty


quickly so for example GPT 4 is trained


uh on data up through April


2023 um and in some instances you wanted


to utilize the relevant knowledge your


proprietary knowledge and it has um no


means of knowing that right out of


box so


enter uh what's become popular uh this


methodology called retrieval augmented


generation or short for


rag it's very


simple it's when you


take um relevant knowledge uh your


proprietary knowledge you inject it into


the prompt along with the user's


original question and then you have the


model use that


knowledge to generate a relevant


accurate answer to the user's question


so for example a popular use case is


building QA systems on top of your


proprietary corporate data so if I want


to find out how long the paternity leave


is at my


company the gp4 is not going to know


that it's um it's buried down in the HR


documents so this is


rag so to break this down um to go a


little bit deeper double click on that


um it works by first generating Vector


embeddings from the user's question then


you typically store your proprietary


data in a vector


database so you would execute a


similarity


search you would find the relevant


documents and then you would construct


the rag prompt and then that's what your


um llm model would use


so let's let's dive into some of those


things so what are vector embeddings so


Vector embeddings are a machine learning


technique to Cluster data by its


semantic


meaning um so llms encode meaning behind


words or


texts um and place it somewhere in the


embedding space or vector


space um and Vector space is a


uh multi-dimensional space so for


example the open AI uh Ada uh model is


uh


1536 Dimensions large so for example we


could take this


string the Ruby programming language is


30 years old and has stood the test of


time is not true and we run it through


an embedding


model and it spits out and it spits out


a vector uh of length uh


1536 so because a human mind can only


visualize 2D or 3D


spaces uh we're unable to show what a uh


1500 dimensional uh space looks like but


if it was reduced down to a 3D space and


you can see that our Blue


Dot is


clustered by the meaning along these um


other


points so this also called


a can you guys hear me


okay and when we're doing semantics


search also known as um Vector search is


we're calculating the distance from your


query to the different points found in


the vector database found in the latent


space and there's different ways of


doing that um you can calculate the


Manhattan ukian uh or cosine similarity


distance


and then now that you've


obtained relevant documents you merge it


as the context into into your prompt


typically the instructions at the top


you would


um


Drive some sort of llm behavior so you


could tell it to reply concise be


concise and straight to a


point and then it would complete the


rest of the answer


so that's


rag this is what naive rag


is so again the query comes in you run


it through the embedding


model um you run similarity search


against your vector


store where your proprietary corporate


documents have been indexed


into and you take those documents you


merge them into the


context package it up as a prompt give


it to the llm and uh derive the


answer so evaluations are very diff are


are very


difficult um typically people rely on


human feedback so if you've used um Chad


GPT you know that every answer comes


with a up vote and down vote and this is


what they use to fine-tune the model and


and figure out how well the their models


are


performing if you're building a system


you can then use gb4


to uh assess or critique itself so you


could say this is how my model performed


this is how my model answered this


question you could send it back to the


gbt 4 and say


is this a satisfactory answer how would


you rate


it um one of the few quantitative ways


to do this and as Engineers we all want


uh a we want to take a scientific


approach um and one of the one of a few


ways is this uh technique called ragas


uh and it stands for rag


assessment um which you measure three


different metrics which are which is


faithfulness


so um making sure that the retrieved


context uh can act as a justification


for the generated


answer uh context


relevance um so making sure that the


context is focused with little or no uh


little to no irrelevant information and


then answer relevance um making sure


that the answer addresses the actual


question um and you run those three


different metrics uh you combine them


and you get a um a a ragas uh


score so there's other Advanced


strategies


where this is just one of them other


Advanced rag strategies where you can


use two different indexes IND


dices um one that just contains


summaries so for example you generate a


summary off of um every single corporate


doc of yours and then the other index


that


contains um


actual


um chunk data and then you first run


similarity search against the summaries


um get the pointers to the more detailed


index yank the data out into a context


package it up into a prompt send it off


get your answer


back


so everyone talks about Job Bots all


these


Enterprises um that have fallen face


into into mud into puddle like Air


Canada GM


um ver building chat Bots because that's


it's it's


sexy but I actually think that when


you're interacting with a human you open


up your prompt for prompt injection


prompt hacking and there's just an


infinite amount of different scenarios


so I actually think that the near term


is running these models behind the


scenes and and


controlling uh your prompts so I'm not


sure that we're going to go at least in


the in the near term towards these open


chat experiences where you're a customer


sales associate and you're chatting with


this AI bot and you say reach out to


this client DRFT this email create a


presentation


deck actually think you're going to have


some sort of a control Panner panel


right we're closing down the set of


tasks the set of responsibilities that


the AI agent to is responsible for to


make these systems more


reliable so I think you're going to have


a control panel with quick links kind of


most common business processes that are


just a little bit more intelligent than


what the current systems are behind the


scenes so enter Lang chain


RB um I'm going to be humbled to say


that it is the most popular Ruby


solution to build um llm uh powered


applications the use cases are uh


General um interfacing with uh


llms um you can build agents you can


build uh rag


systems and our approach is being vendor


agnostic everything changes so quickly


we don't know who the leading providers


are whether it's going to be open


AI uh or Google or a smaller French


startup called


MCH um so we're ready to kind of support


all of them um and we're going to try to


stick with best practice


IES uh and stay on top of the almost


daily emergent


research um and of course in the


greatest Ruby on Rails


fashion um batteries


included so the linkchain RB approach


is's a common interface for llms so you


could just pick your


provider um and and the methods adhere


to the same


signature same with Vector databases


picker


poison um you can add data by just


calling add text and then run similarity


search and then if you do ask that


executes a naive uh


rag so we're going to do some


coding um I'm going to walk you


through uh I'm going to walk you through


a short kind of e-commerce AI assistant


that I've built um I'll try to kind of


explain how it's built and then we'll


run it and see what kind of results we


get so um if you'd like to look at the


repo um just scan this QR code um it's


not necessary um all right


is this uh is this big


enough can you guys hear me


okay okay so let's look at some code um


see just trying to close all uh files


for


now okay


so what we did here is we pulled the


linkchain RB gem in um we pulled a


mainstream gem that validates that an


address is a valid


address um we pulled in the underlying


uh Ruby open AI


gem um I was debugging so of course I


need pry um we have sqlite


SQL uhv EnV and and the mail


gem um and in


our setup DB


file we create a simple


uh SQL database SQL light database so we


have a products table we have an orders


table we have an order items table and


the customer table um and we create


some uh dumy skes with price and


quantity um and this AI assistant is


going to run this kind of tiny


simulated um Ecom Commerce store um and


orchestrate the the business logic


between the different


services and we're calling the store


nerds and threads uh that


um makes graphic nerdy t-shirts for uh


software Engineers that like working


from


home so that's the


background and oh


okay that's that's a lot of


uh it's a lot of keys that I'm going to


have to change shortly


after um all we need is the open AI key


and the database name and we're calling


it nerds and


threads um so we've got very simple


models to represent those tables so


customer has many


orders an order belongs to a customer


and has many order


items an order


item belongs to


order and belongs to a


product and a product has many order


items um so we're going to going to


represent our


services as


tools and a tool is a collection of


functions um so for example we can take


a look at


our um customer management


tool it inherits from linkchain


toolbase uh um and points


to a customer management


Json um that I'll I'll show in a


second and it has just two methods


creating a customer so we create a uh


customer by passing the name and the


email and create it in the database and


send this um Json


response find


customer um which we find a customer by


email and send the hash


back and what the Json does is it


describes these


functions in the open API


spec right so Swagger open


API um and we will be sending this data


we will be sending these signatures or


tool


signatures to


the


llm um and that basically explains it


what these methods are for and how to


use them so in this example you could


see that the create customer would give


it a description that it's a part of the


customer management service and it


creates a new customer with a given name


and then we specify the name and the


email parameters that are


required and same thing for for fine


customer so this is a


tool um and what linkchain RB is doing


here is it's kind of orchestrating that


connection right where you can pass a


tool and it and behind the scenes it


uses this Json to uh match it with the


methods in the class and and uh


orchestrate this function


calling so we have another service


called email


service that uses the Ruby mail


gem um and it it it just it takes the


customer ID and the order ID and it just


sends an email with some order details


thank you for shopping at nerds and


threads we have a Inventory management


tool that you can use to update


inventory with new


quantity or find a product by skew and


get a


hashack same thing in this Json here it


describes the


interface order management


tool that you can use to create an order


Mark the order Mark the order as


refunded or find in the


database a payment


Gateway that has charge customer refund


customer and issue store credit


methods same thing open API spec of


these three different


methods


explaining to the large language model


what they are and what they can be used


for


um and the shipping service that


validates the


address this is kind of nice to have um


creates a shipping label and creates a


return


label so let's um give it a run and and


and see how um well it uh


performs uh so I have saved some of


these instructions in the read me um


because I'm not going to watch you


painfully type out every single


thing so I'll just make sure that that


database doesn't exist is is the


terminal big enough for everyone to see


okay so um I'm going to run the setup DB


to create our


database and I'm going to run our main.


RB


file and let's just check the product


um the database is there the the models


work so now what we're going to do is


we're going to instantiate the llm


client so we're going to say Lang chain


lb ai.


new


key and default options we're going to


use


the chat


completion


Model gbt


turn let me just confirm that I typed


the right


thing chat completion model


name


oh okay so the llm client is


there


so if we just want to send


it a simple


message we can and response back with


hello how can I help you today um we can


also embed data


sorry not not not the uh yeah let's skip


this for


now um so in the instructions here I


wrote out what our


store how our store uh is going to


operate so we're going to say that


you're and we're going to tell the llm


that it is an AI that runs an e-commerce


store called nerds and threads that


sells comfy nerdy t-shirts for software


engineers at work from home you have


access to a shipping service inventory


service order management payment Gateway


email service and customer management


systems and it is responsible for


processing


orders follow these exact procedures


below and we give it steps to execute


the order step by step so create the


customer account if it doesn't exist


check inventory for


items calculate the total


amount charge the customer create the


order create a shipping label and send


an email notification to customers to


the


customer so I will insert these


directions into


um a


terminal and let's instantiate our


assistant blank chain assistant which is


going to use those


instructions it's going to use the


instance of the llm that we've


instantiated and it's going to use


these six different


tools and the thread is


is a thread just keeps track of the


different


messages I'm going to clear the screen


I'm going to paste this in so we have


the


assistant so let's execute the first


order which is myself just purchased


five


t-shirts with


this and my address


is this address in New York not a real


address and not my


address I'm going to clear the screen


and paste the


sy so sending a call to linkchain open


AI looking up customer record sending a


call to linkchain creating customer


record validating address charging


customer updating and


so we will take a


look okay so let's let's see what


happened


so I'm I'm going to skip all these


different calls to the


llm but


we looked up a customer


record it didn't


exist we created


it we looked up the skew to get the


price for this


item we then called the shipping service


to validate the


address and then we charge the customer


for this


amount and I'll show you that it's five


t-shirts at that price amounts to this


amount we then updated


inventory


with um updated quantity


we created the order


record we created a shipping


label which calls the behind the scenes


it calls the the when you create a


shipping label it calls the validating


uh address method


again and then we sent an email


notification


um and if we look


at this is these are all the messages


that were kind of


exchanged


um I I understand it's not super rable


but this is the original prompt system


prompt that we wrote out as the as the


instructions to our um AI


assistant we then sent the message


saying that um I had purchased five


t-shirts and then the this assistant is


is a message that came back from from


the llm and it it it calls these me


methods so it tells us that I want you


to call find


customer with this


email and then we match it with that


tool or that


API we serialize


these arguments and call that method in


that tool with those arguments and send


back the result so our tool actually


returned customer not


found then the assistant said


okay create the


customer this is the name this is the


email we said okay we just did here's


the customer ID


the assistant then


said okay find this product with this


q and then also call validate address


this is known as um uh parallel


functions we said okay here's the here's


the


skew the total we have in the inventory


is 10 here's the price


so if you


take 5 by 24 it's


24.95


the so this is the uh this is this this


tool validated the address and sent uh


true as yes this address is


valid and this tool call ID is is is is


a pointer to the corresponding tool


calls so that's how VM is able to tell


that whatever we send it that output


corresponds to a specific function


call um the assistant then said okay


charge a customer for a total amount of


$14.95 here's the customer


ID we said okay we charged


for that


amount random U ID for a transaction


ID the assistant then said update the


inventory and the new quantity since we


just sold five t-shirts is five so


update this queue with quantity of


five create the


order with these order it items so we


just sold five items of this


SK


ah so I think this was supposed to be a


hash


um okay so the cre


order so the order management create


order call return success true with his


order


ID and the update


inventory Returns the product hash which


contains the quantity but I I forgot to


call do has on


it the assistant then said create a


shipping


label with this address and this


customer ID and there's there's an enom


where it it it can choose from uh


shipping


provider we said okay true here's the


tracking number random U


ID and then we the assistant calls calls


send


email and then this email gets


drafted and then taking all of that the


assistant then just says my order has


been successfully processed here's a


summary customer ID order ID total


charged shipping provider and you got an


email thank you for choosing nerds and


threads so I'm going


to clear this


thread and the reason why I'm doing that


is


because I want to I want to delete all


this uh


uh chat


history so otherwise this chat history


will be sent back to


vlm and recall as as I as I said in


order to make these things


reliable you have to close down the set


of


responsibilities not give them a very


massive prompt because it'll it'll get


overwhelmed and it'll get confused so


these are the limitations and and what


we're trying to do is we're trying to


find that sweet spot where it can


actually be


usable


um and again the this original message


that I


sent I don't think it should be an open


chat interface where some someone runs


this e-commerce store and they just they


they they chat with the sist system


there's


just too many


permutations too many possibilities for


for it go to to go


wrong imagine that this this this this


gets sent by a different system in in in


this format right and and we can


illustrate that this


works 9% n9ine nine times out of


10 so


I'm going to clear this


thread and I'm going to reset the


instructions because that gets set as


the system message so if we look at the


thread


messages


with no extra letter s then it it just


contains that same same same prompt um


again operating procedures for for our


simulated e


store so then my friend stepen comes


in and he's


purchasing um this Q in this


queue I don't believe this is Steven's


uh real Berlin address


and I'll talk through it in a


second end it with uh sending an an


email


okay let's take a look at


this looking up customer record for


Steven looking up the product


skew looking


up the different product


skew create the customer record for this


email this new


email validate this Berlin


address update the


inventory so the original quantity was


five um so Stephen purchased two of


these SKS so it updates the inventory


with the new


quantity and this this CU of the


original quantity was three and Stephen


purchased one of these skes um so the


new quantities two


units charge


them and this was calculated uh


correctly I


believe create the


order create the shipping


label send an


email so let's I'm going to clear the


thread again and we're going to try


slightly different instructions


so we're going to try these return order


instructions so let's take a look at


them the first part is the same you're


operating nerds and threads you have


access to these


Services


Etc follow these exact procedures below


for returning an


order


so look up the order calculate the total


amount refund the payment Mark the order


as


refunded I'm going to insert this these


instructions


and I'm going to set


them let's just


verify that they


were they're in there


and Stephen changed his


mind and he's returning his


order looking up the


order looking up the skew looking up the


other


skew


refunding the payment


Gateway and refunding the the order


which just marks the order as um as


refunded and this is all actually saved


in the


database powered by SQL


light and last


thing let's clear with thread


again reset the


instructions and


we got a new inventory shipment let's


update some


skes so inventory


update


updating this CU updating this queue


updating this


queue and then in at the end it just


tells us the the inventory has been


successfully updated


um you might ask why didn't


we just put these return


steps into


here and again this is this is what I've


kept referring to


is we the these systems are are still


pretty dumb and we don't want to


overwhelm them and we're trying to


figure out that sweet spot where it


could still be productive and reliable


right if you remember my x-axis y AIS we


want to


be in reliable but


focused um quadrant right so focused


interactions so perhaps you have two


agents one that is just laser focused on


completing orders and one that's just


laser focused on returning orders and


that's all they do but they're not


telling you what the meaning of life is


and dealing with all of


it uh and and and and consuming the


massive operate business operating


procedures manual


so I I would love


to check us out and poke Calles in in it


and give me some feedback if you want if


you


can okay so why


Ruby well um I love the language


I've been doing this for a very long


time I love how pragmatic our community


is


um we focus on business


problems and delivering business


value and not over


engineering um we have strong oop


principles I think I think of um such


great books like uh practice Orient


practice uh object-oriented programming


uh in Ruby by Sandy


Mets um a lot of python people come from


research and


Academia um if if you've looked at some


of those projects I'm not saying


everyone does it but I don't see good


patterns I see methods that are uh


hundreds lines long and not tested and I


think we have good mental


models um that um could give us some new


ways of thinking about these agents and


and


llms


um and um I I I tweeted this thing


earlier in the


year um everyone says we're so we're so


far away from python they have all these


amazing libraries and I said um well you


can if rewriting some of them if if if


you're um yearning for a utility in


Python and it doesn't exist in Ruby just


get Chad


gp2 rewrite it um


so uh a ruby developer is sad the


library a python Library doesn't exist


in Ruby they paste the code file by file


uh into Chad GPT ask it to rewrite it


and um with a couple tweaks you can you


can run it and and Ruby developer is


happy now I've done it a couple times it


it's it


works not


bad okay


so some of the things I've learned from


uh running this uh open modest uh open


source project is is you got to be


responsive so when people ask you


questions uh or issue pull requests or


open issues um you got to be responsive


because you you you have them for a


finite attention


span um you got to be friendly you got


to be welcoming uh and you got to be


helpful you got to jump on these


opportunities to to help people and go


out of your way to deliver


value um thank you for all of these


different


humans that have contributed to to the


project


um


and here's our


Discord um where we talk about uh the


project and kind of what we're seeing in


the um AI llm gen AI


space and the last thing that um I want


to leave you with is we do not have it


all figured out yet uh but we're


optimistic thank you


[Applause]


all right any


question thanks for the great talk uh in


your example you've built a system that


for that sends the instructions to llm


for each incoming request and I wonder


because potentially you could have asked


llm once to generate rubby code that per


performs those instructions and I wonder


what tradeoffs are


here yeah so if you didn't connect the


dots or maybe I wasn't uh super


clear this this is what we tried to do


right so we had in nerds and threads


e-commerce AI assistant orchestrating


messages um and there was nothing the


code that that had this business logic


how this e-commerce store should


operate


um


so you know I'm I'm actually very


interested


in this type of


angle to to solving the problem I think


a lot of companies are pursuing um


generating the code and then and then


saving the


code I'm actually curious what would it


be like without that so I'm actually


curious what it would be like


to generate the


execution right and I will ask why do


you even need to save the


code so what we're doing as as as rails


developers is stringing apis together


right API


to the


database mail service this service that


service


um and like I like I've stated


in in in this slide right


um we're we're focusing on on on uh


building business logic so that was my


attempt to to illustrate this concept


that that


can we write out business logic in in


prompts and then can the execution be


inferred and inference is is what the


llm is doing as it's predicting next


tokens that's called


inference thank


you all right guys uh sorry Stephen we


have to move the uh discussions and


questions to the hall uh as we are over


time that's what I heard in there was a


whisper uh let's thank Andrew bondarev


your Ruby AI guy for explaining


this thank you very much


[Applause]