# OpenAI DevDay 2024 | Community Spotlight | Swyx

## Метаданные

- **Канал:** OpenAI
- **YouTube:** https://www.youtube.com/watch?v=wnsZ7DuqYp0
- **Дата:** 17.12.2024
- **Длительность:** 9:40
- **Просмотры:** 12,259
- **Источник:** https://ekstraktznaniy.ru/video/11368

## Описание

Engineering AI Agents

## Транскрипт

### Intro []

hi everyone okay so I'm supposed to bring up this as though as a cooking show and it was already pre prepared um I'm s or Sean I'm natively from Singapore but I've been working and living in the states for the past 15 years um and uh I have a secret I have a couple secret agendas first of all um I was given I was I thought I had 25 minutes then I was given 15 minutes and now 10 and then today I was told it was 9 minutes so I'm going to give you a ton of homework and don't worry you can take photos all these but uh everything's already on my website uh 6i engineering agents um and so I'll just give you a

### Agenda [0:40]

title homelet because my you know I'm just I'm not going to like uh dump things down just uh for you the second thing second agenda I really have is on the 30-year Horizon I want to turn Singapore into an AI engineering Nation uh I want every single one of you to understand how much power you have uh with the foundation models that are available to us today um and to respect that um you know there's so much knowledge is being shared publicly by all the other engineering organizations so I'll go Dive Right into this the reason this talk is awkwardly named is because I wanted to kind of squeeze some form of AI engineering into the talk title um I'm very well known for this essay on the rise of the AI engineer speing out this sort of specialist role for software Engineers building with AI

### Podcast [1:20]

um and particularly uh my the podcast that I've been running with alesio where you at hey um uh we both basically focus on AI engineering covering all the builders and researchers in the space and including one episode with Minister Josephine too on turning Singapore into an AI engineering Nation so highly recommend checking that out if you are interested um personally I've also been building small AI agents um so I have a news um newsletter that reports on the top AI discords reddits and twitters every single day it's the biggest AI newspaper with no journalists and you can also build this yourself so

### Building AI Agents [1:55]

basically I'm here to just condense everything that we've learned on building AI Agents from talking to people from building it ourselves and all that um and um I think the main starting point I recommend to people is from lilan Wang uh she is a very well-known uh writer but also formerly head of uh Safety Systems at open a and she says agents are llm plus memory plus planning plus two use um I personally tend to modify that a little bit um and we'll go into that in a little bit of detail but each of my slides has homework at the bottom and you can obviously um see the sort of slides presentation that the at the end of this talk um so a lot of my thinking about developer tools comes in impossible triangles like this vend diagram of like you start here and you know eventually you start adding the features of competitor tools and where should you

### Developer Tools [2:40]

really pick um and I think having a clear map of what jobs to be done uh for building sort of your LM infrastructure is really important so I tend to recommend having some kind of Gateway some kind of Ops tool some kind of rag framework uh the top ones are all the open source tools that I use um I also want to especially shout out Eugene Shia where are you um who runs feedist AI um uh and so like there's a lot of prominent singaporeans who are actually working on uh this AI so I I highly recommend you checking out Eugene work and fist AI as well the second part

### Memory Knowledge [3:10]

of the stack that is really emerging for agents is memory and knowledge uh not enough people still appreciate chat gbt including memory in February of this year it feels like an eternity ago but uh it's still a very good Agent uh memory implementation and I highly recommend reading the M GPT paper if you're interested in that on the knowledge site uh Vector DB is are very well known but increasingly U of a lot of interest is knowledge graphs I would say the most surprising popular talk at the conference that I run AI engineer conference uh was the graph rag talk again I leave your homework to check it out as well the third part of the stack is kind of the most of vague um area of active research this is planning and multi-agent um here I would recommend people read a couple things from openi one it's the let's verify step-by-step paper which shows the potential process War models um and then secondly the SW F uh project which is a project that's not a framework worked on by Elan and shamal I think who are here in the audience as well um and so highly recommend multi-agent as the single simplest Improvement in um performance grab was also talking about that's their what they aspire to next and at the end of this talk I'm going to show you a demo of how multi-agent actually improves the experience by quite a

### Tools Orchestration [4:20]

bit um together and I think the last part of St is sort of tools and orchestration I kind of jam them together in the same place it's no coincidence that when Lang chain launched the first three things that first three features that they launch with were these three things that eventually every single agent eventually needs to have so the first is a code interpreter like a Sandbox environment like e2b that executes code and runs all the deterministic things that llms cannot run second thing is browser control by being able to surf the Internet to search for results um and to sort of read from using the computer then lastly is the self ask or react Loop where you sort of uh do this sort of cycle of observing and reacting to the external environment um and um these tend to be the building blocks of all agent systems whether it's magentic from Microsoft which I highly recommend checking out or cognition which um I think um you know opening eye works really closely with and also spoke into their architecture at the uh their specific talk in my conference so highly recommend uh checking out all these talks if you're interested and lastly this is my brain down Talk Brain down slide where I'm like basically making the meta point that everything I just told you was like parts of the stack that is well known that you're going to need to build agents I think the meta lesson for AI Engineers is to have some kind of mental map of capabilities because every single time there's an advance in model capabilities there's actually an opportunity for you to build a state of-the-art agent by utilizing those features so that's something I want to really try to show you uh in My Demo today um I'm very inspired by this sort of scene in Iron Man where basically he's kind of talking to his computer and just kind of manipulating directly the the thing that he's working on and I think that the lot of coding should actually move closer towards direct manipulation so I really am interested in showing that off um and I hacked on this a little bit uh in my free time so um I'm going to show

### Game Creation [6:05]

an uh a fork of bolt. new which is an open source sort of text prompt to app creator uh and we're going to make a Space Invaders game this is all standard bolt. newu stuff uh it's basically just going to like based on a very small prompt of make a fastpac space evors game by the way we're using uh today's GPT 40 version be nice if we have like daily uh cuts of GPT 40 right um and uh yeah I mean it like creates a pretty decent app but it doesn't really look like Space Invaders so the second thing I did is I'm going to add a second agent to talk to this first agent and I'm going to use voice because I I'm very obsessed withal voice hey um so uh we have this space Anders um game but uh the aliens are not really sort of coming in waves I want come them I want them to come in waves and to fall down in discreet steps can you do that okay I've asked the code Jenny I think the way that I want to code is just to QA I should play the game call out like what I don't like about they too fast and it should just be working on it in the background so here it is it's been fixed um it's just pretty cool all right but like it's not still not engaging enough I need all my like sort of dopamine hits uh so like every time an alien dies can you drop a special bonus some kind of PowerUp feature um and also like let's maybe add um some sparkling stars in the background to make it feel more like space and then also lastly maybe let us do uh make the aliens look more aliens use alien emojis and make the spaceship look more like a spaceship can you do that great so I just gave a triple prompt it's all squeezed in one there and if you use an agent that has a planning stage you can actually split that out and manage each of those things step by step this is exactly what we're doing inside of here um again I don't have the BWI to give you the source code or anything like that for this uh but it's just kind of building back consider it a visual upgrade a little make uh so yeah it's kind I like the powerups it's very cool um and I think like this sort of human AI collaboration I think is really powerful by the way um any audience suggestions like I can take audience participation this a real life demo make them shoot back can you make the aliens shoot back at me so I can potentially lose oh I just love that I can speak features into existence all I'm doing is kind of QA the game get ready to dodge and it's it makes for such a much better flow it's like it's kind of like we're living in the age of iron is just an opportunity to practice your dramatic defeat scene ready for the laser oh I lost okay well yeah that's uh that's about it um really fun so what is my thank you uh by the way that demo is open uh we also have I've also will be publishing on the Laten space newsletter all the learnings that we've had building the real time API we call it sort of um the uh Missing manual to the opening eye real time API that was the BB I was working on guys um so what's the overall lesson is that you can build so much with this core understanding of the agent stack um I think you as software Engineers don't have to be PhD level researcher to do it um I I feel I think it's so understated how much potential there is in Singapore in all of you building with uh with these Foundation models and it's so underrated how opening eye values you like flying all there San Francisco us team here to talk to you so I hope you take advantage of this opportunity and I hope that we all collectively turn Singapore into an AI engineering Nation thank you very much