{"id":"FCWje0ge8iE","title":"NEAR AI Office Hours #38 w E2B","description":"⭕️Project: E2B - an open-source runtime for executing AI-generated code in secure cloud sandboxes. \nhttps://e2b.dev\nhttps://x.com/e2b\nhttps://github.com/e2b-dev\n\n⭕️Guests: \n-Tereza Tizkova, DevRel\n\nIn this episode of NEAR AI Office Hours, we’re joined by Tereza from E2B to discuss the critical infrastructure needed to run AI-generated code securely. We explore the challenges AI agents face, the role of execution runtimes, and why secure environments like E2B’s sandbox solution are crucial for scalable AI applications.\n\nTereza shares insights into how E2B provides a fast, scalable, and secure cloud runtime for AI-generated code, supporting use cases like app generation, data analysis, and autonomous agent workflows. If you’re building AI-powered applications, understanding infrastructure challenges and solutions like E2B is key.\n\nTimestamps\n00:00 – Introduction to NEAR AI Office Hours & Guest Introduction\n01:03 – What is E2B? Building infrastructure for AI agents\n02:52 – Use cases: Running AI-generated code securely\n06:02 – The challenge of persistent environments for agents\n07:39 – How E2B works: SDKs, cloud execution, and self-hosting\n08:17 – Where AI agents fail today: Reliability, scalability, and security\n13:29 – Making AI agents more reliable: Execution runtimes and feedback loops\n17:05 – Firecracker vs. Docker: Security, speed, and scalability\n26:56 – Lambda vs. E2B: Why choose E2B for AI execution?\n29:30 – Open-source and self-hosting options for E2B\n30:46 – Closing thoughts & new developments from E2B\n\nNEAR is the Blockchain for AI\nWebsite: https://near.org/\nNEAR AI: https://near.ai/\nIntents: https://near.org/intents\nChain Abstraction: https://www.proximity.dev/\nSharding: https://nearone.org/\n\nFor the devs\nGitHub: https://github.com/near \nDev Docs: https://docs.near.org/\nDevhub: https://devhub.near.page/ @NEARDevHub\n\nConnect with socials\nTwitter:   / nearprotocol  \nBlog: https://near.org/blog/\nMedium:   / nearprotocol  \n#NEAR #nearprotocol #NEARAI","channel_name":"NEAR Protocol","published_at":"2025-03-18T00:00:00.000Z","thumbnail_url":"https://img.youtube.com/vi/FCWje0ge8iE/hqdefault.jpg","transcript_text":"hi everyone uh today uh we have a new episode of office hours and with me is Tera from e2b and we're going to talk about infrastructure to run agents SEC uh to run I guess code generated by agents securely right yeah running AI generated code securely y okay so Teresa do you want to introduce yourself and tell a little bit about it Tob yeah uh first of all thank you for invitation I think uh we have met first time at uh AI Event in San Francisco and I really appreciate that because I was uh watching other workshops from neari so thanks for the possibility yeah I'm I'm deel at e2b uh e2b is a startup we are operating from San Francisco I also just opened a check office and we are building uh infrastructure for AI agents uh so I'm always saying we are not AI company we are more like low-level infrastructure boring company uh and today I want to chat about uh where AI agents still fail why is that and how good infrastructure or like good processes in the background uh might help that because I think it's still relevant topic even through there has been like crazy louches like menus Ai and open SDK there is still like the waves where agents are like have more hope than fail enough again and uh it's never like still reliable enough like 100% so I want to go into what are the challenges and how execution runtime might help uh yeah so ITB uh it's the cloud environment uh that you can run AI generated code in I think best demonstrated it's on example uh for example perplexity is using it to be uh let's say you give it a prompt to analyze data to do something more complex uh perplexity has underlying llms that generate the code for that because anytime you need to do something more cool basically the agent of course needs to write code it's like giving agent brain let's say or hands or just like something useful and this code needs to be executed somewhere right and uh of course you could use like some simple like grapple or something like very simple environment and just run it but you usually needs much more for example perplexity has millions of users each needs separate instances so we are building environment where you just run this code and it has actually more more features just uh build right for the llm use cases like that uh other use cases are generating apps right uh like vzer or uh yeah uh so it's also like the same principle you generate code you execute it in e2b sendbox and you have some something like real output like whole apps uh other use cases are like uh running tests and also computer use so this is just the environment we built e2b sandboxes this is the overview of of the use cases and how they look and uh yeah I want to take a step back like why we started building e2b it was in 2023 actually when agents first started it was I think AO GPT launched right and everyone had this like wow moment that like agents are going to be everywhere and do everything and boole trips and do all like your tasks and replace replace everyone so that's when we started building e2b and we knew that agents are going to need a lot of uh different text T right because it changed is the way that software is built it changes a lot of things uh because now suddenly everything will be built for LMS right people now are putting their like documentation into markdown and creating all things like easy for LMS to read Because agents already are browsing internet and uh they need also better security of course because by definition what uh what LMS generates uh is stochastic uh the code is untrusted you don't know what what is going to happen What data it needs to access so now you need to suddenly think about a lot of different challenges uh this is like the diagram from Andre carpa I think everyone has seen that and they made a parallel to like LM being a new Kel and basically everything built around that uh yeah and that's also why we decided to build something that's uh environment or playground that's very special for the llm use case so yeah if you need to if you want to do any questions or jump in just I'm saying feel free to jump in yeah so I guess the the first time I was thinking about these environments was when I was playing with Devon back in the day right and with Devon uh obviously for D that is the primary use case right because it is it is a product that generates code and it it does execute code somewhere uh presumably on some sort of version of it Tob uh you know uh like something similar to it Tob right but when when they want to when they want to expose the code to the user like you know they want to host it so that the user can interact they host it on netlify and I was thinking how um effectively one of the problems that exist in products like that is that a lot of people like I guess in case of perplexity right the code is generated it's executed it's used it shuts down right but with the human in the loop especially the way you say it on the slide it's uh it could be that the code is generated it's hosted but then it seems to stay there for a while I wonder if you guys address that particular use case where the you know where the container needs to exist for a while uh you know whether it's cheap to keep it live or is it is the use case more like you know the code is written code is executed container is immediately shut down oh yeah that's that's a good question because yeah you right that's exactly one of the very specific things for agents and why they fail like they sometimes the process needs to run right and sometimes it runs into error because it shuts down and uh ourand boxes are actually persistent so you can post the sendbox and you can resume it later from the state where it was when you post it uh and you can get back to the sybox file system and memory all the processes uh yeah and in the environment you can also reference like the previous code so that's possible with the sendbox if if that's what what you asking yeah yeah yeah so so effectively um so they can be yeah so the snapshotting is supported right so you can snapshot it put it into the cheap I guess cheap storage wait until the user is ready to talk upon it again yeah that's that that's pretty cool what um I guess um how so so From perspective of we develop right let's say I'm using e2b so my interface is effectively similar to somewhat like some something like AWS right I'm saying Hey I want I want a container it gives me like what SSH access or how do I how do I interact with the with the container so you you interact with the sandboxes we buil through our SDK uh you can use it in python or or JavaScript and yeah and you control you start the sendbox with the SDK and you you control basically everything you can also self host it now it's supported on gcp and working on support for AWS but yeah the SDK basically controls it all you can customize the sbox before and yeah think yeah pretty cool and should I go to the next slide yeah yeah let's do it yeah okay uh yeah we wanted to name this uh where agents fail today because I think uh you see a lot on Twitter like discussions every time there is a new Lounge of Agents someone always says it it fails for them or it's slow or it's going to be expensive so I think the consensus is of course agents are cool but yeah it's still a long way to go right even if it's like really exciting uh I think what doesn't work yet that well in production is the Devin like types of software engineers and I think you might also disagree of course like I haven't tested all but like agree yeah like very autonomous agents right that uh that needs to do like all sequence of tasks one after another that that's like big challenge because there's like more probability to fail of course as you said like it needs to run for a long time keep the context so yeah I think there is like a tradeoff between how reliable the agent is versus how much freedom and autonom it has or and of course like the restrictions there is another tra of like how much you want the agent to do how how much like access to the system you give it versus how secure it actually and I think I see this for big companies especially they are worried how like you know if they can let the like agent for example use computer or even browser so I think there is still a lot to to solve uh and the common challenges we see or hear about from developers so of course hallucinations even through I I feel like the LMS are improving a lot I mean uh yeah like the same same prompts are working better now so of course this is uh like a lot of this is going to be fixed by just better models I think uh but of course it's I feel like agent is the same as like having an intern or like your assistant it needs to like understand the context really well it needs to see the nuances I think this is especially how it differs from humans right because it needs to really catch catch the vibe what you want what exactly you want it can like it has to like read read between the lines so this is still still quite difficult like how to iterate with the humans I think uh multi-agent systems are going to solve this partially I've seen use cases where like they can of course give feedback to each other in production but still it's a challenge and I think another thing is evals that is still like needs to improve and then of course uh like the persistence uh the agents to be fast to be scaled for millions of users and yeah to be secure so that's if you want to add something from your experience like uh from using agents I think it's always interesting to hear no I think with the autonomous development I think for uh for narrow use cases it's probably uh I I have I haven't played with them uh for a couple in a couple months now but I I think last time I tried for narrow use cases they were relatively well but if you try to use a broad one and ask something that is not quite common you know like try to even a simple lean proof or or something in a language that is not as common they uh they hallucinate quite badly still yeah I think it's funny when people try to test like new new launches new agents they always try to find that one use case that doesn't work and I feel like when you try to test it you sometimes can't think of something good but actually then use it in production and you have like most of the work is very very Niche right what people need so yeah I think yeah that might not even be the case right like I presume that 90% of what people do is you know JavaScript and Python and that and that probably just works pretty well at this point yeah I guess but we have use cases like uh we have use cases like running tests in in different languages and for example we have one user that they are building product for scientists and scientists uh write code in r a lot right R is very like specific for like I don't know life science and stuff and they need to so they run this code from R in E Tob and it's very it's very Niche use case because also different types of users right because I feel like who is mostly experimenting with agents now are software Engineers or like tech people but even scientists or like PhD people it's not like the same as Engineers they don't code that well probably maybe not a experience so it's already like different users so I think that's also challenging yeah yeah I guess there would be quite a bit of R in the internet right because even like I guess before before machine learning became before like Python and naai became so dominant R was pretty standard for a lot of I I don't know like how was the like development I feel like now yeah python is popular type script of course yeah okay uh yeah then how to make agents more reliable uh of course of course one way is to make llms better uh but there are also ways you can improve agents and one thing is making them actually like do the things instead of just talking right uh I think at the beginning most agents were just producing texts it was just like chatbots and then they started using more tools like code interpreter of course and producing actual uh results like now for example with deep research they can generate the whole like Report with charts and this is also good when they produce something actionable when they run the code to give you something because then they can iterate on what they produced uh so this also adds the reliability and feedback loop uh security yeah that's what I talked about with e2b so when you have app for end users and if you want to have more than a few users you need to scale it somehow and you need to like is isolate um the instances because you don't want the code produced for every user interact with each other uh obviously but like it's not that easy to make that many instances at scale and still it to be fast and and isolated and then I would say a lot of it is about UI ux because a lot of people had similar idea at the beginning uh building like the same types of Agents like data analysis right but for example with data analysis you can see like how Julius AI made like great product and distinguished from the others even if it's like b2c so it's challenging right but uh you can see that the UI makes big difference especially I think because it's AI so it's good to have like transparency and good good monitoring what's happening how the user should interact yeah and this is the these are some of the problems that e2b solves with the uh Cloud runtime uh so we can run code uh in like secure isolated manner in the cloud in the sendbox uh so again when the code is generated uh the sendbox is started and there all the processes are running you can install packages in the sendbox uh you can access Internet so you can use the sandbox even for like computer use agents uh you can download files from that everything uh like this and yeah and it's good that you can just give it to the agent and that's why I said we are not AI company because it's like separate uh from how the agent is built because we don't care what llm it used right we don't care if it's built with L graph or crew AI or what Tex Tech uses we are just like the environment the infr layer because I I feel like that's that's one thing that developers scare a lot from what we hear is not being loged in for specific Solution that's like opinionated right they always ask they want like low level something that fits their text tag they don't want to like change uh change anything so that's that's good that you can like connect it and be very agnostic I think in general that's good for like llm software because like people already are building like with multiple llms they combine it all they just don't choose their favorite right it's it's like difficult to be just like big Winners so I think it's even more important to be just like agnostic uh okay yeah uh execution run time that's uh what I said basically so here are some of the like uh properties like uh features that you need to have for llms so you need to yeah start the sandbox quickly and uh our sandbox is built with firecrackers by AWS uh it's type of microv VM you have like multiple micro VMS but that's just firecracker is very fast and it's battle tested so we can start it in like I think under under 200 milliseconds uh yeah you can connect it to any LMS you have file system there you can have custom packages uh and yeah you can have persistance and all these features that are like important and you can also sell host so that's how I would summarize probably why the environment is good uh then we have like the second part uh uh what even are the types of like run times uh what you can even do if you just want to start let's say from scratch and uh run your code just simply and then I think we see why microv VMS like firec that we are using uh is that good uh so you can have like very simple uh simple security like just process level uh but of course that's you probably wouldn't use it in production for this because it's just like uh limited security then you have uh containers like Docker and uh people usually ask a lot uh what's the difference between running code in Docker and in and in Tob that's like uh one of the most common questions uh yeah that's I get this at always at like almost every event uh and I think I think there were agentic apps that were using Docker for running code for example gen at least at the beginning not sure if now but they were using Docker to to execute code AS code interpreter and uh yeah so I want to address this like Docker is is fine it also can be good as a runtime but uh it has like limited security because uh the kernel userland interface is very large compared to to VMS uh where the interface is much smaller and uh yeah so it it has better security and yeah and some solutions combine both but of course uh when you run it locally also it's it's difficult to scale them so that's like limitation for Docker uh but yeah of course it doesn't mean that it's not possible or something Docker takes more than 200 milliseconds to spin up presumably I'm saying Docker is taking more than 200 milliseconds to to spin up yeah that I am not sure to be honest I can check after this but yeah probably I mean firecracker is known to be very fast so I guess yeah I get it's not even that fast especially at scale we use Docker today I mean we we run a we have a chat. near. where you can you know you you can execute code and we're running it in Docker and we spend quite quite quite a bit of time to effectively read up on security of Docker containers how to harness them and I think that potentially firecracker is the better alter use it for I guess firecracker is is potentially a better alternative than Docker the on your slide it says tradeoffs security so so does it say uh is the claim that firecracker has better isolation than Docker or Worse isolation yeah I think it has better isolation I think like I need to find the slide yeah but you can see here basically basically like the comparison of the container versus uh virtual machine I think the point is how much you expose like of the kernel and how you isolate uh your processes what is happening from the kernel right because you have containers like Docker and then you have virtualization like microv VM and there you have like extra you are basically protected from the host kernel right I think it's like separated can find and the slide um okay I think it's not here but another question is sorry so with Docker right I can um as a user I can spin up very customized environments right it could be any operating system with any version of you know liaries and what not so with with e2b uh is MI the user configure the environment or is it something that is standardized across all the you know like like OS environment yeah like operating system libraries you know pre-installed software yeah so I think it's two different things like uh operating system uh e2b is using Linux currently and that's the only one I think developers haven't asked that much for for others what but we made add uh other than Linux uh but now it you can run dbn distribution so so that Linux inside the microv VMS uh that's one thing uh with the custom packages as you said yeah so we can we have custom sandboxes uh and you can pre-install packages there so you can customize the sandbox and I think a lot of users not sure what percentage but a lot of people are doing that like you can customize uh you can create basically a Sandbox template uh you generate it and there you can like pre-install what you need what special packages to install if that's then even for a custom uh stand box it's still 200 milliseconds right so so I configure it once in advance and then I can spin up new ones okay yeah that that sounds pretty cool I think it should be possible yeah and you can also customize like CPU and RAM of the sbox template uh and you use it to be CLI to do that yeah also when I was asking about operating system I didn't mean like Linux versus Windows I meant more like uh like a you know I I presume that you know that like uh your actual Hardware is running on Dean or something but let's say I want Arch Linux or or or gento is it something I can do or does it have to be the same operating system as the host sorry I I didn't catch what what you said as example so like can it be a flavor of Linux which is different from Linux running on the H like can I run Arch Linux in my sandboxs yeah that's what I mean uh only Linux is possible now I'm not 100% sure to be honest but yeah I think you only need you only run Linux right now Linux OS but yeah we are definitely like discussing that like uh with users and if if there's like need for other I think there hasn't been that much for this like specific use case with agents but we we might add other probably but that's good question I will find out more yeah yeah we so so the the problem there is that there are like I think it's relatively it's relatively common these days to use auntu right but for example in AWS their standard operating system is their flavor of Linux which is called Amazon Linux I guess and it uses different package manager it uses you know some of the system util is a different right so if someone has dependencies of those right and then you know if a Sandbox can only be B to with apt that's not something they can use right so I guess with Docker I'm pretty sure you can run any Linux uh no matter what host OS is is there I think you should be still able to like uh have any like dependencies you need but like that's very good question so yeah need to find out uh but like you mean your use case would be like someone using already like other ecosystem right and needing some like like using Arch Linux you know Arch Linux is sufficiently different from from Dean yeah yeah not sure okay uh yeah I think maybe you can specify it in your in thebook templ but need to check uh I think I can share after the presentation great okay uh I think yeah let's also summarize so we have also VMS and micro VMS and firecracker is like micro VM and it's like uh specifically fast uh then we have like other types uh of microv VM we have like cross VM uh Cloud hypervisor which is like more general purpose uh and has GPU support so there's like other options than firecracker but again we we uh picked firecracker because it's uh it's very fast it's battle tested and it's built with uh AWS Lambda uh but you can also use like other options so yeah we're also exploring that to build but I think firecracker so far has been Lambda Lambda is built on firecracker I see yeah yeah is it is very battle tested yeah but the question is that's actually a good question right so why like let's say I'm perplexity I need to be running the code why would I use e2b as opposed to Lambda um yeah I think uh because like Lambda is usually how how people start uh or when we have like new customer they usually start building something in-house uh they usually actually very often they build with Lambda functions uh that's like the probably most common way if they want to build inhouse and of course they they can do it uh they can do it like themselves but the problem is like it turns out it it gets like very very complex problem with a lot of in infrastructure to be built uh and you you start to have like a lot of features that you need to add on top of that because as far as I know Lambda also is not like specifically made for agents so you will need like uh some level of Maintenance customization and all these like challenges that I mentioned and you will need to build it yourself so what people most come like tell us is they they started to build build this because they didn't know something actually exists and it didn't exist also long uh some time ago it wasn't like that big market so yeah you can start it like yourself but then probably like how annoying it is to build and uh yeah that that's probably the challenge because I hear usually like the companies don't have like infra team for that or don't want to just focus on making the infrastructure great and I see it that like we we are constantly like working on something fixing bugs right and I I'm not like infra engineer but it's probably very difficult to do inhouse so that's what you mentioned so I guess one of the concerns uh with running something in the cloud is always uh data privacy right you mentioned it would be locally right so so is it open source or how does it work uh well yeah like we can self host and uh we we run into Cloud usually not locally so not sure what uh what you ask like about data so self host would imply that it's running locally on my end right or what is self-hosting yeah yeah that's that's self hosting right but like yeah I'm just saying like not only we don't do only self hosting but a lot of people still run it in the cloud so yeah of course but but for self-hosting so is it is it to be open source or how do I self host oh yeah uh I I think I didn't say but we are completely open source uh including the infrastructure so you can go to our GitHub e2b Dev and for self hosting we have like a complete guide there if you go on our GitHub and find infra repo uh if F guide for self hosting e2b on on Google Cloud so you will need probably terraform and uh yeah Google Cloud CLI and other things and you can do it just there but yeah everything all the code is accessible yeah awesome yeah that sounds great okay we are uh we're getting close to the end of the um of this uh session so is there anything um particular uh any like last U comment what you're excited about in the future what's uh what's the most exciting in the near I'm thinking yeah I wanted to share because yesterday like yeah I think yesterday was the openi launched and they launched some kind of uh computer use agent and we often built build stuff we build like open source let's say side projects to to show how it Tob works so we just launched uh computer use agent with a nice UI that's using the latest open AI model and you can try it for free it's also on our GitHub so we have like a bunch of apps like that we have also like app generation app or data analyses and all these use cases I said at the beginning so if you go to our giab and find open computer use uh yeah we have also one sandbox that is customized just for like uh to have graphical interface and for the LM to to to use computer so I recommend trying that that's like the newest thing we did that that's uh pretty exciting I think awesome yeah that sounds great uh thanks a lot for uh joining us todaya and uh yeah good luck with the U with you know with the e2b future yeah thank you thanks a lot","transcript_synced_at":"2026-11-03T05:00:15.622Z","segments":[{"start_seconds":5,"text":"hi everyone uh today uh we have a new episode of office hours and with me is","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=5s"},{"start_seconds":12,"text":"Tera from e2b and we're going to talk about infrastructure to run agents SEC uh to run I guess code generated by","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=12s"},{"start_seconds":21,"text":"agents securely right yeah running AI generated code securely y okay so Teresa","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=21s"},{"start_seconds":28,"text":"do you want to introduce yourself and tell a little bit about it Tob yeah uh first of all thank you for invitation I think uh we have met first","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=28s"},{"start_seconds":37,"text":"time at uh AI Event in San Francisco and I really appreciate that because I was uh watching other workshops from neari","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=37s"},{"start_seconds":46,"text":"so thanks for the possibility yeah I'm I'm deel at e2b uh e2b is a startup we","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=46s"},{"start_seconds":54,"text":"are operating from San Francisco I also just opened a check office and we are building uh infrastructure for AI agents","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=54s"},{"start_seconds":63,"text":"uh so I'm always saying we are not AI company we are more like low-level infrastructure boring company uh and","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=63s"},{"start_seconds":72,"text":"today I want to chat about uh where AI agents still fail why is that and how good infrastructure or like good","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=72s"},{"start_seconds":80,"text":"processes in the background uh might help that because I think it's still relevant topic even through there has been like crazy louches like menus Ai","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=80s"},{"start_seconds":89,"text":"and open SDK there is still like the waves where agents are like have more hope than fail","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=89s"},{"start_seconds":96,"text":"enough again and uh it's never like still reliable enough like 100% so I want to go into what are the challenges","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=96s"},{"start_seconds":105,"text":"and how execution runtime might help uh yeah so ITB uh it's the cloud","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=105s"},{"start_seconds":114,"text":"environment uh that you can run AI generated code in I think best demonstrated it's on example uh for","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=114s"},{"start_seconds":122,"text":"example perplexity is using it to be uh let's say you give it a prompt to analyze data to do something more","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=122s"},{"start_seconds":129,"text":"complex uh perplexity has underlying llms that generate the code for that because anytime you need to do something","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=129s"},{"start_seconds":137,"text":"more cool basically the agent of course needs to write code it's like giving agent brain let's say or hands or just","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=137s"},{"start_seconds":145,"text":"like something useful and this code needs to be executed somewhere right and uh of course you could use like some","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=145s"},{"start_seconds":152,"text":"simple like grapple or something like very simple environment and just run it but you usually needs much more for example perplexity has millions of users","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=152s"},{"start_seconds":161,"text":"each needs separate instances so we are building environment where you just run this code and it has actually more more","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=161s"},{"start_seconds":168,"text":"features just uh build right for the llm use cases like that uh other use cases","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=168s"},{"start_seconds":175,"text":"are generating apps right uh like vzer or uh yeah uh so it's also like the same","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=175s"},{"start_seconds":183,"text":"principle you generate code you execute it in e2b sendbox and you have some something like real output like whole","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=183s"},{"start_seconds":189,"text":"apps uh other use cases are like uh running tests and also computer use so","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=189s"},{"start_seconds":197,"text":"this is just the environment we built e2b sandboxes this is the overview of of the use cases and how they","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=197s"},{"start_seconds":204,"text":"look and uh yeah I want to take a step back like why we started building e2b it","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=204s"},{"start_seconds":210,"text":"was in 2023 actually when agents first started it was I think AO GPT launched","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=210s"},{"start_seconds":217,"text":"right and everyone had this like wow moment that like agents are going to be everywhere and do everything and boole","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=217s"},{"start_seconds":225,"text":"trips and do all like your tasks and replace replace everyone so that's when we started building e2b and we knew that","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=225s"},{"start_seconds":233,"text":"agents are going to need a lot of uh different text T right because it changed is the way that software is","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=233s"},{"start_seconds":241,"text":"built it changes a lot of things uh because now suddenly everything will be built for LMS right people now are","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=241s"},{"start_seconds":249,"text":"putting their like documentation into markdown and creating all things like easy for LMS to read Because agents","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=249s"},{"start_seconds":256,"text":"already are browsing internet and uh they need also better security of course because by definition what uh what LMS","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=256s"},{"start_seconds":265,"text":"generates uh is stochastic uh the code is untrusted you don't know what what is going to happen What data it needs to","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=265s"},{"start_seconds":274,"text":"access so now you need to suddenly think about a lot of different challenges uh this is like the diagram from Andre","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=274s"},{"start_seconds":281,"text":"carpa I think everyone has seen that and they made a parallel to like LM being a new Kel and basically everything built","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=281s"},{"start_seconds":289,"text":"around that uh yeah and that's also why we decided to build something that's uh environment or playground that's very","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=289s"},{"start_seconds":297,"text":"special for the llm use case so yeah if you need to if you want to do any questions or jump in just I'm saying","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=297s"},{"start_seconds":305,"text":"feel free to jump in yeah so I guess the the first time I was thinking about these environments was when I was playing with Devon back in the day right","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=305s"},{"start_seconds":314,"text":"and with Devon uh obviously for D that is the primary use case right because it is it is a product that generates code","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=314s"},{"start_seconds":322,"text":"and it it does execute code somewhere uh presumably on some sort of version of it Tob","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=322s"},{"start_seconds":329,"text":"uh you know uh like something similar to it Tob right but when when they want to when they want to expose the code to the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=329s"},{"start_seconds":337,"text":"user like you know they want to host it so that the user can interact they host it on netlify and I was thinking how","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=337s"},{"start_seconds":344,"text":"um effectively one of the problems that exist in products like that is that a lot of people like I guess in case of","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=344s"},{"start_seconds":352,"text":"perplexity right the code is generated it's executed it's used it shuts down right but with the human in the loop especially the way you say it on the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=352s"},{"start_seconds":359,"text":"slide it's uh it could be that the code is generated it's hosted but then it seems to stay there for a while I wonder","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=359s"},{"start_seconds":367,"text":"if you guys address that particular use case where the you know where the container needs to exist for a while uh you know whether it's cheap to keep it","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=367s"},{"start_seconds":374,"text":"live or is it is the use case more like you know the code is written code is executed container is immediately shut","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=374s"},{"start_seconds":381,"text":"down oh yeah that's that's a good question because yeah you right that's exactly one of the very specific things","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=381s"},{"start_seconds":387,"text":"for agents and why they fail like they sometimes the process needs to run right and sometimes it runs into error because","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=387s"},{"start_seconds":394,"text":"it shuts down and uh ourand boxes are actually persistent so you can post the sendbox and you can resume it later from","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=394s"},{"start_seconds":403,"text":"the state where it was when you post it uh and you can get back to the sybox file system and memory all the processes","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=403s"},{"start_seconds":411,"text":"uh yeah and in the environment you can also reference like the previous code so that's possible with the sendbox if if that's what what you asking yeah yeah","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=411s"},{"start_seconds":420,"text":"yeah so so effectively um so they can be yeah so the snapshotting is supported right so you can snapshot it put it into","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=420s"},{"start_seconds":428,"text":"the cheap I guess cheap storage wait until the user is ready to talk upon it again yeah that's that that's pretty","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=428s"},{"start_seconds":435,"text":"cool what um I guess um how so so From perspective of we","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=435s"},{"start_seconds":443,"text":"develop right let's say I'm using e2b so my interface is effectively similar to somewhat like some something like AWS right I'm saying Hey I want I want a","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=443s"},{"start_seconds":451,"text":"container it gives me like what SSH access or how do I how do I interact with the with the container so you you interact with the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=451s"},{"start_seconds":459,"text":"sandboxes we buil through our SDK uh you can use it in python or or JavaScript and yeah and you control you start the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=459s"},{"start_seconds":467,"text":"sendbox with the SDK and you you control basically everything you can also self host it now it's supported on gcp and","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=467s"},{"start_seconds":476,"text":"working on support for AWS but yeah the SDK basically controls it all you can customize the sbox before and yeah think","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=476s"},{"start_seconds":486,"text":"yeah pretty cool and should I go to the next slide yeah yeah let's do it yeah okay uh yeah we","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=486s"},{"start_seconds":494,"text":"wanted to name this uh where agents fail today because I think uh you see a lot on Twitter like discussions every time","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=494s"},{"start_seconds":502,"text":"there is a new Lounge of Agents someone always says it it fails for them or it's slow or it's going to be expensive so I","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=502s"},{"start_seconds":510,"text":"think the consensus is of course agents are cool but yeah it's still a long way to go right even if it's like really","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=510s"},{"start_seconds":518,"text":"exciting uh I think what doesn't work yet that well in production is the Devin like types of software engineers and I","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=518s"},{"start_seconds":527,"text":"think you might also disagree of course like I haven't tested all but like agree yeah like very autonomous agents right","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=527s"},{"start_seconds":535,"text":"that uh that needs to do like all sequence of tasks one after another that that's like big challenge because there's like more probability to fail of","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=535s"},{"start_seconds":544,"text":"course as you said like it needs to run for a long time keep the context so yeah I think there is like a tradeoff between","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=544s"},{"start_seconds":552,"text":"how reliable the agent is versus how much freedom and autonom it has or and of course like the restrictions there is","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=552s"},{"start_seconds":560,"text":"another tra of like how much you want the agent to do how how much like access to the system you give it versus how","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=560s"},{"start_seconds":568,"text":"secure it actually and I think I see this for big companies especially they are worried how like you know if they","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=568s"},{"start_seconds":575,"text":"can let the like agent for example use computer or even browser so I think there is still a lot to to solve uh and","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=575s"},{"start_seconds":584,"text":"the common challenges we see or hear about from developers so of course hallucinations even through I I feel","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=584s"},{"start_seconds":592,"text":"like the LMS are improving a lot I mean uh yeah like the same same prompts are working better now so of course this is","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=592s"},{"start_seconds":600,"text":"uh like a lot of this is going to be fixed by just better models I think uh but of course it's I feel like agent is","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=600s"},{"start_seconds":608,"text":"the same as like having an intern or like your assistant it needs to like understand the context really well it","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=608s"},{"start_seconds":614,"text":"needs to see the nuances I think this is especially how it differs from humans right because it needs to really catch","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=614s"},{"start_seconds":621,"text":"catch the vibe what you want what exactly you want it can like it has to like read read between the lines so this","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=621s"},{"start_seconds":628,"text":"is still still quite difficult like how to iterate with the humans I think uh multi-agent systems are going to solve","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=628s"},{"start_seconds":637,"text":"this partially I've seen use cases where like they can of course give feedback to each other in production but still it's","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=637s"},{"start_seconds":644,"text":"a challenge and I think another thing is evals that is still like needs to improve and then of course uh like the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=644s"},{"start_seconds":651,"text":"persistence uh the agents to be fast to be scaled for millions of users and yeah to be secure so that's if you want to","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=651s"},{"start_seconds":660,"text":"add something from your experience like uh from using agents I think it's always interesting to","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=660s"},{"start_seconds":666,"text":"hear no I think with the autonomous development I think for uh for narrow use cases it's probably uh I I have I","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=666s"},{"start_seconds":675,"text":"haven't played with them uh for a couple in a couple months now but I I think last time I tried for narrow use cases they were relatively well but if you try","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=675s"},{"start_seconds":683,"text":"to use a broad one and ask something that is not quite common you know like try to even a simple lean proof or or something","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=683s"},{"start_seconds":692,"text":"in a language that is not as common they uh they hallucinate quite badly still yeah I think it's funny when people try","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=692s"},{"start_seconds":700,"text":"to test like new new launches new agents they always try to find that one use case that doesn't work and I feel like when you try to test it you sometimes","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=700s"},{"start_seconds":709,"text":"can't think of something good but actually then use it in production and you have like most of the work is very very Niche right what people need so","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=709s"},{"start_seconds":717,"text":"yeah I think yeah that might not even be the case right like I presume that 90% of what people do is you know JavaScript and Python and that and that probably","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=717s"},{"start_seconds":726,"text":"just works pretty well at this point yeah I guess but we have use cases like uh we have use cases like running tests","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=726s"},{"start_seconds":735,"text":"in in different languages and for example we have one user that they are building product for scientists and scientists uh write code in r a lot","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=735s"},{"start_seconds":744,"text":"right R is very like specific for like I don't know life science and stuff and they need to so they run this code from","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=744s"},{"start_seconds":751,"text":"R in E Tob and it's very it's very Niche use case because also different types of users right because I feel like who is","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=751s"},{"start_seconds":759,"text":"mostly experimenting with agents now are software Engineers or like tech people but even scientists or like PhD people","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=759s"},{"start_seconds":767,"text":"it's not like the same as Engineers they don't code that well probably maybe not a experience so it's already like","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=767s"},{"start_seconds":774,"text":"different users so I think that's also challenging yeah yeah I guess there would be quite a","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=774s"},{"start_seconds":782,"text":"bit of R in the internet right because even like I guess before before machine learning became before like Python and","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=782s"},{"start_seconds":791,"text":"naai became so dominant R was pretty standard for a lot of I I don't know like how was the like development I feel like now yeah python","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=791s"},{"start_seconds":799,"text":"is popular type script of course yeah okay uh yeah then how to make agents","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=799s"},{"start_seconds":806,"text":"more reliable uh of course of course one way is to make llms better uh but there are also ways you can improve agents and","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=806s"},{"start_seconds":816,"text":"one thing is making them actually like do the things instead of just talking right uh I think at the beginning most","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=816s"},{"start_seconds":824,"text":"agents were just producing texts it was just like chatbots and then they started using more tools like code interpreter","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=824s"},{"start_seconds":831,"text":"of course and producing actual uh results like now for example with deep research they can generate the whole","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=831s"},{"start_seconds":838,"text":"like Report with charts and this is also good when they produce something actionable when they run the code to","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=838s"},{"start_seconds":845,"text":"give you something because then they can iterate on what they produced uh so this also adds the reliability and feedback","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=845s"},{"start_seconds":854,"text":"loop uh security yeah that's what I talked about with e2b so when you have app for end users and if you want to","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=854s"},{"start_seconds":863,"text":"have more than a few users you need to scale it somehow and you need to like is isolate um the instances because you","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=863s"},{"start_seconds":871,"text":"don't want the code produced for every user interact with each other uh obviously but like it's not that easy to","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=871s"},{"start_seconds":879,"text":"make that many instances at scale and still it to be fast and and isolated and then I would say a lot of it is about UI","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=879s"},{"start_seconds":888,"text":"ux because a lot of people had similar idea at the beginning uh building like the same types of Agents like data","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=888s"},{"start_seconds":895,"text":"analysis right but for example with data analysis you can see like how Julius AI made like great product and distinguished from the others even if","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=895s"},{"start_seconds":904,"text":"it's like b2c so it's challenging right but uh you can see that the UI makes big difference especially I think because","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=904s"},{"start_seconds":912,"text":"it's AI so it's good to have like transparency and good good monitoring what's happening how the user should","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=912s"},{"start_seconds":920,"text":"interact yeah and this is the these are some of the problems that e2b solves with the uh Cloud runtime uh so we can","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=920s"},{"start_seconds":929,"text":"run code uh in like secure isolated manner in the cloud in the sendbox uh so again when the code is generated uh the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=929s"},{"start_seconds":937,"text":"sendbox is started and there all the processes are running you can install packages in the sendbox uh you can","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=937s"},{"start_seconds":944,"text":"access Internet so you can use the sandbox even for like computer use agents uh you can download files from","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=944s"},{"start_seconds":952,"text":"that everything uh like this and yeah and it's good that you can just give it to the agent and that's why I said we","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=952s"},{"start_seconds":960,"text":"are not AI company because it's like separate uh from how the agent is built because we don't care what llm it used","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=960s"},{"start_seconds":967,"text":"right we don't care if it's built with L graph or crew AI or what Tex Tech uses we are just like the environment the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=967s"},{"start_seconds":975,"text":"infr layer because I I feel like that's that's one thing that developers scare a lot from what we hear is not being loged","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=975s"},{"start_seconds":983,"text":"in for specific Solution that's like opinionated right they always ask they want like low level something that fits","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=983s"},{"start_seconds":991,"text":"their text tag they don't want to like change uh change anything so that's that's good that you can like connect it","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=991s"},{"start_seconds":998,"text":"and be very agnostic I think in general that's good for like llm software because like people already are building","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=998s"},{"start_seconds":1006,"text":"like with multiple llms they combine it all they just don't choose their favorite right it's it's like difficult to be just like big Winners so I think","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1006s"},{"start_seconds":1014,"text":"it's even more important to be just like agnostic uh","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1014s"},{"start_seconds":1021,"text":"okay yeah uh execution run time that's uh what I said basically so here are some of the like uh properties like uh","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1021s"},{"start_seconds":1031,"text":"features that you need to have for llms so you need to yeah start the sandbox quickly and uh our sandbox is built with","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1031s"},{"start_seconds":1039,"text":"firecrackers by AWS uh it's type of microv VM you have like multiple micro VMS but that's just firecracker is very","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1039s"},{"start_seconds":1047,"text":"fast and it's battle tested so we can start it in like I think under under 200 milliseconds uh yeah you can connect it","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1047s"},{"start_seconds":1056,"text":"to any LMS you have file system there you can have custom packages uh and yeah you can have","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1056s"},{"start_seconds":1063,"text":"persistance and all these features that are like important and you can also sell host so that's how I would summarize","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1063s"},{"start_seconds":1071,"text":"probably why the environment is good uh then we have like the second part uh uh what even are the types of","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1071s"},{"start_seconds":1081,"text":"like run times uh what you can even do if you just want to start let's say from scratch and uh run your code just simply","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1081s"},{"start_seconds":1089,"text":"and then I think we see why microv VMS like firec that we are using uh is that","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1089s"},{"start_seconds":1095,"text":"good uh so you can have like very simple uh simple security like just process","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1095s"},{"start_seconds":1102,"text":"level uh but of course that's you probably wouldn't use it in production for this because it's just like uh","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1102s"},{"start_seconds":1108,"text":"limited security then you have uh containers like Docker and uh people usually ask a lot uh what's the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1108s"},{"start_seconds":1117,"text":"difference between running code in Docker and in and in Tob that's like uh","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1117s"},{"start_seconds":1123,"text":"one of the most common questions uh yeah that's I get this at always at like almost every event uh and I think I","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1123s"},{"start_seconds":1133,"text":"think there were agentic apps that were using Docker for running code for example gen at least at the beginning","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1133s"},{"start_seconds":1140,"text":"not sure if now but they were using Docker to to execute code AS code interpreter and uh yeah so I want to","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1140s"},{"start_seconds":1148,"text":"address this like Docker is is fine it also can be good as a runtime but uh it","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1148s"},{"start_seconds":1155,"text":"has like limited security because uh the kernel userland interface is very large compared to to VMS uh where the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1155s"},{"start_seconds":1164,"text":"interface is much smaller and uh yeah so it it has better security and yeah and some solutions combine both","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1164s"},{"start_seconds":1171,"text":"but of course uh when you run it locally also it's it's difficult to scale them so that's like limitation for Docker uh","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1171s"},{"start_seconds":1180,"text":"but yeah of course it doesn't mean that it's not possible or something Docker takes more than 200 milliseconds to spin","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1180s"},{"start_seconds":1188,"text":"up presumably I'm saying Docker is taking more than 200 milliseconds to to spin up yeah that I am not sure to be","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1188s"},{"start_seconds":1195,"text":"honest I can check after this but yeah probably I mean firecracker is known to be very fast so I guess yeah I get it's","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1195s"},{"start_seconds":1204,"text":"not even that fast especially at scale we use Docker today I mean we we run a we have a chat. near. where you can you","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1204s"},{"start_seconds":1213,"text":"know you you can execute code and we're running it in Docker and we spend quite quite quite a bit of time to effectively","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1213s"},{"start_seconds":1221,"text":"read up on security of Docker containers how to harness them and I think that potentially firecracker is the better","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1221s"},{"start_seconds":1229,"text":"alter use it for I guess firecracker is is potentially a better alternative than","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1229s"},{"start_seconds":1236,"text":"Docker the on your slide it says tradeoffs security so so does it say uh","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1236s"},{"start_seconds":1244,"text":"is the claim that firecracker has better isolation than Docker or Worse isolation yeah I think it has better isolation I","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1244s"},{"start_seconds":1252,"text":"think like I need to find the slide yeah but you can see here basically basically like the comparison","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1252s"},{"start_seconds":1259,"text":"of the container versus uh virtual machine I think the point is how much you expose like of the kernel and how","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1259s"},{"start_seconds":1268,"text":"you isolate uh your processes what is happening from the kernel right because you have containers like Docker and then","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1268s"},{"start_seconds":1275,"text":"you have virtualization like microv VM and there you have like extra you are basically protected from the host kernel","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1275s"},{"start_seconds":1285,"text":"right I think it's like separated can find and the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1285s"},{"start_seconds":1292,"text":"slide um okay I think it's not here but another question","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1292s"},{"start_seconds":1301,"text":"is sorry so with Docker right I can um as a user I can spin up very customized","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1301s"},{"start_seconds":1310,"text":"environments right it could be any operating system with any version of you know liaries and what not so with with","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1310s"},{"start_seconds":1317,"text":"e2b uh is MI the user configure the environment or is it something that is standardized across all","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1317s"},{"start_seconds":1326,"text":"the you know like like OS environment yeah like operating system libraries you know pre-installed","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1326s"},{"start_seconds":1335,"text":"software yeah so I think it's two different things like uh operating system uh e2b is using Linux currently","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1335s"},{"start_seconds":1344,"text":"and that's the only one I think developers haven't asked that much for for others what but we made add uh other","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1344s"},{"start_seconds":1351,"text":"than Linux uh but now it you can run dbn distribution so so that Linux inside the microv VMS uh that's one thing uh with","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1351s"},{"start_seconds":1360,"text":"the custom packages as you said yeah so we can we have custom sandboxes uh and you can pre-install packages there so","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1360s"},{"start_seconds":1369,"text":"you can customize the sandbox and I think a lot of users not sure what percentage but a lot of people are doing that like you can customize uh you can","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1369s"},{"start_seconds":1378,"text":"create basically a Sandbox template uh you generate it and there you can like pre-install what you need what special packages to","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1378s"},{"start_seconds":1387,"text":"install if that's then even for a custom uh stand box it's still 200 milliseconds right so so I configure it","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1387s"},{"start_seconds":1395,"text":"once in advance and then I can spin up new ones okay yeah that that sounds pretty cool I think it should be possible yeah and you can also customize","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1395s"},{"start_seconds":1404,"text":"like CPU and RAM of the sbox template uh and you use it to be CLI to do that yeah","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1404s"},{"start_seconds":1412,"text":"also when I was asking about operating system I didn't mean like Linux versus Windows I meant more like uh like a you","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1412s"},{"start_seconds":1419,"text":"know I I presume that you know that like uh your actual Hardware is running on Dean or something but let's say I want","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1419s"},{"start_seconds":1425,"text":"Arch Linux or or or gento is it something I can do or does it have to be the same operating system as the host","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1425s"},{"start_seconds":1434,"text":"sorry I I didn't catch what what you said as example so like can it be a flavor of Linux which is","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1434s"},{"start_seconds":1442,"text":"different from Linux running on the H like can I run Arch Linux in my sandboxs yeah that's what I mean uh only","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1442s"},{"start_seconds":1450,"text":"Linux is possible now I'm not 100% sure to be honest but yeah I think you only need you only run Linux right now Linux","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1450s"},{"start_seconds":1459,"text":"OS but yeah we are definitely like discussing that like uh with users and if if there's like need for other I","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1459s"},{"start_seconds":1466,"text":"think there hasn't been that much for this like specific use case with agents but we we might add other probably but that's good question I will","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1466s"},{"start_seconds":1475,"text":"find out more yeah yeah we so so the the problem there is that there are like I think it's","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1475s"},{"start_seconds":1483,"text":"relatively it's relatively common these days to use auntu right but for example in AWS their standard operating system is their flavor of Linux which is called","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1483s"},{"start_seconds":1492,"text":"Amazon Linux I guess and it uses different package manager it uses you know some of the system util is a different right so if someone has","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1492s"},{"start_seconds":1499,"text":"dependencies of those right and then you know if a Sandbox can only be B to with apt that's not something they can use","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1499s"},{"start_seconds":1508,"text":"right so I guess with Docker I'm pretty sure you can run any Linux uh no matter what host OS is is","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1508s"},{"start_seconds":1515,"text":"there I think you should be still able to like uh have any like dependencies you need but like that's very good","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1515s"},{"start_seconds":1521,"text":"question so yeah need to find out uh but like you mean your use case would be like someone using already like other","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1521s"},{"start_seconds":1528,"text":"ecosystem right and needing some like like using Arch Linux you know Arch Linux is sufficiently different from from","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1528s"},{"start_seconds":1536,"text":"Dean yeah yeah not sure okay uh yeah I think maybe you can","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1536s"},{"start_seconds":1543,"text":"specify it in your in thebook templ but need to check uh I think I can share after the presentation","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1543s"},{"start_seconds":1552,"text":"great okay uh I think yeah let's also summarize so we","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1552s"},{"start_seconds":1559,"text":"have also VMS and micro VMS and firecracker is like micro VM and it's like uh specifically fast uh then we","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1559s"},{"start_seconds":1567,"text":"have like other types uh of microv VM we have like cross VM uh Cloud hypervisor","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1567s"},{"start_seconds":1574,"text":"which is like more general purpose uh and has GPU support so there's like other options than firecracker but again","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1574s"},{"start_seconds":1581,"text":"we we uh picked firecracker because it's uh it's very fast it's battle tested and it's built with uh AWS Lambda uh but you","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1581s"},{"start_seconds":1590,"text":"can also use like other options so yeah we're also exploring that to build but I think firecracker so far has been Lambda","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1590s"},{"start_seconds":1599,"text":"Lambda is built on firecracker I see yeah yeah is it is very battle","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1599s"},{"start_seconds":1606,"text":"tested yeah but the question is that's actually a good question right so why like let's say I'm perplexity I need to","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1606s"},{"start_seconds":1614,"text":"be running the code why would I use e2b as opposed to Lambda","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1614s"},{"start_seconds":1620,"text":"um yeah I think uh because like Lambda is usually how how people start uh or","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1620s"},{"start_seconds":1628,"text":"when we have like new customer they usually start building something in-house uh they usually actually very often they build with Lambda functions","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1628s"},{"start_seconds":1636,"text":"uh that's like the probably most common way if they want to build inhouse and of course they they can do it uh they can","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1636s"},{"start_seconds":1643,"text":"do it like themselves but the problem is like it turns out it it gets like very very complex problem with a lot of in","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1643s"},{"start_seconds":1652,"text":"infrastructure to be built uh and you you start to have like a lot of features that you need to add on top of that","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1652s"},{"start_seconds":1658,"text":"because as far as I know Lambda also is not like specifically made for agents so you will need like uh some level of","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1658s"},{"start_seconds":1668,"text":"Maintenance customization and all these like challenges that I mentioned and you will need to build it yourself so what","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1668s"},{"start_seconds":1675,"text":"people most come like tell us is they they started to build build this because they didn't know something actually","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1675s"},{"start_seconds":1682,"text":"exists and it didn't exist also long uh some time ago it wasn't like that big market so yeah you can start it like","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1682s"},{"start_seconds":1691,"text":"yourself but then probably like how annoying it is to build and uh yeah that that's probably the challenge because I","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1691s"},{"start_seconds":1699,"text":"hear usually like the companies don't have like infra team for that or don't want to just focus on making the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1699s"},{"start_seconds":1706,"text":"infrastructure great and I see it that like we we are constantly like working on something fixing bugs right and I I'm","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1706s"},{"start_seconds":1715,"text":"not like infra engineer but it's probably very difficult to do inhouse so that's what you mentioned so I guess one","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1715s"},{"start_seconds":1724,"text":"of the concerns uh with running something in the cloud is always uh data privacy right you mentioned it would be","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1724s"},{"start_seconds":1732,"text":"locally right so so is it open source or how does it work uh well yeah like we can self host","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1732s"},{"start_seconds":1739,"text":"and uh we we run into Cloud usually not locally so not sure what uh what you ask like about data so self host would imply","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1739s"},{"start_seconds":1748,"text":"that it's running locally on my end right or what is self-hosting yeah yeah that's that's self hosting right but like yeah I'm just saying like","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1748s"},{"start_seconds":1755,"text":"not only we don't do only self hosting but a lot of people still run it in the cloud so yeah of course but but for self-hosting so is it is it to be open","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1755s"},{"start_seconds":1764,"text":"source or how do I self host oh yeah uh I I think I didn't say but we are completely open source uh including the","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1764s"},{"start_seconds":1773,"text":"infrastructure so you can go to our GitHub e2b Dev and for self hosting we have like a complete guide there if you","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1773s"},{"start_seconds":1781,"text":"go on our GitHub and find infra repo uh if F guide for self hosting e2b on on Google Cloud so you will need probably","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1781s"},{"start_seconds":1789,"text":"terraform and uh yeah Google Cloud CLI and other things and you can do it just there but yeah everything all the code","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1789s"},{"start_seconds":1797,"text":"is accessible yeah awesome yeah that sounds great okay we are uh we're getting close","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1797s"},{"start_seconds":1805,"text":"to the end of the um of this uh session so is there","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1805s"},{"start_seconds":1811,"text":"anything um particular uh any like last U comment what you're excited about in the future what's uh what's the most","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1811s"},{"start_seconds":1820,"text":"exciting in the near I'm thinking yeah I wanted to share because yesterday like yeah I think yesterday","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1820s"},{"start_seconds":1829,"text":"was the openi launched and they launched some kind of uh computer use agent and we often built build stuff we build like","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1829s"},{"start_seconds":1837,"text":"open source let's say side projects to to show how it Tob works so we just launched uh computer use agent with a","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1837s"},{"start_seconds":1846,"text":"nice UI that's using the latest open AI model and you can try it for free it's also on our GitHub so we have like a","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1846s"},{"start_seconds":1854,"text":"bunch of apps like that we have also like app generation app or data analyses and all these use cases I said at the beginning so if you go to our giab and","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1854s"},{"start_seconds":1863,"text":"find open computer use uh yeah we have also one sandbox that is customized just for like uh to have graphical interface","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1863s"},{"start_seconds":1871,"text":"and for the LM to to to use computer so I recommend trying that that's like the newest thing we did that that's uh","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1871s"},{"start_seconds":1879,"text":"pretty exciting I think awesome yeah that sounds great uh thanks a lot for uh joining us todaya and uh yeah good luck","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1879s"},{"start_seconds":1888,"text":"with the U with you know with the e2b future yeah thank you thanks a lot","url":"https://www.youtube.com/watch?v=FCWje0ge8iE&t=1888s"}],"youtube_url":"https://youtube.com/watch?v=FCWje0ge8iE","url":"https://near-pedia.xyz/videos/FCWje0ge8iE"}