Dense Slides: Arrakis: How To Build An AI Sandbox From Scratch - Abhishek Bhardwaj, OpenAI
Source Video
Arrakis: How To Build An AI Sandbox From Scratch - Abhishek Bhardwaj, OpenAI
Method
This deck is slide-only. The existing captured video frame set supplies candidate frames, then local OpenCV rejects sponsor/title/speaker-only frames, crops visible slide surfaces, deduplicates, and saves the cropped slide images.
Cropped Visible Slides

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.97 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Dense multi-column background/timeline slide with small bullets and labels.
Slide text:
Background
Carnegie Mellon
Operating Systems x Distributed Systems.-.
8 Microsoft
Windows Subsystem for Linux
Google - G From scratch Os for pre-Apple Watch Smartwatch:
:.: MlcroVMs - Rust-based ightweight VMs for Linux dev
environments on Chrome Os.
ARQ- Android apps on Chrome Os In containers: Replit
Infra - Proxied through my servkce for your entire session
Arrakis.. RAG based codo chat: whenever you code on Replit:
Sandboxes for Al Agents - code execution and computer use
8Built during my time at South Park Commons

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.99 - Text source: agent_vision.
Slide text:
Why Do We Need AI Sandboxes?
Smarter Models
Ability to call tools during inference
Used in RL training
Smarter Agents
Can debug actions by running Linux commands "ps", "lsof"
Backtrack, replan and work towards the goal with a Sandbox at their disposal
Security
Agent code is no different that Github or StackOverflow code running on your host or production server

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
demo_videoconfidence0.95 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Dense product/UI screenshot with small interface text and terminal output.
Slide text:
. How to Create a ChatGPT Clone. Manus's Computer '
Nuple Guts 1 truty
Executng comrind kdlr' -p thutgot ctare t cd chatgot_clone t& create_tlsik.+p-
Create a ChatGPT clone sheni.
: interface that mimics the functionaity of ChatGPT. Let me get started right: &manus: 1'll hclp you croate a ChatGpT clono. Ill work on doveloping a wob-based chat: _clone &s create_rlask_app. USAGE: ubuntugsandbox: S'cd /hone/ubuntu &s akdir -p chatgpt clone &s cd chatgpt. Creates a nev Flask application' create_flask app fapp_nase]
&manus AeMe Oi Select the most suitable web spplicauon template for a ChatGpT clone Connocted to datasource(1): ubuntuesondbox:~/chatgpt_clone. $. ARGUHENTS: EXAMPLEi. butu'dde create_flosk app ay_app Napc of. the Ftask application to 'create
1'1l be creating a ChatGPT clone using Flask, vhich is the appropriate choice
since we noed both s frontend Intcrface and backend processing for the chat: functionality. Let me set up the project structure and lmpiement the necessary.
. components.
) Creite Ue ChatGPT clone project using tha Flask terplatoA
: Creating the ChaiGPT clone prol+ct using tha Fusk templto.
Creating thie ChatGPT clone Flask project.
0 Using termnat:

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: agent_vision.
Slide text:
Arrakis
Arrakis provides a secure, fully customizable, and self-hosted solution to spawn and manage Sandboxes for code execution and computer use.
It has out-of-the-box support for backtracking via snapshot-and-restore.
Open Source

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.96 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Slide contains small lower panels and dense explanatory text that is OCR-suitable.
Slide text:
D
Backtracking
. Pre-reasoning models - Intuition that a higher order task needs to be broken down by a Planner:
Agent should be able to snapshot progress and restart from last good snapshot versus starting from scratch
.Simple AP!
Python Api Golng Clt Micp Sarver opearbpi compatbio -+ gonerate cionts in sny Languogo
Configurable With Docker tooling'
Dockerfile to configure what runs nside tha Sanabox:

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Architecture diagram with many small labels and connectors that are best handled by OCR.
Slide text:
arrakis-py-sdk _. REST API Linux Server
Arrakis Restserver
arrakis-client REST API:
vWc client Port Forrar. Auto Arrakis MicroVM DA: Sandbox Server Coda Server ws +poy.. Arrakis MicroVM Sandbox YAC' trver ort Foru Auto
Userspace
Kerne! Idxv/kva" /dev/kva
High level architecture of Arrakis

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: advanced OCR
rapidocr-live/bright-screen/contrast. - OCR decision: ready — Dense code-heavy API slide with small endpoint text and code block; OCR is likely better than manual transcription.
Slide text:
Arrakis Restserver exposes the following REST APl. This is also available via a Python SDK, MCP Server and an openapi compliant YAML file
API-Sandboxes,Snapshots,Command Execution,Files
r.HandleFunc("/"+API VERSION+"/vms",s.startVM).Methods("POST")
r.HandleFunc("/"+API_VERSION+"/vms/{name}",s.updateVMState).Methods("PATCH")
r.HandleFunc("/"+API VERSION+"/vms/{name}",s.destroyVM).Methods("DELETE")
r.HandleFunc("/"+APIVERSION+"/vms",S.destroyALlVMs).Methods("DELETE")
r.HandleFunc("/"+API VERSION+"/vms",s.listAlLVMs).Methods("GET")
r.HandleFunc("/"+API_VERSION+"/vms/(name}",s.listVM).Methods("GET")
HandleFunc("/"+API_VERSION+"/vms/(name}/snapshots",s.snapshotVM).Methods("POST")
ndleFunc("/"+API_VERSION+"/vms/{name}/cmd",s.vmCommand).Methods("POST")
dleFunc("/"+API VERSION+"/vms/{name}/files",s.vmFileUpload).Methods("POST")
dleFunc("/"+API_VERSION+"/vms/{name}/files",s.vmFileDownload).Methods("GET")
andleFunc("/"+API VERSION+"/health",s.healthCheck).Methods("GET")

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.99 - Text source: agent_vision.
Slide text:
Arrakis - API
arrakis-py-sdk
arrakis-client
Arrakis Restserver

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Dense multi-section slide with small text; OCR should capture headings and body more reliably than manual reading.
Slide text:
Arrakis: A Quick Tour Of Linux Sandboxing
Linux Execution Modei
A Thread ls the smabest unn of execution. Each Thead has a task struct that's n the Scheduier run queue. Process" conssts of matiple thyeads, Treads" In a process have differeni T..
Why? - My app necos Python38, foo b version 12, your server has Python 3.9, 36, 312 and doosn't have foo Rd. Packago sn opp's dependancios along wth ks coro business logic Run, arbarsr.. Containers: Why? And What?: 全
Containers: Security Story
After al Containers run as native processos on iop of tha kernel Any malcious or buggy procest can attack the kemel and gain root Once t becomes root it can do'sriythirg d want.
Containers: Sandboxing
:JaT containers by rostricting CAPabitos and syscias they can cal Reduce attack surface Con rostrict thg anack surtace by restricting the Unux CAPabltios atowed to a process such as..
Containers Virtualization.
VintusHzation provdes another primaivo to run arbitrary code on your mochlnos Esch Vu has rts own guest user space and guest kerme, Doesn't drocty stsck the host kern
Kvm - Hypevisor In the form of a kernet mnodule that leverages VT-x to run swich CPU context and run VMs Vmu - Taks to kvm to spawn VMs and aso manaes emtuated Linux Virtualization

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.97 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Diagram plus small bullet text and code snippet; OCR is suitable for the detailed labels.
Slide text:
Process 1 Process.2 08x0 4u! xor. ebx, ebx
IThread MThread
ioctl: Syscall Userspace
/dev/kvm: /dev/sda1 Kernel (os)
Execution model of a typical Linux System
: A*Trread' Is the snaBost una of erecuton. Each -Thread' has a task struet that's n thio Schodutor run quoue..
"Processe consists of mutupie thresds. -Thresds" n process huve different TDs but tame Pio, thare page tablo snd other th'ngs. Bound by shared resources Kermel provides prirtegeo' access to hardware. lf not gated, buggy or masclous code can wreak harroc.
Specisal nstructions to switeh to kernel mode / supervisor mode and lnvoke a syscal for prvteged sccess.:

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.99 - Text source: agent_vision.
Slide text:
Containers: Why? And What?

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.96 - Text source: advanced OCR
rapidocr-live/left-72/opencv-adaptive. - OCR decision: ready — Cropped continuation slide with small bullet text; OCR is appropriate for recovering the full content.
Slide text:
100
Mount Namespace Pid Pid:101 Pid 102
/foo:Userspace
Kerne!.
Linux Container with Pid and Mount Namespace
Why? - My app needs Pyhon3.8, foo tb verslon 1.2, your server has Python 3.9, 3.6, 3.12 and doesnt have foo Eb.
ckage an spp' deperdenciet blong with itt core business logic:
On Linior' a Contairer is = Coatctoni of Nomespoces of diferent resources (Process, Mourit, Net etc.). Rur arbltrry uso code on your machhne nsixe a contalner you only ie resources hn your nrmespsce snd not ounside.
Resource siocation hsppens via CGroups ("how much memory or Cpu to abocate ths namespoce??) Outide tho contther you soe ovoryhing insldo your chldren ramespacos?

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: advanced OCR
rapidocr-live/bright-screen/contrast. - OCR decision: ready — Diagram slide with small handwritten labels and nested namespace structure.
Slide text:
Container Pid Namespace Pid 1 Parent Root Namespaces
Pid 100 P!d
Pid 2 Mount Wamespace Pid 3 Pid 101 102 Pid
/foo Userspace
Kernel
Linux Container with Pid and Mount Wamespace

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.97 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Cropped text-heavy slide with bullets and diagram labels suited for OCR.
Slide text:
Mount Namespace Pid 2 Pid. "Pid 101 P!d. 102
/foo.:Userspace
Kernel.
Linux Container with Pid and Mount Namespace'
Ipp nhods Python3.8, foo Eb vorslon 1.2, your sorrer has Python 3.9, 3.6, 3.12 and dootn't havd foo ib.:
kage an app's dependercles along wth hs core business logic:
Run ar bltrary user code on your machine. On Lnunt s Container ks a Colecton of Nsmespsces of differont resources (Process, Mourit, Not otc.).
nside a contahner you only see resources n jour namespace and not ouside..
Cucide the contener you see eveythhg nslde your chldren namespuces
Rtsource alocaicn hoppens via CGroups ('how much rmemory or CPu to sbocire thhs namespace?)
.$..
Contminenk Securkty Story

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Architecture/security diagram with labels and small bullets that OCR can read more reliably than manual transcription.
Slide text:
Containers: Security Story
Root Namespace Userspace
Container
Malicious or Buggy Process Pid Namespace Pid 100 Trusted Process:Pid: 200
:'syscall ioctl Attack Surface Syscall: ioctl
Kernel:
Container processes can attack the kernel
Ater s Containers run ss nativo procosses on top ot thg kernel:
.: Ary mallclous or buggy procass can attack the kerre and galn root..

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Same security diagram with additional bullets visible at the bottom; OCR is appropriate.
Slide text:
Container Root Namespace Userspace
Malicious or Buggy Process Pid Namespace: Pid 100 Trusted Process Pid: 200
syscall ioctl Attack Surface Syscall. ioct!.
Kernel
Container processes can attack the kernel
Ater s Containers run ss nstlve processes on top of the kernel'.
Ary malclous cr buggy process can atsck the kornel and gain rool.: 、
+. Onco it becormes root k can do arything r warts..

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.96 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Agenda slide with multiple small section rows; OCR is the better extraction path.
Slide text:
Arrakis: A Quick Tour Of Linux Sandboxing:
Unax Execution Model
Containerte: Why? And What?'
AinmCotL
Containerr: Sardbadng
ora tuaai l iruao t
Contsinert - Vrtunzston
Uinn Vrtumizadon ::
K·hype
Horestit morigred by fim turta vytes Shocat te dtir by yot: Orty mo Mleroyhs vs Traditoru Vms

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.99 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Dense code-and-policy slide with small text and command snippets that OCR can capture more accurately.
Slide text:
ontainers: Sandboxing
: Joie containers by rosricting CAPsbuities and syscalls thay can cat
Roduco attock surfsco
Userspace
Root Namespace:
Malicious or Buggy Process Container # minijail0 -s /usr/share/minijail0/$(uname -m)/cat.policy -\ /bin/cat /proc/self/seccomp_filter
Pid Namespace Syscall ioctl mmap2: arg2. =a 0x0 munmap: 1 8h07 == 1640 11 9182 == 1640 11 0 == 16xo 11 89172 == 1640 u2do close: 1 # open with O LARGEFILEjO_RDONLYlO_NONBLOcK or some combination.
Reduce attack surface by using seccomp filters'
Con resirfct the stuock surface by restrkcting tho Lirunc CAPsbiltos sowed to a procos such as CAP_NET, CAP.SYS_ ADMIN etc.

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.96 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Dense diagram labels and small policy text.
Slide text:
Userspace
Root Namespace
Malicious'or Buggy Process Container # minijailo -S /usr/share/minijailo/$(uname -m)/cat.policy --\ /bin/cat /eroc/self/seccomp_filter
Pid Namespace: Syscall. P!d ioctl. mmap2:arg2 == 0x0 open: arg1 == 32768 Il arg1 == 0 Il arg1 == 34816 Il arg1 == 2048 L:dowunm # open with O_LARGEFILEIO_RDONLYlO_NOvBLOck or some combination. close: 1
:Kernel.
Reduce attack surface by using seccomp filters
Can restrlct thie attack surface by restrictlng the Lirux CAPabaltes atowed to a process such ss CAP_NET, CAP_SYS_ADMN etc.
: These controt whst syscots a process cani call:: Slmlarty vre can uso seccomp frers to rostict what syscats a procoss can coll
'Minleil' developed by ChrcmeOs provdes an easy lrerface lo do boh.

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: agent_vision.
- OCR decision: ready — Small bullets and multiple diagram labels are OCR-suitable.
Slide text:
Containers → Virtualization

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Architecture diagram with small labels and notes.
Slide text:
Virtual Machine P2 Host Userspace Virtual Machine Monitor (vMA) (qenu / crosvm / firecracker) Create VM Host: Some chent On:
Guest Userspace" P3. running.: the vM Thread:
virtio
net driver. block driver Block Device Net Device'
Guest Kernel. VMRESUAE VMEXIT
Hypervisor /dev/kvn Host Kernel
Linux Virtualization: Key Components.
:: Kym - Hypervisor in the form of a keme module that haveraget VT-x to run tirach CPU contoxt and run VM!..
VMm - Tatka to kvm to spaum Vmt and atso riansges omubhed davicos that servo VM requoss.

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.97 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Continuation of the virtualization diagram with dense small text.
Slide text:
P2 frtual Machine Monitor (VMM) / crosvm / firecracker) Create VM
Guest Userspace: P3: the VM running Thread
net driver: block driver Bock Device Net Device
. Guest Kernel VNRESUAE VMEXIT
: Hypervisor.: /dev/kvm.... Host Kernel.
Linux Virtualization: Key Components.
. To chents VhM tread Is blocked snd rns as *guest VMr, When VM needs soring ft eouts cut via VMEXITS and se sevced snd VAREsUMED. Vum - Tatks to kum io sparn Vms and atso manages erutated devices that serve VM requests Ksm - Hypervisor in the form of a kenol modulo thiat hrverages VT-x to run swich CpU context and rn VMs':
VMEXITS and VhRESUMES are exptrsre, W/e thould try to rminlmlre them

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Diagram plus small bullets; OCR will likely outperform direct transcription.
Slide text:
MicroVMs vs Traditional VMs
HEcrovMs powerod by Rust based VhMs
Sandboxing Emulated Devices.
[emu; Jailed Architecture crosvm
Block Net
Net'
1
Seccomg Seccow
Aou Aq nso eg pnous
: 1. Onty msjor archtecturos and emulated devicos supported
Loss code paths -) Boots Fast.
:.

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Same slide family with small bullets and architecture labels.
Slide text:
MicroyMs powered by Rust based VhiMs
Sandboxing Emulated Devices:
: qemu: Jailed Architecture crosvm
Block Net
Block Net:
Mimi Seccomp Minijail. (seccomp)
: Shouid bo cloor by now' *
Onty major archhectures snd emulated derces supported
2. Less blost -> smaber memory footprint vs trscions! VhuMs. 3. Microvu' - Roflocton of the VuM rathor than the running guost, Less code paths -> Boots Fast

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.97 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Dense bullet text and a central architecture diagram are better handled by OCR than manual transcription in this triage pass.
Slide text:
sowg xooq itey' -': Fast snapshots - Just cump the entire guest memory': -': cioud-hypervsor isn't at the wti'm of Amszon's Frecracker.. Securty as the flrst parameter since mukd-tenant code eoxiecution in m'nd o Had hotphugging and snapshot restora when Arakds started
gVhscr possbty a good in-betwaen cption betweon Contairers and MleroyMs.
Bared on the stcurty gueranees noeded:--、:
Base Project crosvm Desktop focused
Fork Fork
firecracker cloud-hypervisor.
- Serverless loads: - VM Manager daemon - AwS Lambda out of the box. -. Hotplugging devices -. More general Huoddns Ndy - - Snapshot support purpose enterprise VMM

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.96 - Text source: advanced OCR
rapidocr-live/bright-screen/contrast. - OCR decision: ready — Architecture diagram with many small labels is OCR-suitable.
Slide text:
Arrakis MicroVM Runtime
Linux Server
arrakis-py-sdk REST API Arrakis Restserver
REST API
arrakis-client cloud-hypervisor process cloud-hypervisor process
VNC Server Code Server VNC Server Code Server
Arrakis MicroVM Sandbox Arrakis MicroVM Sandbox
Userspace
Kernel /dev/kve /dev/kvm

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.95 - Text source: advanced OCR
rapidocr-live/left-72/opencv-adaptive. - OCR decision: ready — Diagram slide with labels and explanatory text is OCR-suitable.
Slide text:
Arrakis: Storage Using OverlayFS
OverlayFs To. Protect The Rootfs Of Each Sandbox.
..Arakis Sandbox. Is /usr/bin' Arrakis Sandbox' Is /usr/bin
Uerspace Userspact
r overlayfs /usr/in/k riv overloryfs. Tusr/bin/e..
ro rootfs /ar/in/a: ro rcotfs Tusr/bin/a
Kerre! Kerre!
: Shared Rootfs ext9 Lnage
Each sandbox with it's own Rw layer: and a shared RO layer.

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.97 - Text source: advanced OCR
rapidocr-live/bright-screen/contrast. - OCR decision: ready — Code-heavy setup slide is best handled by OCR.
Slide text:
niasabeigis:siyr
Assoon asweboot a sandbox,"initinsidesetsup the overlayfs as detailedbelow.
/bin/busybox mkdir -p $[NEWROOT} /bin/busybox mkdir-p ${woRk} echo"Creating upper and work directories in SwRITABLE Rw" /bin/busybox mkdir -p ${UPPER} /bin/busybox mkdir -p ${NEWROOT_LOWER_RO}
/bin/busyboxmkdir-p ${NEWROOT_WRITABLE_RW]
echo“Mountingoverlayfsto $NEWRooT" SINEWROOT} if[$?-neO ];then /bin/busybox mount-t overlay overlay -o lowerdir=${LOwER_RO},upperdir=${UPPER},workdir=$[wORK} exec /bin/busybox sh# Drop to shell for debugging echo"Error mounting overlayfs!"
else return1 echo "Overlayfs mounted successfully."
fi

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.89 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Cropped architecture diagram with repeated labels is OCR-suitable, but there is no clear title visible in this frame.
Slide text:
vNc Server: rans apoy: vvc Server Code Server.
Arrakis MicroVM Sandbox: Arrakis MicroVM Sandbox:
Cloud-Hypervisor Process cloud-Hypervisor Process:
Disk Device
Per Va Rw. Layer Per vn Rw Layer
Userspace Base Read-Only Rootfs
Kernel. /dev/kvn:: /dev/vm

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: agent_vision.
Slide text:
Arrakis: Networking
1. Tap Devices: Each sandbox gets a unique tap network device
2. Linux Bridge: All tap devices are connected to a Linux bridge on the host
3. Port Forwarding: Automatic port forwarding from the host to the sandbox

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.99 - Text source: none.
- OCR decision: ready — Dense code slide with small text and comments; OCR should read it better than direct transcription.

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.99 - Text source: none.
- OCR decision: ready — Dense code slide with small text and command arrays; OCR is appropriate.

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.99 - Text source: advanced OCR
rapidocr-live/bright-screen/contrast. - OCR decision: ready — Dense code slide with small function and command text; OCR is appropriate.
Slide text:
description string,portForwardDesc string)(portForward,error)
hostPort, err:= s.portAllocator.AllocatePort()
cmd:= exec.Command( "nat", "-t", "iptables", "-A”
"PREROUTING"
"tcp", "--dport",
strconv.Itoa(int(hostPort)), "-j", "DNAT" "--to-destination",
fmt.Sprintf("%s:%d",vmIP, guestPort),
err =cmd.Run()

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: agent_vision.
Slide text:
Arrakis: Updated Architecture

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.99 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Dense code/install slide with small command text and section headings; OCR is appropriate.
Slide text:
86 apt-get clean A:
86:rm -rf. /var/lib/apt/lists/x
# Install Chrome.:
https://dl.google.com/linux/direct/google-chrome-stable_current_amd64.deb 86 \: RUN wget. -0 /tmp/google-chrome-stable_current_amd64.deb # Any failure here will be fixed by the. install -f command below.
dpkg. -i /tmp/google-chrome-stable_current_amd64.deb ll true 86 \
apt-get clean 86 rm -rf /var/lib/apt/lists/*- rm. -rf /tmp/google-chrome-stable_current_amd64.deb 86 \ apt-get update 86 apt-get install. -f -y:86 \
RUN curl: -fsSL https://deb.nodesource.com/setup 22.x -o nodesource_setup:sh 8f # Install Node.js and npm
sudo -E bash nodesource_setup.sh 86 \
rm: -f nodesource_setup.sh 86 \ apt-get.clean: 86 rm -rf /var/lib/apt/lists/x: sudo apt-get:install -y: nodejs: 86

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.99 - Text source: none.
- OCR decision: ready — Dense code/config slide with small identifiers and nested structs; OCR is appropriate.

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.99 - Text source: none.
- OCR decision: ready — Dense code slide with small text and multiple fields; OCR will be more reliable than manual transcription.

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.95 - Text source: agent_vision.
Slide text:
Arrakis: Chrome

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: agent_vision.
Slide text:
Zero Overhead Sandbox Branching
- Zero overhead branching
- Btrfs powered incremental snapshots
- Memory dump compression at snapshot
- Multi stage plans can fail
- Restarting a deep DAG is expensive
- Backtrack + replan + execute much faster at scale
- Can cache key sandbox milestones across plans for reuse

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.98 - Text source: advanced OCR
rapidocr-live/bright-screen/opencv-adaptive. - OCR decision: ready — Small code block and compact labels make OCR the better capture path.
Slide text:
VMM.puso0: VhM.snapshot0 Pershst rw overlayfs iaryer. Recume the VMM
Teec (s aserrer)'srupitoruteta comtistiComnts strint, tzpihtte strl
I+pir?.I+f+ft+rtt+lr+d I+u+at t+ Los++t ra tilt Ioi tt".
Csp, orr'? 'sartetea.tiatuta() piaa,*-, m.apictioi otitw'.+ie.w(cur) 》 Pae't+ tt (irr +) t+is I + irrelhfte.
I! foshre + r+h Iho v +rh If. yupubot (o!ls. tfer focf?:( rrimalt - ma.ipitleat.beresttAt,trsmr'oratn)
Nu. itatat + yattaurel relt
jtateletoluabest: m ath.>eia(tutpastr,.ststeruioiitiioam) i/ Coy the sihtefel eil te the swmuNt eirmteryt'lime we yupaNt aorsh'l ue.rait.
1+ orr.ait (: J).tafet coriimg stetefut eiaa tt saapubet' dirrtery") Irt*+ coyytitkw.ttetetolotitpata, atatwtuthltoit): "4sttio='! ttthfuloliot..1.+y+. a,utat+fulplhprth,.
retws'nit, tat.trrorte raile 1o topy tialotot tik iout dirxtent' t', rr)

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.96 - Text source: agent_vision.
Slide text:
Arrakis: Claude Desktop Vibe Coding
- Create a Google Docs clone where multiple people can join by specifying the name of the doc
- Use NodeJS and run the app on port 5737 in the background.

- Recreated text/layout view: open HTML recreation
- AI slide classifier:
content_slideconfidence0.97 - Text source: none.
- OCR decision: ready — Split prompt-and-terminal demo screenshot with dense small text; OCR is needed for accurate capture.
Classification audit: raw/sources/slide-ai-classification/dense/wsFd22SL1s8/audit.json