← Back to Pulse PULSE. brief
Everything Runs Through Codex Now | Bilawal Sidhu

Everything Runs Through Codex Now | Bilawal Sidhu

Riley Brown63 min2026-08-16 ▶ Watch on YouTube
ℹ️
Partly verifiedA few specific details here couldn't be independently confirmed against the video. The overall summary is sound, but double-check exact numbers or names before you rely on them.
What this video is
⚡ a 64-minute video, readable in 60 seconds

This is an interview with Bilawal Sidhu, a spatial computing and generative AI creator who vibe coded a Palantir-style OSINT surveillance project that hit 2 million YouTube views and is now becoming a startup with a co-founder. The conversation covers how AI agents like Codex, Claude Code, and a new team platform called Buzz are reshaping both his filmmaking workflow (building explainer visuals for projects like his Iron Sights shot-tracking video) and his day-to-day business operations. It also touches on cutting-edge AI video generation research and experiments, plus rising production costs and quality bars in the creator economy.

[01:03] Vibe coded a Palantir-style spatial command center project (WorldView), got 2 million YouTube views, and is turning it into a startup
Key takeaways
+ 23 more takeaways
  • [56:20] Plans to open source the original project this month as a browser-based "spy simulator" (GitHub: bilawalsidhu/gods-eye-view)
  • [57:03] V1 will be open source, but V2 and beyond (full globe scrubbing across 30+ data layers) will be a commercial product due to cost
  • [58:26] Viral reaction included people asking where to swipe a credit card, which he calls true product-market fit
  • [02:08] Says he's "completely codex-pilled," citing Codex's in-app browser as a game-changer versus Chrome attachment flakiness
  • [02:21] Says GPT 5.6 ("5.6 soul") is a bad model for writing
  • [06:24] Says OpenAI was the first company to nail what OpenClaw promised, calling Codex the first AI-powered "super app"
  • [06:53] Identifies four things OpenAI nailed: a frontier agent to talk to, plugins, automations like cron jobs, and an in-app browser
  • [10:01] Argues apps that survive the agent era will be ones that make themselves usable alongside agents, and says Notion is far ahead of Google Docs on this
  • [11:23] Estimates OpenAI's move to Cerebras chip technology could speed responses 5x to 10x, calling the result "borderline Jarvis-like"
  • [14:11] His "Iron Sights" video project uses computer vision models together with Meta Ray-Ban glasses and phone footage
  • [14:21] Fusing both camera perspectives creates what he calls a "spatial shot counter" that measures every hit or miss
  • [14:30] Built the pipeline's explainer diagrams in the same codebase using Claude Code
  • [15:00] Referenced a 127.3K-view X post describing his use of Gemini 2.5 Pro to build the shot counter and an After Effects AR-style HUD overlay script from the footage
  • [17:57] Cites Karpathy giving Opus 5 the opening paragraph of Lord of the Rings with a $10, 1M-token budget; it worked about 2 hours and wrote 5,500 lines of three.js code
  • [18:16] Notes Karpathy's finding that LLMs can't audit their own visual output because they can't natively perceive video, forcing slow screenshot checks that introduce "jank"
  • [25:18] Cites the "Seoul World Model" paper (KAIST AI, NAVER AI Lab, SNU AIS), an ECCV2026 Oral, which uses RAG on millions of Seoul street-view images
  • [25:26] The Seoul World Model generates video over multi-kilometer trajectories without accumulating errors
  • [29:29] Notes some YouTubers now spend $10,000 to $100,000 per video on VFX and motion graphics, versus about $10,000 a decade ago
  • [42:45] Describes Buzz as feeling like Open Claw with a more intuitive interface, but not yet ready to replace Slack
  • [45:02] Says Buzz marks a shift from personal agents to a "team of agents," letting a company build a shared second brain
  • [49:22] Notes OpenAI released "workspace agents" for Slack that run in the cloud and connect to plugins
  • [51:12] His viral 2-million-view project originated from a tweet after using Gemini 3.1, which he found better at spatial reasoning than Claude 4.6
  • [01:00:04] Deliberately publishes detailed "frontier map" breakdowns, including which APIs he used, on his Substack despite advice to keep that proprietary
Shown on screen — grab and go
PROMPTPrompt asking Codex to draft interview questions
Based on conversation ontext with bilawal sidhu and herein the chat reccomend what i should ask him in the vidoe about codex i'm about to film a video.
Read from the Codex chat input box at 9:35; one likely OCR-dropped letter in 'ontext' (context) preserved as read rather than guessed. shown at 9:35
PROMPTInstruction to add Notion doc bio bullets
put these in his notion doc and at the top describe who he is and what he does in bullets
Consistent across three OCR reads between 9:49 and 10:17; spacing restored, wording unchanged. shown at 9:59
PROMPTThe Notion API edit and tabs instruction
please highlight the important things like making the text blue in this document. anything that you think is important that i should include inside the notion doc, just make changes via the api in this doc and then add a section at the end but use the tabs' functionality instead of bullets
Merged from six near-identical OCR reads between 10:42 and 12:57; only crushed spacing was restored. shown at 10:42
Jump to a moment
0:00Intro1:14From OpenClaw to Codex: Agent Setups Then & Now7:23The In-App Browser as a Shared Canvas13:00Agents for Content Creation: Three.js & AI B-Roll18:50Will AI Make Movies? Game Engines Meet Video Models21:56Viral Drone Shots, Omni & Spatial RAG31:25The Hard Parts of AI Video Editing35:52YouTube Analytics on Autopilot42:45Buzz & the Team-of-Agents Era51:05Vibe Coding Palantir & Owning the Movement
Their links, sorted & clickable
🔗 Other links1Spotifyopen.spotify.com
← Back to Pulse Dashboard
Was this brief useful?