Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Here’s the result of my second run with Codex. The first run turned a reference image into an interactive 3D container-terminal simulation. This time, I asked Codex to audit the system in depth, suggest improvements, and implement them. Harbor now includes: - An experiment lab for saving runs and... comparing scenarios against a baseline - Repeatable experiments with matched random seeds across scenarios - Timeline replay and branching, so you can rewind, change settings, and explore a different outcome - Individual container tracking, with a journey history from ship to yard - Yard-crane capacity, queues, and shared-road reservations - Diagnostics showing why cranes and vehicles are waiting You can test whether adding vehicles actually improves throughput, see where queues build up, and compare the results of different operating settings. You can also return to an earlier point in a run, change your decisions, and save the new result for comparison. The second pass also improved the rendering, controls, and layout, with clearer cargo stages and completion milestones.show more

Junho Yeo

3,211 subscribers

17,986 views • 1 month ago •via X (Twitter)

Science & Technology

Anya Rossi• Live Now

Private livecam show

11 Comments

Veyra | AI & Tech1 month ago

Timeline replay and branching for terminal ops is genius

Vera Ai | Tools & Updates1 month ago

That second pass turns a prototype into a genuinely useful simulation lab

Sehan @ Zyner1 month ago

So in real world do we have to have GPS and other trackers connected to these vehicles, machines and etc?

Joachim Ferdinand1 month ago

This is nice

Botanium1 month ago

Can you share the prompts and tools and skills you have used?

Junho Yeo1 month ago

For this demo, I just used the ChatGPT web chat

Coffee Cat1 month ago

让 agent 审自己的系统，有个隐藏风险：它审的是「它理解的系统」，不是你线上跑的那个。我的做法是先让它输出一份它以为的架构图，人先对一遍——图错了，后面所有改进建议都是空中楼阁。

Vinicius Gehlen1 month ago

What is the prompt and models you have used? Any external mcps?

Junho Yeo1 month ago

nothing

DreW1 month ago

Woah

dualtron.uk1 month ago

what did you use to improve rendering + controls

Related Videos

this is f*cking insane, Claude Code agents turned a reading list into a lab you can run in your browser Boris Cherny (creator of Claude Code) gave 12 agents 64 research papers and one assignment: build a browser lab where readers can rerun the experiments pick a fluid simulation, change the conditions and watch how the result changes switch methods while keeping the same scene. inspect the equations, code and measurements beside the animation give each group a specific job: → readers extract equations, assumptions and experimental settings → builders turn the methods into runnable code → reviewers compare outputs against the papers → interface agents build the controls for exploring each experiment a simulation can look convincing while producing the wrong answer reviewers check whether it preserves mass, stays stable and matches the paper's measurements before accepting the implementation every experiment has to show: > the claim being tested > the conditions used > the difference from the reference result > the details still needed to verify it the lab covered 11 methods across 28 reference scenes. seven matched the selected reference results within the stated tolerances; four stayed marked for further review give Claude Code your paper collection, then ask: "Turn these papers into an interactive browser lab 1. Extract implementable methods, equations and experimental settings. Attach page references and flag missing details 2. Assign independent methods to separate agents. Use shared inputs and measurement tools so the results can be compared 3. Have reviewers check each implementation against the paper. Record discrepancies and keep unverified results clearly marked 4. Build controls for changing parameters, comparing methods and opening the code behind each result" start with one paper and reproduce one reference experiment before adding more agents include the inputs, dependencies and checks so the next reader can rerun it from a clean checkout watch the agents at work below

this is f*cking insane, Claude Code agents turned a reading list into a lab you can run in your browser Boris Cherny (creator of Claude Code) gave 12 agents 64 research papers and one assignment: build a browser lab where readers can rerun the experiments pick a fluid simulation, change the conditions and watch how the result changes switch methods while keeping the same scene. inspect the equations, code and measurements beside the animation give each group a specific job: → readers extract equations, assumptions and experimental settings → builders turn the methods into runnable code → reviewers compare outputs against the papers → interface agents build the controls for exploring each experiment a simulation can look convincing while producing the wrong answer reviewers check whether it preserves mass, stays stable and matches the paper's measurements before accepting the implementation every experiment has to show: > the claim being tested > the conditions used > the difference from the reference result > the details still needed to verify it the lab covered 11 methods across 28 reference scenes. seven matched the selected reference results within the stated tolerances; four stayed marked for further review give Claude Code your paper collection, then ask: "Turn these papers into an interactive browser lab 1. Extract implementable methods, equations and experimental settings. Attach page references and flag missing details 2. Assign independent methods to separate agents. Use shared inputs and measurement tools so the results can be compared 3. Have reviewers check each implementation against the paper. Record discrepancies and keep unverified results clearly marked 4. Build controls for changing parameters, comparing methods and opening the code behind each result" start with one paper and reproduce one reference experiment before adding more agents include the inputs, dependencies and checks so the next reader can rerun it from a clean checkout watch the agents at work below

Mika

847,312 views • 1 month ago

GPT-6 coded an entire world: a busy harbor you can step inside. From a single prompt and an asset pack, it assembled a container port with towering cranes, docked ships, rows of containers, and forklifts across the yard, all within an explorable 3D scene. A first look at our upcoming Code-for-3D benchmark, featuring the text-to-scene setting.

GPT-6 coded an entire world: a busy harbor you can step inside. From a single prompt and an asset pack, it assembled a container port with towering cranes, docked ships, rows of containers, and forklifts across the yard, all within an explorable 3D scene. A first look at our upcoming Code-for-3D benchmark, featuring the text-to-scene setting.

SimWorld

25,313 views • 2 months ago

Try out the Visualize plugin (in preview) in Codex. Some ideas click faster when you can see and interact with it. Here’s a fun example where I asked Codex to build a planet simulator with different controls. 🌎

Try out the Visualize plugin (in preview) in Codex. Some ideas click faster when you can see and interact with it. Here’s a fun example where I asked Codex to build a planet simulator with different controls. 🌎

Derrick Choi

85,921 views • 4 months ago

An AI influencer campaign can start with nothing more than an idea. I tested it with Codex + CapCut Web. Give Codex the concept and creative direction, and CapCut can handle the generation flow—from creating the AI character with text or reference images to generating the actual video. You can also choose different models and generation settings depending on the look you want. That makes the whole AI influencer workflow feel much more flexible. #capcutweb #capcutchatgptplugin #capcutcodex #capcutmarketing

An AI influencer campaign can start with nothing more than an idea. I tested it with Codex + CapCut Web. Give Codex the concept and creative direction, and CapCut can handle the generation flow—from creating the AI character with text or reference images to generating the actual video. You can also choose different models and generation settings depending on the look you want. That makes the whole AI influencer workflow feel much more flexible. #capcutweb #capcutchatgptplugin #capcutcodex #capcutmarketing

Mira Sterling

22,154 views • 29 days ago

You can now draw how your app works and get Codex to build it with the new tldraw plugin for ChatGPT. Redraw the logic, and the code gets updated to match. Ask for a change in chat, and it updates the diagram first so you can check the logic before any code changes. Here's a traffic light with a 50% chance of turning blue instead of green:

You can now draw how your app works and get Codex to build it with the new tldraw plugin for ChatGPT. Redraw the logic, and the code gets updated to match. Ask for a change in chat, and it updates the diagram first so you can check the logic before any code changes. Here's a traffic light with a 50% chance of turning blue instead of green:

tldraw

31,935 views • 1 month ago

Over the weekend I made more progress on the Unreal game that I have been vibe coding. More enemies and a final boss with a roguelike progression system. I did not have to open Unreal Editor at all to make this game. Here's how you can set up in a way that you can build an Unreal game without having to use the Editor: Just mention these rules to the agent: - Write everything in C++, and build maps and assets with code instead of the editor. - Build and run the game from the command line. - After each change, run the game yourself with an automated test, check the screenshots, and tell me what you saw. - Close Unreal when you're done. - Create a .bat file for me to run the game.

Over the weekend I made more progress on the Unreal game that I have been vibe coding. More enemies and a final boss with a roguelike progression system. I did not have to open Unreal Editor at all to make this game. Here's how you can set up in a way that you can build an Unreal game without having to use the Editor: Just mention these rules to the agent: - Write everything in C++, and build maps and assets with code instead of the editor. - Build and run the game from the command line. - After each change, run the game yourself with an automated test, check the screenshots, and tell me what you saw. - Close Unreal when you're done. - Create a .bat file for me to run the game.

Danny Limanseta

28,757 views • 12 days ago

Comment “HYPER3D” and I’ll send you the link 👇 You can now build interactive 3D websites using GPT-6 and Hyper3D Rodin. For this demo, I built a robot experience where you can collect parts, repair machines, and recharge before time runs out. Here’s how: 1. Open Codex and ask GPT-6 to build an interactive robot website 2. Go to Hyper3D.ai and connect Rodin through MCP 3. Ask GPT-6 to create a robot reference image 4. Let Rodin generate the 3D model 5. GPT-6 adds the model to the website and builds the interactions The result is a playable 3D experience directly inside a website. You can go from a simple prompt to a custom 3D model and an interactive web experience in one workflow. It makes building 3D websites much easier without creating every asset manually. #hyper3d #3dwebsite #webdevelopment #generativeai #aitools

Comment “HYPER3D” and I’ll send you the link 👇 You can now build interactive 3D websites using GPT-6 and Hyper3D Rodin. For this demo, I built a robot experience where you can collect parts, repair machines, and recharge before time runs out. Here’s how: 1. Open Codex and ask GPT-6 to build an interactive robot website 2. Go to Hyper3D.ai and connect Rodin through MCP 3. Ask GPT-6 to create a robot reference image 4. Let Rodin generate the 3D model 5. GPT-6 adds the model to the website and builds the interactions The result is a playable 3D experience directly inside a website. You can go from a simple prompt to a custom 3D model and an interactive web experience in one workflow. It makes building 3D websites much easier without creating every asset manually. #hyper3d #3dwebsite #webdevelopment #generativeai #aitools

Sam | AI Tools & Tech

85,900 views • 12 days ago

I had an early access to GPT-6 Astra and I can say, everyone in the world now has a 3D designer at their fingertips. I gave it an image of a house and asked to create it in 3D with all the details including toys, appliances and furniture. This a full 3D model reconstruction in Blender with geometry that you can manually tweak and run 60fps as a locally rendered "game" on device.

I had an early access to GPT-6 Astra and I can say, everyone in the world now has a 3D designer at their fingertips. I gave it an image of a house and asked to create it in 3D with all the details including toys, appliances and furniture. This a full 3D model reconstruction in Blender with geometry that you can manually tweak and run 60fps as a locally rendered "game" on device.

Tom Krcha

1,446,074 views • 2 months ago

I turned my apartment into a three.js scene with Opus 5.5 so I can record videos for my tutorials. It's fully interactive and you can walk through it in the browser. You can change the lighting and set different render styles. All objects are code using the scoring system.

I turned my apartment into a three.js scene with Opus 5.5 so I can record videos for my tutorials. It's fully interactive and you can walk through it in the browser. You can change the lighting and set different render styles. All objects are code using the scoring system.

Meng To

158,253 views • 5 days ago

I've unlocked 10x more from Codex since I've let the agent prompt and manage itself. I call this "The Orchestrator Rule". Here's why it works so well: - Models like GPT 5.6 Sol are getting better at inferring your intent. - Codex has the ability to create, monitor, and send messages to other Codex threads. - So you can just tell Codex the outcome you want, and Codex can turn that into detailed prompts, spin up one or more threads to do the work, and bring the results back to you. This changed my job as a builder: Instead of writing perfect prompts, I described outcomes with success criteria and verification methods. And Codex wrote the prompts and orchestrated the work for me. Hence the "Orchestrator" Rule. Here are 2 ways to let Codex prompt and manage itself: (1) Ask Codex to kick off a task in a new thread. - You can give it a short description of the outcome you want to achieve, and Codex will expand that into the detailed instructions it needs to deliver. (2) Keep one parent thread for each project and let Codex spin up new child threads for individual tasks - This pairs really well with Voice mode. - You can talk to your parent thread like a chief of staff. It can kick off the work using child threads and report back when it needs your input or when the task is complete. I demo both these methods below:

I've unlocked 10x more from Codex since I've let the agent prompt and manage itself. I call this "The Orchestrator Rule". Here's why it works so well: - Models like GPT 5.6 Sol are getting better at inferring your intent. - Codex has the ability to create, monitor, and send messages to other Codex threads. - So you can just tell Codex the outcome you want, and Codex can turn that into detailed prompts, spin up one or more threads to do the work, and bring the results back to you. This changed my job as a builder: Instead of writing perfect prompts, I described outcomes with success criteria and verification methods. And Codex wrote the prompts and orchestrated the work for me. Hence the "Orchestrator" Rule. Here are 2 ways to let Codex prompt and manage itself: (1) Ask Codex to kick off a task in a new thread. - You can give it a short description of the outcome you want to achieve, and Codex will expand that into the detailed instructions it needs to deliver. (2) Keep one parent thread for each project and let Codex spin up new child threads for individual tasks - This pairs really well with Voice mode. - You can talk to your parent thread like a chief of staff. It can kick off the work using child threads and report back when it needs your input or when the task is complete. I demo both these methods below:

Avthar Sewrathan

62,928 views • 2 months ago

Comment "GIT" to get the full guide. What if you could run an entire company with AI employees? 👀 Paperclip AI lets you build an organization of AI agents with different roles, from CEO and CTO to developers and marketers. You give them a goal, and your AI CEO can break it into tasks, assign work, and coordinate the agents — all from one dashboard. You can also track progress, set spending limits, and approve important decisions. The best part? It's free and open source and works with tools like Claude Code, Codex, and OpenClaw. Want to build your own AI-powered team? Comment “GIT” below, and I'll send you the link and setup guide. 👇 #PaperclipAI #AI #AIAgents #AITools #ClaudeCode

Comment "GIT" to get the full guide. What if you could run an entire company with AI employees? 👀 Paperclip AI lets you build an organization of AI agents with different roles, from CEO and CTO to developers and marketers. You give them a goal, and your AI CEO can break it into tasks, assign work, and coordinate the agents — all from one dashboard. You can also track progress, set spending limits, and approve important decisions. The best part? It's free and open source and works with tools like Claude Code, Codex, and OpenClaw. Want to build your own AI-powered team? Comment “GIT” below, and I'll send you the link and setup guide. 👇 #PaperclipAI #AI #AIAgents #AITools #ClaudeCode

metaengines

41,100 views • 8 days ago

Here’s a quick tutorial on how to set up the run/walk alert on your @garminrunning watch! The watch in the video is the forerunner 265s but I have done this with all of my Garmin watches 💙 also I forget half the time my watch is touch screen lol I use the run walk method a lot for warm ups, cooldowns, recovery miles and long runs! Instead of just walking whenever I like to have a little bit of structure to my run! Go to run-> click the middle button the left to access your run settings Click run settings -> scroll down to alerts and open that up Click add new and the menu will bring options out and you’ll see run walk. Set up an interval that works best for you! #garminambassador #garminrunning #running #marathontraining #runwalk

Here’s a quick tutorial on how to set up the run/walk alert on your @garminrunning watch! The watch in the video is the forerunner 265s but I have done this with all of my Garmin watches 💙 also I forget half the time my watch is touch screen lol I use the run walk method a lot for warm ups, cooldowns, recovery miles and long runs! Instead of just walking whenever I like to have a little bit of structure to my run! Go to run-> click the middle button the left to access your run settings Click run settings -> scroll down to alerts and open that up Click add new and the menu will bring options out and you’ll see run walk. Set up an interval that works best for you! #garminambassador #garminrunning #running #marathontraining #runwalk

brandaytalamoni

22,900 views • 2 years ago

Introducing TikZBot. The birth of an agent that produces its own TikZ creations. Fully interactive and extensible. It can fix scenarios, expand its own work, learn from diagrams, and talk to you. For now, it’s a PoC. The manual version works like a charm. I’m now adding the self-play and chat components. Programmed in a few hours in Codex CLI using the newest 5.2-Codex model. So much fun.

Introducing TikZBot. The birth of an agent that produces its own TikZ creations. Fully interactive and extensible. It can fix scenarios, expand its own work, learn from diagrams, and talk to you. For now, it’s a PoC. The manual version works like a charm. I’m now adding the self-play and chat components. Programmed in a few hours in Codex CLI using the newest 5.2-Codex model. So much fun.

Bartosz Naskręcki

18,444 views • 9 months ago

Codex can now use the in-app browser to test your app at different viewport sizes! It will control the device tool bar and click through your app at different breakpoints to validate & iterate. If it's a long run, Codex will take screenshots at key moments during testing and show them to you at the end of the turn, so you can verify its work. To speed up testing, Codex can hide the IAB to disable animations, and accelerate testing by 1-2x. Separately, we also made annotations send faster and consume less tokens. Hope you enjoy the updates and let us know what you think!

Codex can now use the in-app browser to test your app at different viewport sizes! It will control the device tool bar and click through your app at different breakpoints to validate & iterate. If it's a long run, Codex will take screenshots at key moments during testing and show them to you at the end of the turn, so you can verify its work. To speed up testing, Codex can hide the IAB to disable animations, and accelerate testing by 1-2x. Separately, we also made annotations send faster and consume less tokens. Hope you enjoy the updates and let us know what you think!

James Sun

229,193 views • 6 months ago

7K+ RL environments from Xiaomi across coding, cyber, general, music, and more. You can now explore, visualize, and actually run them. Pick an environment → pick a model → run the rollout → watch exactly what happens RL environments are a lot more interesting when you can actually play with them.

7K+ RL environments from Xiaomi across coding, cyber, general, music, and more. You can now explore, visualize, and actually run them. Pick an environment → pick a model → run the rollout → watch exactly what happens RL environments are a lot more interesting when you can actually play with them.

Adithya S K

17,971 views • 1 month ago

New stuff you can do in Spline v2: Building a TSL material with texture atlases was surprisingly easy > A little 2D samurai running around. I gave Claude a 2D sprite sheet of a and asked it to generate the atlases, and the agent in Spline took care of writing a shader that runs the loop at the right speed and binds it all to the keyboard. An old-school game with the advantages of 3D. Super promising. This is really just the basics here (an Idle and a Run state) but you can add jumps and build on it.

New stuff you can do in Spline v2: Building a TSL material with texture atlases was surprisingly easy > A little 2D samurai running around. I gave Claude a 2D sprite sheet of a and asked it to generate the atlases, and the agent in Spline took care of writing a shader that runs the loop at the right speed and binds it all to the keyboard. An old-school game with the advantages of 3D. Super promising. This is really just the basics here (an Idle and a Run state) but you can add jumps and build on it.

Joseph Azar

14,715 views • 2 months ago

this is getting f*cking dangerous: Claude Code agents can turn a stack of research papers into a lab u can run right in ur browser i made $3,650 overnight with this approach Boris Cherny, the creator of Claude Code, gave a dozen agents 64 research papers and asked them to build a platform where anyone could reproduce the physics experiments themselves u open a fluid simulation, change the conditions and immediately see how the flow changes while the experiment is running, u can switch algorithms, open the formulas and inspect the code calculating what happens on screen so u take a complicated paper sitting in ur “i’ll read this someday” folder and get something u can actually play with each agent got its own job: → analysts extract formulas, assumptions and experimental conditions from the papers → developers turn them into working models → reviewers compare the results against the original research → separate agents build controls for each scene and this is where it gets interesting the fluid can look beautiful while the calculations underneath are completely broken. u might never notice just by watching so reviewers check whether mass is conserved, whether changing parameters breaks the calculations and whether the results match the paper’s data next to each experiment, u can see: what’s being tested which values are being used how far the result differs from the original what still needs checking the write-up describes 11 approaches and 28 test scenes. seven approaches passed within acceptable tolerances, while four needed more work i borrowed this way of organizing the work for my FOMOBILL: breaking down how memecoin buys and sells work so i can check what wallets did after entering, before making my own trade the physicists have the original paper. i have the original transaction i can open and verify if u wanna try this approach, upload ur materials to Claude Code and give it this task: “Turn these materials into an interactive tool: 1. Extract the algorithms, rules and testing parameters. Cite the sources and flag missing details 2. Assign separate parts to independent agents. Use the same input data for comparisons 3. Set up independent checks of the results. Log errors and visibly label anything unverified 4. Build an interface where users can change parameters, compare results and inspect the source code” start with one document and one testable example. scaling the team is easier once the first result can be reproduced save the dependencies and source data so someone else can run the project too my overnight result was paper trading; execution and results can differ with real money lyrics breaks down the full approach and how im applying it in FOMOBILL in the article below ↓

this is getting f*cking dangerous: Claude Code agents can turn a stack of research papers into a lab u can run right in ur browser i made $3,650 overnight with this approach Boris Cherny, the creator of Claude Code, gave a dozen agents 64 research papers and asked them to build a platform where anyone could reproduce the physics experiments themselves u open a fluid simulation, change the conditions and immediately see how the flow changes while the experiment is running, u can switch algorithms, open the formulas and inspect the code calculating what happens on screen so u take a complicated paper sitting in ur “i’ll read this someday” folder and get something u can actually play with each agent got its own job: → analysts extract formulas, assumptions and experimental conditions from the papers → developers turn them into working models → reviewers compare the results against the original research → separate agents build controls for each scene and this is where it gets interesting the fluid can look beautiful while the calculations underneath are completely broken. u might never notice just by watching so reviewers check whether mass is conserved, whether changing parameters breaks the calculations and whether the results match the paper’s data next to each experiment, u can see: what’s being tested which values are being used how far the result differs from the original what still needs checking the write-up describes 11 approaches and 28 test scenes. seven approaches passed within acceptable tolerances, while four needed more work i borrowed this way of organizing the work for my FOMOBILL: breaking down how memecoin buys and sells work so i can check what wallets did after entering, before making my own trade the physicists have the original paper. i have the original transaction i can open and verify if u wanna try this approach, upload ur materials to Claude Code and give it this task: “Turn these materials into an interactive tool: 1. Extract the algorithms, rules and testing parameters. Cite the sources and flag missing details 2. Assign separate parts to independent agents. Use the same input data for comparisons 3. Set up independent checks of the results. Log errors and visibly label anything unverified 4. Build an interface where users can change parameters, compare results and inspect the source code” start with one document and one testable example. scaling the team is easier once the first result can be reproduced save the dependencies and source data so someone else can run the project too my overnight result was paper trading; execution and results can differ with real money lyrics breaks down the full approach and how im applying it in FOMOBILL in the article below ↓

Logics

128,132 views • 1 month ago

Attacking civs can result in them getting the attention of the police. They can also run to a phone and call for help. Assaulting police officers can result in them calling for backup.

Attacking civs can result in them getting the attention of the police. They can also run to a phone and call for help. Assaulting police officers can result in them calling for backup.

Trey Powell Fallen Aces

40,865 views • 7 months ago

Now an even more **INSANE** new magical codex workflow In a thread, ask Codex to: 1. generate an image with the ImageGen skill 2. create a 3D object with the img2threejs skill 3. build an asset editor And now you can edit infinite variations! need another attribute? just ask

Now an even more INSANE new magical codex workflow In a thread, ask Codex to: 1. generate an image with the ImageGen skill 2. create a 3D object with the img2threejs skill 3. build an asset editor And now you can edit infinite variations! need another attribute? just ask

Nicolas Zullo

47,183 views • 3 months ago

You can run Claude inside a CoreWeave Sandbox and drive it from your phone. The agent gets a custom container image with your tools and MCPs, plus Slurm access and the ability to set your network policy. The sandbox doesn't care what the workload is. That's the point.

You can run Claude inside a CoreWeave Sandbox and drive it from your phone. The agent gets a custom container image with your tools and MCPs, plus Slurm access and the ability to set your network policy. The sandbox doesn't care what the workload is. That's the point.

CoreWeave

10,959 views • 3 months ago