After a week of business travel, I was on a flight from San Francisco to Beijing, facing 14 hours of boredom, when an idea suddenly flashed through my mind: if I did not buy the in-flight Wi-Fi and used only the large language model and coding agents already installed on my computer, what interesting things could I make? I had actually thought about this much earlier: if I took a laptop back to ancient times, or if the apocalypse brought all public services to a halt, what could I do with AI?
I opened the overhead bin and pulled out my computer without hesitation.
First, let’s see what I have
- A company-issued MacBook Pro with an M3 Max chip—and luckily, I had not checked the power cable
- LM Studio
- GPT-OSS-20B installed in LM Studio
- Claude Code and Codex
- Some development environments and packages I had installed before (I was not entirely sure what was there or what would be useful)
The challenge
Both Claude and Codex were connected to my company’s LLM Gateway. First, I needed to connect them to GPT-OSS in my LM Studio.
With LM Studio’s UI, finding the gateway API was quite easy. I then entered my question and the API information into LM Studio’s Chat function. I first tried configuring Claude Code’s model gateway, but LM Studio’s answers did not seem quite right, so I decided to look at Claude’s help information. I tried running claude –help, claude -help, and claude help manually, but none of them displayed the help information; instead, the browser jumped to the company authentication page.
To save time, I tried Codex first. codex –help successfully displayed all the configuration options. I sent my question and the output of codex –help to LM Studio Chat, and got the answer I needed: I should configure it in ~/.codex/config.toml. The configuration went smoothly too!
Great. Now let’s do something interesting!
Exploration 1
Before starting, I told Codex that the current environment was completely offline and that it must not try to download or install any new packages. I asked it to explore the programming languages and dependencies already installed locally. Everything ran normally!
What should I make next? I had no ideas, so I asked Codex to tell me. It suggested text games, a local Web API, a command-line todo manager, and some relatively boring things. Clearly, during a 14-hour flight, a game would make better company.
A pinball game suddenly came to mind: use a paddle to catch and bounce a ball to score points. So I got started. Like the text games and command-line tools it had just mentioned, a command-line version might be easier to implement. I started simple to validate the idea. Sure enough, without any further prompting, it quickly completed a basic command-line pinball game. It only had basic bouncing and scoring, though, so I tried turning it into a complete game and eventually got a two-player version.
Next, I tried switching to a more intuitive Web version. After two or three rounds of prompt iteration, that worked too!
Exploration 2
Besides games, could I make something with more practical value? Suppose I went back to ancient times or the end of the world. A complete knowledge-base system for learning would be extremely valuable. Setting aside the accuracy of AI-generated content for now, I tried making a simple encyclopedia website.
Building the website was harder than I expected. My subjective impression was that the Codex + GPT-OSS-20B combination seemed more “laid-back” and was prone to simple mistakes. Its laziness showed up like this: when asked to generate 10 entries, it would generate 5 first; only after I asked again would it fill in the rest. When fixing broken links, it would check only some of them, and I had to emphasize “check all of them” several times before it handled everything. A simple example of an error: links between HTML pages were often wrong, leaving a large number of broken links.
Still, after several rounds of conversation, it did manage to generate a simple encyclopedia page.
Some thoughts
If I had time to prepare and could take a computer and an LLM back to ancient times, I would choose the largest and most advanced model the computer could support. It would be slower, but its output would be more accurate. But how would I solve the power problem? Ask the LLM to generate a way to produce electricity?
By this point, more than an hour had passed, and I had not thought of anything more interesting for the moment. It was time to put the computer down and take a break. If you have any good ideas, feel free to leave a comment—I can keep experimenting.