Epic is still a heavy user of that language at the lower parts of its stack, but there are other, friendlier/more modern languages in pretty widespread use too. Depends on the team & sub-application.
It seems that much like Aider, you use separate models for creating code edits and validating them. That's a win in my book. It seems Claude Code does not do that, which is part of the reason it racks up a (relatively) large bill for long work sessions (that and the cost of sonnet-3.7).
I bounce back and forth between Aider, Claude Code, and Simon Willison's LLM tool ("just" a GOOD wrapper for using LLMs at the CLI, unlike the other two which are agent-y.) LLM is my favorite because I usually don't need/want full autonomy, but Claude Code has started to win me over for straightforward stuff. Plandex looks cool enough to throw into the rotation!
My main concern at this point is that I use a Mac and as far as I understand it Docker containers can have pretty poor performance on the Mac, so I'm wondering if that will carry over to performance of Plandex. (I don't use Docker at all so I'm not sure if that's outdated info.)
> It seems that much like Aider, you use separate models for creating code edits and validating them.
That's right. To apply edits, Plandex first attempts a deterministic edit based on the edit snippet. In some cases this can be used without validation, and in others a validation step is needed. A "race" is then orchestrated with o3-mini between an aider-style diff edit, a whole file build, and (on the cloud service) a specialized model. I actually wrote a comment about how this works (while maintaining efficiency/cost-effectiveness) a couple days ago: https://news.ycombinator.com/item?id=43673412
And on the Docker question, it should be working well on Mac.
- Plandex is more agentic—it can complete a complex task, updating many files, all in one go.
- Changes are applied to a sandbox by default rather than directly to project files, helping you prevent unintended changes.
- Plandex can automatically find the context it needs in the project.
- Plandex can execute commands (like installing dependencies, running tests, etc.) and auto-debug if they fail.
- Plandex should be more reliable on file edits—it uses an enhanced version of aider's diff-style edit that is resilient to multiple occurrences, but it also has validation, a whole file fallback, and on the cloud service, a custom fast apply model is also added to the mix. Will be publishing benchmarks on this soon.
My master's thesis[1] was half research, half dev project, exploring how we can continue to fully fuse traditional RPGs with computers. This goal is my life quest, my life's work.
I think virtual tabletops (VTTs) as they currently stand are barking up the wrong tree[2]. I want a computer-augmented RPG to allow the GM to do everything he does in the analog form of the game. On-the-fly addition of content to the game world, defining of new kinds of content, defining and editing rules, and many other things ... as well as the stuff VTTs do, of course. The closest we've gotten in the last 30 years is LambdaMOO and other MUDs.
The app I made for my thesis project was an experimental vertical slice of the kinds of functionality I want. The app I made after that last year is more practical and focused on the needs of my weekly game, in my custom system; I continue to develop it almost daily.
I'm itching to tackle the hardest problem of all, which is fully incorporating game rules in a not-totally-hardcoded way. I need rules to be first-class objects that can be programmatically reasoned about to do cool things like "use the Common Lisp condition system to present end user GMs with strategies for recovering from buggy rules." Inspirations include the Inform 7 rulebook system.
[2] Anything that requires physical equipment other than dice and a regular computer is also barking up the wrong tree. So no VR, no video-tracked physical miniatures, no custom-designed tabletop, no Microsoft Surface... Again, just my opinion.
I'm working on something similar. I'm building a MUD with an LLM playing the role of GM. Currently it just controls NPCs, but I eventually want it to be able to modify the game rules in real time. My end goal is a world that hundreds of players can play in simultaneously, but has the freedom and flexibility of a TTRPG (while still remaining balanced and fair).
That's really cool, Elias. I keep seeing people try to put LLMs into the role of the GM. But I think you're doing something new and important by working to have the rules available to it.
Is your project available anywhere? Best of luck!
If you're interested, because I kept seeing "LLM as GM" projects, I got curious about how well it would work to have LLMs as players instead. So I made this:
It's a training ground for GMs to practice things like spontaneous description, with 4 AI players that get fed what each other say so they act in a reasonably consistent manner. It's not perfect, but I've gotten some good use out of it.
how do you feel about Talespire? it allows pretty fast on-the-fly map-making as long you’re not dealing with significant vertical distances, although it’s got very little in common with LambdaMOO. but MUDs generally seem to be MMRPG precursors at this point, unless there’s an underground community I’m unaware of.
I feel the same way about Tailspire as I feel about pretty much every other VTT. They might do okay, even pretty well, at combat maps and/or character sheets, but I want is the whole game world in the computer. Maps are just a fraction of what I need as a GM. I need data on economics and population numbers and power structures. And I need computation over all those things.
For instance, my game rules include an economic subsystem, which takes in the production of goods and services at hundreds of in-game cities, and computes prices for over a thousand player-purchaseable goods. The "second app" that I referred to above allows players to (among many other things) purchase stuff at the market nearest their current location and have those items go straight into their character sheet. If the "item" is actually an animal, a hired mercenary, etc. then a different subsystem generates a new NPC with the right statistics and attaches the player to it as owner/liege.
I could write an extension for a VTT that talks to my economic system over an API, and throws items up on screen, lets players purchase them, moves them into their character sheet using the right function calls in the VTT's extension library, etc. But every step of the way, I would be fighting to cram this subsystem into the VTT's conception that gameplay begins and ends with maps and char sheets.
I am full-time building LLM-based NPCs for a text-based MMORPG. Been doing a lot of work recently on allowing progression through scenarios with them where the rules are in a class, and the LLM takes care of communicating user intent to the rules engine, based on free-text, and writes back to the user with the results.
That's sweet! I think LLMs have incredible potential for descriptions and for NPC behavior, and I really like that you have this bridge between freeform intent and a rules engine. I'd like to pick your brain about it - I'll send an email.
You just got a signup :) Free plan, I'll admit. I don't need or want anything other than email notifications, and the free plan for that is very generous. Thanks for building this.
I've had tremendous trouble trying to get my iPhone SE to show up as a USB device. Never did figure it out, but the macOS app iMazing has served me well as a workaround.
Ahh, these are beautiful. Thanks 'merothwell for linking.
I was in Monument Valley for the first time ever this year; if you like these photos, try and pay it a visit. You won't be disappointed.
VWWHFSfQ, you may already know this, but: I recommend this talk by Armin Ronacher (Flask creator) on how Python's implementation internals contribute to the difficulties of making Python faster.
One case study Ronacher gets into is the torturous path taken through the Python interpreter (runtime?) when you evaluate `__add__`. Fascinating stuff.
