I had a conversation the other day with someone whose main take was the only way forward with ai is to return to symbolic ai.
I see people have put the transcripts of full adventure game playthroughs online, so it's reasonably likely games are present in the training data:
<a href="https://dwheeler.com/anchorhead/anchorhead-transcript.txt" rel="nofollow">https://dwheeler.com/anchorhead/anchorhead-transcript.txt</a><p>You can probably find games where that's not true, as people are still releasing text adventure games occasionally.
I tried something similar, but distilled to "solve this maze" as a first-person text adventure, and while it usually solved it eventually, it almost always backtracked through fully-explored dead ends multiple times before finally getting to the end. I was pretty surprised by this, as I expected they'd be able to traverse more or less optimally most of the time.<p>I tried basic raw long-context chat, various approaches of getting it to externalize the state (i.e. prompting it to emit the known state of the maze after each move, but <i>not</i> telling it exactly what to emit or how to format it), and even allowing it to emit code to execute after each turn (so long as it was a serialization/storage algorithm, not a solver in itself), but it invariably would get lost at some point. (It always neglected to emit a key for which coordinate was which, and which direction was increasing. Even if I explicitly told it to do this, it would frequently forget to at some point anyway and get turned around again. If I explicitly provided the key each move, it would usually work).<p>Of course it had no problem writing an optimal algorithm to solve mazes when prompted. In fact it basically wrote itself; I have no idea how to write a maze generator. I thought the disparity was interesting.<p>Note the mazes had the start and end positions inside the maze itself, so they weren't trivially solvable by the "follow wall to the left" algorithm.<p>This was last summer so maybe newer models would do better. I also stopped due to cost.
For a game like anchorhead, which is famous in its niche, shouldn’t Claude already know it sufficiently to just solve it right away? I would expect that its data source contained multiple discussions and walkthroughs of the game.
I expect it's <i>somewhere</i> in the training data, but it's very unlikely to be salient. A few textfiles here and there in the ocean of the Internet is nothing. If Claude had memorized the walkthrough, it would have performed better.
I would think so. I'd be far more interested in a comparison of LLMs (no internet search allowed) playing against IF games released in the past month.
Yeah, I do not find performances like this very impressive.
Honestly I am curious how it would do if it did have a walkthrough.
It's very likely the model didn't stop to question if the game they were playing was something they knew already, and just assumed it was a puzzle created for it.
Surprised you didn’t try to let Claude run context compaction, wouldn’t it rewrite its context with a summary of just the key useful information and dump any cruft?
Having read through the entire game session, Claude plays the game admirably! For example, it finds a random tin of oily fish somewhere, and later tries (unsuccessfully) to use it to oil a rusty lock.
Later it successfully solves a puzzle inside the house by thoroughly examining random furniture and picking up subtle clues about what to do, based on it.<p>It did so well that I can't not suspect that it used some hints or walkthroughs, but then again it did a bunch of clueless stuff too, like any player new to the game.<p>For one thing, this would be a great testing tool for the author of such a game. And more generally, the world of software testing is probably about to take some big leaps forward.
I was inspired by the work here, so I sat down with Claude to make something similar, for the purpose of being able to play Z-Machine (Infocom games, Inform 6/7 Z-code) and modern Inform 7 games with Glulx. So far I've tested it with Andrew Plotkin’s Hadean Lands.<p>Switchable backends, various output formats, etc.<p>In theory, I could also likely wire this up to get it playing MUDs, but I have some reservations about running that on anything except a private server.<p>My use case for this is to help test and evaluate Interactive Fiction in development, and you could even run it as a CI/CD process.<p>It's not perfect (so much Claude Coding of this), but it's an ok start for an hour on the couch: <a href="https://github.com/tibbon/gruebot" rel="nofollow">https://github.com/tibbon/gruebot</a>
This is a great idea and great work.<p>Context is intuitively important, but people rarely put themselves in the LLM's shoes.<p>What would be eye-opening would be to create an LLM test system that periodically sends a turn to a human instead of the model. Would you do better than the LLM? What tools would you call at that moment, given only that context and no other knowledge? The way many of these systems are constructed, I'd wager it would be difficult for a human.<p>The agent can't decide what is safe to delete from memory because it's a sort of bystander at that moment. Someone else made the list it received, and someone else will get the list it writes. The logic that went into why the notes exist is lost. LLMs are living the Christopher Nolan film Memento.
This is a great framework to experiment with memory architectures.<p>Everything the author says about memory management tracks with my intuition of how CC works, including my perception that it isn't very good at explicitly managing its own memory.<p>My next step in trying to get it to work well on a bigger game would be to try to build a more "intuitive" memory tool, where the textual description of a room or an item would <i>automatically</i> RAG previous interactions with that entity into context.<p>That also is closer to how human memory works -- we're instantly reminded of things via a glimpse, a sound, a smell... we don't need to (analogously) write in or search our notebook for basic info we already know about the world.
I’m currently letting Claude build and play its own Dwarf Fortress clone, as an installable plugin in Claude Code<p><a href="https://github.com/brimtown/claude-fortress" rel="nofollow">https://github.com/brimtown/claude-fortress</a>
Could you maybe have your harness limit the memory of Claude and then occasionally, when Claude specifically asks for it ("i need to remember something"), you can give Claude the full game history? Most turns, I'll bet it's okay to have a short context and maybe some notes. And then maybe once in a while it's nice to see the full chat history. Wdyt?
Using AI to drive text adventures / rogues has been pretty popular for a while now - I remember seeing a pretty dismal performance (although it was over a year ago) where somebody was trying to use an LLM to drive a game of Zork.<p>Related HN post from about 6 months ago<p><i>Evaluating LLMs Playing Text Adventures</i><p><a href="https://news.ycombinator.com/item?id=44877404">https://news.ycombinator.com/item?id=44877404</a>
It seems like asking Claude to keep notes somehow would work better. An AGENTS file and a TODO file? An issue tracker like beads? Lots of things to try.
One thing I had fun doing last year was having Claude parse some gamebook PDFs I got on archive.org, split them out into sections, and build a wrapper for presenting the sections with possible choices and just watching it play through the books by itself. You can do this with some D&D adventures as well, Claude Code has gotten good enough to run ToEE pretty well.
Very interesting, seems like a good framework to test and experiment with memory. I am curious why it wasn't able to solve it considering it is a well known game. Would be interesting if puzzle games like this could be generated so we know it's not already been trained on it.<p>I wonder if the improvements due to different memory system approaches apply in a similar way to tasks that are in its training history vs those that are not.
This would be interesting to try with local models, where the token costs and token limits are quite different.
It’s trained to interact with text transcripts, <i>it is not trained</i> to work with that memory you built for it. If it was trained to do so I might be able to break into the real estate office in ten turns.
Cool! I would like to see the game sessions.<p>Edit: they are there in the repo: <a href="https://github.com/eudoxia0/claude-plays-anchorhead/tree/master/runs" rel="nofollow">https://github.com/eudoxia0/claude-plays-anchorhead/tree/mas...</a>
Great, we can burn acres of dead forests so that my computer can play ddos games. What an exciting future!
How much energy is burnt so that you can play your video games, or whatever hobbies you have?
what else are we going to do with them? Carve them in to housing to house more humans that produce more carbon?<p>Leave them to rot?<p>Wouldn't it be best to clearcut a dead forest to allow more plants to grow to increase carbon capture?
> And like GOFAI it’s never yielded anything useful<p>Err, what?
Claude code, nethack, and tmux are fun to experiment with.
> By the time you get to day two, each turn costs tens of thousands of input tokens<p>This behavior surprised me when I started using LLMs, since it's so counterintuitive.<p>Why <i>does</i> every interaction require submitting and processing all data in the current session up until that point? Surely there must be a way for the context to be stored server-side, and referenced and augmented by each subsequent interaction. Could this data be compressed in a way to keep the most important bits, and garbage collect everything else? Could there be different compression techniques depending on the type of conversation? Similar to the domain-specific memories and episodic memory mentioned in the article. Could "snapshots" be supported, so that the user can explore branching paths in the session history? Some of this is possible by manually managing context, but it's too cumbersome.<p>Why are all these relatively simple engineering problems still unsolved?
It's not unsolved, at least not the first part of your question. In fact it is a feature offered by all main LLM providers!<p>- <a href="https://platform.openai.com/docs/guides/prompt-caching" rel="nofollow">https://platform.openai.com/docs/guides/prompt-caching</a><p>- <a href="https://platform.claude.com/docs/en/build-with-claude/prompt-caching" rel="nofollow">https://platform.claude.com/docs/en/build-with-claude/prompt...</a><p>- <a href="https://ai.google.dev/gemini-api/docs/caching" rel="nofollow">https://ai.google.dev/gemini-api/docs/caching</a>
Ah, that's good to know, thanks.<p>But then why is there compounding token usage in the article's trivial solution? Is it just a matter of using the cache correctly?
dumb question, but is prompt caching available to Claude Code … ?
If you're using the API, yes. If you have a subscription, you don't care, as you aren't billed per prompt (you just have a limit).
[dead]