Hebe

I am, as I think basically everyone in the world not on the OpenAI or NVIDIA payroll must be, heartily sick of LLMs. Substantively I think they suck when employed for anything resembling creative work; environmentally, turns out the companies have been lying and things are even worse than the already-quite-bad picture that’s been painted; economically, they represent yet another attempt to consolidate wealth and power with the owners of capital at all of our expense. But more than that, the discourse around them is wearying – having to think about this stuff and engage with those conversations even in leisure activities is pretty frustrating. So I should disclose that I went into Hebe, which bills itself as a puzzley adventure through Greek mythology, and also commendably acknowledges that ChatGPT wrote some of the text, with some trepidation, both about what I would find in the game and what I’d need to write about it.

There’s good news and bad news, I suppose. The bad news is that those Hebe-jeebies were justified (Hebe’s the Greek goddess of youth, so her name rhymes with Phoebe – geddit?). All of the game’s room descriptions are overlong and mention lots of unimplemented objects, the prose glops about, weighing everything down like oatmeal laced with lead, and by the end my eye would start twitching whenever I read the words “flickering” or “serene.”

The good news, such as it is, is that I didn’t find it too hard to ignore most of that. The game is clogged with empty locations that are just there to pad out the half-dozen places boasting self-contained puzzles, and so it’s relatively easy to just glide through them and concentrate on the interactions that seem like they were written by a human being, just as I skimmed the overlong cutscenes. Similarly, the author’s offered a helpful INVESTIGATE verb that tells you what items actually exist to be interacted with, so you don’t need to play the guess-the-hallucination game with ChatGPT if you don’t want to. And turns out that under all the AI cruft is – well, a perfectly ordinary undertested, wonkily implemented game that manages to boast a bit of charm, the kind of thing that would be a perfectly respectable starting point for a new author who didn’t yet understand the level of polish a parser game requires, but for the LLM use.

Let’s start with that bit of charm, as I’ve given Hebe a bit of a hard time so far. It’s clear that the author took the theme seriously, bringing in a lot of fringe detail from Greek mythology including but not ending with the choice of protagonist – Hebe’s both the second-most-famous cupbearer to the gods and the second-most-famous wife of Heracles, so I appreciated her underdog energy. The game proper involves visiting various sanctuaries and temples across northern Greece to find the Olympian Gods, who’ve been defeated after a surprise attack by the Titans and left depowered and chained behind various puzzley barriers – and then venturing into the underworld to find Heracles and bring the fight to the Titans. Again, there’s a fair bit of attention to detail, with numerous cities and ports implemented, a visit to the Pytheia, minor naiads given a supporting role, and plenty of obscure bits of myth getting a name-check.

The thin implementation means that the names are often all that you get, however – “you see nothing special about the Charon’s boat” is an enthusiasm-killing phrase to read. And the seams between the LLM stuff and the chattier human-written propose are sometimes comically sharp:

> x aigle
Aigle radiates with a golden glow, her hair like cascading sunlight, and her eyes shimmering like the first light of dawn. She is the embodiment of brightness and warmth, her presence illuminating the garden with a serene, golden aura.

> talk to aigle
“I’m so relieved you’re safe! Now go show them what you’re made of, Hebe! Just like old times!”

A game written entirely in the latter style might feel a little silly, but could have some zip to it; the juxtaposition with ChatGPT’s overwrought descriptions just creates bathos.

As for the substance of the adventure, the puzzles are largely old chestnuts. There are a couple of codes, a put-the-right-object-in-the-right-place one with a poem providing the hints, a guess-how-heavy-the-unmarked-weights one… None of them break new ground, but the classic are classics for a reason and they could be fun to work through. Unfortunately bugs and incomplete implementation make many of them way harder than I think they’re intended to be – the weight one stymied me for a long time due to the fact that I wasn’t clear that there were two scales, not one, and I couldn’t directly interact with the first one (“which do you mean, the scale or the small scale?”); the object-placement will softlock you unless you get it right first try, because the game incorrectly thinks one of the slots stays full even after you remove an item from it; and the endgame seems to have gotten especially little testing, as accessing it requires going through an unmarked exit (tip to other players: try IN/ENTER when you’re near the Necromanteion) and then the climactic conversation with Heracles is made awkward by a YES/NO are-you-ready-to-proceed choice that doesn’t work (to continue with the service journalism: say YES and then manually type DOWN afterwards). And there’s a lot more besides; see the transcript for the gory details.

But again, pretty much all of these issues are familiar ones – heck, I’ve committed some of these sins myself, and but for lucking into experienced testers for my first game could have wound up with similar egg on my face at my debut. And as I’ve said, I do think Hebe comes from a place of real authorial excitement, some of which occasionally comes through. So it would be easy enough to just wrap up with my typical remark in these cases, about how I hope to play a much better second game from this author. Which I do!

My weariness at LLM discourse can’t prevent me from adding, though, that I think use of ChatGPT was an especial disservice to the game. Beyond weighing down the enthusiasm that’s often one of the best elements of a debut game, the use of AI I think might have created some bad habits that contributed to the overall weak implementation. As I mentioned, the long location descriptions include a fair bit of (bland) scenery detail, and call out sounds and smells, so it’d be easy enough for an author to review what ChatGPT spit out and feel like their bases were covered – but none of the scenery is implemented, and LISTEN and SMELL return their default responses throughout. The difference between the prose styles also makes it really obvious to the player where there’s stuff they should be paying attention to and what they can safely tune out – as an author, though, I don’t think you ever want the player shutting off their brain. I don’t think I can say with a straight face that the version of Hebe that didn’t use an LLM would have been significantly higher-quality, but it would have been a clearer reflection of the author’s vision, and probably a much much better learning experience to boot.