You don't need it to be fair, you need it to be fun. Most video games play with randomness so that the distribution serves the gameplay, sometimes bypassing completely the RNG to balance frustration/satisfaction.
GPT can learn do to that: not providing true randomness, but the illusion of it to serve the game.
Players will detect drama being injected into their randomness (unless it is very gently and rarely done), and over time they will start to be able to predict dice rolls [edit: or even learn to shape their responses to trick the model into generating the rolls they want.] That will ruin the game.
So that's not removing randomness, it's just grade inflation. I suppose it's also secret range narrowing, since an 3% improvement from 92% to 95% is actually no improvement at all at 1.1 * (random) capped at 100%.
When you think of the content of the books in Baldur's Gate and the sheer amount of human hours poured into something just to give background to the game, this is going to save indeed a lot of time.
But on the other hand, just like open world was a terrible curse with hundred of games feeling empty, I'm afraid we will see a wave of games with a lot of generated but eventually uninteresting content because this is so easy to do.
After all, good content is coherent, and filtered. It's not just about producing a lot, it's what you don't put in the game, and what you do to glue things together.
I sure GPT will eventually be better at it than humans, but for now, having a human veto is necessary to have a good noise/signal ratio and because GPT is "good enough", being lazy is tempting.
But yeah, I can confirm GPT4 is already quite good at it and will only improve from this point.
This is going to be very interesting years.
Edit: I just made it write a full video game scenario, using BG as a competitor for basis. At first it was average, but with directions, it refined it and became quite good. Some kind of steam punk guild called the Navigators reading the wind, living in floating islands and a girl rediscovering the earth bellow the clouds and rebuilding old aliances with lost civilizations, plus a politician leader trying to prevent that and even a mentor to help her. Not bad.
I would play this, and even read it.
You can see the inspiration from "The Pillars of the Earth", "Gunnm", "Baldur's Gate" and "nausica" proving again it's mostly assembling known stuff.
But a cake is mostly assembling known stuff and it's still good. The "last of us" is just a good take on something that has been done a thousand times.
What's missing is overt structure. More or less the same reason as to why AI can't draw hands - there is no "understanding" of the fact that a hand is a rectangle with five sticks coming out of it in a specific arrangement. For a human, that is all that's needed to recognise a hand, i.e.
This is a (very crude) hand. This registers as a hand for you, because it has the structure of a hand. This is less similar to a photorealistic hand than a 6-fingered AI hand, objectively, however it fits the human understanding of what a hand is supposed to be, because it has the right kind of shapes in the right relations to one another. In the same vein, if we could get the AI to understand the structure of a story, it should be able to DM session by session, starting from a high level abstract idea and refining it based on player actions, incorporating player input into the game. I'm not sure if hierarchical abstract reasoning should be forcibly added to the model, or if it can be emergent in a specific configuration, but if we figure out how to build an AI with that, then it should be able to produce something a lot more sensible.
I had a fun conversation with GPT-4 about your ASCII hand. You're absolutely right, it had a very difficult time and thought it was a city skyline, game or puzzle diagram, and a scale, but after a few hints, it got it:
"Considering the hint that the image is part of a human being and the diagonal line is similar to the vertical lines, it appears that the image might represent a hand, specifically an open hand with fingers splayed out. The vertical lines represent the fingers, and the diagonal line represents the thumb, which is extending out at an angle from the square (representing the palm of the hand). The image is a 2D, stylized representation of a human hand.
With the provided hints and context, the interpretation of the ASCII art as a hand seems more plausible. If you have more hints or information, please share them, and I'll be happy to refine the interpretation further if necessary."
This technology appears to show great promise, but is clearly still in its infancy.
I never have a hard time to write things with words for the average stuff, which is what the AI is good at. It's always faster for me to type it. And for the hard stuff, AI cannot help (yet).
For programming having help for the average stuff saves me time. But for words? That's my natural way of expressing thoughts, why would I need speed up with the most natural thing in the world?
But then I saw some of my friends use it, and I understood.
Most people suck at writing.
In fact, they suck at communicating in words in general. I started to pay attention, and realized that most of the discussion is not even proper French or complete sentences most of the time. The tone, the gestures, grunts and context are all used by every day people to express essential basic ideas.
And we rarely express anything beyond basic idea in days to days activity, no matter how smart we think we are, life is mostly mundane stuff.
And unlike me, they didn't have to write a ton of paperwork in their life, or articles, they have no training.
They struggle with the medium like a fish trying to climb a tree.
For them, GPT is a savior for any formal writing task.
Since this is DARPA, this shows they are interested in rust, which mean we will probably have strong toolchain certifications coming up eventually, making rust even more fall in the category of "the language you want to use for serious stuff".
If you have a secret to use in your python program, such as a password or an API KEY, you can use keyring to store it easily and securely.
This uses Keychain on macOS, Freedesktop Secret Service/KWallet on linux and Windows Credential Locker on Windows. So decrypting is managed by the OS usually it's done when the user log in or at first access, and is transparent to your program.
This avoids a pitfall I see too often where people hard code their password in their script because it's the simple thing to do.
Keyring is nice because it's only an inch more effort than that. It's the best ratio "ease of use" / "security" there is. You manipulate it like a key-value store, don't bother about encryption, prompting the user for decryption or anything. And you don't risk committing your secret or letting linger at the wrong place. Plus you benefit from your OS security team effort.
You can also use "os.environ" to read the key from env vars in production and only use keyring if it's not there.
Can confirm. This week a physicist friend came by to give it a try. He asked some questions related to his domain. I was fully expecting GPT to bullshit its way through it, from experience with GPT 3. But to my surprise, and to his, it got quite decent answers.
because they could be viewing all their private communication/media on their device and that causes a lot of anxiety cause they won't confirm or deny if they have access.
All of the above, plus, maybe, don't let people see you put in your pin.
don't take security for granted. Be careful.
I guess the lesson I learned is you need to pay attention to security.
GPT can learn do to that: not providing true randomness, but the illusion of it to serve the game.