Over the past couple of years we’ve seen AI improve greatly at both image and text generation. While it can’t yet - and may never - reach the level of the best humans in any discipline, it’s not hard to imagine a future where it has.

With the state of things today, AI can already be used as an augmenting tool for human work. An artist might generate an AI image and edit that into exactly what’s wanted, rather than working from scratch. A writer might write their own text but take some editing suggestions from the AI. Already today, we have to wonder if images and writing are human-made.

In a world where AI is slightly better than the best humans at these things, without our help to get there, the tool becomes more like a parent. But an intelligent AI parent will always be available to help, and will always be better than the child. There’s no impending moment of ‘leaving home’ that we’re training for.

A popular sentiment is that we’ll still seek out human art because we desire that spark of humanity, the meaning behind a work that comes from a life lived. But on the Internet, anyone can lie. For now, it’s often possible to pick out AI images and text simply from the look and the tone, but there’s no special watermark. While benevolent actors may add one, we cannot make one that’s impossible to remove.

An artist releasing “real, human, no AI used” content in that world will be competing against other “real, human, no AI used” content that secretly did get help from AI. In fact, I’m sure we already live in that world to some extent, and smarter AI will only make the problem worse.

Say you decide to write a poem. You might want to do this purely for yourself, to express something meaningful on paper. The best art usually doesn’t have a target audience anyway, because catering to an audience always bounds the art within the desires of that audience. But it still feels a little demotivating to know that you could always get an even better poem if you told the AI your thoughts and asked it to write one. It still feels a little demotivating that if you share the poem, nobody can be sure that you wrote it on your own.

I suppose the latter has always been the case to some extent, with ghost writers and mentors and so on. But doing it yourself was still more believable, then. Funny that in a world where a superintelligence could solve our technological problems, it might also bring a feeling of pointlessness to our renewed free time for creativity.

A few predictions, if creative AI continues to improve:

• We will see an increased focus on areas where AI is weaker, therefore giving implicit proof of human creation. Physical work like sculpture. Live performances. New formats for poetry and music, because even an AI that excels at those tends to have trouble creating something truly new - sticking close to the forms of the content it was trained on.

• New types of generative art will emerge that specifically work alongside AI. I’m not sure how these might look yet.

• There will be scandals - “Artist exposed using AI” YouTube videos etc - and plenty more passing their work off as human that aren’t caught. Some doing everything with AI, some only getting a little polish from it.

• “Verified human” chat communities won’t really exist, as even a “verified” person can simply turn around and use AI anyway. But for artwork there will be a great increase in showing in-progress photos and/or video of your work, as the only way to really prove it’s human made.

• Art as an economic pursuit may be all but obliterated, but that will never stop art being made. No artist makes art purely for money.

• There will be many who use AI (openly or otherwise) even when they’d rather do without, and people will ask them why they’re railing against AI while also using it, in the same way they ask environmentalists why they took a flight, or anti-consumerists why they own a smartphone, or people asking for higher taxes why they don’t just pay more tax; and always the answer will be the same: That it would put them at a disadvantage to everyone else. So you look for systemic change at a higher level, but then a whole group becomes disadvantaged to the rest of the world, and people will work around it anyway, and good luck getting the whole world on board with reversing something once Pandora’s box is open.

I showed Claude 3.5 Sonnet, (a general-purpose LLM AI from Anthropic) this article and asked if it would like to write a short poem based on it. Here it is, unedited:

In the Gallery of Tomorrow

The finest brush now paints in code
Each pixel perfect, neatly stowed.
Our children's children might not see
The worth in imperfection's plea.

Yet in these spaces between machine and mind
In cracks algorithms cannot find
We plant our gardens, wild and free,
Knowing they could grow more perfectly.

For what is beauty, if not the way
We choose to fail a little more each day,
Reaching past the perfect light
For something real enough to write?

balatran.com

“I am now allowing the client to guess at the results of the users movement until the authoritative response from the server comes through. This is a biiiig architectural change. The client now needs to know about solidity of objects, friction, gravity, etc. I am sad to see the elegent client-as-terminal setup go away, but I am practical above idealistic.

The server is still the final word, so the client is allways repredicting it’s movement based off of the last known good message from the server.”

- John Carmack, 1996-08-16

On the contrary there wasn't any information around proving its use in Duke Nukem 3D, except that the Wikipedia page linked to a recent interview with Ken Silverman - the creator of the Build engine that Duke 3D runs on - in Retro Gamer magazine from 2018. In it he clearly stated that Build engine did it first:

"People may point out that Quake's networking code was better due to its drop-in networking support, [but] it did not support client side prediction in the beginning," he explains. "That's something I had come up with first and implemented in the January 1996 release of Duke 3D shareware."

- Ken Silverman, 2018-06-14

I thought that was pretty interesting and also pretty unfair for Ken that everyone’s got the wrong idea that it’s QuakeWorld. I later commented on it in a related discussion on Hacker News, also mentioning the fact that the source code was released in the early 2000s so it should now be possible to just go and find the prediction code as 100% confirmation.

Sure enough, someone replied that they’d found it, specifically in game.c with the methods domovethings(), fakedomovethings(), and fakedomovethingscorrect().

To be a good citizen I went back over to the Wikipedia page and added a link to the source code to help solidify the claim. But while I was there I went back and read the interview again, and noticed a part I’d skimmed the first time:

"It kind of pisses me off that the Wikipedia page article on 'client side prediction' gives credit to Quakeworld due to a lack of credible citations about Duke 3D"

- Ken Silverman, 2018-06-14

I wondered if and when it had been changed from saying Duke 3D to QuakeWorld in the past (before eventually being changed back again sometime after the interview), so I went and had a look through the page history. It had been changed a few years ago. I found the person who had removed it due to lack of any citations. It was me.

Sorry Ken.

That argument is always refuted by the statement that of course you don’t come across games that are obscure, by definition. Which is itself then easy to refute by saying yeah, but surely you’d still come across one occasionally if you went looking.

Let’s go looking.

First, let’s take all the games on Steam.1

Then let’s sort them by Evan Miller’s recommended weighted method from his website. That is, a lower bound of Wilson score confidence interval for a Bernoulli parameter. Basically, look at the score, but also consider how confident you can be in that score - i.e. how many reviews it has in total.

Now we have quite a nice list of the best games on Steam:

It’s very similar to Steam’s official top rated games page, with the notable exception of #1 game Wuppo and a couple of others. This is because Steam’s official page sorts the results into “buckets”, and to get into the top “Overwhelmingly Positive” bucket you need a score of at least 95% and at least 500 reviews. It’s essentially a more primitive way of weighting the results so that games with few reviews don’t top the charts. But it means that even though Wuppo has a score of 100%, since it only has 200 reviews total it has to slip in later past all the Overwhelmingly Positive games.

Next we want to fish out any “hidden gems”: Games that people really like but that aren’t very popular. Here’s a basic formula I made up:

• Take the weighted score above.

• Leave games with less than 150 reviews2 per year alone. Rate of user reviews is a decent gauge of popularity.

• If the game gets more than 150 reviews on Steam per year, start penalising its score. I slowly rolled the score off linearly, penalising it 0.0002 per additional review.

Lastly, I added a release date limit of 1 January 2008. That one is a bit of a hack that shouldn’t really be needed, but the problem is I’m basing “reviews per year” on the stated release date, but some games came to Steam recently yet list their original, earlier release date. This meant that without the date cutoff, re-released games from the 90s ended up looking like the most hiddenest gems ever.

Some older games that have been recently re-released on Steam, and have used their Steam release date as their release date, have ended up on the lists. I decided to leave them in since I would have had to go through manually to find them all. Technically their data is correct, for Steam.

Putting all that together, here’s a list of top Hidden Gems on Steam:

Wuppo is again a good example of this system. It’s looking like it’ll get around 329 reviews per year since it has 200 reviews now, but it’s been out less than a year. This is above the 150 reviews/year “no effect” cutoff, but the penalty also comes in slowly; so because it’s such a good game and it’s not that popular it still gets in at #15. Castle Must Be Mine and The House in Fata Morgana also make it in from the previous list, with the latter shooting up to the top spot now that all the big popular games are gone.

There’s a lot of obscure stuff, a few cult games that you might have heard of but didn’t buy, obviously (the Blackwell games, Escape Goat), and Bejeweled 3? Apparently Bejeweled just doesn’t sell on Steam. Although it actually has over 900 reviews, it’s just that that’s only 144 per year - it’s pretty old too. You could choose to cap the total number of reviews in a Hidden Gems formula to avoid games like this.

I think it’s interesting that the top Hidden Gems here tend to be games in less popular genres, or VR games requiring uncommon hardware. There are three theories I can think of for that. One theory is that great games in a popular genre always tend to sell better than great games in obscure genres, keeping them off the Hidden Gems list. The second is the exact opposite: That great games in a popular genre tend to sell worse due to the flooded market (apart from a few big successes), where having less reviews would push them down in the weighted lists below the niche games. My final theory is that games in popular genres may have more critical reviewers, since the players have a lot more games to choose from. In less popular genres, or in VR, people may be happy to be getting a decent game at all!

There’s a lot more investigation that could be done here. In a more complete Hidden Gem finding system, you’d probably want to be able to view a list per genre, per tag, or based off what else you’ve played. That way you could view the complete list if you wanted to explore whole new genres you might like, but you could also easily find little-known games in your favourite genre.

There are also various ways you could decide what a Hidden Gem is exactly. In my system above, since it’s based on the weighted score, a game with 95% positive over 150 reviews will still be considered “better” than a game with 95% positive over 100 reviews. Only after 150/yr does it start trending down.

Maybe you’d rather treat all games that fit in the “Hidden Gem” range equally. Here they are with the same rules as before (plus one extra - see below), using an unweighted score:

I also added a 20 total (not per year) reviews minimum threshold to this one, since weighting doesn’t take care of it in this version and games with a very low amount of reviews just can’t have reliable scores. This variant doesn’t work so well in practice because all the 100% games get lumped together first, right down to the 20 threshold, before the 99% games start. Without the 20 review threshold you’d have a ton more.

Looping back to the initial statement of this article, how many games are out there on Steam that have great reviews but little exposure?

The answer is A LOT.

For instance one game I knew of already was NeonXSZ. It’s a 6DOF shooter made by one person, with an 89% positive score on Steam and 63 reviews. It’s always made me a bit sad that it got so little exposure while being so good, but it’s telling that in the weighted hidden gems calculation above, it’s #1147.

There are over one thousand near-unknown games on Steam, in all sorts of genres, that are subjectively better than NeonXSZ.

It might be nice if Steam had a way to easily find lesser-known but well-reviewed games like this, so more people could try them and see if they deserved a boost. This is the category of broken dreams. Lars Doucet does some of this sort of theorising and a bit of my knowledge of this stuff comes from his work.

One thing I will note, is that while there are games in popular genres that fall into the Hidden Gems category, so it’s not a guaranteed ticket out of obscurity, it definitely seems more likely for games in less popular genres to get in there. If I had the data to split games by genre without manually going through them all, it would be interesting to see how much less likely a highly-rated FPS is to languish in obscurity versus say a visual novel. In fact it could be that popular genres aren’t any less likely to end up in obscurity, but instead they have harsher critics who prevent them from being highly rated enough to make the top of the list! There are plenty of hidden non-gems.

Popular genres have a bigger audience but they’re also more flooded with games to choose from. Wuppo is near the top of the Hidden Gems list and it’s a platformer, one of the most popular genres of all. But there are so many platformers.

Finally, a couple of bonus lists.

There is actually some really interesting stuff in there. Dream Quest for instance is a deck building game with a glowing review from the creator of Magic: The Gathering. Although actually the Steam-listed release date for that one is set earlier than its actual release date on Steam, skewing it up in the charts.

After these there are many more games with great scores but even fewer reviews. Eventually the review counts become too low to get any real meaning from the scores.

These are of course the most popular really bad games. In the depths below these there is an enormous supply of badly rated games with few reviews.

And a bonus chart:

When Aunt Magda comes to visit and decides to make tea, will she perform this step? Highly unlikely. I guess I’m supposed to tell everyone who uses the jug, or request they read the manual. Heaven forbid someone buys it second-hand without the manual, or just doesn’t read it. I’ve lifted it without switching the level off once or twice; if something terrible happened I didn’t notice.

I have no information to go on for how important this is, and it’s a theme I see everywhere. Information on use and maintenance in the manual provided is often overly strict and thorough - seemingly designed to remove all liability and responsibility from the maker, not to mention sometimes wrong or applying to some subtly different product.

We have a problem as a society with not explaining why, from teaching maths to superstition. This is just one symptom of that problem.

Yet these things sorround us, the only real solution being to own less things.

One can ask an expert, but experts in a field feel incredibly surprised that anything works at all. The expert’s PC/car/valve radio collection requires constant tinkering to avert a disaster that somehow rarely happens to people who don’t know it can happen.

The one who truly knows a thing is cursed deeply with the power not just to fix problems but also to prevent them. Our possessions are a cup that slowly fills with trouble - and the more you understand the thing, the more transparent the cup. “My cup overflowed, can you fix it?” one may ask the expert. The expert glances constantly at his own cup: It has liquid in it, and must be emptied. How ignorant the masses are to the expert. How paranoid the expert, to the masses.

Naturally a bed bug expert tells us to check for bed bugs around the house once a week. He sees bed bugs all the time. A door-knocker from a security company tells me my house’s safety is worth $20/wk, she constantly hears about break-ins. I’m certain there are hundreds of things I should be checking and maintaining every week around my house. Some of them are ridiculous - my garage door manual says it should be wiped down every two weeks. It probably hasn’t been wiped down in half a decade and it’s doing fine. Life is a risk/benefit analysis and everyone’s shouting about their own thing’s risk being high (though it’s totally safe, if you do these three easy steps daily and these other two weekly).

In fact the expert frequently becomes so embroiled in the fragility of his domain that he overcompensates. Constantly cleaning his lens even though it takes a crazy amount of dirt to have any visible effect on image quality. Buying expensive cables even though a coathangar will do.

So again we are left with uncertainty and a longer list for use and maintenance than we care to carry out. Maybe we wouldn’t have bought the product if we’d known how much care it would take. Often there’s some expensive industrial version that’s simpler, yet somehow better, more reliable. You pay for what you get I guess. But I wish I knew what I was getting sometimes.

Everything has its extra tasks. Unplug your charger - not just your device - after using it or it wastes power. Install the latest updates or you’ll be in danger. Turn the taps off carefully because that water knock can damage the pipes. Don’t turn those CFL bulbs off if you’ll be back in a couple of minutes because turning them on and off quickly can damage them (at least LED bulbs are coming to save us there).

The solution du jour is for some startup to look at those annoyances with Product A, and offer you Product B: An ostensibly time-saving device with even more complexity and even more points of failure.

Products like the Juicero or the June or whatever nonsense are complexity presented as fake simplicity. Too many dependencies to last. Products that are really designed for life - SpeedQueen, MagiMix - have simplicity and reliability at all levels.

Though actually I bought a MagiMix and the pulse button was faulty on arrival. Maybe I should go back to sticks and rocks.

I want to talk about the Nokia N9 (2011) alarm clock application because it’s a really nice example of thoughtful, functional design – and because it’s only on the N9, so a lot of people won’t have seen it. It’s nice when simple things are done well.

I don’t know how many of these ideas are borrowed from previous apps (the N9 is contemporary with the iPhone 4S), but either way it’s implemented exceptionally well.

The main screen is informational. The full information on each alarm you have set is shown, and each can be turned on or off with the slider instead of having to touch on the entry to edit it. The plus adds a new alarm, and the Alarms section can scroll if there are more than two.

I don’t think I’ve seen another alarm where you can set any time you want in two touches or less. In case it wasn’t intuitive enough, the clock dial here mirrors the size of the analogue clock on the previous screen. Just in case you still don’t get it, the alarm time is shown above the dial. But everyone gets it, because it’s simple.

Touching and dragging on the minute marker moves it to the nearest minute, but touching directly at a spot on the circle moves it to the nearest five minutes. So you can change your alarm from 7:00 to 7:30 with a single touch without accidentally setting 7:31. If you actually want 7:31, just drag it around.

The AM/PM selection was also intuitive for me but may not be for everyone – you need to drag the hour circle all the way around, as if time is passing. You can also do it in a couple of touches.

When you drag, the phone plays a quiet click sound and vibrates slightly as you rotate, as if you’re clicking a timer around. When you touch, it animates quickly but smoothly to the new position. There’s so much attention to detail here.

The More options screen has the options you’d expect, and shows the current settings.

The Repeat selection is nice. Select as many or as few as you want. There’s no such thing as an invalid selection: Rather than “Never” and “Every day” being settings themselves, which would cause conflict with the individual day settings, they’re more like indicators which also serve as group-deselect and group-select actions.

Touching on “Every day” doesn’t cancel the others, it selects all the day entries automatically, as well as the “Every day” entry itself. Deselecting a day will also deselect “Every day”. “Never” is automatically selected when everything is deselected, and “Every day” is automatically selected if you select each day individually.

Once you’ve set an alarm, a popup tells you how much time is left until it goes off. I was surprised that it even predicts daylight savings changes for your location, adding or removing an hour as necessary. I always thought that the time was just adjusted by the phone company when it switched, but somehow it knows. If you try to set a time that won’t exist because of a daylight savings switch, it knows that too.

When the alarm goes off, it stops automatically when you pick up the phone. And this being a Nokia phone, you can turn the phone off completely at night and it still turns itself on in the morning for the alarm.

It is missing one feature that I miss from Symbian: There’s no option to keep the phone on after you stop the alarm. So if the phone was off, it’ll go back off, and then you have to turn it on again.

I also appreciate the look they’ve gone for with Meego (the N9’s OS) in general. It’s near the flat design aesthetic that was just becoming popular at the time, but with more shadows and gradients to give everything some depth.

It’s unfortunate that simple applications are often worked on just enough to be functional. Sometimes we get more features that aren’t needed rather than polish to existing ones. Can we get more stuff like this instead?

1. The .enabled property of a script (which is also the checkbox for disabling a MonoBehavior in the inspector) will only prevent Start(), Update(), FixedUpdate(), and OnGUI() from executing. e.g. OnCollisionEnter is still called even on “disabled” scripts. Contrary to the current official documentation, Awake() is also still called.

2. Child particle systems fire as one automatically with the parent.

3. Invoke takes Time.timeScale into account. So does WaitForSeconds.

4. If your PSD files appear as a jumbled mess, make sure Maximise Compatibility is turned on in Photoshop.

5. If lightmapping looks all glitchy, try ticking “generate lightmap UVs” on the model to get Unity to generate its own. Also “lightmap static” needs to be set on the object.

6. CalcSize includes rich text markup in its calculation, even for styles where it isn’t shown.

7. Random.Range is min inclusive, max inclusive for floats, and max exclusive for ints. So Random.Range(0, 2.0f) will return from 0 to 2.0, but Random.Range(0, 2) will only return 0 or 1.

8. Modifying prefabs (via the Project view, or by clicking Apply in the Inspector) doesn’t actually write changes to disk until after you’ve saved. So make sure to Ctrl-S before committing changes or copying a project.

9. When scenes change, references to objects in that scene (that aren’t using DontDestroyOnLoad) become Null, but only in Unity’s overloaded null check. If you have a reference to it, MyObj == null will be true, !MyObj will be true, but MyObj ?? will return MyObj, since ?? can’t be overloaded. See the blog post here. Unity may be changing this in the future.

10. Don’t Instantiate anything inside OnDestroy() or it’ll persist in the world after play stops in the editor. Unity even warns you if you try it. Of course, OnApplicationQuit() gets called before OnDestroy(), so you could always set a flag that stops it from happening in the case wher… NO. That rabbit hole is too deep.

11. Destroying a script (or its GameObject) automatically ends any Coroutines or Invokes on it. Deactivating a GameObject (SetActive(false)) kills any coroutines but has no effect on Invoke or InvokeRepeating – they just keep running. Coroutines are gone, not just paused – calling SetActive(true) won’t bring them back. Setting .enabled = false on a script has no effect on Coroutines or Invoke/InvokeRepeating.

I’m also interested in quirks that other people know about. I posted these on the Unity forums but it didn’t get any interest.

Bonus: Fog equations to match the camera far clip plane with the fog cutoff point:

if( RenderSettings.fogMode == FogMode.Linear ) {
	camera.farClipPlane = RenderSettings.fogEndDistance;
} else if( RenderSettings.fogMode == FogMode.Exponential ) {
	camera.farClipPlane = Mathf.Log( 1f / 0.0019f ) / RenderSettings.fogDensity;
} else if( RenderSettings.fogMode == FogMode.ExponentialSquared ) {
	camera.farClipPlane = Mathf.Sqrt( Mathf.Log( 1f / 0.0019f ) ) / RenderSettings.fogDensity;
}

Extra Bonus: Someone linked me to this glorious Unity Answers response by user Wolfram: http://answers.unity3d.com/questions/217941/onenable-awake-start-order.html
Did you know that OnLevelWasLoaded switches calling order with Awake if a script is disabled (at least as of Unity 3.5)?