Little Learning Machines, PS5 Pro, Roblox AI and More | AI and Games Newsletter 11/09/24
I dig into last weeks big case study reveal.
The AI and Games Newsletter brings concise and informative discussion on artificial intelligence for video games each and every week. Plus summarising all of our content released across various channels, from YouTube videos to episodes of our podcast ‘Branching Factor’ and in-person events like the AI and Games Conference.
You can subscribe to and support AI and Games right on Substack, with weekly editions appearing in your inbox. The newsletter is also shared with our audience on LinkedIn. If you’d like to sponsor this newsletter, and get your name in front of our audience of over 6000 readers, please visit our sponsorship enquiries page.
Hello all,
back with another edition of the newsletter, and this week we have a bumper issue dedicated to the cracking indie game ‘Little Learning Machines’ by Transitional Forms.You may recall last week that Little Learning Machines sponsored this here newsletter just as it dropped a big content update - all for free might I add. Last week we expanded on that collaboration with a deep dive into the making of the game itself, including a dedicated episode of AI and Games on YouTube, plus an extended interview with some of the developers on the Branching Factor podcast.
So for this issue I want to dig into it some more and give you an insight into why I was excited to collaborate with the games developers Transitional Forms here on AI and Games. Plus give you plenty of reason to check out the game while it’s got a lovely 40% discount on Steam.
But of course as with any edition, we have plenty of content to dig into, with announcements, news stories and other little bits of excitement. Let’s get into it!
Announcements
Before we dig into the main stories, a quick round-up of important announcements AI and Games related and otherwise!
AI and Games Conference 2024
Another week, another update on what’s happening with the AI and Games Conference, running just two months from now on November 8th at Goldsmiths, University of London. Our call for submissions is now closed, and I want to take a moment to thank everyone who submitted to us. We received enough submissions to run a 4 day conference if we so wished - next year perhaps? For now we’re now knee deep in the review process right now, and we will be notifying applicants in the coming weeks.
Plus a reminder that early bird ticket sales close this week! At the time of writing we are running out of all ticket types - industry, student/indie, and professional - as we’ve had a fantastic reception thus far. We’re looking forward to welcoming developers from Avalanche, Ubisoft Reflections, Sony Interactive Entertainment, Splash Damage, Red Kite Games, Creative Assembly, Hangar 13, Kythera AI and many more.
Other Announcements
I’m pleased to announce I’ll be speaking at this years ‘Next Level’ event at King’s College London: a day of talks looking at “the exciting, surprising and playful future of games from AI researchers, game developers, digital artists and more.” It’s free to attend, but you need to register. Head to the website to find out more.
Upcoming Events: I’m attending these, and speaking at most of’em. You should check them out.
2024 iGGi Conference, York, UK, September 11th and 12th
GamesIndustry.biz HR Summit, London, UK, September 18th
NEXUS Games Conference, Dublin, Ireland, September 25th and 26th.
‘Large Language Models (LLMs) for Game Designers’ with Gamaste.
Paris, France, October 7th. Lyon, France, October 9th.Next Level 2024, King’s College London, UK, October 25th.
AI and Games Conference 2024, Goldsmiths, University of London, UK, November 8th.
Supporting AI and Games on Substack: A quick note to say thank you everyone who signed up as a paying subscriber of AI and Games after my posting here last week. It’s very kind of you! Stick around for this months (sponsor-only) digest coming next week!
AI (for Games) in the News
A couple of the recent stories that caught my attention, and may also be of interest to yourselves:
Sony Unveils the PS5 Pro, with AI Upscaling [VGC]:
The worst kept secret in video games was announced yesterday: the updated version of the PlayStation 5, the PS5 Pro, will come to market on November 7th. It includes a bunch of new features, including a 2TB SSD for storage, a larger GPU and advanced ray-tracing, plus the ability to ‘boost’ PS4 games for better performance. In amongst all of this, is also the AI upscaling technology PSSR.
Now we actually discussed this back in April when the rumours first cropped up, and more notably documents were leaked of the underlying AI technology. PlayStation Spectral Super Resolution (PSSR) is similar to the DLSS technology deployed in the NVIDIA graphics cards, and is a technique that uses machine learning technology to upscale from 1080p to 4K resolutions.This is a quite a decent upgrade, but you will most certainly have to cough up the cash for it, with the console having a staggering $700 / £700 / €800 price tag - and that’s without a built-in disc drive either! When I discussed this back in April I mused who this was aimed at. Particularly when you consider how lacklustre the PS5’s slate of 1st party titles has been this year. But this is clearly targeting the most ardent of PlayStation fans. Sony already seem to be making it difficult for new fans to join their ecosystem (having increased the price of the base PS5 models in 2022, raising it again in Japan last month, and now the controllers are more expensive too), and this is just doubling down on gouging their existing fan base.
But hey, all you parents out there: now’s a good time to source second-hand PS5’s on eBay and Facebook marketplace!Roblox Continues it’s Expansion into Generative AI Tools [Roblox]:
The game ecosystem that everyone loves to forget, Roblox continues to be one of the largest and most pervasive gaming platforms for younger demographics. It’s also arguably the platform that has most aggressively pursued generative AI tools for end users. At their latest developer conference they highlighted a new generative model for 3D assets using text, video and 3d prompts - shown in the video below. This sits alongside existing work in their code assistant and texture generator they released in 2023.
SAG-AFTRA Secures AI Agreements with Studios Big and Small [GamesIndustry.biz]:
The SAG-AFTRA strike regarding protections for performance actors - including risks of their data being exploited using AI - has seen some movement in the past week. 80 studios have signed tiered-budget or interim agreements to protect actors, all of which are working on upcoming titles. Meanwhile Tencent’s Lightspeed LA was the first ‘big’ studio to sign up the week prior.Investors are Getting Squirrely About AI’s Future Prospects:
While I’m steeped in this space, I’m always an AI sceptic, and have argued we should approach the current AI hype with caution Largely in part, because the big generative boom - while impressive - is in many respects fundamentally limited due to how the technology works. Open AI has more or less had to scrape the entirety of the internet - laws be damned - to build GPT: a system that on its best day is still unreliable and flimsy.As discussed in an opinion piece by Michael Hiltzik in the LA Times (link), and a more meatier analysis by Brian Merchant over at Blood in the Machine (above), investors are starting to take notice. A lot of money has been pumped into building tools that, while impressive in certain use cases, are far from reliable in many others, and that is contrary to the narrative that many a large scale AI company has presented.
Upcoming Games
I figured let’s start keeping an eye on recent and upcoming releases, and point stuff out that I’m interested in, just in case people aren’t aware of them.
The Casting of Frank Stone [PC, PS5, Xbox Series X|S - Sep 3rd]
The latest story-driven game by Supermassive Games, famous for the interactive slasher Until Dawn (which is getting a remaster/remake in October). I highlight this game, despite it’s weird name, given it takes place within the Dead by Daylight universe - a game that I have a strong affinity for. So I’m keen to check this out now that I’m back from vacation. It appears to be doing some interesting stuff in letting players explore different narrative branches as a replay feature, an idea that I’ve longed for with games of this ilk for many years.
Dumb fun by the looks of things. Warhammer 40,000: Space Marine 2 [PC, PS5, Xbox Series X|S - Sep 9th]
I recently put the original Space Marine from 2011 - which was more or less ‘hack’n’slash meets 3rd person shooter with the Warhammer 40K license’ - into my backlog and the new entry looks even better. With all the big weapons, dumb dialogue from morally-questionable-British-religious-zealot-space-nazis (ooft, say that three times fast) and shooting ugly aliens that you want.Marvel Vs. Capcom Fighting Collection - Arcade Classics [PC, PlayStation 4, Switch - Sep 12]
One of the best fighting franchises of all time finally getting a modern re-release. Sure, I own an arcade cabinet with all of these games on it - X-Men: Children of the Atom, Marvel Super Heroes, X-Men vs Street Fighter, Marvel vs Street Fighter, Marvel vs Capcom, Marvel vs Capcom 2 and The Punisher - but it’s not exactly portable. While there is no Xbox release at this time, it will be coming to the Series X and S in early 2025.
The Big Story: The Making of Little Learning Machines
So here we go, this weeks big story all about Little Learning Machines. I was really excited to collaborate with Transitional Forms to help get this game out there and showcase what it’s all about, as it is certainly one of the most interesting examples of applying AI in gameplay for quite some time.
Little Learning Machines is a puzzle game, in which you travel to a collection of islands that have tasks that need to be completed. In order to complete those tasks, you rely on your ever-growing collection of ‘Animos’: cute and colourful little robots that will step out into the world and collect crystals, chop down trees, plan seeds, water flowers, and even pet dogs - because that’s a key requirement for a videogame in 2024.
But the problem is that by default, your Animos don’t know anything: they either just stand around idle, or wander around randomly. Players must train the Animos to solve problems, and this is what makes Little Learning Machines really interesting. It’s not just about solving puzzles, it’s about training your Animos using AI to figure out how to solve them for you.
Now as mentioned already, we had a dedicated case study for this game drop on the site last week. But I wanted to spend this weeks newsletter unpacking the game some and discuss what I find so interesting about it. Read on for more info, and then I’ll provide all the links at the end for you to dig into it some more.
Early Impressions
As we discussed in the recent episode of Branching Factor, I stumbled upon this game last summer while attending the AI and Games Summer School in Cambridge. As discussed in the vlog I released at the time, I was invited to attend the school to moderate a panel and generally be a professional nuisance. Given my visibility and profile in the game/AI community, the organisers have asked me to attend the past couple of years given a) we’re all friends but also b) I chat with the students to foster connections, and make them feel welcome in the space (particularly given some of them know who I am already).
One of the those attendees was Dante Camarena from Transitional Forms - the lead programmer on Little Learning Machines. Dante came over to chat with me at the closing reception of the event, and asked if I would be interested in looking at the game he was working on, and so myself and a few others clambored around a laptop - trying to make sure we didn’t spill any of our drinks on it!
At this time, the game lacked a lot of those final tweaks and features that really helped the experience shine - a point that art director Nick Counter discussed in our case study episode - but the crux of the idea was there, and it was working. Setting up an Animo in the cloud, getting it to train on problems indicative of the challenge you want it to solve, deploying the trained Animo in the world to complete the task. Even in this still rather rough form I could see the potential of that loop of observing, training, and then deploying. Given it was and still is a rather refreshing application of AI in games when compared to the current market.
Machine Learning as a Game Mechanic
As discussed in the case study about the game, the idea of training AI to complete tasks within a game isn’t new. Creatures from 1996, and Black & White from 2001 explored this idea in more rudimentary ways. They both trained neural networks to create in-game behaviour using learning algorithms of some sort. These are in fact two notable absences from the AI and Games case studies. While I’ve yet to find definitive evidence of Black & White’s inner workings (if you have, leave a comment), I simply refrained from covering Creatures given my friend (and conference co-organiser) Alan Zucconi made a great video discussing it a couple years back, which you can find below.
But it’s worth stating the obvious, these games are old! The target demographic for Little Learning Machines wasn’t born when either of those games hit the market (heck, my youngest brother was a toddler when Creatures came out). And as cool as they were, they have really not aged very well: the controls and user experience in both of those games is very clunky, not to mention it can be difficult to get them to run on a modern PC (it’s possible, but it requires a lot of hoops to be jumped through).
I think it’s also important to differentiate Little Learning Machines from these other titles given it exposes so much of the underlying mechanics of it. Rather than simply providing objects for interaction and then allowing the player to engage with the non-player characters using them, we have a much larger influence on how the Animos are trained given the ‘cloud’ area of the game is effectively an AI gym. From this gym we can define what they can see, what their priorities are, what they should be worried about, and in turn this leads to them generating interesting behaviours.
As mentioned in the case study, the thing that really stuck out to me when I was introduced even to the early version of the game at the summer school, was how it was highly evocative of the sort of things you would see in AI textbooks. The grid structure, the discretised actions, the objects required to interact with puzzles, it all felt incredibly familiar. I say this given we often use these ‘simple’ representations as a mechanism to help aspiring AI scholars to understand the underlying mechanisms of the problem space. I would even given students practical exercises with little maze environments in which they had to figure out what actions the AI agent would take in a given scenario. But here it’s now a game mechanic. It would be interesting to try and use this in the classroom, were it not for the fact I quite my academic career a couple years back!
And this brings me to my final point. That while it’s great to see a new game come along that revisits this idea of training AI chracters for gameplay. Perhaps more critically, Little Learning Machines is also tackling a challenge that neither of those previous games explored: it tries to teach end users how the underlying technology works.
Teaching Players About AI
I can tell you from experience that trying to teach people about artificial intelligence can be a challenge. I’ve spent close to 15 years of my professional career (ha!) trying to do that. And it was with this experience that I was frankly impressed by the task Little Learning Machines set out for itself. In that the overall design actively seeks to educate its players about the fundamentals of what the underlying technology is trying to achieve.
As mentioned already, the core mechanic of the game is to interact with the underlying learning algorithm - which as discussed in the case study is the Proximal Policy Optimisation (PPO) algorithm. The game doesn’t tell you directly what PPO is, but it explains at a high level what it does. It’s a reinforcement learning algorithm, meaning that your Animos are going to attempt to solve the problem you put in front of it tens of thousands of times (what you see in-game is a sort of summary of where the Animo is at that time in its quest to learn). They will learn to try and avoid accruing negative reward, which come about either from doing things that will harm it (like drowning), or that you have specifically identified as bad. In the meantime they are also seeking to maximise positive rewards, all the good things you have told them they should be doing (collecting crystals, petting dogs etc.). So even trying to solve the simplest of puzzles in the game, it’s forcing the player to engage with the fundamentals of reinforcement learning: you effectively paint the reward landscape of the problem space, without realising that’s what you’re doing.
But it’s more than just that, you get to dictate the observations - what the Animo can see in the world. Plus even the Animos you select have different underlying neural network architectures. Some of this is exposed to you via help screens and the UI, and some of it isn’t as readily apparent. Overall I think it is a mixed success, given while the fundamentals translate across really well, I think some of the more demanding subject matter may prove to be a stretch for those without an established AI background.
But of course it’s important to caveat this by recalling this is a game aimed at younger audiences. If it does enough to get the basics in the minds of young people, then the mission is most certainly accomplished.
Find Out More
To find out more about Little Learning Machines, be sure to check out all of the links below, both from Transitional Forms themselves, as well as my recent publications on the site about the project.
Catch the full episode of AI and Games that explains how it works above, or via the write-up on the site.
Check out the episode of the Branching Factor podcast in which I sit down with 3 members of the development team to discuss their experiences in more detail.
Find out more about Little Learning Machines via the official website.
Check out the official 'learning companion’ made by the development team for educators to use
The post-mortem at the GDC AI Summit this year that goes into detail on the technical underpinnings.
And of course, the game is currently 40% off on Steam until September 17th!
Wrapping Up
Thanks for reading this latest edition of the AI and Games Newsletter. Funnily enough, now that I’m back from vacation, we’re actually going back to a more regular schedule! The monthly digest will drop next week, in which I’ll be talking a bit about games such as Fire Emblem: Three Houses and Humanity, my thoughts having just finished reading Reflections of Alan Turing: A Relative Story, and *that* character reveal in Alien: Romulus.
In our next issue of the newsletter, we’ll be back with a long overdue discussion of the SAG-AFTRA strikes and critically how the technology has evolved to a point that this has become such a hot talking point.
In the meantime, y’all take care of yourselves, and I’ll be back!