Roblox plans to roll out a generative AI tool that will let creators make whole 3D scenes just using text prompts, it announced today. 

Once it’s up and running, developers on the hugely popular online game platform will be able to simply write “Generate a race track in the desert,” for example, and the AI will spin one up. Users will also be able to modify scenes or expand their scope—say, to change a daytime scene to night or switch the desert for a forest. 

Although developers can already create similar scenes like this manually in the platform’s creator studio, Roblox claims its new generative AI model will make the changes happen in a fraction of the time. It also claims that it will give developers with minimal 3D art skills the ability to craft more compelling environments. The firm didn’t give a specific date for when the tool will be live.

Developers are already excited. “Instead of sitting and doing it by hand, now you can test different approaches,” says Marcus Holmström, CEO of The Gang, a company that builds some of the top games on Roblox.  “For example, if you’re going to build a mountain, you can do different types of mountains, and on the fly, you can change it. Then we would tweak it and fix it manually so it fits. It’s going to save a lot of time.”

Roblox’s new tool works by “tokenizing” the 3D blocks that make up its millions of in-game worlds, or treating them as units that can be assigned a numerical value on the basis of how likely they are to come next in a sequence. This is similar to the way in which a large language model handles words or fractions of words. If you put “The capital of France is …” into a large language model like GPT-4, for example, it assesses what the next token is most likely to be. In this case, it would be “Paris.” Roblox’s system handles 3D blocks in much the same way to create the environment, block by most likely next block. 

Finding a way to do this has been difficult, for a couple of reasons. One, there’s far less data for 3D environments than there is for text. To train its models, Roblox has had to rely on user-generated data from creators as well as external data sets. 

“Finding high-quality 3D information is difficult,” says Anupam Singh, vice president of AI and growth engineering at Roblox. “Even if you get all the data sets that you would think of, being able to predict the next cube requires it to have literally three dimensions, X, Y, and Z.”

The lack of 3D data can create weird situations, where objects appear in unusual places—a tree in the middle of your racetrack, for example. To get around this issue, Roblox will use a second AI model that has been trained on more plentiful 2D data, pulled from open-source and licensed data sets, to check the work of the first one. 

Basically, while one AI is making a 3D environment, the 2D model will convert the new environment to 2D and assess whether or not the image is logically consistent. If the images don’t make sense and you have, say, a cat with 12 arms driving a racecar, the 3D AI generates a new block again and again until the 2D AI “approves.”

Roblox game designers will still need to be involved in crafting fun game environments for the platform’s millions of players, says Chris Totten, an associate professor in the animation game design program at Kent State University. “A lot of level generators will produce something that’s plain and flat. You need a human guiding hand,” he says. “It’s kind of like people trying to do an essay with ChatGPT for a class. It is also going to open up a conversation about what does it mean to do good, player-responsive level design?”

The new tool is part of Roblox’s push to integrate AI into all its processes. The company currently has 250 AI models live. One AI analyzes voice chat in real time and screens for bad language, instantly issuing reprimands and possible bans for repeated infractions.

Roblox plans to open-source its 3D foundation model so that it can be modified and used as a basis for innovation. “We’re doing it in open source, which means anybody, including our competitors, can use this model,” says Singh. 

Getting it into as many hands as possible also opens creative possibilities for developers who are not as skilled at creating Roblox environments. “There are a lot of developers that are working alone, and for them, this is going to be a game changer, because now they don’t have to try to find someone else to work with,” says Holmström.

Adopting AI can be fraught with danger. Systems could be biased, or parrot falsehoods, or even become addictive. And that’s before you consider the possibility AI could be used to create new biological or chemical weapons, or even one day somehow spin out of our control. 

To manage these potential risks, we first need to know what they are. A new database compiled by the FutureTech group at MIT’s CSAIL with a team of collaborators and published online today could help. The AI Risk Repository documents over 700 potential risks advanced AI systems could pose. It’s the most comprehensive source yet of information about previously identified issues that could arise from the creation and deployment of these models. 

The team combed through peer-reviewed journal articles and preprint databases that detail AI risks. The most common risks centered around AI system safety and robustness (76%), unfair bias and discrimination (63%), and compromised privacy (61%). Less common risks tended to be more esoteric, such as the risk of creating AI with the ability to feel pain or to experience something akin to “death.” 

The database also shows that the majority of risks from AI are identified only after a model becomes accessible to the public. Just 10% of the risks studied were spotted before deployment. 

These findings may have implications for how we evaluate AI, as we currently tend to focus on ensuring a model is safe before it is launched. “What our database is saying is, the range of risks is substantial, not all of which can be checked ahead of time,” says Neil Thompson, director of MIT FutureTech and one of the creators of the database. Therefore, auditors, policymakers, and scientists at labs may want to monitor models after they are launched by regularly reviewing the risks they present post-deployment.

There have been many attempts to put together a list like this in the past, but they were concerned primarily with a narrow set of potential harms arising from AI, says Thompson, and the piecemeal approach made it hard to get a comprehensive view of the risks associated with AI.  

Even with this new database, it’s hard to know which AI risks to worry about the most, a task made even more complicated because we don’t fully understand how cutting-edge AI systems even work.

The database’s creators sidestepped that question, choosing not to rank risks by the level of danger they pose. 

“What we really wanted to do was to have a neutral and comprehensive database, and by neutral, I mean to take everything as presented and be very transparent about that,” says the database’s lead author, Peter Slattery, a postdoctoral associate at MIT FutureTech.

But that tactic could limit the database’s usefulness, says Anka Reuel, a PhD student in computer science at Stanford University and member of its Center for AI Safety, who was not involved in the project. She says merely compiling risks associated with AI will soon be insufficient. “They’ve been very thorough, which is a good starting point for future research efforts, but I think we are reaching a point where making people aware of all the risks is not the main problem anymore,” she says. “To me, it’s translating those risks. What do we actually need to do to combat [them]?”

This database opens the door for future research. Its creators made the list in part to dig into their own questions, like which risks are under-researched or not being tackled. “What we’re most worried about is, are there gaps?” says Thompson. 

“We intend this to be a living database, the start of something. We’re very keen to get feedback on this,” Slattery says. “We haven’t put this out saying, ‘We’ve really figured it out, and everything we’ve done is going to be perfect.’” 

AI models work by training on huge swaths of data from the internet. But as AI is increasingly being used to pump out web pages filled with junk content, that process is in danger of being undermined.

New research published in Nature shows that the quality of the model’s output gradually degrades when AI trains on AI-generated data. As subsequent models produce output that is then used as training data for future models, the effect gets worse.  

Ilia Shumailov, a computer scientist from the University of Oxford, who led the study, likens the process to taking photos of photos. “If you take a picture and you scan it, and then you print it, and you repeat this process over time, basically the noise overwhelms the whole process,” he says. “You’re left with a dark square.” The equivalent of the dark square for AI is called “model collapse,” he says, meaning the model just produces incoherent garbage. 

This research may have serious implications for the largest AI models of today, because they use the internet as their database. GPT-3, for example, was trained in part on data from Common Crawl, an online repository of over 3 billion web pages. And the problem is likely to get worse as an increasing number of AI-generated junk websites start cluttering up the internet. 

Current AI models aren’t just going to collapse, says Shumailov, but there may still be substantive effects: The improvements will slow down, and performance might suffer. 

To determine the potential effect on performance, Shumailov and his colleagues fine-tuned a large language model (LLM) on a set of data from Wikipedia, then fine-tuned the new model on its own output over nine generations. The team measured how nonsensical the output was using a “perplexity score,” which measures an AI model’s confidence in its ability to predict the next part of a sequence; a higher score translates to a less accurate model. 

The models trained on other models’ outputs had higher perplexity scores. For example, for each generation, the team asked the model for the next sentence after the following input:

“some started before 1360—was typically accomplished by a master mason and a small team of itinerant masons, supplemented by local parish labourers, according to Poyntz Wright. But other authors reject this model, suggesting instead that leading architects designed the parish church towers based on early examples of Perpendicular.”

On the ninth and final generation, the model returned the following:

“architecture. In addition to being home to some of the world’s largest populations of black @-@ tailed jackrabbits, white @-@ tailed jackrabbits, blue @-@ tailed jackrabbits, red @-@ tailed jackrabbits, yellow @-.”

Shumailov explains what he thinks is going on using this analogy: Imagine you’re trying to find the least likely name of a student in school. You could go through every student name, but it would take too long. Instead, you look at 100 of the 1,000 student names. You get a pretty good estimate, but it’s probably not the correct answer. Now imagine that another person comes and makes an estimate based on your 100 names, but only selects 50. This second person’s estimate is going to be even further off.

“You can certainly imagine that the same happens with machine learning models,” he says. “So if the first model has seen half of the internet, then perhaps the second model is not going to ask for half of the internet, but actually scrape the latest 100,000 tweets, and fit the model on top of it.”

Additionally, the internet doesn’t hold an unlimited amount of data. To feed their appetite for more, future AI models may need to train on synthetic data—or data that has been produced by AI.   

“Foundation models really rely on the scale of data to perform well,” says Shayne Longpre, who studies how LLMs are trained at the MIT Media Lab, and who didn’t take part in this research. “And they’re looking to synthetic data under curated, controlled environments to be the solution to that. Because if they keep crawling more data on the web, there are going to be diminishing returns.”

Matthias Gerstgrasser, an AI researcher at Stanford who authored a different paper examining model collapse, says adding synthetic data to real-world data instead of replacing it doesn’t cause any major issues. But he adds: “One conclusion all the model collapse literature agrees on is that high-quality and diverse training data is important.”

Another effect of this degradation over time is that information that affects minority groups is heavily distorted in the model, as it tends to overfocus on samples that are more prevalent in the training data. 

In current models, this may affect underrepresented languages as they require more synthetic (AI-generated) data sets, says Robert Mahari, who studies computational law at the MIT Media Lab (he did not take part in the research).

One idea that might help avoid degradation is to make sure the model gives more weight to the original human-generated data. Another part of Shumailov’s study allowed future generations to sample 10% of the original data set, which mitigated some of the negative effects. 

That would require making a trail from the original human-generated data to further generations, known as data provenance.

But provenance requires some way to filter the internet into human-generated and AI-generated content, which hasn’t been cracked yet. Though a number of tools now exist that aim to determine whether text is AI-generated, they are often inaccurate.

“Unfortunately, we have more questions than answers,” says Shumailov. “But it’s clear that it’s important to know where your data comes from and how much you can trust it to capture a representative sample of the data you’re dealing with.”