A look at Oasis, a generative AI take on Minecraft

Welcome to Gametime, a weekly newsletter recapping the past week of games on LinkedIn, what’s new in the puzzles space, and more — brought to you by LinkedIn Games Editor Paolo Pasco . Click 'Subscribe' to join the community and be notified of future editions.

A few weeks ago, the AI startups Decart and Etched released Oasis , an application that emulates the game Minecraft. The player can only play for a few minutes, during which time they can do many of the actions you’d expect from Minecraft — move, mine blocks, obtain items, and start to build structures.?

The notable thing about Oasis: none of the gameplay behavior is coded into the program. Instead, the gameplay is generated from a model trained on, according to the website, “millions of gameplay hours.” As the player pushes buttons, a transformer performs the necessary calculations and generates the next frame, and the frame after that, and the frame after that. In other words, there’s no hard-coded process for, say, “jump when the player pushes the space bar.” Instead, the model recognizes that, based on the gameplay it’s seen in training, whenever the space bar is pressed, the player’s view of the world changes in this specific way, and changes the current view of the world to match. In a sense, the model uses the gameplay data to derive a rough sense of gameplay physics and controls. According to a post announcing Oasis, Decart and Etched present Oasis as “an impressive technical demo,” and add that they “believe this research will enable an exciting new generation of foundation models and consumer products.”?

I tried Oasis out recently, albeit in sessions lasting a few minutes each (presumably to avoid strain caused by a large number of people using the site at once, visitors are placed in a queue. Once they make it to the front of the queue, they can try out Oasis for five minutes before being kicked from the server). I found the gameplay to be imperfect, but interesting. As anyone familiar with AI-generated content would expect, the model was prone to hallucinations. Often, I would place a few blocks and go somewhere else, only to return and find the blocks have disappeared, changed, or (in some cases) turned into a fully built house.

Some commenters noted that playing with Oasis felt like watching someone have a dream about Minecraft. Objects would appear and disappear once they leave your field of vision, graphics have a fuzzy and surreal quality to them, and the model is prone to hallucinating new objects and structures. Nevertheless, the effect is impressive, especially considering the gameplay is being generated frame by frame. For me, the benefit of the demo came not from treating it as an actual game, but as you would an art installation — a showcase of a technical project, as well as an interesting, if not immediately practical, recontextualization of a familiar property.?

There’s a long way to go if Decart and Etched truly intend on using Oasis as the backbone of an actual game. In the meantime, the project has attracted some critics, who point out that Oasis is a resource-intensive way (both computationally and environmentally) to create a buggier version of an existing game. On this point, the developers frame Oasis as the first step of many, envisioning a future in which game content can be generated by the user on a whim. In an interview with the MIT Technology Review, Decart CEO Dean Leitersdorf says, “What if you could say ‘Hey, add a flying unicorn here’? Literally, talk to the model. Or ‘Turn everything here into medieval ages,’ and then, boom, it’s all medieval ages.” According to the developers of Oasis, the major factor holding back this level of interactivity is hardware — with faster and more powerful chips to make it easier to produce real-time interactive video, they see large developments in the performance of Oasis.

How feasible is this? Currently, the use case of Oasis is scraping massive amounts of Minecraft gameplay to generate an approximation of Minecraft gameplay. Some people are skeptical as to the model’s ability to generate new content, such as the “flying unicorn” mentioned. Without heaps of training data including examples of what the player wants, will the model be able to deliver? Can Oasis create content beyond the already existing games it sees in its training? As development work continues, and as hardware improves, we may see answers to these questions very soon.

?? Pinpoint: The Monday, November 4 astronomical objects puzzle with clues [“Cloud,” “Star,” “Moon,” “Planet,” “Black hole”] sparked a lot of discussion. I saw (and was tagged in) a lot of comments to the effect of “a cloud is not an astronomical object.” In response to that, I’ll point out the second definition listed by Merriam-Webster, which includes “an aggregation of usually obscuring matter especially in interstellar space.” Likely the most famous example is the Oort Cloud , believed to be a cloud of icy objects surrounding the Sun, which scientists believe to be the origin of many comets. Other clouds, like high-velocity clouds (HVCs), also exist in astronomers’ vocabularies. Is it the most obvious definition of cloud? No. Is it fair game for the first clue in a Pinpoint? I’d say yes, especially since earlier clues are intended to be more difficult.?

Play today’s Pinpoint?

?? Crossclimb: I want to spotlight a few comments that I enjoyed:

Too funny. This reminds me of a few weeks ago, when I used my stopwatch app to time myself solving a cryptic crossword. I also got distracted mid-solve, and had forgotten about the stopwatch by the time I finished solving. Checking my phone recently, I was surprised to see the stopwatch still going, at 237 hours, 48 minutes, and 33 seconds. I promise I have solved puzzles faster than that. Tomorrow will be quicker.

This speaks to one of my favorite things about trivia games — when an answer comes from something that just happened in your day-to-day life, it can be immensely satisfying.?

For context, this was the Monday, November 4 puzzle ending in two plural terms for rodents (MICE/RATS). Rob, I hope you use this as an excuse to keep all the mice and rats you want now.

Play today’s Crossclimb?

?? Queens: The Thursday, November 7 puzzle showed that a larger puzzle doesn’t necessarily mean a harder puzzle. I saw many times in the sub-30 second range, and I managed to finish in 10 seconds. Let’s take a look at the puzzle:

The large purple region is a good place to start. Note that the purple region takes up the entire topmost row, so we know the Queen in the purple region must be in that row (otherwise, it’d be impossible to place a Queen in the top row). Likewise, the purple region takes up the entire leftmost column, so we also know the Queen in the purple region must be in that column. There’s only one square that’s both in the topmost row and the leftmost column — the square in the top-left corner. Therefore, the purple Queen must go there.

Now, note that the only available squares in the second row are in the blue region. Using similar logic to the previous step, we can X out every square in the blue region that’s not in the second row.

This leaves only one available square left in the third row, so we can place our first Queen! Note that we could have arrived at this faster (the gray and green regions are both fully contained within the fourth and fifth rows, so we can X out every square in those rows outside of those two regions, leaving the one square left in the orange region). If someone’s share text from that day says their first three colors were ?????, they probably started with that strategy.?

Play today’s Queens

?? Tango: The puzzle from Sunday, November 3 had a relatively high average solve time, taking the average solver 3 minutes 39 seconds. Let’s take a look at the board for that day. We can place a few Moons by blocking potential three-Sun-in-a-row spots:

Where to go from there? This strategy is a little tricky to spot, but it’s very useful. Look at the square highlighted in red below.?

We can definitively place a symbol in this square, but it requires a little bit of “if X, then Y” thinking. In particular, if this square contained a Sun, that would mean the rest of the empty cells in that column would have to be Moons (since there’d be three Suns in that column). However, this would force the three Moons to all be next to each other, which isn’t allowed by the rules. Therefore, since having a Sun in that square would break the puzzle, the square must contain a Moon. After that placement, we can use the X and = signs adjacent to that square to make further progress. (Note: there’s one more column where you can use this strategy. Can you spot it?)

Play today’s Tango

Want to be notified when new editions are released? You can sign up for reminders for each of our games here .?

This week’s topic: Rest in peace, blanks in Pinpoint puzzles (2024-2024)

If you’ve been playing Pinpoint over the past week, you’ve probably seen that a few puzzles have blanks, like the one below:

This was an experiment, intended to help out with the “word association” style of Pinpoint, which is the type of Pinpoint that traditionally plays the hardest. As a result, we found out a few things: win rates did go up as a result, but the downside was that for many commenters, it was just either distracting, confusing, or just too easy. We hear you: from now on, word association Pinpoints will go back to having no blanks. This means the puzzles will go back to being a little harder, but you got this!

How do you feel about the change back?

Share your thoughts in the comments below??

Know someone who would enjoy Gametime? Share it with them directly or your network by clicking the “Share” button below!