In the leaked financial projections, OpenAI's compute spend in 2024 was fairly even between the two ($4B inf, $3B train). Also:

1. there is a lag here: the inference done in 2024 would mostly be on models trained in 2023 with a smaller budget
2. inference is growing faster than training

Neat! Up / down arrow seem to be inverted? I'd expect up to accelerate, down to reverse, and for the pointy end to be the front of the ship.

The bullets have the same initial speed no matter the ship's velocity. This makes shooting while moving quite ineffective, and overall feels wrong.

Yeah, a prime example of using a bad UI to generate artificial difficulty. There is a 3p alternate UI (hosting the NYT puzzles, but that was taken down last month). connections.swellgarfo.com

It at least allows visualizing the full solution before committing, even if not very fluidly.

But it wasn’t quite a positive growth coefficient, and didn't go viral. The player base I got has been surprisingly sticky though. The playership is still at 50% of the initial peak, with no market and only minor updates. (All I’ve added is an account system and some rudimentary statistics.) 9/9

But in practice it was hard to get traction, so maybe the polish was wasted. Twitter had turned useless (engagement <5% of what it used to be), Reddit didn't have a good genre-specific subreddit. A HN submission got good visibility, and the game was clearly being passed around in group chats. 8/n

My goal had mainly been to finish some game project in 2024. The game was honestly good enough to release after a few weeks, but the feedback from my wonderful playtesters both in person and at the Quarter to Three forums was good enough that I thought it worth polishing for an extra month. 7/n

I prototyped with the original Wordle word list, but switching to my own hand-curated list was like night and day. Well worth the boring grunt-work, even if sifting through possible English words for 20 hours wasn't what I expected the dev work for this project to be. 6/n

While the basic mechanism didn’t need tweak, there was a ton of value in curating the word list. What this game needed was a really good variety of words, but like 99% of them had to be recognizable as real words, and that's not the case for most online word game word lists. 5/n

After a couple of games, it became clear that the game would be better if the player was given at least one word to start with (but not its location), and after an hour I decided that allowing the player to ask for more hint words would be a good way for players to adjust the difficulty. 4/n

The very first idea I had was to create a word grid, subdivide the grid into colored areas, remove the letters, and task the player with reconstructing the grid. I tried it out in a spreadsheet, and it worked pretty well right from the start. 3/n

In April, I got a random idea of combining a word game with the aesthetics of an advanced Sudoku variant (the kind where Cracking the Cryptic has to start coloring the grid to keep track of the logic). 2/n