legend

I think you dropped this.

They're the hardest part of the equation!

I agree that model won't work. The logistics is too difficult to track. I also think open source models may need special rules from models/services that are sold. But for the sold service setting, I think a better model is just opt-in consent. Set up a market place where artists/photographers, etc can sell their work a fixed price. I think a lot of artists would consider that. Especially because as tech gets better, generative images will become a more useful tool for artists and they'll want it to work well. But figuring out an economic model isn't really my specialty, so that might not be great either. But in any case, I'm pretty sure people can find a way to make this work. Giving up because you cannot just steal would be utterly ridiculous

Not that surprising. A considerable amount of my research involves interaction between AI and people!

This is either incorrect or so reductive that it conveys no meaning. You can use a random search algorithm to "train" an image generator and you'll have a very bad time. The methods being used here are *vastly* more powerful. "Generative AI" is not the same "AI." Generative AI has numerous limitations that will prevent it alone from ever realizing the dream of what we want "AI" to be, but not all AI paradigms have those same limitations, they're just harder problems to solve, and the progress in generative AI methods can still play a *part* of a bigger system or at least inform future work. It is not.

Crunchy Roll

These tools don't exist without this training data, and the tool is being sold. If you're going to make a tool and profit from it, we ought to compensate and get consent from people whose work you used to build the tool. Beyond that, people *do* have to be very careful when they use references images. There's a huge amount of existing law working out when something is sufficiently novel. These image generators are not just building from a reference idea though. Their objective is to out right replicate the training data. I mean that very literally. The objective function being optimized to train the models is optimized optimally when it can perfectly replicate training data. Now we also strive to build models and method such that we don't *only* replicate exact instances in the training data, but also generalize for new inputs (in this case, a language prompt). But the fact remains that the algorithm and method works well as intended when it *can* reproduce the training data exactly when prompted accordingly. And as things stand now, many of these image generator models tend to fall back very quickly to the training data even when prompted with slightly different inputs. That is, they're doing a better job memorizing than generalizing. That's right. Hopefully my above comments help.

I got up to the spider scene. I think I'm done.

Great video!

I don't think so. The rules are pretty manageable, especially if you played the video game whereas Gloomhaven feels vastly more complex, if nothing else but because of all the management of each unit. It's possible it's a party make up issue. The game, to its credit, has *a lot* of different ways to build your character, but then maybe isn't well tuned for all party set ups. And unlike the video game, you can't go grind something else, or configure things in some crazy set up well outside the usual combat box to achieve victory. That said, with the third and fourth bosses we played yesterday we tried the house rule of halving their AP and they were kind of push overs. However, it was to an extent that seemed more than just because they had fewer moves. Their hits just weren't nearly as strong as the first two bosses that hit like a truck *and* kept taking moves from their huge AP pool. So maybe some bosses are well tuned and others not? I dunno. It might be worth trying a mid way point to reduce to 3/4ths AP instead of half. Although my wife tends to like this half rule and is much happier with the last two boss fights because she doesn't want an intense challenge And she also dislikes just how long the boss turns go if you give them the full AP because it makes it boring to just have the boss go on and on.

I think his level is fine because there is still actually a lot of act 1 left, but I was level 6 when I started Act 2.

After five whole fucking years since the kickstarter, the Divinity Original Sin board game finally arrived this month! It's a purely coop game for 1-4 players. (If you play solo, you control two characters). My wife and I have been playing through the game in 2 player mode and have finished many sessions so I thought I'd give my thoughts on it here. Overall, the game is fantastic and though it was a long wait, there is a lot of good stuff here. 1. The artwork is stellar. 2. The materials have excellent quality -- especially if you got the bonus metal coins and boss figurines. 3. The core game mechanics of the video game have largely translated beautifully to the board game 4. I really enjoy the mix of exploration, combat, and light story telling. 5. For a game as complex as this, the setup and tear down is pretty good. We've mostly been keeping the core materials out (thanks Wyrmwood gaming table!) but it's still so much better than something like Gloomhaven where you have to devote your day to setup and tear down and custom organizers for the 8 billion parts. The box itself is very well laid out and keeps everything organized when you put them back. You will never have a problem where you feel like the box doesn't quite fit like it did when you first got it. 6. The game starts with a nice tutorial section and clearly labels what you should learn for that and then wait and learn the rest as you go. It works well. 7. The enemy rules are easy to follow. 8. The game does a great job allowing for out-of-combat skill checks that use properties unique to your origin. The bad: 1. The game is super stingy about gold and I don't get it. We got gold the first two sessions (tutorial and first main session) and never since in something like 7 additional sessions (sessions are broken up into locations and give clear stop points if you want to pause. We usually play 1-2 sessions a day). This is not from a lack of exploring or losing combat, it's just not given out. And if you lose a fight, you often lose *all* your gold. It seems like the only way to make gold is to sell stuff to the merchant between sessions. But... 2. The merchant rules are tedious. You have to *pay* a substantial amount of gold (equal to your party level) just to visit the merchant and you get a small random draw of two decks to choose from what to buy. If you want to see a few more random items, you have to pay again. Selling only gives you half price and it would be easy to lose most of your gold just trying to find an item you want. This sucks and doesn't make buying items fun. I highly recommend you house rule this to something like pay no gold to see an item deck and one consumable deck and they pay if you want to see additional decks. Or something else less tedious. 3. The boss fights are horribly balanced for 2 players, at least the first 2 we played were. I think they balanced for 4 players and didn't have time to do proper balancing for fewer players. Outside of bosses, the way the game works for balancing is minions basically get one move + attack action per turn. For fewer players, the game scales down by spawaning fewer enemies, which naturally keeps the action economy balanced and this works quite well and those regular encounters are good. Bosses, however, get their own action point system like the player. But the game doesn't adjust the action points the boss gets for the number of players. The only difference is how its health is *slightly* lower for fewer players. This is horrible, because it means the boss as an action point economy balanced for 4 players. When it's the boss' turn, they just go on and on and on and can often wipe the party in one turn before you can do anything. Ultimately, the boss has * 2-4x the party's cumulative health (meaning summing each players health together) * Attacks that do much more damage than a player can do with attacks or skills, even with synergies * More armor than any party member, which often makes player attacks wiff unless they can do direct (armor bypassing) damage * Many more actions per turn than the total turns of the party <- this is the real killer It's just makes the boss stupid. The only way for two player parties to win with rules as written is very lucky dice rolls. On the last boss we did from the haunted house expansion, the boss party wiped us in the first turn before we could even take a single action. While stacking all those things against the player is bad, the fundamental issue here is unlike the minion fights for which action economy naturally scales by spawning fewer minions for fewer players, the game does *not* scale the action economy of bosses for fewer players. I haven't tried this yet, but we're planning a house rule where we scale the action points the boss gets by the number players assuming it's written for 4 players (e.g., half it for 2 players, as in our case). That might make it more reasonable. So overall, fantastic game with a few problems that can probably be fixed with house rules, with the boss issues for two players (or one player) in particular being the real sour point.

I’m very open to biological computers of any sort and do hope we someday augment our brain. I’m not sure how much utility we will have in pairing them until we have a better understanding of the biology though. Maybe some things but without a deeper understanding it seems like it will be hard to engineer with it.

Better to start with an instance where you think an effect might exist to see if it does. If it does, then you can narrow down further on what kinds of 3D video games would satisfy and which wouldn't. If you threw in a bunch at once, that would be a lot more variables to control for (or just muddy the result if you don't)

Tech-wise, the community needs a replacement for CPython (the programming language/interpreter), or Python needs to be massively overhauled. It's a terribly slow language that creates barriers and prevents all kinds of research avenues from being pursued. It only works for things now because all the big AI libraries used in Python are actually written in C++ with bindings for specific kinds of operations in Python. The current trend is to allow compiling python code into native code at runtime, but this has numerous limitations. Something like Mojo lang might be one way to escape this problem, but we'll see. Research-wise we need to move out of the realm of static dataset fitting. Intelligent organisms don't learn by fitting nicely curated datasets. They learn interactively with an environment (sometimes with an interactive teacher to help) and there is tremendous power in being able to do so, but we haven't really cracked how to do it especially well (despite several big milestones, it remains hard). In particular, most ML methods rely on "iid" assumptions about data that don't hold when you're interacting with an environment and the kludges we have around that are, well, kludges. Dataset fitting might still play a role in the story for pretraining well defined things, but if it's all we can ever do, we'll never succeed at replicating the kind of intelligence found in biologically intelligent organisms. Aside from getting out of dataset fitting, we need to improve on control methods (how to act) especially at very long time horizons (e.g., a person can plan for the next few seconds or next few years. AI control methods... not so much). We need a way to learn casual world models (which kind of depends on learning interactively with an environment to do interventions) that facilitate planning. We also need to figure how to efficiently store and recall previously learned skills without forgetting them and be able to apply and reason about them compositionally. There are other things I could list, but those are the big ticket items. It's also kind of the same set of things that we were largely in the dark on 20 years ago. Mostly, in that time we've only learned how to do function approximation (that is, fit a function to data) well (as long as we have an iid data assumption). Doing function approximation well *is* a wildly significant and transformative advance, but it's still just one small slice of the problem.

This might hold for other Nintendo games, but I'm not sure how many. Mario Odyssey is kind of unique IMO in just how chill of a game it is to play. The game isn't really hard, so there isn't frustration, yet it somehow manages to remain engaging and just a joy to play.

Well then...

You can still recruit her even if you took the quest. I'm not sure how you're missing it. Just choose non-violent responses in dialog.

Put some light armor on him, not robes. Gale is great, but wizards tend to get better at higher levels and die very easy at low levels. Make sure you're at a minimum attacking with a cantrip and not physically. You need to get Halsin, who is part of the main quest, assuming you don't kill him or fight for the bad guys

Having a bigger party will help for sure. Remember that you can aways travel back to camp and swap out who's playing with you. Astarion will join you, but I'm guessing your party is already full (4 people at a time). Tell someone else in the group to wait at camp (ignore their moaning) and then ask Astarion to join. Do *not* get rid of Astarion. Unless you're heavily leaning into a very specific the RP, I would advise against getting rid of anyone. They all have complex and interesting stories you'll miss and evolve a lot as characters.

Don’t do it bacons way. He has a very specific hatred for Druids and worked hard to find a way to kill them all without other side effects. Just don’t kill Kagha. If you are just going to repeat the same decisions you made before I don’t know why you restarted. You’re going to be back in the same situation.

I'll also say this, if you go into this game trying to play it the "obvious" way, you will have a much harder time. Go outside the box. If it seems like it should work, it probably will. Also, shove is one of the best abilities. Always look for ways to shove enemies off of things. You can end entire bosses in one move this way. Again, if it seems like it should work, it usually will.

In a game like this, never, ever use one save file. Get the biggest rotation of save files they'll let you have. (On PC, I don't think there was any limit, so I have... a lot of saves!) If there is anything you should save scum, it's making sure you get all the party members in act 1. Dark Urge is the best. Custom character but fully integrated into the plot. Play to "resist" if you want to be a good guy. It's what I did and it has a very satisfying conclusion to that story.

Linux is fantastic and if this were a sane world, more games would target it instead of relying on Valves nearly magical work with Proton. Windows isn't the target because it's a good OS. It's the target because it's what most people have.