RandomRanger
Just build nuclear plants!
No bio...
User ID: 317
RandomRanger
Just build nuclear plants!
2d ago
Ah but the typical AAA game is not Baldur's Gate III.
Mass Effect Andromeda. Or Dragon Age Veilguard. Or Concord. This is what I'm thinking of: https://x.com/celestesangels/status/2003911076988260714
https://x.com/deadlock/status/2015887428964266047
An underlying issue is that the people who can't write good dialogue surely can't write good prompts or lora/finetunes for AIs either.
Still, some fun can definitely be had in a version of LOTR where Saruman is shilling NordVPN: "50% off with code ISTAR15". I saw an AI make a great pun about Isengard's no-logs policy.
Trying to replicate the peak of human literature should not be the target for AI gaming, better to focus on things that only it can do in terms of reactivity and dynamism to create new kinds of fun.
RandomRanger
Just build nuclear plants!
3d ago
It'd be super cool to have access to a Google Genie like world-model game, perhaps with an AI 'dungeon master' overseeing a larger storyline or controlling game mechanics. In a more freeform mode, you can type in something and it just happens (apparently this was too fun and interesting for the public demo of Genie 3 but it exists in principle, since it literally just generates everything you see).
More of a longer-term thing though since world models are quite costly to run in real time.
I think the inference side of things is profitable. That's precisely why they need to make it more efficient for gaming purposes...
The general trend is that businesses looking to make money will spend a lot more than consumers looking for fun. Coders have a huge appetite for the highest-quality tokens. Consider the GPU shortage, the RAM shortage, compute is being reallocated from consumer command to commercial demand via pricing... Gamers are not prepared to pay that much money, there's already a lot of unhappiness about GAAS, a move up beyond 60 USD games.
It would be stupid for Microsoft to lower their margins by reallocating compute from lucrative Azure to less-lucrative Minecraft gaming. They need to keep margins high or raise them by reducing cost of production. And they should definitely be able to get better results at lower prices with a dedicated minecraft AI, even if it's just a finetune. It's like the Chinese paper where they finetuned an AI to play Genshin Impact, solve puzzles, complete hours long missions... Presumably that's quite expensive to run since it's a full video model that plays like its a human. But Microsoft could easily make a smaller text model that gets data directly from the game, maybe it calls stronger models for particularly difficult building tasks.
RandomRanger
Just build nuclear plants!
4d ago
LLM's getting more deeply integrated into games?
I think this is a good idea. It's not like many AAA games are acclaimed for their dialogue, characters and writing, people literally joke about how crap their writing is. Let people have conversations with in-game characters, why not?
Open source communities have gone out of their way to set up general-purpose AIs to play Minecraft with you in the crudest ways imaginable and it kind of works. Microsoft literally owns Minecraft and they have a ludicrous amount of compute. They could make a minecraft-specific AI model, special servers where the player (players?) could be warlords with whole armies they direct and manage. The sky is the limit. This is a GAAS subscription goldmine just waiting to happen if they can cut down the inference costs, which they should definitely be able to do with a specialist model.
The real problem Microsoft has is dysfunctional culture. It's really not that hard to make Halo Infinite and have it be actually good. They have the money but not the necessary organizational skills. How hard can it be to make Windows 11 run smoothly enough for people to risk their computers and 'upgrade'? Windows 10 was OK...
RandomRanger
Just build nuclear plants!
6d ago
The distinction matters for policy and diagnosis. If moral decay is primary, then the intervention is cultural: restore warrior values, punish softness, reform gender norms. If institutional decay is primary, then the intervention is structural: fix training pipelines, improve logistics, reform command selection, rebuild industrial capacity. These point in very different directions.
But why are there issues with training pipelines, why was industrial capacity lost, why are there all these crap commanders who deflect responsibility?
In the US, training pipelines are inevitably going to be recruiting lower quality leaders since Americans are getting increasingly less interested in military service (hence a shortfall of recruits). The military is an inherently dangerous, demanding, stressful role. Lives are at stake. You don't get paid that much either compared to civilian jobs. There's a shortage of patriotism and dedication to ideals that gets high quality recruits to join and stay in the military for a long time. Soldiers fight best when they believe in something beyond dental plans and a free education.
There's also a culture issue in selecting commanders. It's a very bureaucratic, boxticking process which selects for suck-ups and master blame dodgers. Not necessarily people who know how to take risks and fight well:
Industrial capacity - this was jettisoned by greedy executives looking for higher margins and opportunistic cost-cutting by Pentagon officials. Around the 1970s and 80s there was a shift in corporate culture towards short-termism and a lack of capital investment in boring industries like steel and shells. Optimistic planners assumed there would be no need to fight slow, attritional wars with large quantities of munitions (despite these being the most important kind of war).
The defence contractors are themselves decadent and slack internally. I watched this long video from a guy who worked there: https://youtube.com/watch?v=FIONXPbIkVo?list=LL
The TLDR of it is that management at Lockheed and NASA was toxic and grossly incompetent, they'd basically treat contractors like fourth-rate 'citizens', make them do all the work and then take credit for it, they'd do absolutely retarded nonsense like try and scale up tech from the Apollo era launchers, they'd lie about what was tested when and by who to keep the flow of money coming in, they'd say that every little change (from the idiots at NASA) required them to start over... Idiot managers would scream at the technical employees for asking questions to cover up their own worthlessness. If you knew what you were doing and tried to obey the rules you were the enemy. DEI made everything a lot worse. The reporting/whistleblowing system was wholly ineffective. This guy sure is disgruntled (and perhaps presents himself as too holier than thou) but, considering the ridiculous stretch of time it took to develop Orion, I'm inclined to believe him.
Institutions are downstream from culture. It isn't neccessarily that a decadent state must be effeminate and wimpy (they can just be addicted to civil wars like Byzantium) but a dysfunctional, short-termist, self-interested culture will naturally degrade key institutions needed for military power. Effeminacy is only one failure mode.
Anyone else think that the new Grok 4.2 is a little underrated? People on twitter seem to be going 'it's bad'. I can see where they're coming from but it has some value add too.
The good: It can oneshot a couple of things where Opus 4.6 just burns through all its thinking tokens and dies. It codes in a much more restrained way too. 100 or 200 lines where Opus would make a huge extravaganza of code.
The bad: They didn't open it up to API and its no good for creative writing, pure STEMcel... The features surrounding the model are barebones, I can't seem to just copy in 10 files of context and have 4.2 edit them inline. We're back in the 'Below are the exact inline edits you can copy-paste' era and that's shit.
Interesting that they've chosen a different path with their '4 copies talking it out' approach, compared to everyone else and their 'big model go brr' approach.
Not really... I tried to look into this and there's a huge continuum of 'missiles'. ATACMS is a missile, Tomahawk is a missile, GMLRS is a missile all in that general ballpark but wildly different to eachother in quality and performance, let alone PRSM or the air launched stuff... Same with their Chinese equivalents. So when this random Chinese factory says 'oh we can produce 1000 a day' nobody really knows what they're talking about. Are they talking about missiles or just components? Probably the latter.
The Chinese military is pretty secretive and there's also a deadening layer of propaganda.
But there's also been a huge build-up in production capacity in the last few years, so maybe high tens of thousands per year once the new factories all come online? As for scaling, producing advanced missiles is quite difficult in terms of machining, it's a bit like advanced engines which are hard for anyone to produce at scale. But China does have an enormous industrial base so they should be capable of making the cheaper sort at phenomenal scale.
More options
Context Copy link