#it's hard to simplify these concepts down to the scale where most people could intuit them without basically writing a crash course

[demigirlboss2girlrotboogaloo]

I have this Ars Technica article pinned in my clipboard now because of how often I've had to cite it as a sort of accessible primer to why the oft-cited numbers on AI power consumption are, to put it kindly, very very wrong. The salient points of the article are that the power consumption of AI is, in the grand scheme of datacenter power consumption, a statistically insignificant blip. While the power consumption of datacenters *has* been growing, it's been doing so steadily for the past twelve years, which AI had nothing to do with. Also, to paraphrase my past self:

While it might be easy to look at the current massive AI hypetrain and all the associated marketing and think that Silicon Valley is going cuckoo bananas over this stuff long term planning be damned, the fact is that at the end of the day it's engineers and IT people signing off on the acquisitions, and they are extremely cautious, to a fault some would argue, and none of them wanna be saddled with half a billion dollars worth of space heaters once they no longer need to train more massive models (inference is an evolving landscape and I could write a whole separate post about that topic).

Fundamentally, AI processors like the H100 and AMD's Instinct MI300 line are a hedged bet from all sides. The manufacturers don't wanna waste precious wafer allotment on stock they might not be able to clear in a year's time, and the customers don't wanna buy something that ends up being a waste of sand in six months time once the hype machine runs out of steam. That's why these aren't actually dedicated AI coprocessors, they're just really really fucking good processors for any kind of highly parallel workload that requires a lot of floating point calculations and is sensitive to things like memory capacity, interconnect latencies, and a bunch of other stuff. And yeah, right now they're mainly being used for AI, and there's a lot of doom and gloom surrounding that because AI is, of course, ontologically evil (except when used in ways that read tastefully in a headline), and so their power consumption seems unreasonably high and planet-destroying. But those exact same GPUs, at that exact same power consumption, in those same datacenters, can and most likely *will* be used for things like fluid dynamics simulations, or protein folding for medical research, both of which by the way are usecases that AI would also be super useful in. In fact, they most likely currently are being used for those things! You can use them for it yourself! You can go and rent time on a compute cluster of those GPUs for anything you want from any of the major cloud service providers with trivial difficulty!

A lot of computer manufacturers are actually currently developing specific ML processors (these are being offered in things like the Microsoft copilot PCs and in the Intel sapphire processors) so reliance on GPUs for AI is already receding (these processors should theoretically also be more efficient for AI than GPUs are, reducing energy use).

Regarding this, yes! Every major CPU vendor (and I do mean every one, not just Intel and AMD but also MediaTek, Qualcomm, Rockchip,and more) are integrating dedicated AI inference accelerators into their new chips. These are called NPUs, or Neural Processing Units. Unlike GPUs, which are Graphics Processing Units (and just so happen to also be really good for anything else that's highly parallel, like AI), NPUs do just AI and nothing else whatsoever. And because of how computers work, this means that they are an order of magnitude more efficient in every way than their full-scale GPU cousins. They're cheaper to design, cheaper to manufacture, run far more efficiently, and absolutely sip power during operation. Heck, you can stick one in a laptop and not impact the battery life! Intel has kind of been at the forefront of these, bringing them to their Sapphire Rapids Xeon CPUs for servers and workstations to enable them to compete with AMD's higher core counts (with major South Korean online services provider Naver Corporation using these CPUs over Nvidia GPUs due to supply issues and price hikes), and being the first major vendor to bring NPUs to the consumer space with their Meteor Lake or first generation Core Ultra lineup (followed shortly by AMD and then Qualcomm). If you, like me, are a colossal giganerd and wanna know about the juicy inside scoop on how badly Microsoft has screwed the whole kit and caboodle and also a bunch of other cool stuff, Wendell from Level1Techs has a really great video going over all that stuff! It's a pretty great explainer on just why, despite the huge marketing push, AI for the consumer space (especially Copilot) has so far felt a little bit underwhelming and lacklustre, and if you've got the time it's definitely worth a watch!

I don't care about data scraping from ao3 (or tbh from anywhere) because it's fair use to take preexisting works and transform them (including by using them to train an LLM), which is the entire legal basis of how the OTW functions.

[mechanicalinertia]

Content-Free Erpdert 2020-Oh-One: The Roles We Must Play in the Great Game

Okay. Let’s talk about my Bubblegum Crisis 2069 RPG. This isn’t a postmortem, though it feels like one. Think of it as laying out the prospects for the project’s necromantic revival. Think of me as Frankenstein, looking for a good ol’ bolt of lightning to give his hacked-together corpse-man life.

So:

1. The original reason I stalled out on the RPG was Megatokyo. Basically I liked The Sprawl’s citybookish supplement November Metric, specifically Benjamin Kouppi’s mini-citybooks for Brussels, Lagos, and Miami, so much that I figured the least I could do was highlight different parts of Megatokyo - how GENOM’s vision of the model megacity clashes with the reality of daily life, the street finding its own uses for things, the usual cyberpunk stuff - in a similar manner. Only then I realized I really, really wanted to avoid Orientalist stereotypes, and then I realized that really, really restricted what I could actually do in the various districts - or would a district-based approach really work? Why not just layer down the fundamental concepts, the aesthetics of Megatokyo? Probably because I didn’t know what the specifics of a cyberpunk megacity would be, realistically, and I didn’t want to just imitate CP2020′s Night City book - I wanted to create anchors for the players to tie their campaigns to. Something freeform, yet concrete, and yet I just couldn’t gin up the creative juices to write it in a satisfactory way. Well, alright, I thought, I’ll just move on. I’d written plenty of lore already, right, so there was nothing wrong with moving on.

But I’m not very good at moving on. See, it was at that point that, in light of all of BGC2069′s inspirations, I started to rethink the first half of the game - namely, its mechanics.

2. I made it very clear from the start that what I was doing was a hack of the RTAL rules for CP2020 and the BGC RPG. Kick out all the unnecessary stats from Fuzion - just scrub Fuzion from the record entirely. Use the pared-down gameplay from Hunter-Seeker, but ignore the godawful setting from that same hack - it reeked too much of Neuropolitan, which in turn reeked too much of Snow Crash, and that wasn’t the sense of the lore I was going for. And for the most part, I think the basics succeed in imitating CP2020 gameplay and combat without too many extra moving parts.

Oh, but when I started bolting moving parts on, that was when things got complicated. There’s no doubt I cast my net too goddamn wide, trying to alter the flawed systems of the original 2020 in the style of various fan-supplements from the internet of days long past. One of the few people who looked my abomination over said that he didn’t like the way lore and mechanics were blended together in those sections - too confusing, he said. And yeah, I was trying to twist CP2020′s looser ends and make them my own. But I overdid it. Gun-printing? Cybernetics and humanity loss? Hacking? All these things are Cyberpunk staples, yes, but I a) made them too damn complex (and in the case of the hacking’s vision of the ‘Noosphere’, potentially inaccurate), and b) they’re not Bubblegum Crisis staples.

By this I mean that original BGC eschews cybernetic augmentations (except for the bizarre and hyper-pulpy AD Police Files OVA, whose canonicity is dubious at best), elaborate hacking, and unhardsuited combat in its action scenes, or really a lot of excessive gun-porn, and it does just fine. So if I had to do it all over again - and something tells me I ought to - I would minimize these elements. Streamline the whole project. Focus on BGC’s strengths.

3. Well, what are those strengths? Anime-Superheroic-Cyberpunk is a great genre cocktail, but where do those mental Venn diagrams align? Well, in the Knight Sabers. Normal(ish) people in a cyberpunk future (not professional black-trenchcoat operatives in the least) with melodramatic problems solved by the application of high-flying, high-tech mecha-violence. That’s the bottled lightning that is Bubblegum Crisis, and as much as I hate to admit it, that’s what people want to play in this particular universe. If people want to play a different kind of cyberpunk campaign - one where hardsuited combat against Boomers with a pinch of anime-esque melodrama isn’t the norm - they’re gonna play Shadowrun or CPRed or The Sprawl or The Veil or Interface Zero (ech) or even Eclipse Phase if they’re desperate. So if we take ‘play as the Knight Sabers or a similar team’ as our modus operandi - well, what does that tell us?

4. Well, again, that ‘bareskin’ combat should be minimalized, cybernetics largely ignored, and hacking simplified for heat-of-the-moment electronic warfare (hacking a Boomer on the fly, not cracking a massively complex corporate mainframe). It doesn’t preclude using Interlock-ish rules (Stat-Skill-D10 vs. Target Number) per se - it doesn’t mandate a simpler system like BESM or OVA or something similarly rules-light. I’d argue that, because mecha-porn is so fundamental to BGC as its own thing, that if anything the game needs a slightly crunchier rules system just to build those mecha.

5. Or does it? Don’t get me wrong, I’m very proud of the mutation I cobbled together of Mekton Zeta Plus. However, I’m not as proud of the hardsuit-construction system I made based off of that system. It more feels like something the GM would use to make hardsuits for the players, not letting the players build their own suits.

Furthermore, I was sort of kicking the can down the road by giving GM’s such an extensive toolkit. It would be more desirable on the players’ end, I think, to just give them a monster-manual full of pre-built Boomers and maybe a few larger mecha, so the only mecha-building the GM or players are doing is their hardsuits / Motoslaves at the beginning of the campaign, then maybe a monster-of-the-week supermech like the DD once or twice. Likewise, I hadn’t bothered to actually make the rebooted versions of the Sabers and other characters to interact with PC’s, or for the players to inhabit, or hell, just to serve as inspiration for their own superheroes. So. Gotta do that.

6. I keep itching for some sort of roleplaying ‘training wheels’, some sort of system to encourage, if not directly reward, players to play around with inhabiting their characters. Like, if you had people who were interested in ‘hard’ roleplaying in their TTRPG, but didn’t really know how to do it. I keep thinking about the ‘clocks’ in The Sprawl, how they determine a sense of danger and urgency and generally are designed to keep the players moving forward through the beats of the story. I think that’s really cool, but at the same time it’s a very PBTA-tastic system - would it work for the ‘bareskin’ side of the game where the characters don’t have much more than a pistol and their wits to uncover what they must do once they get into hardsuits? One could make two very different games, here. The ‘Bareskin’ game, which is all about feelings and plots and story beats, and the ‘Hardsuit’ game, which is much more classic bang-bang-punch-punch stuff.

It’s not as intuitive, though, to design those sorts of things. And then I wonder if I’m taking the game too far away from it’s CP2020 roots, trying to make the game be something it isn’t. I’ve got this great grab bag of inspirations, but to make a BGC game truly itself - that’s hard. Because I’m not sure what trends I should be chasing, here.

7. But I’m getting myself muddled, here. I need a plan of action.

So:

A. Modularize the game. Core Rulebook largely ignores gun-printing, complex (deep) hacking, cybernetics, etc, and eschews them in favor of getting the players in their character’s shoes and hardsuits at a reasonable clip. I can put the complex stuff in its own pseudo-expansion and play around with it as necessary. Likewise, Mekton Infinity is great for complexity and granularity, but players are more gonna be coming from something more like Lancer than Mekton Zeta. So they’re gonna want some pre-built inspiration even if the endgame is to have players build their own suits.

B. Build the ‘sourcebook’. Focus a little less on the lore, focus more on where lore and mechanics intersect, namely in things. NPC’s to hang out with, mecha to battle.

C. While I’m at it it probably couldn’t hurt to make the general rules of heavier mecha combat more granular. Car chases and dogfights and things like that, alongside the more infantry-scale of hardsuit combat. Help people play the game with just some crude printed miniatures and nothing else.

D. Write the GM’s advice thing, see if that helps nail down what kind of game BGC is supposed to be. You know, what makes a BGC game unique, what should draw the players in, contrast it with other mecha-cyberpunk-anime-superhero games.

E. Then, I think, whether or not the game needs to segregate ‘bareskin’ and ‘hardsuit’ gameplay - and subsequently whether or not I need to add more rules like that - will become much clearer.

Whoo. That was a lot of text. We’ll see if anything comes of it, I guess.