There isn’t anyone who cares about sc2 enough to run a bot network on its social media sites. Blizzard also sure doesn’t care enough either. The skills to do that are rare and better spent making $75/hour on projects that actually matter. Nobody has time for the sc2 forums. Reddit, maybe. SC2 forums – definitely not. It’s a good theory because a lot of the posters here and on reddit definitely do behave a lot like bots (bad reasoning, low effort, discontinuity between posts, etc) so I had the same thought. But there is no API to interaact with discourse (the forum software) and developing one just isn’t woth the time. Nobody cares about this place nor reddit. I could see some youtubers caring enough to maybe promote their reddit content with bots, but there are services for that already. You can literally buy upvotes and comments. How often is that used? Hard to tell but I am sure you could find suspicious posts with a statistical analysis. It’s an interesting question but no real practical uses. I could see an academic doing it to get their phd or maybe activist against dead internet, or maybe a youtuber rivalry.
EDIT: yeah so you are talking about the nvidia sell off. It’s boomers who don’t understand AI tech. Sabine Hossenfelder had a good video about it. There is a natural limit to the accuracy that emerges in the training time and this is true across every known model. Deepseek’s claims are improbable because it would be a steep departure from past trends. Even if Deepseek’s algorithm is more efficient, it would still be faster to do it on a GPU unless of course the algorithm is totally incompatible with GPU architecture and I’ve seen nothing to indicate that. I would be surprised if their training model is open source. It’s probably the client that runs the model that was already trained that is open source. There is no way China would allow the open-sorcing of an algorithm like that especially when they’ve work hard to develop it as a response to CPU and GPU embargos. So we can’t say if the algorithm is incompatible with the GPU architecture or not (probably not). So it’s a lot of superstitious reasoning not based on the actual science and engineering of AI.
AI research has been the majority of my time as a software developer. I’ve worked on a program that finds equations to represent systems, to find alternative ways of computing the same answers (to avoid patents) and to find approximate solutions to known problems that are faster than the known methods. That was a big part of my life a couple years ago. To this day, I have an AI running on my gaming PC (since I am in Texas helping a friend) and it optimizes centrifugal impellers for car turbochargers. On the CNC machine I prototyped, I plan to build one and install it on my car. While relaxing here, I’ve been working on a new version of the CPU block designing AI that made my current CPU block. The hardest part of an AI project is to define the solution space in a way that makes it searchable and that requires a deep understanding of the actual system you are modelling – you simply don’t know what the ideal parameters are for a given application, and want to optimize the parameters. The program has to have a way of understanding how to not create nonsense parameter combinations as a method for iterating the design based on past testing.
Well, these algorithms can’t be parallelized on a GPU because the way the GPU is so fast is by doing a lot of the same calculation in parallel, while these algorithms require steps in series. On a GPU, you can calculate A, B and C at the same time, but if C depends on B and A depeds on C then you can’t use a GPU. That’s the gist of it.
AI models are usually matrix multiplication and mass amounts of it so GPUs are very good at it. They use a gradient descent algorithm like simulated annealing where values in matrices are randomly tweaked and if it improves the result then the tweaks are kept and if not they are discarded. When you have a chain of 1,000 matrices per solution, the probability of finding a tweak that improves the algorithm is very low, so this has to be millions of times to find each improvement.
If deepseek is incompatible with the GPU architecture, it would have to do no matrix math nor any parellelizable calculation. It’s very unlikely that it wouldn’t benefit from GPU acceleration and that means the GPU sell off is dumb. You could say that maybe their training model is better and that that means a lower demand, but it’s very unlikely due to the natural law I was mentioning. There isn’t a name for it and it isn’t proven, but it’s there in the data. Proving that such a limit exists is probably possible, and it would probably be an adaptation of the Shannon coding theorem. Compression models as well as data encoding models used in telecommunications have been thoroughy explored and we know the limits of what’s possible. Shannon’s theorem tells you how well you can predict a future character based on past characters which is how data compression works. That’s basically what language AIs do and it almost certainly has a limit similar to the Shannon coding limit.
Anyway I haven’t been following the news because I am busy as heck down here. We’re talking 11 hour days on my feet – sheetrock, tile, framing, installing windows, plumbing, you name it. It’s a full remodel of 2 bathrooms. So anyway I thought this chick I knew was in a relationship with another chick, but it turns out my friends girlfriend is just a friend-girl. My friend let her stay at her house to help her out. I ask her, so, you trying out something new, as I tilt my head at the friend-girl. She almost chokes and says no I am just helping her out and she’s really smart and deserves a second chance. She has had a hard life but she is super smart. She tought herself how to read at the age of 7. I am thinking Y I K E E S. Her spelling to this day is still atrocious, she continues, and I am thinking oh no this girl is dumb as a rock and my friend is just lonely and needed some company. So anyway my friend is helping me tile a bathroom floor and you spread out the mortar and place a tile, clean the edges with your finger and and place levelling clips under the tile which pinch neighboring tiles together to keep them level. Well anyway each time I used my finger to wipe away the excess mortar I would flick it into a pile of mortar and it would make a noise. So I am going about my business and she says “that noise is so satisfying” and I am like what noise. She says that noise as I do it again. Well it has a rather striking resemblence to a clapping noise of a rather prurient nature. I wasn’t really paying attention to what she was doing as I was zoned in, so I look up at her and say what is that supposed to mean. Is that some kind of hint. She sticks her tongue out, bites it, looks up into the air and shakes her head before saying what do you think it’s supposed to mean. Her hair was waving as she tilted her head. She then runs off giggling. I am just sitting here thinking what an odd time to make a move. I am sweaty, covered in mud and dirt. I guess the mortar flicking was a trigger. Whatever. You don’t look a gift horse in the mouth, if you know what I mean. Well anyway I’ve got my hands full down here until this Saturday. We were having so much fun I stayed another week and it will be hell to pay when I return home but oh well. She has to go back to being lonely as she hangs out with her new roommate. How long until the roommate overstays her welcome?
This girl is a bit of a tomboy and so she has a lot of random tools, but mostly automotive. I am in her garage and there is a random small block chevy crate engine up on an engine mount. She bought an old hot rod to work on it, lost interest, sold the hotrod but still has the engine she bought. So anwyay she doesn’t have any sheetrock tools nor any tiling tools, so I am now buying 300 and 150 dollar tools so I can have this finished for her before the week is over. I tell her if she wants to keep them, she can have them, but if she doesn’t want them to please resell them for 75% the value and send the money back to me. It’s another tool for the collection but it’s also space in your garage.
Meanwhile the roommate is complaining that things are taking too long. So my friend tells her she should pitch in by banging out some concrete. Well this girl decides to do the whole thing. Puts in insulation, levels the wall, puts up wallboard. All on her own, she completed a process that took me 3 days. You might think wow that’s amazing but it looks like an orangutan was let loose with a hammer. It’s shaping up to be the worst tile job I’ve ever seen, and I’ve seen some really bad ones. I show this to my friend and she doesn’t want to be mean to her guest so she’s like well maybe if you show her why it’s not going to work, like you showed me, then she will understand when we rip it all out and redo it. I am like oh man this girl has adopted this other girl and treats her like a child. So anyway, she invites her to come out and the girl refuses, and now she’s a recluse and probably has a grudge against me. Oh well. Anyway, where do you think the name Slam’er came from. Yep, I’ve laid so much pipe across so many states they ought to sent me my plumbing license by now.
ax^3 + bx^2 + cx + d = y
This can be parallelized because you can compute each term separately. The terms have no co dependence until they are added together. This would be a good use of SSE instructions.
Now if it were,
(aX+bX+Cx)^3 = y
Then there wouldn’t be much to parallelize because there are operations dependent on previous operations. You can think of it like operations in chains vs branches. If the logic branches, it can be parallelized. If the logic chains, it can’t be parallelized.