11,108 karma · joined December 5, 2016
People formed their opinions on LLMs in late 2022/early 2023 when ChatGPT was first released and taking the world by storm and then never updated them after nearly 4 years of improvements.
Hell, just being able to search the Internet did MASSIVE things for reducing hallucinations.
Of course, maybe I'm wrong. Maybe you DO need to use a calculator to quickly tell me what 8 * 6 is. But I would be highly surprised if you got a degree in math and still needed a calculator for that.
And has since killed it for desktop gaming CPUs.
Which makes some sense, I suppose. It was pretty expensive to use, which made it uncommon. These days, GPUs are so power hungry that'd you need a MASSIVE power supply to do it, unless you're SLI'ing low-end GPUs, which is kind of a waste. You're better off with a mid-grade GPU.
...they would still have fumbled it.
3dfx failed for a single reason: Failure to design for the future.
They had a hyperfocus on brute force rasterization and fill rate that became detrimental. They dropped the ball on 32-bit rendering because they thought having a higher fill rate was more important. They dropped the ball on implementing hardware transform because they considered that the CPU's job.
And then they blew up their own business model buying STB Systems and manufacture cards themselves, rather than selling chips to other manufacturers. That ended up being a gift to NVIDIA who adopted the opposite strategy of providing chips, reference designs, drivers, etc. and let a large number of board manufacturers compete to sell them.
And then there's the fact that NVIDIA was allowing themselves to create new architectures, while 3dfx was basically just saying "Let's make the same GPU, but with more texture units". By the time the Voodoo5 came out, NVIDIA was already on the GeForce 2.
3dfx lost because it kept optimizing for yesterday. NVIDIA understood where the industry was going and 3dfx didn't, and there's little reason to think AI would have gone any differently. If 3dfx had somehow survived another 20 years, I'd expect them to arrive at the AI boom with the world's fastest rasterizer.
The high quality sprites are 64 pixels across per tile. With each tile being 1 meter, 1 pixel is 1.56 cm (About 0.63 inches) across. By today's standards, that's a very low texture resolution, though they can get away with it when your camera is so far away.
But if the game was 3D, you could expect players to put the camera closer to your objects. You'd need significantly higher texture resolutions. And you'd need a texture for each part of the object.
Meanwhile, consider a 3x3 object. That's 192x192 = 36,864 pixels. Assume it has a 60 fps, 2-second animation loop. That's 4,423,680 pixels. With 8 bytes per channel and 4 channels (RGBA), that's about 17.7 MB, which allows you to store 56 objects per GB.
Which doesn't sound like a lot, but we're making some huge assumptions. Not all objects are 3x3 tiles. Most don't have 120 frames of animation.
And as I mentioned in another comment, you don't need to store ALL the game's art in VRAM at once. Only what's on the screen. And you can certainly fit an entire screen's worth of animations in a 2 GB GPU.
The game doesn't consume that much VRAM. It plays on 4 GB GPUs just fine. The GPU only needs to keep resident what's currently on the screen. Even including every animation frame, that's 1 GB tops.
Hell, the minimum requirements are a GeForce 750 Ti, or Radeon r7 360. Those are 2 GB GPUs.
And those people are often bad at math or grossly undervalue their time.
Some people drive 10 minutes out of their way (20 min round trip) to save 5 cents/gallon. Even if you have a 20 gallon tank, that's only saving $1.00. I wouldn't take 20 minutes out of my day just to save $1, even 15 years ago when I was living paycheck-to-paycheck on $10/hr.
They're definitely not. At least, certainly not any time within the next 10 years.
Encourage building upwards to increase housing supply. This can be done via tax breaks or subsidies. Especially encourage multi-use buildings to make cities walkable. Stores/restaurants at the street level with apartments above.
Alternatively, they keep the house, but get either a reverse mortgage or a cash out refi to survive until they die, at which point their kids can't afford to buy it from the bank.
No you don't. 20 minutes at a DCFC will get you to at least 80% for most EVs.
You need several hours to charge it at home, but it's gonna be at home for several hours anyways. With an L2 charger, you can go from 0-100% overnight.
EVs end up with unique brake problems because they get used so little.
> Then there's tire dust particulate considerations with the heavier EV with all that torque.
This part is slightly true. EVs are heavier and will wear out tires faster, and if you actually use the extra torque, that'll contribute.
Sadly, I might have to settle for the Porsche 718. Polestar 6 got me excited, but that project is basically dead, especially with Polestar being effectively banned starting with model year 2027. If the US government got its head out of its ass regarding Chinese cars, I'd REALLY like a BYD Denza Z Spider.
The sad thing is, the Venn Diagram of EV enthusiasts and sports car enthusiasts is almost two separate circles.
...was also incredibly slow.
A pitiful 70 horsepower motor assisted by a 44 horsepower motor that combined to an anemic 104 horsepower [0] on a car that weighed ~2,800 lbs.
I had to borrow my in-laws' car once, which is a first gen Prius, and I found the extremely sluggish performance to be outright frustrating. I felt like I had to floor the accelerator to reach highway speeds in any reasonable amount of time.
[0] Not a typo. The numbers don't add to 114 because they can't both deliver peak power at the same time.
Otherwise, you're gonna need an electrician to put a 240v outlet or dedicated EV charger.
Something I've been yelling from the rooftops.
Housing can either be affordable, or it can be an investment that's bought, rented, and sold for the sole purpose of profit. It's not possible for it to be both.
People expect their house to appreciate faster than inflation, but all that does is rob the next generation of home ownership.
Indeed. For me, I find the ecosystem around AI/LLMs works better on Linux than Windows, but Windows is my main OS since I'm a gamer. Being able to run GPU-accelerated AI in a VM is huge for me.
In my case though, I just use WSL which does an amazing job.
> What kinds of things can a guest running undesirably applications (viruses, malware, LLM escaping a sandbox, etc) get up to with shared GPU access?
The most obvious answer is a DoS. If my malicious VM is sharing a GPU and has full access to it, I could simply tell the GPU not to run a victim VM's workload, or manipulate it in some way. I might not be able to pivot to having a shell on their VM, but I could at least read/write their data in VRAM. If it contained secret data (custom model, or secret data being processed by AI), I could easily steal it.
The problem of course is that a restaurant that refuses to work with deliveries then ends up getting significantly fewer orders than ones that do. It creates a bit of a racket. Restaurants feel forced to cooperate, or lose significant business.
As a customer, I get around a lot of the extra costs because my credit card gives me the Dash Pass for free as a little bonus, and Costco sells $100 DD gift cards for $80, so I'm effectively getting every deliver at 20% off.
One word: variety.
My local sushi place isn't going to have enough delivery orders to justify having delivery drivers on staff.
Heck, I remember years ago (Was it the late 90s?) when McDonald's tried to do delivery.
And while in the past, the doubling of transistors DID double performance, it's no longer true.
Exactly.
Jev has no moat, and the incumbents will devour their lunch if Jev actually starts gaining traction.
Dash cams would effectively become illegal.
Someone just wanting to record themselves and some friends doing something would need to ask explicit permission from everyone in the background.
It's just not a tenable option.
In the USA, a significant amount of tax-funded infrastructure is actually built out by contractors, which means there's profit going to a corporation, and there's often a cynical idea that the contracts go out to whoever will send a kickback in the form of a campaign contribution to whoever makes the decision on which contractor to use.