Bassman2003
Veteran
The speed is related to the spend. And they are spending...
is AI gonna take our jobs?
- Thread starter DBP
- Start date
Bassman2003
Veteran
I think the new development is interesting and looks to be positive. I do not know where this leaves the hardware makers. While the new Chinese AI model shows it can be done on lesser hardware, the big boys want the compute power to install AI everywhere in realtime. So the need to ultra fast hardware still seems to be there to some extent. But I might not be realizing that the Deepseek deployment is already realtime everywhere... This is complicated stuff and I am sure a lot of folks are taking notice.
ahalpert
Major Contributor
I think the question is: does demand increase if the cost goes down? If it's cheaper/easier to use AI, does it get used in more places? If so, then the hardware demands may be stay the same pre-DeepSeek and post-DeepSeek.
But if there's a limit to how much AI will be used and we were gonna employ it everywhere we possibly could anyway, then these efficiency breakthroughs will mean that NVIDIA gets to sell 10% as many chips. My guess is the answer lies somewhere in the middle. Which is bad news for NVIDIA even though they'll still make a fortune.
ahalpert
Major Contributor
if you have to change the law to accommodate AI, that suggests that it's currently breaking the law
NorBro
Major Contributor
but laws are systems made up by flawed humans that have changed over the course of history. they can drastically vary by only miles apart, existing over pieces of land that are separated by imaginary lines, and aren't always recognized thoroughly or accepted - and, of course, many are broken frequently.
Bassman2003
Veteran
Yes, what is bad for Nvidia is great for Meta etc... The "deployers" of AI do not want to buy 100,000 Blackwell units from Nvidia if they do not have to. If this Deepseek development is the real deal, then I do think the near term bubble has been popped. Nvidia will still have demand for their new Blackwell unit but maybe not at the numbers first thought and the amount of players in the space. It is all kind of made up demand/importance anyway right? Sort of like people making NFTs of their Myspace page to hang on the wall of their Metaverse home
ahalpert
Major Contributor
Yes but sometimes you have a law rooted in principles of universal fairness. And sometimes somebody with a lot of money pays your politicians to change the law so they can legally steal
ahalpert
Major Contributor
I think that's true. Meta and Apple are up 2% today. NVIDIA is down almost 17%.Yes, what is bad for Nvidia is great for Meta etc... The "deployers" of AI do not want to buy 100,000 Blackwell units from Nvidia if they do not have to. If this Deepseek development is the real deal, then I do think the near term bubble has been popped. Nvidia will still have demand for their new Blackwell unit but maybe not at the numbers first thought and the amount of players in the space. It is all kind of made up demand/importance anyway right? Sort of like people making NFTs of their Myspace page to hang on the wall of their Metaverse home
"Nvidia (NVDA) and Broadcom (AVGO) led chip stocks down, while the tech-heavy Nasdaq dropped 2.6%. Hyperscalers Microsoft (MSFT), Google (GOOG), and Amazon (AMZN) also fell."
ahalpert
Major Contributor
yeah that's the logic of market bubbles. it doesn't really matter if you're making anything substantive as long as you get people excited long enough to leave them holding the bag
Bassman2003 said:
ahalpert said:
The $5 million number, though, is highly misleading, according to Bernstein analyst Stacy Rasgon. “Did DeepSeek really ‘build OpenAI for $5M?’ Of course not,” he wrote in a note to clients over the weekend.
That number corresponds to DeepSeek-V3, a “mixture-of-experts” model that “through a number of optimizations and clever techniques can provide similar or better performance vs other large foundational models but requires a small fraction of the compute resources to train,” according to Rasgon. But the $5 million figure “does not include all the other costs associated with prior research and experiments on architectures, algorithms, or data,” he continued, adding that this type of model is designed “to significantly reduce cost to train and run, given that only a portion of the parameter set is active at any one time.”
Raymond James’s Pajjuri thinks investors should also think about inferencing. Training is the process of showing a model data that will teach it to draw conclusions, and inferencing is the process of putting that model to work based on new data. Pajjuri argued that “as training costs decline, more AI use cases could emerge, driving significant growth in inferencing,” including for models like DeepSeek’s R1 and OpenAI’s o1.
ahalpert
Major Contributor
Isn't that like saying the true cost of chatgpt is the entire budget of OpenAI from its founding until the release of the latest model? In that case, DeepSeek would still come out way ahead.
But I think the bigger picture is that they are able to train and run the model using less memory, far fewer GPUs, and much less compute than chatgpt. There's no question that it's a cheaper approach. It may not be as good. I've read estimates saying it gives 90% of the performance. But I've also read estimates saying that it's on par or that it's superior.
The efficiency breakthroughs are interesting. They remind me of how organisms in nature evolve time- and energy-saving workarounds. It only activates part of its model at a time. They had an efficiency breakthrough in the training methodology as well but I can't locate the details.
Looks like DeepSeek was just hit with a massive cyberattack. I'd be shocked if the malicious actors weren't American.
Broadly speaking, my hope is that if AI can't be monopolized then the powers that be will lose their incentive to shred the social contract in service of its profits because the profits won't be there in the same way.
So I gave AI a spin recently in the form of the various online image generators, and I am not worried about real cameras going away. I say "spin" because they all feel like you're in a casino turning a roulette wheel every time you click. There's so little control over the details, even when following the best practice guides and writing the most optimized prompts for each one, you just have to accept whatever random result comes out of the machine. While these images aren't as smeary as I remember them being a few years ago, they still don't look believable. I guess a quintillion flops of whatevers aren't enough to do anything consistently on the next go 'round. Video is essentially a sequence of still images. When a real subject is in front of a lens, any turn or new pose instantaneously results in an true-to-life image. People like consistency - getting to know someone or something they're looking at. I don't think blurring and upscaling a hodgepodge of library textures is a way to ever achieve this. Even if they tack on a few more algorithms that check for backwards hands and extra fingers, the uncanny valley isn't going to be bridged.
I think the hype for it all goes back to the casino analogy, it's thrilling at first to have some brand new gizmo light up for you, and it might remain popular with the less "picky," but I felt my excitement fade after discovering how narrow the scope of creative possibilities gets when you leave most of it up to chance. I've heard people who make music say similar things about working with a real DAW compared to the song generators - you can't get it to sound exactly like you hear it in your head. Although, maybe there aren't that many people with the type of intent that demands precisely what you envision reflected in your "work."
Is a set or a costume more expensive than a new graphics card? Do the credit token subscriptions cost more than a guitar or analog synth? It plays out the same screen and speakers in the end, so who cares, let it ride.
I think the hype for it all goes back to the casino analogy, it's thrilling at first to have some brand new gizmo light up for you, and it might remain popular with the less "picky," but I felt my excitement fade after discovering how narrow the scope of creative possibilities gets when you leave most of it up to chance. I've heard people who make music say similar things about working with a real DAW compared to the song generators - you can't get it to sound exactly like you hear it in your head. Although, maybe there aren't that many people with the type of intent that demands precisely what you envision reflected in your "work."
Is a set or a costume more expensive than a new graphics card? Do the credit token subscriptions cost more than a guitar or analog synth? It plays out the same screen and speakers in the end, so who cares, let it ride.
ahalpert
Major Contributor
It's interesting to me that so many people offer extreme takes like how the US has lost the AI race or something like that (or how AI will put millions out of work, etc). For example, NVIDIA dropped roughly 17% yesterday. But they're still up 8% over the last 6 months!!! They're still up 2,000% from 5 years ago! The S&P is still up over 2% from a month ago. People are calling this a stock market crash. It is not.
I don't know where this is all going. And the Deepseek development upends a lot of assumptions. But for the most part it should only tweak our predictions for the future, not alter them completely.
I think the news is much worse for OpenAI than for NVIDIA. It shows that they're not so special or unique. And the efficiency breakthroughs should lead to cost savings for the customer. Sora won't be able to charge $200/month after a Chinese competitor offers a similar model for $20/month.
I don't know where this is all going. And the Deepseek development upends a lot of assumptions. But for the most part it should only tweak our predictions for the future, not alter them completely.
I think the news is much worse for OpenAI than for NVIDIA. It shows that they're not so special or unique. And the efficiency breakthroughs should lead to cost savings for the customer. Sora won't be able to charge $200/month after a Chinese competitor offers a similar model for $20/month.
Bassman2003
Veteran
I am very interested now in seeing how this all turns out. Hamburger has good points and experiences in line with mine. It might be a flash in the pan for visuals, as it just takes too much effort to make it replace reality while it kills it in database gymnastics.
Do I want to use it, not at all! I see Gmail has all of this assistance stuff in email and it reminds me of a nat buzzing around at this point...
Do I want to use it, not at all!
I see Gmail has all of this assistance stuff in email and it reminds me of a nat buzzing around at this point...
NorBro
Major Contributor
I think of the AI quality like I do about video camera quality.
This stuff is just getting going, it took digital video like 40 years to get to where it is now.
Some have doubts about how good the AI could really get but so far it just keeps getting better and better.
Let it cook...
This stuff is just getting going, it took digital video like 40 years to get to where it is now.
Some have doubts about how good the AI could really get but so far it just keeps getting better and better.
Let it cook...
ahalpert
Major Contributor
Film still surpasses digital in numerous ways which is why the top dogs still use it. I think it's helpful to analyze what goes on beneath the hood of the competing systems to make predictions about where they'll end up.
For example, AI works in a kind of probabilistic way ("autocomplete for everything" as some say) and only works well with things it's seen before. I keep trying to upscale or correct focus on shots with artworks and it never works because the machine doesn't know what the artwork is supposed to look like.
Last edited:
NorBro
Major Contributor
Numerous sounds like a lot. How many ways could there possibly be?
I don't think I would notice the difference at this point.
I saw a 2024 list of movies shot on film a few weeks ago (end of the year) and I only recognized 2 names out of like 30.
That doesn't mean much since I'm not really a movie person, but I feel like sometimes someone might choose to shoot on film as part of the artistic statement (b/c almost no one would see the difference in the end result).
ahalpert
Major Contributor
I notice the difference even just watching TV shows from the 90s vs today. There's a more organic quality to the color. But the larger point is that you made an analogy between digital video and AI. If there's an eternal difference between video and film, there may be an eternal difference between AI and real photography.