You jest, but at this point YouTube’s recommendation system sees even less complicated than this. Just random videos from years ago will one day become blessed by the algorithm and receive millions of views.
YTs algorithm now just seems to amplify very, very small fluctuations in view counts. 3 people just happened to watch the same video at the same time? It must be trending! Shove it in front of everyone. It’s like Google has used all of machine learning to recreate taking a derivative.
Always cool to see new cloud GPU offerings, but isn't Lambda Labs' cloud offering roughly equivalent (in the case of V100s) or substantially cheaper (in the case of A100s)?
Thanks for the comment! Yes, Lambda Labs recently lowered pricing, so they're now roughly equivalent to us, and they beat us on some configurations.
I haven't used them, so please fact check me, but it seems like the machines come with directly attached storage. So, if you're using an 8x V100 and want to switch to a 1x RTX 6000, you'd have to spin up a new server and manually migrate your data over.
We built our platform with networked storage. You can spin up a CPU-only instance for $0.027/hour (<$20/month), upload your data, convert it into a GPU instance to train your models, and then convert it back. We frequently see users converting servers from 8x A100s (to train workloads) back to 1x RTX 4000s (to run inference). This kind of flexibility saves people time, which equates to money given how expensive ML developers are now.
(Our networked storage model also enables people to shut off their VMs and save money)
I'm sure Lambda Labs is working on something similar, but it seems they are doing dedicated servers based on how they advertise.
I think we also have a higher variety of GPUs (10 SKUs with us vs 4 SKUs). This lets people switch from between, say, an NVIDIA A6000 to A5000 to A4000 to truly "right-size" their compute so they don't pay for anything they don't need.
Cost-wise, we also have better long-term pricing like GeForce 1070s for $100/month in Boston or $150/month in Singapore Equinix SG1 - which is really good pricing for an APAC city in my opinion (https://console.tensordock.com/order_subscription), and we're working on a marketplace to let compute suppliers list their compute on our platform to get closer to cheapest for those who really care about cost (https://www.tensordock.com/product-marketplace).
Lambda Labs has (slow, low IOPS) cloud filesystem to persist data between instances. Attached storage does not persist but is high bandwidth and high IOPS, which is a necessity if training small-medium sized models.
The government has an implicit return on investment for the entire economy, through the power to levy taxes. Why should the government own stock in specific companies? The only likely outcome for that route is the government playing favorites even more so than at present
Why should tax dollars be sent to a private company without anything but a promise in exchange? The promise of "jobs" isn't enough, as evidenced by just about every single other time that these programs (including the tax cuts) happen. Further, how are we going to hold these companies accountable for their promise? What comes to mind immediately is when we gave telecoms a bunch of wire up rural America, and they purchased and consolidated the cell providers instead.
If we just write them a check, they'll do whatever they want with it. They need to be held accountable, and ownership percentage is a way to do that.
Alternatively, the gov could take this money, start up a corporation of their own, and make their own fab. The language in part of this bill is that they are trying to ensure that older tech and DoD stuff is made in house. So let's make it in house.
We don't necessarily need to say "we're buying intel or nvidia", but we can make the money available to any company in exchange for the ownership percentage. That eliminates the "playing favorites" issue, I would think.
Let's say Intel takes the deal and trades some ownership percentage for some funds to build more fabs. Suddenly the US government has a multi-billion dollar reason to provide political support for whatever is good for Intel. It's inherently favorable to incumbents who have the market share to make a favorable deal, regardless of whether or not those incumbents are best able to build world-class fabs.
The government has plenty of demand-side tools to on-shore development without hurting competition or playing favorites with large incumbents. The DoD and DoE are massive semiconductor purchasers and have a lot of leverage in the market who can adjust their procurement strategies to promote American interests. That's not entirely without the opportunity for corruption and grift either, but at least there's more accountability.
I mean, I think we basically agree that handouts to large semiconductor companies without anything in return is a bad idea.
> Suddenly the US government has a multi-billion dollar reason to provide political support for whatever is good for Intel.
That's one of the reasons we're doing this in the first place. National security is in the mix as well. What would actually change [edit: beyond citizens gaining ownership and some say in what happens]?
If we just write them a check, they'll do whatever they want with it. They need to be held accountable, and ownership percentage is a way to do that.
You could also just consider the grants to be loans that never need to be paid back as long as some set of concrete requirements is met by the company.
The federal government previously gave the telecom monopolists billions to build out nationwide fiber-optic internet infrastructure... and they just took the money and ran.
I don't see any language around how we're going to recover those funds, if they are not spent in ways that align with the requirements laid out in the bill. Did I miss that section? It is a pretty big bill.
I do see a section that would allow the suspension of new funds, if it's found that these companies aren't upholding the agreement. I don't think suspension of new funds is sufficient.
> don't see any language around how we're going to recover those funds, if they are not spent in ways that align with the requirements laid out in the bill
Implementation is left to the agencies. We're only seeing the Congress appropriating money. Contract specifics are being developed.
I mean, when the US government gives out money for student loans, the loan may be on very favorable terms, but it still expects repayment. It's only in business where we hand out money and don't expect it to ever be paid back.
I would agree that student loans shouldn't be that way either, but, we shouldn't be handing out money upfront without guarantees that the incentivized behavior will actually happen - the two things that come to mind are the 90's fiber rollout and the Foxconn subsidies in wisconsin, both of which the company took the money and then never actually built the infrastructure.
If there are going to be outright subsidies, they certainly need to be contingent on meeting milestones, and not just "pretty please build a plant, here's a big check to encourage you but no strings attached".
Think about how the US did the auto bailouts as well - we got a big chunk of equity that we eventually unwound on favorable terms, but we didn't just write a check and be done either. Same thing, maybe if TSMC wants a big grant for expansion, maybe the US should get a big chunk of equity (at guaranteed and favorable terms) and then if milestones are met the equity is again wound down at favorable terms such that the benefit of the "favorable terms" on both sides amounts to the subsidy.
As I understand it, even though the Foxconn deal was ill-conceived and wasted tax payer money via planning for the deal and funding foolish road projects, the deal does have milestones built in that resulted in not receiving the credit for at least the first few years. I also think the amounts in the deal have also been further diminished after failing to meet the original milestones.
That said, I was looking into it today and it looks like they have actually been meeting some sort of revised hiring target recently [1]. I have heard from some peers that the jobs they are hiring for are basically make-work jobs, but the deal doesn't care what they accomplish -- only how many "jobs" they create.
It's still shady crap which wasted tax payer money (not to mention abused eminent domain), but I don't think in any way involved a big check with no strings attached.
If there's some externality in the way that someone is using public air, like burning garbage in their backyard, then they should be expected to pay for it
Economies of scale don’t just have to be areas where unit marginal costs decrease as volume increases, they can also be areas where fixed costs dominate marginal costs, which is the case for many software firms because marginal costs are usually very small
I’m working on making robots to collect bottles and redeem the California refund value. Basically so you could bring your bottles to a big mechanical sorter, dump your bottles in, and get maybe like 2c per bottle paid to you, then 3c pays for the cost of transport. Currently based out of Mountain View. My background is in machine learning and I’m muddling through the electronics and mechanics, would definitely like to have help with that stuff or really just find other people interested in this intersection. If interested send me an email at contact@treasurerobotics.com
Might try experimenting with changing the reward. Instead of money, every 25 bottles means a tree is planeted or something. The 2cents a bottle isn't much of motivation for regular people.
Easy for me to suggest as I don't have to figure it out, but maybe something to consider.
I just recycle my bottles here in CA but does CA not already have these? We had machines for the bottles/cans in Vermont. Interested in helping if I'm missing something.
