Software, there are 100s of ai hardware startups. It’s easier relatively to build a domain specific chip, but building a compiler stack that integrates with a major framework and gets good perf from your chip is incredibly difficult. Getting people to switch from nvidia? Even harder
Yeah for inference this is true, there could be a viable subset of models. You’re not going to build a viable business on inference though. It’s super cheap already and plenty of hardware can do it ootb with an existing framework as you’re saying. The $$ for selling chips is in training, and researchers trying new architectures are not going to wait for a port of their favorite model in a custom DSL or learn a new language to start prototyping now. You can port models forever, but that isn’t an ecosystem or a cuda compete. OpenCL + AMD != a from scratch company
This 95% of models are statically computable thing really shows how much he is trivializing this problem. I’d be interested to see his SW stack compile MaskRCNN. His ISA is massively under-defined and people will not change their model code to run on this accelerator unless his performance beats cuda significantly and even then they still won’t - usability matters more than performance every time. In the end you need a compiler, and it needs to be compatible with an existing framework which is not trivial at all, since they are written in python.
It isn’t that there are suboptimal patterns, it’s just the more expressive your language can be at runtime, the less you can reason about statically. An example is data-dependent control flow. If you can’t reason about what branch your code is going to take statically (without your runtime data) it is harder to generate fast code for it.
Something that’s lost in these arguments a lot is that you can often get large models to cough up an exact sample from their training set. It’s fair to say that a human cannot do that to the degree a model can.
Police should be okay being watched. They should be held to a higher standard than citizens, because they have the legal authority to use force lawfully.
Yeah I mean we also have more crime too, so I feel like we’d need to normalize this somehow with like incarcerated/crime rate or something. Given I think nonviolent offenses should be less severely punished than they are today.
I agree the US incarcerates too many people, does the comparison to China make sense? How can we trust their numbers especially when they disappear people.
The irony is meta has none of these things. It’s the least bureaucratic place I’ve ever worked. Whether or not that’s a good thing I’ll leave up to the reader.
I really love this idea! One critique is I only saw twelve recipes on my phone. I know you said you’re adding more but wanted to mention in case it’s a bug.
Powershell is incredibly powerful with | and %{} for transforming collections and SUPER simple .net integration. You can write commandlets in c#. It was one of the bright spots of my time at Microsoft.
I’m pretty appalled by some of the comments here. If someone scams people it doesn’t matter how complex the scam was or how stupid the victim was. It is a crime. We aren’t talking about robbing a convenience store here. 40 billion dollars - people’s life savings were completely put up in flames in the name of “innovation”. This is out of control. The reason these sentences need to happen is because retail investors need to be properly informed of the risks, and if leaders of crypto companies continue to say it’s stable and safe they should be prosecuted when it turns out not to be the case.
I’ve worked at a couple, I’ve probably been an outlier but I’ve really enjoyed them as long as you’re ruthless about finding projects that you enjoy. There is a ton of opportunity to work on pretty much anything you want at a large corporation. I’ve mostly focused on r&d hardware and compiler projects.
I get that temperature comes from the connection to physics but this seems like an overly complex explanation to a relatively simple formula. If you have a softmax and you multiply the input by some factor, if that factor forces the exponent to approach zero you’re approaching the uniform distribution, and if you make the exponent approach infinity you get the arg max. Curious if others feel the same way.
I work in a FAANG and I e always loved my job (switched teams and companies a few times) if you don’t love it anymore, find a project you do love, that’s the glory of a big company tbh. This was a pretty surface level article. There isn’t any industry on earth where you can be an IC making 600k except tech. At a FAANG 600k is reaching your bank account(stock is real comp)
My point is it is also non trivial, using privacy again (a much simpler problem IMO) the US utterly failed here. The EU tried but I don’t think we can call their legislation a success either. Biases are present in every statistical model, the idea the government will magically come up with easily verifiable metrics that need to be met is pure fantasy. (If such metrics are even possible) enforcement of something like that is probably impossible anyway due to IP concerns.
I respect her as a researcher and all but the constant reminder that there are biases in AI is utterly pointless. This is going to go the way of privacy - companies simply do not give a shit. They will pay lip service in some way or another for a little while, but growth and desperation if revenues fall will result in completely dropping any feigned morality and they will attempt anything to survive.
I watched about half of the video. Tbh his excuses about not properly counting sales tax to me just sound weird. if you have a business you need to hire someone who has expertise in this area if you grow to the size where it matters. I also find how he only shares his error rate and not the absolute numbers a little sus. I highly doubt the government would be auditing him if there wasn’t a lot missing.
Meanwhile the rest of hn is ready to jump on the “I hate the government, it’s so dumb” bandwagon when it’s actually really hard to create a system that works as well as it does. I think it should be simpler and clearer for everyone but it’s a hard problem.
So you’re suggesting we just let the financial and auto industries fail? There’s a cost to that too because once again, these industries provide value to society. Crypto provides no added value and wouldn’t even solve either of the problems you mentioned. And yes we could horribly replace our currency with it but there are zero benefits to doing that over the current system. The fact is it is not a currency and it is not a business providing value, so you are purely speculating on sentiment, hence it’s a ponzi.
Stocks are different because there’s actually an underlying business beyond “waiting for the next guy”. For example in a non-ponzi security there is long term value as a business grows. No such thing for shitcoins. Nobody is saying sentiment doesn’t matter in the actual economy, it’s made up of humans, of course sentiment matters. It’s “is sentiment the only thing that matters?”, and in the case of crypto the answer is yes.
What you’re saying makes sense but I actually agree with the intent. They’re basically saying don’t destroy the government/society in the name of shareholder value. I feel like there should be a way to say this without saying “because our portfolio value!” But that’s the justification they want to try.
It’s really interesting how factorio problems are pretty much the same as multiprocessing/distributed systems problems. It just shows how universal and fundamental the idea of “work” and “workers” are. An engineer on my old team came up with a credit system to solve this problem, the receiver issues credits to the sender, which only sends when there is a credit available - this ends up with the same better failure mode that the article discovers and allows you to overlap communication with work.