dontreact

It’s not a very fair system if only certain participants can afford to avoid taxes. If these schemes were doable for most folks and not only the very rich then no one would have an issue with the system (assuming that it provided enough revenue).

Yeah exactly this is why there’s hype. It’s the best model that you can use for free easily

Umm I want to pay zero and run locally

If there is a better mode I can easily integrate using python I’m all ears. For what I’m building quality is most important

Quantized training doesn't work. Quantization-aware training does, and that is when you use floats but adjust the math to compensate for the fact that you are going to quantize at inference time.

There is absolutely no one who has managed to do anything close to these models on a single CPU when it comes to _training_, so that is why no one is considering the possibility.

There is so much hype around federated learning but often the hard and insurmountable part of this is federated labeling.

For example for your cancer use case, you have to convince multiple hospitals to feed the system labels and this is a very very tall ask.

For healthcare it’s also not clear how to get a regulatory clearance if you can’t actually test the performance of the federated deployments.

So while federated learning solves some problems generated by an unwillingness to share data, it doesn’t solve all of them. Describe the use cases of your product carefully.

I'm curious for those who have actually tried it, what they think of the performance. A lot of the threads here are just talking about the country-limited rollout.

People on Twitter are jumping to the conclusion that this allows you to read someone’s mind but maybe all it does is allow to you to reconstruct something similar to what they are looking at.

For example if you recorded the activity of every cell in the retina, you would be able to do good image reconstruction but no one would interpret this as mind reading.

It’s possible that this is a similar story but instead of the retina, the decoder is just getting all of its signal from early visual cortex and therefore it’s not doing anything that a well placed camera pointing the same way as the subjects eye couldn’t do.

Yes, in principle it’s possibly doing more than that and using signal from eg inferotemporal cortex, but how much? There is no evidence it is doing anything beyond simply leveraging the retinotopic organization of primary visual cortex here.

This result is not meaningfully different from things that have been done with MRI 10 or so years ago. Why do you think it’s getting there? I see this field not moving forward in any meaningful way.

How do you create 1-2 floors of shared space without forcing some people out of their desks?

I am sitting at a Google desk and there are 5/6 people not here in the desks in my vicinity. I would prefer fewer desks and to actually get to sit next to people rather than being scattered among a sea of mostly empty desks due to people WFH.

Scale AI does this by getting high quality labels. They also have built some tooling to inspect your labels using your models: https://dashboard.scale.com/nucleus

I think it's useful when you don't know what to search yet but are looking to solve a problem. For example, I am planning a wedding and I asked it to put together a rough budget for me and this helped me to prioritize where to spend my energies on the vendor search.

I would bet money that the answer here is no

Sure, I’m just saying 20% time is still a thing though

It's not true

Source: I work at Google

Serving a request to one of their APIs requires orders of magnitude more compute than your typical web service

This is BS. ML targeted ads were the first application to make money and that happened a long long time ago (maybe 15 years at this point?). Nowadays there are a lot of products where ML is an essential feature.

Yes it’s even harder as you have to figure out which loudspeakers to trust!

Please explain how you could use pathology slides for ads? Does not seem possible.

The title is not correct and not supported by the article. The article clearly states that Google has not obtained any medical records.

“Ultimately, even through negotiations, we were unable to find a pathway that we legally could do and ethically should do,” Simon said. “And the partnership dissolved.”

In normal situations you don't need speech recognition, but to drive as well as a human in all possible situations you do. What if there is an emergency and there is a loudspeaker giving drivers instructions on how to turn around?

They certainly have their fair share of negative aspects that I wouldn’t want to integrate.

To be more concrete one specific thing I would like to see more of in the US is some sort of change to the political process that would allow more engineers to get into positions of power. Right now it’s all lawyers, and I think this is a big reason why the US is not good at large infrastructure projects anymore (California high speed rail, building more nuclear plants etc.)

I don’t believe that this voting bloc is actually voting in its best interests. The power rests more with the people who are able to manipulate it.

Wealth isn’t zero sum but power over law (as defined by who has influence over the rules we live by) is.

When you have concentration of power due to concentration of wealth, the society will become unstable and undemocratic.

So even though wealth isn’t zero sum, wealth inequality is a problem that needs to be addressed for long term stability.

It’s hard problem that can only be solved through attempts and iteration. The fact that a wealth tax is difficult is not a good enough reason to not try, because the problem of wealth inequality is so severe now.

Yes, every external force that has influenced the US is unique in some dimension.

To me the biggest difference between China and Japan is that China is much more likely to challenge the US among all dimensions of power, whereas my understanding of Japan in the 80s is that they simply started to be slightly competitive in an economic dimension.

So I do think that there is a higher likelihood that China’s rise opens up more debates about certain deep assumptions we have in the US than Japan’s did. And some of these may be about the primacy of private property. Private property is not a law of physics but it is a very deeply set assumption we have. The idea that housing is an investment more than it is a human necessity is something that may be worth questioning more and more. If investor owned housing in the US keeps trending up (already around 20%) then at some point it will likely be helpful to have another successful system to point at and learn from for solutions.

All I am saying is that it's a vast oversimplification of an economic and political system composed of over 1 billion people to simply say "it's capitalist". The overarching point here is that as China rises hopefully we in the US can learn something from their system and integrate it into ours. The idea that we can't do so because "they are a capitalist country and so are we" is reductionist and I was trying to point out ways in which that was the case.

There is quite a lot of fear of China being ahead or getting ahead, and it is being exploited by tech companies to secure funding from the department of defense: https://t.co/7by2iroWWD