themoose5

Came to this thread from the WAN show topic and wanted to add my 2c.
 
First a little background, I work professionally as an AI engineer. While I don't do direct model development I am very versed in the processes around model development. Also I don't work for google or yt so I don't have specific inside knowledge of their processes or models.
 
On the surface I think Linus' hypothesis he presented on the WAN show could be very close to the truth of how this word ended up on the automatic filter. If in fact google/yt is using a model or collection of models for automated comment filtering it's architecture is probably going to be some thing pretty common.The key here really isn't the architecture of the model but what training data was used to create that model. Specifically where google/yt got that training data from for non-English content. If their training data contained a significant amount of examples containing the Chinese characters in question that lead to a negative sentiment following their use it is easy to see how the model would learn to flag these characters as offensive. 
 
This seems to be backed up by the fact that only the use of these specific Chinese characters result in a comment being removed but the use of the romanized version does not receive the same treatment. If this phrase was added by someone manually I think it would be a safe assumption that the person adding them would add in all the variations of the phrase in Chinese characters and their romanized counterparts and not just the Chinese character version. On the other hand if this was something learned by an AI model using data pulled from outside sources all it would take is having enough examples in the training data for the model to learn the Chinese characters of this specific phrase. That learning would not necessarily transfer to learning the romanized version as well, resulting in what seems to be the current behavior. 
 
Training data review and analysis is something that is still kind of a sore spot in the field of AI. Training data quality has a large impact on the final performance of a model and there are numerous examples of how AI models can be biased by their training data. Most likely what happened in this case was the non-English language training data was not reviewed well enough to catch these issues before deploying.

Came to this thread from the WAN show topic and wanted to add my 2c.
 
First a little background, I work professionally as an AI engineer. While I don't do direct model development I am very versed in the processes around model development. Also I don't work for google or yt so I don't have specific inside knowledge of their processes or models.
 
On the surface I think Linus' hypothesis he presented on the WAN show could be very close to the truth of how this word ended up on the automatic filter. If in fact google/yt is using a model or collection of models for automated comment filtering it's architecture is probably going to be some thing pretty common.The key here really isn't the architecture of the model but what training data was used to create that model. Specifically where google/yt got that training data from for non-English content. If their training data contained a significant amount of examples containing the Chinese characters in question that lead to a negative sentiment following their use it is easy to see how the model would learn to flag these characters as offensive. 
 
This seems to be backed up by the fact that only the use of these specific Chinese characters result in a comment being removed but the use of the romanized version does not receive the same treatment. If this phrase was added by someone manually I think it would be a safe assumption that the person adding them would add in all the variations of the phrase in Chinese characters and their romanized counterparts and not just the Chinese character version. On the other hand if this was something learned by an AI model using data pulled from outside sources all it would take is having enough examples in the training data for the model to learn the Chinese characters of this specific phrase. That learning would not necessarily transfer to learning the romanized version as well, resulting in what seems to be the current behavior. 
 
Training data review and analysis is something that is still kind of a sore spot in the field of AI. Training data quality has a large impact on the final performance of a model and there are numerous examples of how AI models can be biased by their training data. Most likely what happened in this case was the non-English language training data was not reviewed well enough to catch these issues before deploying.

Came to this thread from the WAN show topic and wanted to add my 2c.
 
First a little background, I work professionally as an AI engineer. While I don't do direct model development I am very versed in the processes around model development. Also I don't work for google or yt so I don't have specific inside knowledge of their processes or models.
 
On the surface I think Linus' hypothesis he presented on the WAN show could be very close to the truth of how this word ended up on the automatic filter. If in fact google/yt is using a model or collection of models for automated comment filtering it's architecture is probably going to be some thing pretty common.The key here really isn't the architecture of the model but what training data was used to create that model. Specifically where google/yt got that training data from for non-English content. If their training data contained a significant amount of examples containing the Chinese characters in question that lead to a negative sentiment following their use it is easy to see how the model would learn to flag these characters as offensive. 
 
This seems to be backed up by the fact that only the use of these specific Chinese characters result in a comment being removed but the use of the romanized version does not receive the same treatment. If this phrase was added by someone manually I think it would be a safe assumption that the person adding them would add in all the variations of the phrase in Chinese characters and their romanized counterparts and not just the Chinese character version. On the other hand if this was something learned by an AI model using data pulled from outside sources all it would take is having enough examples in the training data for the model to learn the Chinese characters of this specific phrase. That learning would not necessarily transfer to learning the romanized version as well, resulting in what seems to be the current behavior. 
 
Training data review and analysis is something that is still kind of a sore spot in the field of AI. Training data quality has a large impact on the final performance of a model and there are numerous examples of how AI models can be biased by their training data. Most likely what happened in this case was the non-English language training data was not reviewed well enough to catch these issues before deploying.

My guess is this is on Verizon for not completing their due diligence. As long as Yahoo didn't actively hide the contract terms from Verizon it's the purchasing company's responsibility to be aware of all the contract terms that the company they're buying are bound by. If Verizon missed this in the Mozilla contract then they have no one to blame but themselves.
 
Though that doesn't stop them from trying via a lawsuit...

That has to be one wicked exponential curve to jump from 16GB to 16TB

I love how advertising corporations are trying to sell the impact to their business as something that is anti-consumer and will hurt end users. Most end users will probably be happy about this not hurt by it. Honestly I don't think anyone in history has ever been hurt because they didn't see a product advertisement. 

I love how advertising corporations are trying to sell the impact to their business as something that is anti-consumer and will hurt end users. Most end users will probably be happy about this not hurt by it. Honestly I don't think anyone in history has ever been hurt because they didn't see a product advertisement. 

More than anything VR suffers from a lack of selection of software over a lack of quality in hardware. If there were more awesome games and applications for the expensive hardware people would be shelling out even at the current price point. Because it's still gen 1 and there isn't quite the developer support there yet it's hard for many people to justify the large hardware expense it takes to have the VR experience.
 
As the mentioned in the article, the price drop expands the market out to people that are almost ready to buy the headset but don't see the value quite yet with the current hardware price and available software. I personally haven't tried VR but many people who have seemed to love the experience. I'm going to need to wait to see some really great software available personally before I go plunking down some hard earned cash on a headset.  

Basically MIL-STD-810 is a military grade design standard for electrical hardware. It's something that you learn from day 1 in the industry. ISO 9001 is a business standard completely unrelated to MIL-STD. Its quite difficult to confuse the two things but he did.
 
Maybe it's context, but it was one of them head-in-hands situation..
 
Watching someone in such a high authority position, who should understand the industry inside out, squirm and blag his way through.

The last scrapyard wars are a while back now and i think it s more than fair to say that most of us really enjoy that series.
Personally I think the logical step now would be to extend SYW to a threeway now. I was thinking Linus+Luke vs. Jay+Nick vs. Bitwit (Kyle)+Paul, although we obviously don t know if this is at all possible.
As a theme I d really like to see something new like extreme overclocking (hood nitrogen cooling or something like that) and maybe a slightly higher budget.
 
What are your ideas? Do you even want another SYW?

I'm looking seriously into selling my 21:9 for a 1440p 16:9 just for the additional gaming support and not having to rely on 3rd party applications for the games I play.  

This is kind of what I was worried about with UW. It's becoming more popular but it hasn't reached the point where most games support it naively. 
 
I think I'm just going to go with a larger 1440p 16:9 display based on the replies from everyone. 
 
Thanks!!

This is kind of what I was worried about with UW. It's becoming more popular but it hasn't reached the point where most games support it naively. 
 
I think I'm just going to go with a larger 1440p 16:9 display based on the replies from everyone. 
 
Thanks!!

At launch Overwatch didn't support 21:9 Blizzard had to patch it in at some point later.

PREACH! I have that same keyboard for my rig at home and I love it! I ended up buying the same one but with MX Brown switches for the office so I didn't drive my co-workers totally crazy but still had some of the tactile feed back that I love so much about the MX Blues!
 
I don't think I could ever go back to a membrane keyboard now.

Yeah, they're loaded because they can afford an extra $50/year electricity bill. 
 
Let's assume the system is used at full load for 4 hours a day 365 days a year (which is pretty fair/generous assumption). Now the average electricity cost is $0.13/kw/h. So, 365days/yr * 4hrs/day = 1460 hours. A Fury X based system pulls about 400w at full load ( http://www.anandtech.com/bench/GPU16/1533 ). So a system with two Fury X's should be around 650w~. Now assume the PSU has a 90% efficiency, and your power consumption under load is about 720w. So math time: 
720w * 1460 hrs * $0.13kw/hrs / 1000 = $136/year.
 
Now if we were to use a two 980Ti, then it would be 40w~ less, which would be $129/year. 
 
Or if you were to use a single GTX 1060, then it would be 300w~, which would be $57/year. 
 
So yeah, enthusiasts must be loaded because they can afford to spend an extra 10 hours worth of minimum wage a YEAR. Also, anyone who would buy AMD anything must be a moron...I mean, $7/year is really bank breaking stuff. 

I always found the argument of "I paid for the game I should be able to cheat" ridiculous. For one it ruins the game for other people who also paid for the game and aren't cheating. Just because you paid for a game does not give you full reign to do whatever to other players. If hackers want to hack a game and mess around with it locally or on their own private server then fine but not on public matchmaking servers. 
 
The blizzard permaban is great IMHO. Would much rather see permabans than their games ruined by cheaters and script kiddies.

I always found the argument of "I paid for the game I should be able to cheat" ridiculous. For one it ruins the game for other people who also paid for the game and aren't cheating. Just because you paid for a game does not give you full reign to do whatever to other players. If hackers want to hack a game and mess around with it locally or on their own private server then fine but not on public matchmaking servers. 
 
The blizzard permaban is great IMHO. Would much rather see permabans than their games ruined by cheaters and script kiddies.