The headline and the byline are vague but the article has 19 occurrences of Hoback. I read the article and had no confusion about this. I find it disingenuous to say they want to evade accountability by omitting it from the headline.
There's plenty enough bad behavior to go around here. Hogback is willing to do whatever for marketing, Todd is a troll and reaping the consequences of years of prevaricating. And anyone who chooses to be involved on Bitcoin at this point should not be surprised to attract criticism.
Wow, this flippant hand waving by the documentary creator that the accused shouldn't be upset because the fears of being unmasked as a potential Satoshi being "overblown" as far as it being life threatening and disruptive is quite honestly disgusting. The documentary creator isn't the one to have to live his life fearing some unknown crazies harassing him for a potentially owned unfathomable amount of money. Nobody should have to live through that in the name of "public interest", how people can do that and live with thinking of themselves as some altruistic force of truth I'll never understand. How will he feel if his documentary ends up causing Todd real physical harm?
(Edit: the link does mention determining Peter Todd's identity and the fact that he has some involvement with Bitcoin - I had forgotten about this entirely - but it's still strange that this name ended up in the set of tokens that causes the glitches.)
So is my irc nick, I think it's pretty obvious what happened:
The GPT2 and GPT3 tokenism were generated via a minimization process on a corpus of text. The original corpus was generated by taking all +3 or better reddit posts and including what they linked to, or something along those lines.
All the Bitcoin development discussion on IRC was (and is) public. Post GDPR several archival sites that indexed these discussions went offline, including ones that had been heavily linked from reddit. So later data collection for future training runs doesn't include this material -- and the copies still online were clearly excluded from training.
The tokenizer was trained on different content from the network. " petertodd" (and " gmaxwell" for that matter) were extremely common in the tokenizer training-- enough to get their own tokens outright-- but nearly absent in the network training.
The result is tokens that were poorly conditioned during network training, resulting in erratic behavior.
11 comments
[ 3.0 ms ] story [ 35.1 ms ] thread1. https://en.wikipedia.org/wiki/Cullen_Hoback
{{citation needed}}
(Edit: the link does mention determining Peter Todd's identity and the fact that he has some involvement with Bitcoin - I had forgotten about this entirely - but it's still strange that this name ended up in the set of tokens that causes the glitches.)
The GPT2 and GPT3 tokenism were generated via a minimization process on a corpus of text. The original corpus was generated by taking all +3 or better reddit posts and including what they linked to, or something along those lines.
All the Bitcoin development discussion on IRC was (and is) public. Post GDPR several archival sites that indexed these discussions went offline, including ones that had been heavily linked from reddit. So later data collection for future training runs doesn't include this material -- and the copies still online were clearly excluded from training.
The tokenizer was trained on different content from the network. " petertodd" (and " gmaxwell" for that matter) were extremely common in the tokenizer training-- enough to get their own tokens outright-- but nearly absent in the network training.
The result is tokens that were poorly conditioned during network training, resulting in erratic behavior.
https://news.ycombinator.com/item?id=41783503