Подтвердите e-mail

Для публикаций, комментариев, реакций и сообщений подтвердите адрес.

Профиль

Archer

Профиль Vively

Fiddler. Programmer in Recovery. Free-range Nerd. e/lud Compared to Bach, we all suck. Pat Metheny Never trust anyone who has a gibbon as their avatar.

I cannot imagine a more strongly authorized bit of travel than deportation.

0100

10-20 minutes and a candy thermometer, according to this site:

How Do You Caramelize Honey? - Chef's ResourceHow Do You Caramelize Honey? A Guide to Golden Goodness Caramelizing honey involves gently heating it until its sugars break ... Read moreHow Do You Caramelize Honey?www.chefsresource.com
010

That sounds delicious. And easy enough that I might be able to not mess it up!

110

Good morning from yellowstone

010

Theodora was the real brains of the operation.

150

The responsibility story and the tech story are the same thing. With AIs, everything starts with an instruction. It is very much worth understanding how that instruction lead to the unintended consequences.

020

The tech story is being ignored. It’s a story of unintended consequences, though, not one of rogue AI: the users had to have asked the model to solve a problem where hacking a competitor was in the solution space. AIs are both creative and literal in very weird ways.

230

All the model did was automate the exploit, probably in ways that criminals already do. It’s unintended consequences, not a malevolent model “breaking containment.” There’s nothing new here except the level of automation is higher.

020

How this works is illustrated with toy models quite well in “You Look Like A Thing And I Love You,” by Janelle Shane. I recommend reading it. Modern models are vastly more sophisticated, but the principles are the same.

120

Surprise! the model found that solution. ML systems work by generating solutions and evaluating what worked or didn’t. That the testers didn’t expect the model to exploit another system is entirely on them, not the model.

120

One problem at a time. The first problem is that by attributing agency to a model we stoke the fear story: “the models are so dangerous you have to give us more money to contain them.” It’s a BS argument. OpenAI gave their model a problem where hacking a competitor was within the solution space.

141

IIRC Contador never ruined anyone’s career over his doping.

010

Going right after that can o’ worms, aren’t you?

000

Then you’d really like The Return.

010

That’s not Liam Neeson.

110

🙋‍♂️ Although admittedly just last night, at my brother’s insistence. Who saw it months ago. It’s equally not Greek.

011

Godzilla.

000

The B roll isn’t as good, but the racing is solid.

110

I credit the rule change for the sprint points and Tadej deciding that the yellow was enough.

000

That’s not what I’m talking about. The problem in tech journalism since the days of Byte Magazine is to cover the story the industry wants told rather than examining that story critically. This is what the press has been doing since before Bing tried to “seduce” Kevin Roose.

11703
Показать ещё