HackerTrans
トップ新着トレンドコメント過去質問紹介求人

MacsHeadroom

no profile record

コメント

MacsHeadroom
·10 か月前·議論
From the murderer's text messages:

> Roommate: Why?

> Robinson: Why did I do it?

> Roommate: Yeah

> Robinson: I had enough of his hatred. Some hate can't be negotiated out.

Besides everyone around him testifying that he'd been left-wing ever since dropping out of college, his admission that he found Charlie's relatively moderate speech "hateful" strongly suggests the murderer was quite far to the left.
MacsHeadroom
·10 か月前·議論
This isn't a culture argument or even a sound/good argument for anything. Americans are just wealthier and you can't compare like that.

When you compare groups of students within the same country and adjust for both household income and intelligence you find that (again, even within the same intelligence brackets and income levels) some ethnic groups simply study more while others spend more time on things like unprovoked violence.
MacsHeadroom
·3 年前·議論
No, you can't. Property exists. You can replace the word intellectual with "personal," "private," "public," etc. But removing it does not work. Property is not a legal fiction. Who the property belongs to is.
MacsHeadroom
·3 年前·議論
Crypto is rallying. Bitcoin is higher than it has been in over a year. It's been trending up for a while though.
MacsHeadroom
·3 年前·議論
How many copyrights are worth half a billion dollars to register from 50-60 years? Presumably zero are worth registering for $50B to protect from 60-70 years.

This hardly seems perpetual in practice.
MacsHeadroom
·3 年前·議論
"It is better to let 100 criminals go free than to imprison 1 innocent man." — Benjamin Franklin
MacsHeadroom
·3 年前·議論
>Is it still useful or not?

Is what still useful? A LoRA is about as good and useful as a full fine tune. If you have unlimited storage space to store them or unlimited compute to make them then I would still prefer full fine tunes. But the difference is marginal and generally not worth the storage space or increased compute costs for individuals.
MacsHeadroom
·3 年前·議論
Modern peft methods with LoRA actually do reduce training time by orders of magnitude.

Here's an example of 20 seconds per epoch on a single consumer GPU: https://github.com/johnsmith0031/alpaca_lora_4bit/issues/7#i...
MacsHeadroom
·3 年前·議論
LoRA finds a subset of the original weights (about 1%) which can be trained to achieve about the same result as training the whole model while using 100x less compute.

Original weights frozen = Rather than modify the original model, the training results are saved to a small file of only a few MB.

In practice this means you can fine tune a 30B parameter model on a consumer GPU in a couple of hours. Without LoRA you would need to run multiple expensive data center GPUs for days or weeks.
MacsHeadroom
·4 年前·議論
PoW is needed for the "trustless" qualifier. Other schemes like PoS can be very low trust but not fully trustless. They also tend towards ever increasing centralization.