tmalsburg2

9 ms·

61.
by tmalsburg2 3y ago
Do LLM not also think when they encode the prompt? If Karpathy's explanation is accurate, longer prompts should also help even if they don't contain additional information, just by virtue of giving more time to think.
62.
by tmalsburg2 3y ago
> - Training isn’t done at 4-bits, to date this small size has only been for inference. Wasn't there a paper from Microsoft two weeks ago or so where they trained on log₂(3) bits? Edit: https://arxiv.org/pdf/24
63.
by tmalsburg2 3y ago
I learned a lot about this from the book "The design of the Unix operating system" by Maurice J. Bach.¹ It's an old book and many details deviate from actual present-day Linux, but it nonetheless gives a great overview of th
64.
The Shen Programming Language(shenlanguage.org)
242 points by tmalsburg2 3y ago|84 comments
65.
by tmalsburg2 3y ago
> Not for electricity production. A quick Google search suggests ~350g CO2 per KWh in Germany vs. ~400g in the US. Looks plausible to me. Germany uses coal but also has a very substantial share of renewables. Another quick search also sh
66.
by tmalsburg2 3y ago
And yet CO2 per capita is much lower in Germany than in the US. So Germany must also be doing something right.
67.
The Hare Programming Language(harelang.org)
3 points by tmalsburg2 3y ago|1 comments
68.
by tmalsburg2 3y ago
I didn’t accuse anyone. I just laid out the possibilities. Reading comprehension.
69.
by tmalsburg2 3y ago
+1 for restic. I tried various solutions and restic is the best by far. So fast, so reliable. https://restic.net/
70.
by tmalsburg2 3y ago
> Alaska Airlines doesn't trust Boeing enough They don't trust Boeing enough or their customers don't trust Boeing enough and it's a PR measure to calm the passengers' nerves. Could also be a bit of both.
71.
by tmalsburg2 3y ago
The answer is in the second paragraph.
72.
by tmalsburg2 3y ago
My understanding (IANAL) so far was that licenses apply to the code and binaries derived from the code. Why do you think that GPL also applies to a from-scratch rewrite? There are cases where a rewrite was done precisely to circumvent lice
73.
Toyota-owned automaker tampered with safety tests for 30 years(cnn.com)
143 points by tmalsburg2 3y ago|91 comments
74.
by tmalsburg2 3y ago
That depends on the country and source of the data. In Germany, pretty much all buildings are tagged with house numbers.
75.
by tmalsburg2 3y ago
How can it be dense and wordy at the same time? One is the opposite of the other.
76.
by tmalsburg2 3y ago
It’s good for beginners (also a great reference). However, stuff gets serious pretty fast, and the reader is expected to put in serious time and effort.
77.
by tmalsburg2 3y ago
Source?
78.
by tmalsburg2 3y ago
No, I do not think that all lawyers are dumb. What a bizarre thing to say. Why derail an interesting conversation in such an aggressive way?
79.
by tmalsburg2 3y ago
If this phenomenon is so obvious, why is this story the top post on HN? And why did it take a picture of the mirror scenario to make people aware of this issue? Hell, the article even implies that this an issue only with images of mirrors w
80.
by tmalsburg2 3y ago
Completely valid nitpick.
81.
by tmalsburg2 3y ago
It’s unlikely, true. But precisely that makes it so dangerous. If such a picture is presented as evidence in a murder case, the possibility that it is telling the wrong story will be discounted and someone may go to prison for the rest of
82.
by tmalsburg2 3y ago
Rolling shutter could in principle explain this, but it’s probably too fast to capture these three poses which must be temporally rather far apart.
83.
by tmalsburg2 3y ago
This effect likely resulted from aggressive noise filtering. You can probably reduce the amount of filtering in the settings.
84.
by tmalsburg2 3y ago
Imagine such an image being used as evidence is a court case. E.g. showing someone pointing a gun at another person when that actually never happened.
85.
by tmalsburg2 3y ago
Not clear at all. Ilya leaving would look really bad for OpenAI. Altman needs him to stay, and he keeps the door wide open for that in the statement. Question is: How much is he willing to offer (money and otherwise) to make that happen?
86.
The Transistor: a 1953 documentary, anticipating its coming impact on technology(youtube.com)
2 points by tmalsburg2 3y ago|0 comments
87.
by tmalsburg2 3y ago
I doubt it. It’s too damn costly computationally.
88.
by tmalsburg2 3y ago
What you say is correct but it doesn’t explain why the model fails to capture the semantics of “is”.
89.
by tmalsburg2 3y ago
This is a useful observation, but it doesn’t explain the particular example given in the article.
90.
by tmalsburg2 3y ago
Why do you think that tokenization is to blame on this particular limitation?

More ›