DominikPeters

7 ms·

61.
by DominikPeters 3y ago
It's plausible that mistral trained on GPT-4 output and therefore has similar mannerisms.
62.
by DominikPeters 3y ago
It's not just cutting. The answers were obtained by taking still photos and inputting them into the model together with detailed text instructions explaining the context and the task to the model, giving some examples first and using c
63.
by DominikPeters 3y ago
There are also these posts, pointing out many inaccuracies: https://www.lesswrong.com/posts/NRrbJJWnaSorrqvtZ/on-not-get... https://www.lesswrong.com/posts/7iAABhWpcGeP5e6SB/it-s-proba...
64.
by DominikPeters 3y ago
The need for responsive design. In 1990 you could make a pixel perfect design, that looked great on the 640 x 480 display that everyone used for everything. Today your application (i.e. your website) needs to look good on phones, tablets, l
65.
by DominikPeters 3y ago
It also knows how to make download links for files in its file system, when the prompt doesn't explain how this can be done.
66.
by DominikPeters 3y ago
> While per-token inference costs for fine-tuned GPT-3.5 is 10x more expensive than GPT-3.5 it is still 10x cheaper than GPT-4! Not quite accurate; finetuned 3.5 is only 4x cheaper than GPT-4. Cost per million output tokens from https:&
67.
by DominikPeters 3y ago
His comments more broadly make no sense to me. “Over time, the quality of the applicant pool increased to the point where we really didn’t have [give legacies a boost]” - how does that work? If the applicant pool gets better, then if anythi
68.
by DominikPeters 3y ago
There’s also the stage where the court decides whether to take on a case which requires only 4 justices to vote for it, and we don’t know what happened at conference, eg how forcefully Alito was arguing. Thus, a 7-1 majority does not make a
69.
by DominikPeters 4y ago
I thought the answer was terrible. It’s wrong in saying you must buy a smart tv to get a reasonable display panel (this comment section does provide good non-smart options), condescending to the question, needlessly verbose (sounded like ea
70.
by DominikPeters 4y ago
Is there a larger example output pdf available, to be able to compare to latex?
71.
by DominikPeters 4y ago
It would have been helpful to show some example generations of the model, unless I've missed them.
72.
by DominikPeters 4y ago
It would be helpful for me if there was clearer public communication about how recommendable the flu vaccine is in any given year. It varies a lot because some years they've correctly guessed the circulating strains and in other years
73.
by DominikPeters 4y ago
I would recommend using the lwarp package for turning large latex documents into HTML. Pretty much all other converters attempt to parse the tex files, which is an almost hopeless task. Lwarp has a different strategy: it redefines all macro
74.
by DominikPeters 4y ago
The blog post says: > The model is often excessively verbose and overuses certain phrases, such as restating that it’s a language model trained by OpenAI. These issues arise from biases in the training data (trainers prefer longer answer
75.
How to shorten red light clearance times without reducing safety [video](youtube.com)
4 points by DominikPeters 4y ago|0 comments
76.
by DominikPeters 4y ago
Here you can browse the list of all reasons: https://scotrail.datasette.io/scotrail?sql=++++select+Transc...
77.
by DominikPeters 4y ago
In case there is no Condorcet winner, there is actually a more principled way to randomize over alternatives than just choosing each with the same probability, known as maximal lotteries (e.g., https://pub.dss.in.tum.de/bran
78.
by DominikPeters 4y ago
The resulting voting rule (Borda except if there is a Condorcet winner) is commonly known as Black's rule ( https://en.wikipedia.org/wiki/Black%27s_method ).
79.
by DominikPeters 4y ago
It sounds like Danes over 50 can get a free vaccine, and those under are prohibited from getting one. Why can't we have a middle ground where young people can get a vaccine if they really want to (and e.g. pay for it themselves)? This
80.
by DominikPeters 4y ago
This seems to be a sample containing the relevant parts. https://cdn.standards.iteh.ai/samples/36631/c0883203ea25445c...
81.
by DominikPeters 4y ago
A4 paper is precisely 210x297mm (according to the standard), so the aspect ratio deviates from 1:sqrt(2) a little. Following that aspect ratio, the size would be 210x296.985..mm, so the rounding is really very slight.
82.
by DominikPeters 4y ago
Apparently this comes in a package with a constitutional amendment that enshrines marriage as exclusively opposite-sex, which will make legalising same-sex marriage harder down the road. https://twitter.com/kixes/status
83.
by DominikPeters 4y ago
I have plenty of emails in my inbox that don't use both. I think you're off by several orders of magnitude.
84.
by DominikPeters 4y ago
A more important reason that McDonald's has become slow is that they've switched to just-in-time assembly of burgers. They used to have an inventory of burgers, and for most orders, you ordered and the person taking your payments
85.
by DominikPeters 4y ago
LPL makes few videos with locks where he can't do anything. To me it's not a red flag but more likely a signal that he hasn't found an exploit (yet).
86.
by DominikPeters 4y ago
I always thought this is a bug!
87.
by DominikPeters 4y ago
12" was 0.92kg. New MBA is 1.24kg.
88.
by DominikPeters 4y ago
Looks like in 2019 there were 200-250 million airplane trips starting or ending in Germany. So maybe the surprise is how little average people travel rather than how little they rail.
89.
by DominikPeters 4y ago
I agree that ticketing is terrible. But note that you can take the train from Amsterdam to Barcelona or even Madrid or Sevilla using essentially only high-speed lines (the only segments that aren't high-speed are Amsterdam to Lille and
90.
by DominikPeters 4y ago
That seems wrong. New high speed lines are being built and have recently opened (e.g. Berlin-Munich, soon Stuttgart-Ulm). The number of passengers using high speed trains (ICEs) is steadily increasing: from 75 million trips in 2009 to 99 mi

More ›