lrei

5 ms·

1.
by lrei 10mo ago
tech debt generator - another job AI is taking away from humans!
2.
by lrei 1y ago
Warning: This is AI generated, probably a low end model as some of the content is outright nonsense eg: """ concept of MoE is quite prevalent (refer Outrageously Large Neural Networks: the Sparsely-Gated Mixture-of-Experts L
3.
by lrei 3y ago
No, that's not what I meant. I meant that in its reinforcement learning phase, GPT saw examples of "fix this text" style requests and was rewarded for doing a good job. That's different from seeing examples of typos and
4.
by lrei 3y ago
Yup totally plausible. Things like word (token) dropout and inserting random uniform noise into embeddings or just edit distance perturbations to the tokens are all well known but still Figure 1 looks extremely impressive.
5.
by lrei 3y ago
GPT-4 was clearly trained to fix typos and handle not well written written requests. That much is visible directly from just using it within chatGPT UI in normal usage and fits common user scenarios (eg fix my bad draft). We know it was tra
6.
by lrei 3y ago
Transformer is the architecture. "Generative Pretrained" is just a term made up by the author to mean what everyone called for decades before and will call for decades after "Language Modelling". It was just a new way of
7.
by lrei 4y ago
These ToS, CSAM, moderation, etc are very orthogonal to the app itself. Why conflate them? Most non-hackernews people are completely unaware of them, so it won't be a factor in their choice of platform. Are we even sure other apps aren
8.
by lrei 4y ago
Yikes! good luck, hope it works out. This is scary. Gmail is still, by far, the best email app. Something like this could happen to me and, I suspect, a lot of other people.
9.
by lrei 5y ago
Exactly the same reason why I use Iosevka and 79chars. However I use WQHD monitors and fit 4-5 splits.
10.
by lrei 5y ago
I don't know what being in "tools for thought space" means but if it just mean "Roam or similar user", I can give my personal answer: I set up 2 or 3 templates when I started. The most used one I updated once since
11.
by lrei 5y ago
Must've been quite the statistical masterpiece to disentangle it from the World Wars and the Great Depression. Off course extrapolating conclusions from early 20th century education and career progression to the 21st will be an even gr
12.
by lrei 5y ago
Research shows that doing the recall is the important part. The effect is by now well accepted. Off course if you do the cards yourself you can guarantee both their quality and their relevance to you. Ive heard of research pointing that lea
13.
by lrei 7y ago
I would if I could - I have an external GPU at home. Unfortunately Apple is (not without reason) angry at nvidia so they dropped support for Nvidia in Mac OS. I’d have to use Windows which is a big no no for me. Obvious pytorch can’t suppor
14.
by lrei 7y ago
Probably, they were the type that had no problem killing. Someone of a different type was in charge of justifications. The latter would’ve also been more the type that wrote, now historical, documents.
15.
by lrei 8y ago
This is true of every field of science and even social science I’ve known. Also the same terms have different meaning in different fields. Also the standard use of a term within a subfield can be different from its usage in another subfield
16.
by lrei 9y ago
I have plenty of experience with Mercedes, Audi, volvo, jeep and BMW cars - driving, renting, test driving them or being the passanger. I would say Mercedes is at least 5 years behind the competition at approx same price point. Disclaimer:
17.
by lrei 9y ago
Cool - I didn't know this project. I might take your offer on the free with survey next time I work on a research project with economists or finance people. (If the offer is still available) In a eU research project we used this http:
18.
by lrei 9y ago
trivial processing power but news organizations tend to want traffic on their websites where you can see ads and click on them and other content. Not sure for how long they will even support RSS let alone make an effort to support websocket
19.
by lrei 9y ago
There is a free tier with the number of API requests limited.
20.
by lrei 9y ago
Correct. Main author of the project is Gregor Leban: https://scholar.google.co.uk/citations?user=5pAxBWsAAAAJ&hl=... The original crawler (newsfeed.ijs.si) paper is from Trampus, Mitja and Novak, Blaz: The Internals Of
21.
by lrei 9y ago
I used it only for research (1 for media bias analysis related to the refugee crisis and 1 at the intersection of social/media and economy). I know the people who started EventRegistry (same department/institute as me) so I didn&#
22.
by lrei 9y ago
I got access for free. You should really ask them (they are responsive) I think it depends on org type/size - but the figures I heard unofficially were in the netflix (per user) price range.
23.
by lrei 9y ago
Afaik there is no meta data scrapped (technically I think some HTML meta tags are scrapped but not sure if used). There is machine-learning in the annotations - categories rely on a (cross lingual) text classifier, entities rely on matching
24.
by lrei 9y ago
Hi - I've used this API several times. Ask me anything (though there have been been changes since I last used it)
25.
by lrei 10y ago
> Code is only easy to replicate when they give you or publish the code No. Graduate ML students can implement the papers they read w/o a reference implementation - just search github. As I said, I implemented PV w/o the refere
26.
by lrei 10y ago
> I'm surprised (and concerned) that there haven't been any retractions or errata published for the paper cited there is a subsequent paper by Mikolov and Mesnil at ICLR with the correct results. And code to replicate them. Ret
27.
by lrei 10y ago
> about machine learning: how hard it is to actually trust results I find the opposite true: code is easy to replicate and the datasets for algorithm comparison are open (e.g. imdb used in the PV paper). If you show very good results (es
28.
by lrei 10y ago
Energy is energy, a calorie is a calorie in gasoline, food or electricity. However when talking about nutrition labels, those are generally an estimate of calories potentially extracted by humans (which I think is what you mean with nutriti
29.
by lrei 10y ago
Eat (drink) 4000 calories of gasoline. wont make you fat. point is: only calories you can convert into "human fat" will make you fatter. that process can not happen without certain chemicals (hormones) in your body, like insulin m
30.
by lrei 10y ago
Good one.

More ›