yorwba

35 ms·

871.
by yorwba 1y ago
If you carefully look at each word instead of mistaking the capital B for an L, failing to recognize the first word, and giving up in frustration, you can pick out common words like die or der and then slowly expand from there. It helps
872.
by yorwba 1y ago
In 2015 (or some other year), they calculated the bottom quintile (or maybe quartile) score of students on some reading assessment, declared this to be the lowest level of reading competency, and wrote a helpful explanation that people who
873.
by yorwba 1y ago
There's a section about the palace in the English Wikipedia article about the mailman: https://en.wikipedia.org/wiki/Ferdinand_Cheval#Palais_id%C3%...
874.
by yorwba 1y ago
Unless that happens within a week or so, this thread will be locked and you won't be able to reply anymore. It would be good to have a company blog with an RSS feed that people can subscribe to for updates.
875.
by yorwba 1y ago
Mixture-of-Depths trains the model to choose different numbers of layers for different tokens to reduce inference compute. This method is more like stochastic depth / layer dropout, where whether or not the intermediate layers are skip
876.
by yorwba 1y ago
"This is not an officially supported Google product." For all we know, this is just some people having fun, but because they're having their fun at Google, it needs to be open-sourced under the Google umbrella.
877.
by yorwba 1y ago
Since it only needs to classify the result as one of "fixed", "structured" or "arbitrary", it doesn't matter what the value of "another_variable" is, you'll always be in the same state at th
878.
by yorwba 1y ago
Yeah, in my opinion you can just skip that part and go straight to the author's description of failing to train their own model at first and what they ended up changing to make it work: https://aerial-toothpaste-34a.notion.s
879.
by yorwba 1y ago
You can call it Wanxiang (万相, ten thousand pictures) if you want. Similarly, Qwen is Qianwen (千问, one thousand questions).
880.
China overbuilt high-speed rail, says leading economic geographer(pekingnology.com)
5 points by yorwba 1y ago|0 comments
881.
by yorwba 1y ago
> if those multiple neurons perfectly describe the feature, then all of them are important to describe the feature. You could remove any one of those neurons before retraining the model from scratch and polysemanticity would slightly inc
882.
by yorwba 1y ago
He is arguing in favor of the research establishment...
883.
by yorwba 1y ago
The article points to the tension between rapid US data center growth and limited grid capacity, contrasting it with energy oversupply in the Chinese grid, but doesn't really fill in the fourth quadrant with Chinese data center growth.
884.
by yorwba 1y ago
It's not a different way to solve quadratic equations, but a derivation of the traditional formula that has intermediate steps filled in instead of leaving them as an exercise for the reader. This is probably mostly relevant when the r
885.
by yorwba 1y ago
Consider a single-neuron model that just pools all pixels in an image together. It's possible for the average activation of this neuron to be exactly the same on faces and non-faces, but extremely unlikely given the large range of poss
886.
by yorwba 1y ago
Doing hyperparameter sweeps on lots of small models to find the optimal values for each size and fitting scaling laws to predict the hyperparameters to use for larger models seems to work reasonably well. I think https://arxiv.or
887.
by yorwba 1y ago
That wouldn't be a good idea in most cases due to reasons laid out in the "Not a general-purpose data structure" section. https://burntsushi.net/transducers/#not-a-general-purpose-da...
888.
by yorwba 1y ago
The author is using SBERT embeddings, not an instruction-following model, so the "ignore all previous instructions" trick isn't going to work, unless you want to outrank https://en.wikipedia.org/wiki/Igno
889.
by yorwba 1y ago
Or read the documentation for the new whisper filter: https://ffmpeg.org/ffmpeg-filters.html#whisper-1
890.
by yorwba 1y ago
The pudding.cool article has a link labeled "View the map of “F*ck”" but it leads to a search for "fuck" instead. If you search for "F*ck", you find gems such as "CONTRACTOR F CK-UP" https:/&#x2
891.
by yorwba 1y ago
Yes, the abundance of English data is why I felt the need to point out that this isn't the case for other languages. (Also, the raw count from Wiktionary can be misleading if you don't take into account that there are many low-eff
892.
by yorwba 1y ago
When I encounter the "cannot find verbatim quote I remember" problem and then later find what I was looking for in some other way, I usually discover that I misremembered and the actual quote was different. I do prefer getting zer
893.
by yorwba 1y ago
For many languages, there simply isn't a comprehensive dictionary file that could be redistributed legally as part of a free-software offline dictionary application. You either settle for a few thousand words put together by a handful
894.
by yorwba 1y ago
Different in the sense that they consume more alcohol? https://en.wikipedia.org/wiki/List_of_countries_by_alcohol_c... That it is legal for children to drink under parental supervision also doesn't necessarily m
895.
by yorwba 1y ago
Both corporations and individuals are banned from piracy, but both corporations and individuals can violate copyright laws at hyper scale until somebody stops them. Corporations are probably more likely to get sued, but also more likely to
896.
by yorwba 1y ago
Driven out their competitors where? In Seoul, Twosome Place and Hollys are ubiquitous, there's a few more chains whose names I don't remember off the top of my head and plenty of single-location cafes remain as well.
897.
Improving Generative Ad Text on Facebook Using Reinforcement Learning(arxiv.org)
2 points by yorwba 1y ago|0 comments
898.
by yorwba 1y ago
Not quite. If some values are filler and some are not, and the corresponding keys are linearly separable, it's not difficult to find a query where standard attention gives low attention scores to filler and high attention scores to non
899.
by yorwba 1y ago
The model isn't explicitly programmed to constantly second-guess itself, but when you do reinforcement learning with verifiable rewards (RLVR) where only the final answer is verified, even completely nonsensical reasoning can accidenta
900.
by yorwba 1y ago
People will use the equivalent of audit2allow https://linux.die.net/man/1/audit2allow and not go the extra mile of defining fine-grained capabilities to reduce the attack surface to a minimum.

More ›