spion

20 ms·

241.
by spion 4y ago
I used to think so too, until I tried Rust. By comparison, JavaScript (and by extension, TypeScript) is still lacking fundamental features and the library ecosystem situation is pretty bad. I wish there was a modern language that took all t
242.
by spion 4y ago
llama.cpp needs 40GB for the 65B model (due to int4 quantization) RamNeeded(other_size) ~= 40GB * other_size/65B
243.
by spion 4y ago
It is https://github.com/ggerganov/llama.cpp/blob/master/utils.cpp...
244.
by spion 4y ago
Use `--top_p 2 --top_k 40 --repeat_penalty 1.176 --temp 0.7` with llama.cpp
245.
by spion 4y ago
I like the story, but I think people will notice pretty quickly as almost everyone reviews their photos right after taking them (so they can compare them with what they see in reality)
246.
by spion 4y ago
Its not the first time. Chess was the first time.
247.
by spion 4y ago
I tried some of the examples from the article: > Me: Does the phrase “John is too stubborn to talk to” typically describe a situation where John doesn't talk with anyone? > ChatGPT: No, the phrase "John is too stubborn to ta
248.
by spion 4y ago
That's the beauty of it - you don't have to choose, the network chooses for you. If you set up your layers such that the first input layer has one input for each possible token (lets say total of 80K weights), and the next layer c
249.
by spion 4y ago
You can try the dance problem with this prompt prefix To answer the following problem, work through it by reasoning step by step and writing that reasoning down, making sure steps are not conflicting with previous steps. Only after you&#x27
250.
by spion 4y ago
Ask it to work through the problem first and write down intermediate steps and only write the answer at the end. You should get better results than "wrong answer, then trying to justify it"
251.
by spion 4y ago
Try this prompt: For the following question, do all the calculations first and write them out. Only then answer the question based strictly on the result of the calculations. What is heavier, 1000 tons of helium or a cubic m
252.
by spion 4y ago
There is some commentary in the GPT-3 paper https://arxiv.org/pdf/2005.14165.pdf (figure 3.10 and table 3.10) note: I may have extrapolated a bit more than strictly correct from those two bits, but the accuracy being a
253.
by spion 4y ago
Not just connect something. For example, at a certain size, GPT models start to "learn" how to do basic arithmetic (addition), even for numbers with multiple digits they've never encountered before. It might look like a small
254.
by spion 4y ago
I like the token window bit. I don't really like the probability bit, because it kinda alludes that OpenAI just built a huge probability map of all N-grams (N=8000) and called it a day. Which incidentally would also imply that a lot of
255.
by spion 4y ago
We do have "successful" (but primitive) AI systems already, at least of the kind that we consider difficult to align (value optimizers / maximizers), which have a considerable impact of the world. An example would be social m
256.
by spion 4y ago
valtio is a modern mobx-like implementation that uses `useSyncExternalStore` (via useSnapshot - https://valtio.pmnd.rs/docs/api/basic/useSnapshot ). Its still not quite as nice as MobX, as passing larger objec
257.
by spion 4y ago
Counterpoint: I've worked on a big application that used MobX and had no such issues.
258.
by spion 4y ago
That example makes a lot of sense once you realize how JSX works in SolidJS (and its very different from React): 1. The function is called only once at component construction, and never again. Unlike React, it is _not_ called for every re
259.
by spion 4y ago
I don't believe they're added after every message. I think this may be one of the reasons why Bing slowly goes out of character. As for ChatGPT, I think the technique is slightly different - if I had to make some fun guesses, I&#x
260.
by spion 4y ago
A language model is trained to predict the next word in a sequence of words. Thus the "programming" is just a prompt that primes the model in a certain way. There is no difference between user input and prompt - the training data
261.
by spion 4y ago
The funny thing is, something like the Metaverse might actually be helpful to construct multimodal models - it could provide higher fidelity training data.
262.
by spion 4y ago
I like starbucks dark roast too. For me, it requires quite a bit of work to make it work, but when you succeed, my oh my its phenomenal.
263.
by spion 4y ago
I'm going to regret reacting to this, but here we go - Nobody cares about microgram scales. 1/10 of grams, however, are important, especially if you are dosing for espresso, where the amount of coffee affects the flow rate. - Hari
264.
by spion 4y ago
The fundamental bit I was aiming at here is that running gradient descent or similar training algorithms on a thing doesn't mean you will understand the end result fully - and we don't understand it fully. Its entirely possible th
265.
by spion 4y ago
The article itself admits (after covering the basics) that we don't really know. I thought it would be interesting to post some additional cool papers written by folks trying to figure out what's going on.
266.
by spion 4y ago
The answer to this is: "we don't really know as its a very complex function automatically discovered by means of slow gradient descent, and we're still finding out" Here are some of the fun things we've found out so
267.
by spion 4y ago
> But then another customer still wants gamma@1, they are allowed to do that! But they also want your new features. In this case, you still have the option of working with them to help them migrate to gamma@2, if the cost of maintainin
268.
by spion 4y ago
I'm assuming a customer-driven culture where you work for your customers needs. In the case of libraries, teams using the libraries are customers. If you're the maintainer of alpha and your customer needs beta, your customer needs
269.
by spion 4y ago
Note: steps 1-4 will likely be done in somewhat reverse order (3,4 first, then 1,2, then building `beta`)
270.
by spion 4y ago
> How many internal libraries does your "separate services" contain? You service A depends on library alpha@1, your service B depends on library alpha@2. All happy now. Introduce another layer, your service A depends on library

More ›