crackalamoo

7 ms·

31.
by crackalamoo 2y ago
I think standardizing scientific terms and the like is a very minor concern when it comes to preserving endangered languages. Most science is done in English anyways, and that's unlikely to change.
32.
by crackalamoo 2y ago
I think the commenter you're replying to might have been referring to pictograms like many Chinese characters, not icons that you have to figure out. Or maybe not
33.
by crackalamoo 2y ago
Yeah, that's the ambition. More generally, ML is good at approximating many NP-hard problems efficiently, so I wonder if it will be a more practical alternative to quantum computing for things like molecular simulation.
34.
by crackalamoo 2y ago
Can machine learning accomplish the same? See e.g. Microsoft MatterSim https://www.microsoft.com/en-us/research/blog/mattersim-a-de...
35.
by crackalamoo 2y ago
You can do both. Preserving the corpus and building the LLM probably gives the best chance for future generations.
36.
by crackalamoo 2y ago
But of those 126,301 people who have edited in the last 30 days, some of them have edited more than one article. In fact, some have made up to millions of edits (lifetime), which disproportionately increases the total. At least 5000 people
37.
by crackalamoo 2y ago
As far as I can tell, OpenAI really did invent RLHF in its current form. https://arxiv.org/pdf/1909.08593
38.
by crackalamoo 2y ago
It probably doesn't matter much either way, but I use vanilla HTML/JS
39.
by crackalamoo 2y ago
This is true in principle, yes. In practice, the way this usually works is by converting inputs to bits and bytes, and then computing the result as a digital circuit (AND, OR, XOR). Doing this encrypted is very slow: without hardware accele
40.
by crackalamoo 2y ago
See also CrypTen, Meta's library for privacy preserving machine learning: https://github.com/facebookresearch/CrypTen . This isn't fully homomorphic encryption, but it is multi-party computation (MPC), which h
41.
by crackalamoo 2y ago
Is an electron conscious at some level? If we use a quantum particle instead of a pseudorandom number generator in ChatGPT does it become conscious? This seems extremely unlikely to me. Do non-quantum, classical effects of the brain, all th
42.
by crackalamoo 2y ago
> Today it might be valuable to be 6/10 at a bunch of things. But if AI tools can be 6/10 at those things too, it’ll become more valuable to be 10/10 at one or two things. This is true, but I don't think it makes bein
43.
by crackalamoo 2y ago
> an arbitrary person does something small that turns out to be on the critical path, and without it the big thing would not get completed. A great example is the discovery of the Thermus aquaticus bacteria in hot springs, seemingly pr
44.
by crackalamoo 2y ago
Very interesting. Certainly good to know for the debate on AI stealing vs repurposing artists' content. Which version of stable diffusion is this? I would think a newer, less overfit model would be much less likely to make these mistak
45.
by crackalamoo 2y ago
> It’s not like language keeps inventing new logic at a rate we can’t keep with. Maybe it does. Maybe, to a smart enough model, given its training on human knowledge so far, the next logical thing after "Sure, here's a technica
46.
by crackalamoo 2y ago
I'm not exactly sure how you would define 99.99% or 70% correct for protein folding, but AlphaFold is extremely accurate. The original AlphaFold paper had an rmsd of 0.96 Å at 95% residue coverage, which is less than the diameter of a
47.
by crackalamoo 2y ago
I don't think we're at a plateau. There's still a lot GPT-4 can't do. Given the progress we've seen so far with scaling, I think the next iterations will be a lot better. It might even take 10 or even 100x scale, bu
48.
by crackalamoo 2y ago
This is true. One definition I've heard for AGI is something that can replace any remote worker, but the definition is ultimately arbitrary. When "AI" was beating grandmasters at chess, this didn't matter as much. But we
49.
by crackalamoo 2y ago
> I'll be surprised if this event occurs before 2005 or after 2030. I'm not truly confident AGI will be achieved before 2030, and less so for ASI. But I do think it is quite plausible that we will achieve at least AGI by 20
50.
Show HN: Algorithm to Predict Romance/Indian Languages from Latin/Sanskrit(harysdalvi.com)
4 points by crackalamoo 2y ago|1 comments
51.
by crackalamoo 2y ago
I wonder if this discrepancy can be explained by the equation of time? https://en.wikipedia.org/wiki/Equation_of_time
52.
by crackalamoo 2y ago
Very interesting, I'll have to look into AWGN and the Fourier transform. I guess in the trees blocking the flashlight example that's not at all AWGN. Although while we care about the relative amplitudes in AM, AWGN would make this
53.
by crackalamoo 2y ago
Basic Newtonian physics is pretty efficient to compute, but afaik some more complex physics like fluids is faster with network inference. There are probably a lot of cases where network inference physics is faster.
54.
by crackalamoo 2y ago
Yes, phase noise exists, but I would think that in practice amplitude noise is greater. In physics, when a wave passes from one medium to another, its frequency is supposed to stay the same. Even if this isn't perfectly true in the rea
55.
by crackalamoo 2y ago
This kind of thing is surprisingly common: languages and cultures may change drastically over time, but the DNA of people living in a region tends to change much more slowly. If the DNA is indigenous, we often have to look at Y DNA and mito
56.
by crackalamoo 2y ago
Nice job! Were you able to use something like the OpenAI API for custom training or did you end up going for an in-house solution?
57.
by crackalamoo 2y ago
Yes, it seems like it. Although I would imagine the features and architecture of the model still take some physics into account. You can't just feed weather data into an LLM, after all.
58.
by crackalamoo 2y ago
I'm currently learning Go, and I recommend it just to have a new perspective. It's a very opinionated language that I think almost forces you into writing good code. It's also a joy to write because once you fix any errors, e
59.
by crackalamoo 2y ago
Interesting approach. Diffusion models can be useful. But as the technology stands, I don't think artists who don't use AI will be replaced by artists who do, or by AI itself, anytime soon.
60.
by crackalamoo 2y ago
This makes intuitive sense to me. It looks like o1 is a significant improvement in terms of performance, but maybe not the underlying architecture. Adding chains of thought and prover/verifier games to what OpenAI already had in 4o see

More ›