pizza

22 ms·

421.
by pizza 2y ago
failspy’s or mlabonne’s models. Or just look for any model with ‘abliterated’ in the title. Eg try failspy/meta-llama-3-8b-instruct-abliterated-v3 though of course bigger models will probably be better
422.
by pizza 2y ago
I think this assumes a lot of efficiency and rapid equilibration in the reallocation of labor. There are real costs and inefficiencies in the job search process and constraints limiting which new jobs these now-displaced workers can actuall
423.
Tree Attention Explainer(twitter.com)
2 points by pizza 2y ago|0 comments
424.
by pizza 2y ago
Seems to get stuck loading the response for me
425.
by pizza 2y ago
- rich - einops - pytrees - torchinfo - joblib - symbex - nbdev
426.
Lilly says weight loss drug cuts heart failure risk by 38% in trial(reuters.com)
6 points by pizza 2y ago|11 comments
427.
by pizza 2y ago
This is an interesting case, I think the sequence of events was this: - Someone started a thread related to this topic 3 weeks ago - Bing indexed that page a while back - OP found this thread, and then posted on the same thread, today - Cha
428.
by pizza 2y ago
Question for music theory experts who are also coders: does it feel like understanding a programming language, or something else? Admittedly I've never really made any serious independent effort to learn it, or to retain whatever I tri
429.
by pizza 2y ago
Lilian Weng's blog - a goldmine - has an in-depth post on this: https://lilianweng.github.io/posts/2024-07-07-hallucination/ . She leads safety and alignment at OpenAI so might be worth checking out :^)
430.
by pizza 2y ago
That's the autoregressive aspect. The decoder aspect is that the last layer converts representations into output sequences (and the generation happens autoregressively, one at a time). Similarly at the last layer an encoder outputs a r
431.
by pizza 2y ago
Would this help? https://en.wikipedia.org/wiki/Allen%27s_interval_algebra
432.
by pizza 2y ago
Let's compromise and go with kibigrams
433.
Nvidia SM89 (4090) Instruction Set Architecture(kuterdinel.com)
7 points by pizza 2y ago|0 comments
434.
by pizza 2y ago
(in mice)
435.
GLP-1 gene therapy reduces weight by 22% & positive results of long-term therapy(inf.news)
4 points by pizza 2y ago|5 comments
436.
by pizza 2y ago
yup, plus if the program is already running in the bad state you don't even have to restart it to investigate, just do strace -p <pid> -c to attach
437.
by pizza 2y ago
Not on Intel Macs
438.
CoLoR-Filter: Conditional Loss Reduction Filtering for Targeted LM Pre-Training(arxiv.org)
2 points by pizza 2y ago|0 comments
439.
by pizza 2y ago
You'd have to make refusal a high rank subspace, and that seems like it could be quite difficult. One alternative approach I've seen is to make refusal behavior more likely to just output the end-of-speech token.
440.
by pizza 2y ago
People find groups they want to join in proportion to the ease of finding those groups (which is in proportion to the size of those groups) and the degree to which they agree with them. You could try to change those factors, but would it be
441.
by pizza 2y ago
Ah, I'm not the creator just an admirer hehe
442.
by pizza 2y ago
https://einx.readthedocs.io/ :)
443.
by pizza 2y ago
Eh, I don't think that's their position: > nothing in this paper should be interpreted as claiming that large language models cannot display emergent abilities; rather, our message is that some previously claimed emergent abili
444.
by pizza 2y ago
You’re looking for group equivariance - there is some work in group equivariant neural networks. You might find the very recent paper on Clifford Group-Equivariant networks interesting - it applies to not only Euclidean spacetime transforma
445.
by pizza 2y ago
I don’t think that you are wrong but what do you mean by language that can’t covered by p(y|x)?
446.
by pizza 2y ago
can't be modeled is probably not what I should have said, rather I meant more like the error for a globally optimal model is still high
447.
by pizza 2y ago
I think there's nuanced distinctions between: - something that can't be modeled because there's no training data - something that can't be modeled because it's fundamentally stochastic - something that can't be
448.
TikTok hit with malicious malware that's taking over accounts – don't open DMs(tomsguide.com)
1 points by pizza 2y ago|1 comments
449.
by pizza 2y ago
Something like 1/3rd of Google searches have never been asked before. Maybe something similar could be said of Stack Overflow questions? There could still be quite a need for human expertise. Tbh, without significant coaxing, LLMs tend
450.
by pizza 2y ago
Stack Overflow did try this but imo their LLM wasn't that good

More ›