msp26

16 ms·

211.
by msp26 3y ago
Performance on long contexts just isn't very good right now. Relevant paper: Lost in the Middle: How Language Models Use Long Contexts https://arxiv.org/abs/2307.03172
212.
by msp26 3y ago
It's for your own safety /s
213.
by msp26 3y ago
They're incredibly good at data extraction. You can start from zero and have a decent pipeline in hours.
214.
by msp26 3y ago
Great work :)
215.
by msp26 3y ago
Very nice! I've been looking for more of this kind of information. Some additional stuff: -A jailbreak+local builder to see exactly what function calls look like when they actually go into OPENAI's model. Note how many aspects of
216.
by msp26 3y ago
I hate that I have to use this boated app for the sole purpose of changing my DPI. It's obnoxious.
217.
by msp26 3y ago
Just inject random 'diverse' keywords in the prompts with some probabilities to make journalists happy. For an online generator you could probably take some data from the user's profile to 'align' the outputs to the
218.
by msp26 3y ago
>Reasons behind the project: I have been working with OpenAI LLMs recently and I like to get my data output in structured JSON. I realised it's rather good at taking a schema where properties are compressed within an instruction, an
219.
by msp26 3y ago
Any tips on building robust LLM software? And yeah I agree on how important test cases are. Having some sort of objective benchmark for judging how effective prompts are is really useful.
220.
by msp26 3y ago
I don't think a frontend for tiktoken is worth a monthly subscription
221.
by msp26 3y ago
Don't advertisers care that they're wasting money? Or are they just unaware?
222.
by msp26 3y ago
Excellent. I hope to see more games following suit away from the bloated messes of fandom/fextralife. At times they feel unusable on mobile. And I don't think a single user has wanted to see the embedded twitch streams when they&#
223.
by msp26 3y ago
Almost every paragraph has 2 references to 'safety'. At some point it feels a bit redundant, we get it.
224.
by msp26 3y ago
Ah thanks for clarifying
225.
by msp26 3y ago
(Sorry if I ramble I've spent some time on token counting recently and have some opinions on it) It's neat that it's client side only but it's quite feature lacking compared to https://tiktokenizer.vercel.app&
226.
by msp26 3y ago
Wouldn't that be caused by an OCR issue? No matter how primitive I doubt a TTS model would mistake I for 1. The book was probably poorly scanned, OCRed, and then had automated audiogen.
227.
by msp26 3y ago
It's been done before. "Textbooks are all you need". https://arxiv.org/abs/2306.11644
228.
by msp26 3y ago
I know it's not true but part of me believes that meta intentionally held back the 34B to make the community work on improving 13B and 70B model usage.
229.
by msp26 3y ago
Do you still use few-shot prompting with a fine-tune? Or does it make little difference?
230.
by msp26 3y ago
The article is shit; I just wanted to encourage more discussion on the topic in the comments.
231.
by msp26 3y ago
> The solution can simply be “get better at extroversion.” I agree completely as a former shut-in. I just forced myself to be as extroverted as possible for a while until it came more naturally. I still prefer to be an introvert and have
232.
by msp26 3y ago
I disagree that LLMs are always better for things like classification. But the barrier to entry is so low with them that it's usually worth trying to use an LLM before training some other model.
233.
by msp26 3y ago
Making good tools with language models takes time. There are nuances that aren't immediately visible without working on them. Unless there's a mismatch in the schema/messages, function calling is very reliable at providing JS
234.
by msp26 3y ago
If it's anything like Glaze, it's a complete meme. That took a lot of compute to glaze an image (with very visible artifacts) and was bypassed with some trival image processing on a shitty laptop.
235.
by msp26 3y ago
>This works well for me except the 15B+ don't run fast enough on a 4090 I assume quantized models will run a lot better. TheBloke already seems like he's on it. https://huggingface.co/TheBloke/CodeLlama-13B
236.
by msp26 3y ago
8x is a big gap. I feel that it might be better to just use few-shot prompting and supply some examples (not applicable to every use case ofc) every prompt to steer the output. But it's worth experimenting with at least. Edit: no funct
237.
by msp26 3y ago
All I want is llama-2-34b (seriously what's taking so long on this specific model) but this is interesting too I guess.
238.
by msp26 3y ago
Tachiyomi makes Android significantly better for comics/manga than iOS. It's one of the reasons I sold my iPad for a galaxy tab s7+. Every alternative I tried on iOS was just shit in comparison, paperback especially.
239.
by msp26 3y ago
>A minor weakness with LoRAs is that you can only have one active at a time Uh this isn't true at all, at least with auto1111.
240.
by msp26 3y ago
>The real kicker of this demo was that they passed the LLM-generated response dialog in to a deep-fake model trained on each character's original voice actors (since the corpus in Skyrim is vast, this ended up quite accurate). The v

More ›