yunyu

7 ms·

61.
Clarity Reader: LLM powered depth-first reading for complex documents(github.com)
54 points by yunyu 3y ago|9 comments
62.
by yunyu 4y ago
With ever growing context windows, you can upload your entire codebase (or at least the public API) in the prompt and have the model reason off of that.
63.
by yunyu 4y ago
It turns out that accurately predicting text using a model that's a couple orders of magnitude smaller than the inputs requires understanding what the words mean.
64.
by yunyu 4y ago
There's nothing stopping you from identifying the code, running it, and passing the output back into the context window.
65.
by yunyu 4y ago
BERT can process 512 tokens. LLAMA and FLAN-UL2 can process 2048 tokens. GPT-4 can process 32768 tokens, and is much better at ignoring irrelevant context. These general models can be fine tuned with domain specific data with a very small n
66.
by yunyu 4y ago
LLaMa isn't instruction tuned
67.
by yunyu 4y ago
I assume you haven't tried Alpaca (which hasn't been released), only Llama. See the instruction fine tuning section in the article.
68.
by yunyu 4y ago
FLAN-UL2
69.
by yunyu 4y ago
Hellaswag is also a deeply flawed benchmark, I wouldn't read too much into it: https://www.surgehq.ai/blog/hellaswag-or-hellabad-36-of-this...
70.
by yunyu 4y ago
A simple bag-of-words classifier requires some feature engineering, parameter tuning, continuous retraining for generalization, and actual thinking on the engineering end. Simply tossing a small amount of data at DeBERTa v3 (for classificat
71.
by yunyu 4y ago
I assume most people needing fractional scaling on Mac are doing it with hidpi displays, where differences in text rendering aren’t visible and layout correctness is way more important. I have a 4k display scaled to 5k resolution to get mor
72.
by yunyu 4y ago
Model parallelism is possible for inference and you can host extremely large models across multiple accelerators (see DeepSpeed ZeRO), but the inference speed is several times slower compared to just having all of it on one accelerator due
73.
by yunyu 4y ago
FLAN-T5 11b is the most competitive public model that has undergone instruction finetuning. It has better few shot and instruction following performance than much bigger models like GPT-NeoX: https://huggingface.co/google&#x
74.
by yunyu 4y ago
The claim that current LLMs are not that good at deep thinking might be true in general , but it is definitely not the best or correct explanation for the rot13 task being described in the article (or other simple character manipulation ta
75.
by yunyu 4y ago
This is obviously the case if you apply much narrower criteria, most benchmarks of existing large datasets aren’t for character level tasks. That said, the synthetic noise section should be extremely interesting if not fully representative
76.
by yunyu 4y ago
The ByT5 paper shows significantly better resiliency to typos and noise in internet scale benchmarks, as well as better reasoning capabilities with tasks like punctuation and spacing correction: https://arxiv.org/pdf/21
77.
by yunyu 4y ago
There's a way simpler answer than this Type I Type II thinking stuff. Most LLMs like GPT are not trained on the level of individual characters – they process input and outputs on the level of subword units that compose multiple charact
78.
by yunyu 4y ago
Most LLMs are crippled at tasks regarding character level manipulation because of how the inputs are tokenized (words and numbers are split into common substrings rather than the original characters, 101 -> “10” + “1” vs “1” + “0” + “1”)
79.
by yunyu 4y ago
Venmo? Cash app? Zelle? I use one of these every day
80.
by yunyu 4y ago
The edge cases are more common at elite institutions, which are the typical admissions mania targets. At the top ~10 university I attended, there were plenty of people (myself included) falling in the poor HS GPA bucket (mostly due to extra
81.
by yunyu 4y ago
Just fine tune the model? Prompt leakage and bypass is only a problem with zero/few shot inference, and is trivially solved by removing the need for a prompt.
82.
by yunyu 4y ago
Zero Asians have ever benefited from any affirmative action policies in US college admissions, regardless of their socioeconomic status
83.
by yunyu 4y ago
I have used very little algebra/calculus/pre-college content, but have used a lot of linear algebra/discrete mathematics. Would this be a common exception to the rule?
84.
by yunyu 4y ago
You can update the "graph edges" with content longer than the window by fine tuning: https://beta.openai.com/docs/guides/fine-tuning
85.
by yunyu 4y ago
How many people would buy a Tesla that costs 3x as much?
86.
by yunyu 5y ago
PyTorch
87.
by yunyu 5y ago
A quarter million dollars is what a typical new grad offer looks like nowadays
88.
by yunyu 5y ago
RCS doesn’t need a separate app either. You can have SMS and RCS conversations in the same app regardless of if you’re using Google Messages/Samsung Messages/etc
89.
by yunyu 5y ago
If you look closely at the profile photo that the post is "Posted by", the distorted background makes it clear that it's a deepfake
90.
by yunyu 5y ago
That Ryzen chip has 20% slower single thread performance, which is what really matters for desktop workloads

More ›