Shouldn’t be too difficult to replicate I think. Want to try it on RP, @laion_ai
?
@hardmaru
-
Replicating model on RP with LAION
By
–
-
Bard Demonstrates More Personality and Attitude
By
–
Yeah… Bard has more of a personality and attitude.
-
Domestically-trained LLMs vulnerable to prompt engineering for harmful content
By
–
Even domestically-trained LLMs can be prompt-engineered to generate text that crosses the line.
-
Model card withheld from external publication
By
–
Model card in the Appendix: “withheld from external publication”
-
PaLM 2 Technical Report Missing Information Disclosure
By
–
They didn’t even disclose this info in “PaLM 2 Technical Report” https://
arxiv.org/abs/2305.10403 -
Importance of Disclosing Model Size and Training Data
By
–
In this particular case, I think it’s perfect fine to tell people how large a model is, and what size is the training data it was trained on. It’s the absolute minimum information one should convey (say, in a model card) anyways, to communicate about the ML model to its audience.
-
ML Trade Secrets Impossible to Keep Due to Research Culture Sharing
By
–
As I’d said before, it is impossible to keep trade secrets in ML, especially at big companies, due to ML research culture. Researchers from all the competing labs all hang out with each other and share all sorts of info about their work, as hackers would. https://
x.com/hardmaru/statu
s/1620318664997339137
… -
PaLM 2 Leak: Parameters and Tokens Revealed, But Key Details Missing
By
–
There’s much talk about Google’s PaLM 2 details being leaked. But seems the only thing “leaked” is unsurprising info: the model has 360B params trained on 3.6T tokens, rather than important details like dataset and architectural tricks. What’s the big fuss?
https://
cnbc.com/2023/05/16/goo
gles-palm-2-uses-nearly-five-times-more-text-data-than-predecessor.html
… -
Tiny Language Models Generate Coherent Text With Simple Vocabulary
By
–
TinyStories: How Small Can Language Models Be and Still Speak Coherent English? A tiny 10M-param transformer can generate paragraphs of coherent text and reason, when trained on synthetic stories limited to only words that a 3-4 year-old would understand.
-
Altman has become the villain of the community according to critics
By
–
"Altman really has become the villain of the community"
https://redd.it/13jhduh