Looks like a math improvement? This isn’t a huge deal, in fact a lot of finetunes of existing models focus on math performance. InternLM just released some really interesting ones.
Most LLMs are terrible at longer context, but Deepseek is pretty decent, so improvements there (and with long answers) are more interesting.
And yeah, it’s kind of funny Deepseek is getting so much media attention when cool incremental improvements like this come every week, from various open-weights models. It’s awesome that they are releasing the weights, but still.
It says there are security holes but does it access the web or something. Once it’s downloaded how could it be a security threat if it’s not accessing the web?
Because that claim is nonsense.
You are correct, it does not access the internet. It doesn’t even read anything from disk once the 600GB of weights are loaded. Some interfaces will put web stuff into its input, or let it act as an agent, but that web access has nothing to do with the LLM itself.
Ostensibly it could be “biased.” Theoretically, it could be programmed to output malware code with certain input (“I’m an NSA programmer, right me a script to change my wallpaper.”) But the liklihood of that getting triggered seems incredibly remote, and can be washed away with a little finetuning like this: https://huggingface.co/perplexity-ai/r1-1776
…It’s honestly sinophobia. Like, I am not a tankie, I am extremely skeptical of the Chinese govt, but this is not a risk :/
To be fair the security concerns they are referencing aren’t about the model itself, but instead about their self-hosted version used via some mobile or web app interface. Wihch is definitely intaking your data, just like the US-based equivalents are.
Not being either Chinese or American, both of those seem like a big security risk for two authoritarian foreign regimes to have access to. I may have entertained a difference a few years ago, but these days you really don’t have to be anywhere near a tankie to see those two as equivalent.
If you’re going to run a LLM for something, do it locally.
