And what if they can very convincingly and quickly reply?
Why would being able to write different emails help? Wouldn't they just reuse the same email over and over again, maybe just changing the addressee's name?
Also, now I wonder what the typical response rate is for phishing emails? Are they already getting too many responses to keep up with or are they already managing just fine without bots?
I think LLMs might be better for scams carried out over instant messaging services like Discord because that probably requires greater attention and flexibility on the part of the scammer.
Still, has the number of people that could conceivably be scammed over an email or instant messaging-based scam increased? That is, will LLMs, by virtue of their response quality, responsiveness, or some other salient quality, increased the pool of people who could fall for these scams? If the answer is no, then I don't see how increased scale and throughput for human-quality text generation will really matter much.
Think of it as a large scale A/B test of circumvention techniques against both human and automated scam detection systems.
Done right, it can effectively keep getting better on its own. Then add phone scamming and various other interactive scams once it is good enough. TBH, it is pretty scary where this will inevitably head.
Turn this around, and have an LLM that replies to scammers and keeps them busy.
There's already https://en.wikipedia.org/wiki/Lenny_(bot) , but this is an area where AI could actually be a useful addition.
https://huggingface.co/models?sort=modified&search=65b+ggml
Note that 33b llama models show up under "30b" or "33b" due to an old typo from Meta.
I also look at what people are choosing to host here, though Kobold is mostly centered around the roleplaying crowd: https://lite.koboldai.net/#
Or you can also see here https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderb...
Thats very interesting.
The infamous Pygmalion 6B is a GPT-J finetune, predating the LLM craze. Yet its decent in its roleplaying niche.
But the LLaMA 13B version, with instruct finetuning, is massively better, even with dataset errors that allegedly messed up its performance. In fact, a chat with Metharme 13b, where it made some very introspective logical jumps, was my first real LLM "Wow!" moment.
And Airoboros-Chronos 33b is leagues ahead of that.
If someone in that forum has a 3090, and trains LLaMA 33b on that dataset + a instruct dataset off huggingface... Yeah, that would be terrifying.