I don't think AI is bringing anything that is not on the table in terms of the basic structure of propaganda. Seeing 10% of the specific propaganda you see on the internet in numbers of 10s of news, and seeing it on 100s but still at 10% doesn't really change something. The means of production is evolving fast. I would agree to that. But there is always one secret sauce that we usually don't produce under competition that gives propaganda an edge. Think about the feeling of sincerity between traditional and social media. Nobody was sure 360p videos, 140 character tweets, etc. would replace the traditional media. People found sincere feelings combined with quick dopamine in it. Same thing here, it's not the means to an end that brings the real change. Although it is catalyzing a substantial change for sure.
The difference is now there is the ability to do it at scale, and convincingly.
Bots on social media that can argue with you, share experiences, other forms of astroturfing, etc. Bots that can directly respond to your questions.
The difference IMO will be that previously making interactive propaganda was very difficult and expensive to do in a convincing manor. And also literally anyone can now ask for "a CDC article describing the dangers of [literally anything]" and have an article so convincing even diligent people won't be able to immediately recognize.
Currently unless you know what you're looking for, GPT4 is very convincing. I think even the hn crowd is already seeing bots on hn/reddit/etc and not realizing it.
AFAIK the previous models couldn't convincingly continue a conversation chain, they were limited to mostly individual posts/comments since they were rather obvious when trying an actual conversation. It's fully interactive now.
And it's not just mega-coorperations/nation states, soon everyone will have access to something vastly better then GPT2.
---
I've been in the GPT-3 beta, and we have had community trained GPT-2 models for a while as well. Those seriously don't compare to GPT3.5/4.
Those severely underperform compared to even LLaMa+finetuning now, and GPT3.5 is a mile ahead of that, and GPT-4 leaves everything in the dust. Yes we've had it, and yes they're often able to pass the "casual glance turing test", but they suuuuuck in comparison to GPT-3.5.
What I'm trying to say is that propaganda/astroturfing is no longer limited to simply being a post/comment. It can now be an entire chain.
----
For example, "I was attacked by [person]'s dog"
You ask "Oh man did you get hurt?",
They respond " Luckily, I didn't get seriously injured, but it was a pretty scary experience. The dog managed to bite me on my left arm, but I was wearing a thick jacket which protected me from any deep punctures. I still got some bruises and scratches, but nothing too severe. I'm just really grateful it didn't escalate further.
I talked to the owner, and they apologized profusely, assuring me that they would be more careful in the future. It turns out that the dog has had some behavioral issues in the past, but they've been working with a trainer to help correct them. "
GPT 2 could never, ever even compare to that. This response IMO would pass as human to 99% of people who weren't specifically looking for it.
I'm not sure how effective such bots will be. People have tried having masses of actual humans defend their country in the comments of news articles and the like. That often seems to backfire instead of help.
AI is able to craft the type of highly persuasive rhetoric that has literally changed the world in the past 20 years or so at scale for pennies on the dollar.
A more complex natural language computer plugged into all social media can learn and adapt at a very high pace. If we also get it to hallucinate at at least 25fps then the whole thing can become murky. It could impersonate people on video calls and stuff along those lines. This tech looks like it will bring at least some changes in many aspects of our lives.
IMHO, this is less surprising discovery about him considering this probing is around 300 years ahead of his time, and Da Vinci already has proof for concepts 500 ahead of his time.
>Very cool. I've been doing some "prompt design" myself, how did you land on the temperature 0.001?
The point is to use the least generative methods to parse the HTML.
> It's kind of clunky to iterate on a prompt. If you change one thing about the prompt to get a better result for one aspect of the desired output, it can negatively impact the quality of another aspect.
There is an implementation for testing prompts internally when you set it up on local. This way you can get an idea about which parts a prompt and restrictive hyperparameters are failing you.
With that being said, you can also design a unique custom parser by changing the prompt for different subsets of data.
>If you wanted to make the ratios more accurate. You might add something like "Make sure the ratios of ingredients are correct.
> Which is better, but many times something else gets worse (e.g. the list of intructions to actually cook the thing). Curious what process you followed when designing the prompts.
This is really interesting. I would love to hear more about it if you would like to open an issue under the repository with something like: "Using additional instructions to improve parsing of preset parsers", I may add a preset parser on that issue to showcase.
> I think someone could build a business around "prompt design". Small tweaks make such a huge difference that a tool that helped you improve prompts would be really valuable.
