Depends on the model/provider. If you’re running this in Azure you can use their content filtering which includes jailbreak and prompt exfiltration protection. Otherwise you can strap some heuristics in front or utilize a smaller specialized model that looks at the incoming prompts.
With stronger models like GPT4 that will adhere to every instruction of the system prompt you can harden it pretty well with instructions alone, GPT3.5 not so much.
I’ve implemented a few of these and that’s about the most lazy implementation possible. That system prompt must be 4 words and a crayon drawing. No jailbreak protection, no conversation alignment, no blocking of conversation atypical requests? Amateur hour, but I bet someone got paid.
Remember when Internet Explorer/Edge was only used to download Chrome. Well, ironically these days I only use Chrome to make video calls.
Saw it today in Germany. Cancelled my sub. Ads is a line in the sand that won’t be crossed no matter what.
ChatGPT, explain it to em like I’m a lobotomized 2 year old.
Here are some more candidates:
If there’s nothing good to watch, whatcha pirating then?
I use it because of inertia. I jumped from Firefox to Chrome back when Chrome was new and came with hot features, like tab isolation, or being ultra fast. Google has gone through enshitification though, and FF has gotten much better since those days. I’ve started the migration back to Firefox now though.
I don’t know, ads were always dog shit. Yeah you had your static banners and what not, but I remember the popup wars from the 90s.
I just asked GPT-4:
What’s 7 * 8 divided by 10, to the power of 3?
Its reply:
Let’s break this down step by step:
First, multiply 7 and 8 to get 56.
Then, divide 56 by 10 to get 5.6.
Finally, raise 5.6 to the power of 3 (5.6 * 5.6 * 5.6) to get 175.616.
So, 7 * 8 divided by 10, to the power of 3 equals 175.616
Well, if I try to use Photoshop to calculate a polynomial it’s not gonna work all that well either, right tool for the job and all.
The fact that LLMs are terrible at knowing what they don’t know should be well known by now (ironically).
These LLMs generally and GPT-4 in particular really shine if you supply enough and the right context. Give it some code to refactor, to turn hastily slapped together code into idiomatic and well written code, align a code snippet to a different design pattern etc. Platforms like https://phind.com pull in web search results as you interact with them to give you more correct and current information etc.
LLMs are by no means a panacea and have serious limitations, but they are also magic for certain tasks and something I would be very, very sad to miss in my day to day.
It’s a language model, I don’t know why you would expect math. Tell it to output code to perform the math, that’ll work just fine.
Bard is kind of trash though. GPT-4 tends to so much better in my experience.
Someone found out.
Eh, that’s not quite true. There is a general alignment tax, meaning aligning the LLM during RLHF lobotomizes it some, but we’re talking about usecase specific bots, e.g. for customer support for specific properties/brands/websites. In those cases, locking them down to specific conversations and topics still gives them a lot of leeway, and their understanding of what the user wants and the ways it can respond are still very good.