School is out. Fewer kids asking if for homework help
Meanwhile because school is out, they seek out character-based chatbots like character.ai, probably one of the reasons there was a server outage recently.
They completely banned it at my job, I’m willing to bet some companies are banning it.
Especially frustrating because we work very closely with Microsoft and have a team specifically for helping our clients develop applications with AI.
I can see where they're coming from in terms of security, but that sounds a bit harsh.
At least where I work, we're told basically "use it if you want, just be prepared for it to be wrong and double check everything it tells you" which sounds a little more reasonable IMO
Ironically, security isn’t a concern at all. With our relationship with Microsoft, we can use the Azure OpenAI API which, despite my own strong personal distrust in Microsoft, still meets all of our privacy and security standards. We trust Microsoft as much as we do our internal teams.
The concern is mostly legal/copyright related, as they’re worried any code or documents that come out of it could be considered copyrighted.
Wow that's a very protective stance to take! Good reason though. There's probably enough people out there actually copy + pasting the code that gets spit-out to have to worry about that. I try to only use ChatGPT to figure out syntax or get a high-level understanding of an approach to doing something, without having to scour the web for documentation.
Not only is it banned from my work, they have banned anything even close to AI. Even deepL is blocked and I need to translate things for my job daily. Google and Bing aren't blocked so I could still use their AI if I was going to use AI but that's not what I'm even trying to do lol.
DeepL is just the best at translating business related language things. Google does a decent job at it anyway...
I've been using it a lot less recently because GPT-4 has just been spitting out gibberish code. They really nerfed it.
I find that recently the effort needed to get the "right" answer is much more than in the past for gpt-4. That's my impression. At the end I am finding myself more often going back to google, stack overflow, manuals, medium...
I believe they distilled the model to much for performances, or the rlhf is really degrading the model performances
The fact that you have to double-check the answers anyway is a pretty big sign that it’s not useful for finding information.
I used mainly for programming support, brainstorming on techs and libraries, and to refine technical documentation. But support for programming is becoming a bit of a pain honestly...
I use it for help with relatively simple scripting at least once a month and it's become noticeably dumber recently. I find it's a lot more obnoxious to interact with and generally just isn't as useful as it was a month or two ago.
Same experience. It is also increasingly ignoring instructions, always ends up with similar answers whatever the request is formulated. It's a pity
Could it be as simple as the official ChatGPT apps being released for iOS and Android, so more people have moved to the apps?
there just isn’t much ChatGPT can help me with that I can’t do better myself and with less frustration. I can’t trust it for factual information or research
This mixed with the constant reprimanding and moral instruction just makes it so frustrating. I'm not asking for no filter, but it's gotta be more lenient if they want it to be a good and useful tool. I am so over reading "It is important to remember..." because it misunderstood a prompt.
I think I have one of the few use cases that it is still useful for. I write textbook materials, the ability to have chatgpt write a short piece of writing saves me a lot of time. For example, 'write an 80 word passage for a 2nd grade A1 esl learner about visiting grandma. Use the present simple tense.'
Or, make ten cause and effect sentences involving animals that live in the forest.
But even with those very simple prompts, I have to check it and usually ask it to change some vocabulary.
However, if I ask it to do anything with logic or even slightly complicated, it can't do it, so I always wonder what the heck people are using it for.
For example, 'list all the elephant and piggie books and fly guy books in order of AR level'
It can't do it. Thats a pretty basic task, but it gets it completely wrong.
I see this becoming more of an advanced “auto-complete”. It really shouldn’t be authoring anything, but instead work with software to make suggestions on how to improve human generated work.
Also, for software development it is a minefield. They train the AIs on code from GitHub and other projects and then suggest it back to users in violation of the license the code was built with.
That's exactly what a LLM is, an advanced autocomplete using a huge training database to determine the probabilities. It never was anything more, even with the unexpected characteristics from further development. The models that are being fine-tuned for their training (stepping them back to a more narrow field LLM) do a lot better than a general purpose one.
The generic ChatGPT is far too error-prone and limited compared to the many variations of other GPTs out there. It was a fad for those who weren't going to fine tune a use that worked well or are doing actual research in better tactics. How many who are knowledgeable on computer systems have moved to smaller locally installed versions that work just as well or better?
What local models are you using that are better? Not trying to argue, honest interest
There are a number of them now, but I've put the Vicuna 13B one on my Windows side before. Trying to get it on Ubuntu so it can use the GPU, but it's being difficult. Look up TheBloke on github, they have a large selection that can be used through the text-generation UI coding.
I may have misspoke saying "better", as it looks like it's a few percentages below on comparisons. I thought I had seen some varieties of local compared that rated higher though, such as on AI Explained's channel.
Thanks! I tried vicuna, but I didn't find it very good for programming. I will keep searching :)
I didn't either, actually. It seems to me that where LLMs excel is in situations where there will be a large consensus of a topic, so the training weights hit close to 100%. Anyone who has read through or Googled for answered for programming in the various sources online has seen how among the correct answers there are lots of deviations which muddy the waters even for a human browsing. Which is where the specialized training versions that hone down and eliminate a lot of the training noise come in handy.
Unofficial ChatGPT community to discuss anything ChatGPT