• 0 Posts
  • 10 Comments
Joined 1 year ago
cake
Cake day: June 13th, 2023

help-circle



  • They’re definitely reducing model performance to speed up responses. ChatGPT was at its best when it took forever to write out a response. Lately I’ve noticed that ChatGPT will quickly forget information you just told it, ignore requests, hallucinate randomly, and has a myriad of other problems I didn’t have when the GPT-4 model was released.




  • There a great Wikipedia article which talk about it. Basically AI has always been used as a fluid term to describe forms of machine decision making. A lot of the times it’s used as a marketing term (except when it’s not like during the AI Winter). I definitely think that a lot of the talk about regulation around “AI” is essentially trying to wall off advanced LLMs to the companies who can afford to go through the regulation paperwork while making sure those who are pushing for regulation now stay ahead. However, I’m not so sure calling something AI vs LLMs will make any difference when it comes to actual intellectual property litigation due to how the legal system operates.


  • This project might not be exactly what you’re looking for due to the limited amount of prebuilt models, but this is an interesting project nonetheless. It seems to run on a variety of hardware (even smartphones), however, you’ll need to compile your own models if there isn’t a prebuilt model available. Luckily at least Vicuna is included as a prebuilt model. There’s another model included called RWKV-Raven which is actually an RNN instead of a transformer that approaches its level of performance. Seems pretty interesting.



  • Bing Chat seems to be severely limited not only in its functionality but also the context it can “remember” as opposed to ChatGPT despite them both using variations of the GPT4 foundational model. Bing will usually give me either inaccurate answers or ones that don’t relate to what I’m asking about. The browsing plugin for ChatGPT performs much better but unfortunately OpenAI has disabled it recently due to it linking to outside sites (which I believe means they would have to pay certain sites, such as news sites, a fee for linking in certain countries). Overall their browsing plugin worked pretty well but would still get distracted by various links on a page (a subscription link or FAQ for example it would erroneous visit). Their recently released GPT 3.5 browsing plugin (in alpha) actually seemed to do a better job browsing and would get less distracted than the GPT4 version. Anyway this was a bit of a rant. One last thing to note, despite OpenAI disabling browsing, you can still browse the web using a third party plugin (beta feature) such as “Mixerbox”.