I have to redo about 80% of my output in GPT and that’s only for fairly easy text output which should be far easier than coding, for example. I can’t imagine the headaches coders must go through trying to use GPT…
The more I use ChatGPT the more I realise that it’s completely unreliable, and quite useless actually, and only suitable for pre-school or high school level output.
For example, I asked it to to reconstruct a text file that had added modifications in specific sections, exactly as written, with no omissions, by putting the modification in the right place of the text. It was basically just a cleanup. It couldn’t do it. Each version was missing text, some of it big chunks of text, other sections were summarized, or had rewritten content to a point where I had to start all over again.
When I tried to perform the task in smaller segments of less than 400 words each , it still couldn’t do the task accurately.
I then asked it to analyse the file and the generated output to find all missing sections and text in the output. It provided a list of obvious failures, many were either misrepresented, shortened, or entirely omitted from the reconstruction provided earlier. Gpt then promised to rewrite everything “from the original draft, line by line, with zero alterations or summarisation”.
It still couldn’t do it. We tried again. And again. Again and again and again it couldn’t complete a fairly basic task. And it’s been bad for about a year… I don’t know what the “great” updates are when it doesn’t show in the quality of the output. And it doesn’t matter what model you use. I actually have better results in other LLMs although GPT is suppose to be superior for the tasks I use it for.
And the same goes for questions and answers. GPT is basically just a glorified search engine that thrawls through anything found on the internet, good quality or not, with no structure to assess whether the data is accurate or just made up, or someone’s pointless opinions.
It’s obvious that there’s something wrong with the technical team at OpenAI who work on the training, algorithm, fine-tuning - whatever - and the same failed strategy and thinking to prepare the work keeps being repeated, which implies that someone in the team is recycling the same mistakes, the same wrong approach to structuring the LLM in a loop.
It’s clear there is a need for changes with new people, new skillsets added to the team.
Am I wrong?