Hacker News Reader

Comment by rany_ | Hacker News Reader

Somehow this makes me less impressed and excited by GPT-4. I was expecting it to be better than Bing Chat.

One thing to note is that Bing Chat isn’t the straight GPT-4 model, it’s heavily prompted to show you an answer in a format that Bing Chat wants it to be in. Notice how most lines in a Bing Chat response are direct summaries of lines from the source with a link to the source? I’d argue that while doing that is the morally correct thing to do, it does hamper the usefulness of you aren’t looking for website sourced results.

It’s like you’re given a bathtub full of water and you can only use the water through a straw. It’s great if you’re thirsty, but what if you really want to take a shower…

This is also really good because everything it saids can be easily referenced and crosschecked. Definitely a big reason I switched over to bing.

Just thinking about the headlines that Bing Chat made in the past few weeks made me go "Ouch"

Same.

I've found it incredibly useless.

I'm scared something might happen with GPT3/playground. They might start adding restrictions for medical or marketing/psychology(I use it to defend myself from exploiters).

In January, it seemed like something happened to ChatGPT like it was nerfed. It made me panic and I ended up spending hours asking GPT3 questions because I was afraid it would be gone soon.

> In January, it seemed like something happened to ChatGPT like it was nerfed

I asked it some questions this morning that came back surprisingly crappy compared to the exact same prompt just a few weeks ago.

Had a similar reaction. I think there is a lot more going on here... Is it possible that some VIP told them to "turn it down" out of fear of losing their own business model overnight?

Would make a good movie

> I use it to defend myself from exploiters

how do you use it for this?

Using gpt3, describe yourself in detail then end the prompt: "This is how marketers exploit me:"

or "this is how people exploit me:"

swap exploit with other things like 'flatter'.

(Note that OpenAI is going to keep this data on you, might be worth using a different persons name)

Bing Chat appears to specifically, be an 'early' (to quote the MS press statement) GPT-4. I read that as implying that it was only partially trained, and so better than GPT-3.5, but still well below the final GPT-4 model.

I've read that just as Bing using the GPT-4 foundation model, which finished training in August 2022, according to OpenAI. Microsoft doing the tuning themselves would explain why this aspect wasn't on par with ChatGPT. Microsoft previously tuned MT-NLG for their Indian Sydney test, so they would have some experience.

Reply on news.ycombinator.com