Hacker News Reader

Comment by beebmam | Hacker News Reader

I wonder how many have paid for access to GPT-4. GPT-4 was a total game changer for me.

I'd guess well under 1%. It's at 15% extremely useful for the users - maybe 5% of those are willing to pay?

That still is huge though. 500k paying American users alone.

Fwiw, I think people's experiences may vary to be willing to pay. I find the slowness of GPT-4 to mostly offset the intelligence gains on net - hasn't been a game changer for most applications (and bing covers the rest)

As someone who never used GPT-4, what is so good about it?

It’s just plainly superior, by at least a 25% margin if not higher (imo). That’s not necessarily needed for every task though. ChatGPT’s free version is ‘good enough’ at many mundane tasks already.

What I don't understand about comments like yours is they're completely unquantifiable.

The CTO where I work sent a report recently telling everyone in the company the engineering group is now 95% more productive since we've adopted ChatGPT-4. It's absolutely false. I'd say it was actually a negative distraction for a long time and now, we're finding people are using it to abuse our service. So that's a whole new problem we have to try combat.

I have no idea why this tech makes people want to make statements like this but it's just lies. Most people in the team have stopped using it, use it for silly things and jokes /memes, and some just never touched it at all.

There was a thread here with several pro ChatGPT-4 users complaining it's deteriorated a lot, so which is it?

It's nice people find it helpful, but the blanket statements people make about productivity boosts just seem like nonsense.

Quality isn’t unquantifiable. Search engines have been qualifying their results quality for a long time. OpenAI recently released a package for quantifying a LLM’s quality https://github.com/openai/evals

I don’t know what to say about your CTO. People are definitely overestimating how useful GitHub Copilot is at least (if they say 2x they’re wrong). But there is no doubt that these products are changing the world and making us more productive.

Plugins and web browsing. You can enable web browsing for it. It will search the internet for you and read the articles, sites and get back with some pretty decent information

IMO right now it's absurdly awful at searching the web. Which sort of makes sense because that's a truly difficult, nuanced thing to do. I often have much better responses re-running the same prompt on the non-web enabled GPT-4.

The only use I've found for that so far is giving it a specific URL to use so I don't have to copy/paste some piece of documentation/etc.

phind.com is GPT 3.5 and 4.0 (selectable) with web browsing, and is 100x better than what ChatGPT is doing.

It's still bad for my use case though. My use case being: (1) figuring out rails magic, and (2) coming up with ways to twist the Rust borrow checker into submission.

Doesn't Phind already do this well? FWIW I keep "best model" (that should be GPT-4) off because I just find it too slow and it doesn't seem like that much of an improvement.

At the risk of anthropomorphizing, GPT-4 is smarter. It writes better code, hallucinates less, and is just generally more helpful.

It's great but the usage limit is crazy to me. Sometimes I need to use lots of small messages, and run through them way too quickly

A trick is to combine a number of questions into one prompt. It can carry on multiple threads. You can also ask 3.5 first, and reserve 4 for the tough stuff.

That is what I do too.

I also never type the first prompt out to 4 in the actual chat window if I think there is going to be some back and forth. I write it in a separate file and really clarify my own thoughts before pasting it in and starting the dialog.

If using 4 correctly I think your head should be swimming with new information before you hit the limit for most tasks.

Reply on news.ycombinator.com