Hacker News Reader

Comment by ugjka | Hacker News Reader

> Riding with Altman, I can almost hear the ringing, ambiguous chord that opens “A Hard Day’s Night”—introducing the future. Last November, when OpenAI let loose its monster hit, ChatGPT, it triggered a tech explosion not seen since the internet burst into our lives. Suddenly the Turing test was history, search engines were endangered species, and no college essay could ever be trusted. No job was safe. No scientific problem was immutable.

ChatGPT craze has wore off for me, because of constant hallucinations when you ask for something slightly more esoterical. And I can't justify paying 20$ for GPT-4 to have more convincing hallucinations

View on HN

I might be a mediocre developer but I find ChatGPT invaluable for my job. It finds potential problems in my code, suggests solutions, helps with libraries I don’t know about, helps in drafting solutions when given a high level description of the problem. It makes mistakes but I don’t expect it to be perfect. As a non native speaker, it also helps me with my English (didn’t use it for this comment).

> didn’t use it for this comment

It's worrying that many people are starting to use it as their default. The day that everyone uses an AI to sanitise and write all of their speech, which is in itself an expression of your personality and experience, is the day that the soul is sucked out of all human communication.

Just to add, I think AI tech is great in of itself, I just don't see the value in everyone using it to turn into some amorphous nondescript communication medium. There's a lot of value in imperfections and quirks.

> There's a lot of value in imperfections and quirks.

Value for whom? For someone whose intent is to track you?

Value for relationships and the general value in diversity in life. Are you proposing that all digital communication should be normalised, across dialects, languages, etc, to remove any trace of individuality to avoid tracking?

Hallucinations are awesome sometimes. It hallucinates great APIs that you can than take and implement them. I also use it as an early audience to build up my arguments, as dialectics is kind of a refined hallucination. It's also a very good idea summarizer and doesn't hallucinate too much in that case. It is great for NLP tasks that were hard previously, too. I remember spending two months to implement a decent aspect based sentiment analysis model and now I get a better version with the turbo-gpt model without any training and there's no big hallucination problem over there. You may have too high expectations from a statistical model to be an expert in uncommon or esoteric matters, as there's too many of those and they may look similar enough to other ideas that the model finds it more likely that there exists an explanation that it didn't see than that there isn't one on the web. Therefore the utility of hallucinated APIs. My problem with their web service is that I pay 20$ and I still have to complete awful CAPTCHAs on a daily basis and GPT4 doesn't fix the smaller model issues, just amplifies it's strengths.

People bitch a /lot/ about the hallucinations AI produce.

They're definitely a problem if you're working in academic or engineering fields, and it'll probably take a lot of work to make them less so.

However, I think the hallucinations have kind of been helpful for the limited use I've had of chatGPT: Getting bored/writer's block and asking it to help generate some ideas I can play off, based on a limited description of the situation and context. There's no copypasting involved, but holy crap chatGPT already outperforms most "Talking to another human being" experiences I've had for the same purpose.

I can justify paying $.02 per 1k tokens and using it via the command line llm[1] tool, though. That is a far better deal.

[1] https://github.com/simonw/llm

Saying that ChatGPT sucks after trying only the free version is like saying that pizza sucks after trying only the frozen pizza because you don't want to spend $20 on a pizza in a good Italian restaurant.

I think he's saying he can't justify keeping the subscription.

I'm in the same boat - whenever I think it would be faster to use chatgpt it usually ends up being a waste of time and flow breaker. And it got worse over time. At some point a few months ago I realized I haven't used it once in a month, so why keep paying ?

Copilot is way more useful to me.

What model, pray tell, do you think Copilot runs on, if not a fine-tuned version of GPT?

It's not just about the model, it's about the integration. Copilot is easy to use because it's basically automatic in vscode. So it isn't a 'flow breaker'. That is in some ways its primary benefit: it gives you something similar to searching for the solution a problem without having to leave the editor. (At the loss of being able to see the context of where the solution comes from, of course.)

This is a thread about hallucinations.

>ChatGPT craze has wore off for me, because of constant hallucinations when you ask for something slightly more esoterical. And I can't justify paying 20$ for GPT-4 to have more convincing hallucinations

Integration has nothing to do with hallucinations.

Your response and the person you responded to were discussing copilot vs chatgpt. Both may "hallucinate". Copilot is more useful (to the GP and to me) - it's domain specific and well integrated, despite occasionally producing pure crap.

I also pay for copilot and don't for chatGPT.

Yes, we were discussing Copilot and ChatGPT, two products run on essentially the same model with no clear difference in their ability to hallucinate (or not).

ChatGPT is a product built with models. So is copilot.

GPT and llama are model (families).

Yes and GPT powers copilot?

I'm really straining to understand your point here. Are you implying Llama is the model underneath Copilot?

You should never conflate products with models, it leads to invalid comparisons between models and products. Which for some reason has been the norm. How many times have you heard someone ask "How much worse is Llama than ChatGPT?"?

The point is that while both ChatGPT and Copilot are powered by the same model, they are different products, and one can be useful for one task where the other isn't. Just because someone doesn't find ChatGPT useful doesn't mean they can't find Copilot useful, since there are differences between the two products even if they are powered by the same model.

Copilot is a tool that does intelligent copy paste with an interface that doesn't break flow - smaller models can support that.

ChatGPT tries to be a knowledge oracle/problem solving via conversation and it's not up to that task.

Replying from Italy right now, $20 easily buys you two pizzas, just sayin'.

two _better tasting_ pizzas

FWIW I pay for GPT-4 and it actually hallucinates much less. I feel it was trivially worth the upgrade and would sacrifice other services before ChatGPT (I'd rather get API access and use my own frontend of course but I doubt they want to let me).

You can pretty easily access GPT-4 via the API in a chat interface using SlickGPT, and that was just the first result on Google. That is pay-per-request (technically per token), not a subscription.

Yes I know there are lots of such apps, the trick is getting API access in the first place.

I think they've opened up GPT-4 with 4k tokens for a month or so now, and I wasn't in the preview. I still don't have access to the 32k token model.

Sometimes it’s a matter of communication style with the LLM, some communicators mesh much better.

If you think the info is general enough to be embedded, there is probably a skill to finagle it out.

Definitely noticed a degradation of quality over time.

Nothing beats ChatGPT 4 Week1, not even ChatGPT4 Today.

You can just specify the 0301 model in your API calls.

Bing Chat, which uses GPT-4, is free.

asks for edge

All you need is Edge user agent and it works fine.

Use an extension like https://addons.mozilla.org/en-GB/firefox/addon/enable-bing-c...

Reply on news.ycombinator.com