Hacker News Reader

Comment by stanete | Hacker News Reader

For me the unreliability is not a problem because I’m not using ChatGPT to search for information but rather to help me think and solve problems. ChatGPT can make incredibly good questions and can create relationships between concepts that it would take me hours to make on my own. So ChatGPT just become my private coach and a really smart actual assistant.

View on HN

Indeed, this is an assisted brain storming tool, not automated problem solving where you're supposed to trust the result wholesale.

For work I used ChatGPT to write a very specific programming problem and it wasn't a solution you'd ever copy/paste...the lack of context to the surrounding system means that's impractical 99% of the time outside of toy problems.

But it was still super useful for getting my brain working and suggesting a really good basic structure it would have taken me 3-4 failure cycles to get to.

The same exists in DALL-E/GPT for corporate art generation, writing 'essays', or stories or w/e. It's almost never producing the end product (besides toy problems).

So, if it can't do that in the first place, why evaluate it as if that's what it's supposed to be?

This is how I've used chatgpt as well. A primer for brainstorming that can wander on any tangents you are after

I’ve found that the more context I give to it, the better it is to help me solving a problem. If I spend a bunch of minutes giving info about what I’m doing, why I’m doing it and what I personally know about the problem space, it will usually help me to either explore more or to arrive at the solution faster.

When I talk about coaching I meant that I told ChatGPT to answer only with questions. In a way, ChatGPT is good at having socratic conversations. Sometimes it gets out of character but it depends mostly on your answers or questions.

Keep in mind that...

> The model is able to reference up to approximately 3000 words (or 4000 tokens) from the current conversation - any information beyond that is not stored. (https://help.openai.com/en/articles/6787051-does-chatgpt-rem...)

So if you overflow that it'll loose context.

This is how I am using it too. It completely blows away the procrastination aspect of starting a task. I just start chatting about it and get the ball rolling.

In line with this, something I feel like I’ve observed is that you can split people into those that like/need to hash things out with others and those that either don’t or aren’t good at it. It’s more than rubber-ducking. It’s collaborating on exploring a cognitive space (ideas, understanding, etc).

Personally, I’m one of these people, and I’ve struggled with how much others either don’t work this way, don’t want to or don’t know how. I think it’s powerful and that finding good collaborators at this way of thinking is a factor for general success.

Interesting to think about how AI can provide a general base level for this kind of thinking.

This. And AI assistants will mostly help people who are good at collaborating. Think of Jarvis and Tony working together to solve a problem. Now think of you and ChatGPT filling a Value Proposition Canvas or preparing the questions for a user interview. ChatGPT’s current limitation is that it doesn’t have enough context about the problem space. So to get good results I have to talk to it for a bunch of minutes to explain everything. But I’m sure they’ll fix this pretty soon so I could feed it additional context (a bunch of documents or the code of my project).

Yea, the prospect of domain specific AIs and even workplace/product/family/personally specific AIs that are learning/re-learning on the full history of interactions feels rather formidable at this point.

I've already forgotten where I saw it (probably here), but it'll be interesting to see how quickly we get into a period like iPhone circa 2010 with the mantra "there's an AI for that".

I've always put chat bots through the ringer since the smarterchild days, only to be disappointed.

ChatGPT gave me answers I would have never gotten regarding extremely technical concepts that the wikipedia articles would never surface, and would need hours of research papers or formal academic work or maybe just working in the field and being in those communities.

I think it is ironic that Google is sitting on a better conversation AI that is just too busy finessing its own engineers trying to escape.

This just reminded me ChatGPT is around and I could just ask it how to do something in Hashicorp Vault...and it gave me exactly the prompt I needed to do the rest.

Reply on news.ycombinator.com