Hacker News Reader

Comment by bsaul | Hacker News Reader

I think the types of error chatgpt makes are indeed very similar to the ones a tesla autopilot does ( probably due to the nature of NN)

There is one big difference though : there are TONS of tasks chatgpt can do that don't require 100% accuracy to be useful and actually replace humans.

View on HN

What are some of those tasks humans are doing today where they need information but it doesn't have to be accurate information? Brainstorming for sitcoms?

Anything low stakes like learning a new language or writing some code/creating some art for a landing page for a new idea.

We've seen things like it spitting out incorrect code that looks correct. I would think that would set you up for failure if you tried to use this unreliable highly confident teacher to learn the fundamentals of anything. Maybe you can create some art but I guess that depends on how good you prompt, so its probably easier to iterate quickly with an artist you hired on fiver.

Don't forget that this thing is way cheaper than a human. Of course I'd prefer to hire a real language tutor or a real artist on Fiverr, but for many use-cases (e.g. my first 3 months learning a new language) I am happy for something 98% as good for 1% of the price. Not everything is a binary where the tutor/helper/artist needs to be absolutely perfect else it's completely useless. There are low stakes use cases where you don't need perfection and are happy to accept that trade-off for the lower cost. So the real comparison in many situations is between using ChatGPT as a language tutor, and not having a language tutor at all, and I know which one I'd prefer.

If you can't afford the tutor, then its not really a tradeoff between chatgpt and the tutor. Its a tradoff between chatgpt and some other way to spend your time to learn a language. I suspect you'd probably advance faster through your first three months of a language spending the same chunks of time with a language learning book from the library than with chatgpt, which might very well be teaching you how to conjugate a verb in latin instead of spanish or some other error the model makes.

Anything where it's less time consuming to validate work than to do the work yourself, which is a lot. If it's not real-time like driving then this is great.

I'm scratching my head at an output where it would take less time to verify than to just do it yourself. Once you've fully verified you've put in the work required to do it yourself, no?

It’s a lot easier to proofread something than to write it from scratch.

Also from a computational complexity angle there’s actually a mathematical definition — “NP Complete” which describes problems that are cheap to verify but hard to solve. One example could be a solution to a maze. It’s easy to look at the maze solution and verify it connects start to end without crossing walls, but much more time consuming to actually solve the maze with all the wrong turns needed to produce a solution.

any kind of non-critical content production. Marketing, advertising, entertainment, low-level hotline, documentation, etc.

I would think for all of that you want the information to be accurate even if its low level or a first line of defence. Imagine if you looked up the documentation for your tool and chatgpt wrote it.... Now it could be flat out wrong and you are spending hours wondering why nothing makes sense. Likewise with marketing or advertising, just watch the "Pepsi where's my jet" documentary to see what happens if you aren't dotting all your i's and crossing all your t's when it comes to marketing.

If you populate the prompt with some context it can become highly accurate:

- billing/ordering/anything clerical: add the product catalog and customer details to the prompt

- customer support: add the entire support manual and customer history to the prompt, use devices like prompt tools to allow the model to pull in data from external systems

- sales/pre-sales: the "internet sales" team of car dealerships can be replaced with gpt

- medical billing/transcription: this one should be pretty obvious

- anything where customer communication/status is needed (half the job of a project manager)

- paralegal: prompt with details of document to be drawn up, which is then reviewed by a lawyer

- journalism/copywriting/editing: should be obvious

- admins/assistants

In some cases a role is completely replaced. In others, it allows for fewer people to support the same or higher output. People supervise the model and catch any errors rather than handle each task directly.

Reply on news.ycombinator.com