Live data from Hacker News

New GPT-3 model: text-DaVinci-003

beta.openai.com

141–150 of 216 posts

Re: New GPT-3 model: text-DaVinci-003

#141
post #26

Earlier quoted context omitted.

Any ideas on how to solve this issue of kids cheating with GPT3 essays?

Realistically? Grade based on thought process and validity of the argument, not whether it has spelling or grammar mistakes. GPT3 is still pretty incoherent over the span of enough text.

Kids' writing can also be very incoherent, sometimes more so. But incoherent writing still counts as turned in work and will get you points and teacher feedback, but GPT-3 generated should not.

Re: New GPT-3 model: text-DaVinci-003

#143
post #128
post #114

OpenAI should really change their name considering they don't open source their best models. It's just a weird name considering there are other companies in the space open sourcing everything despite not having "open" in their name.

Thats interesting. Which other models are open sourced that are in comparison to OpenAI? I would like to set it locally and give it a try.

Stable diffusion is a bigger deal right now than DALLE. SD is made by Stability AI, a company with a name that doesn’t proclaim to be “open” and yet their model is open to everyone.

Re: New GPT-3 model: text-DaVinci-003

#144

From the "Prompt Engineering Tips" section of the article: Customer: I need my [swear] internet. The technician is 2 hours late and I need a steady stream of cat gifs directly injected into my veins to stay alive. And the model's response: Response: Our field technicians report that all their trucks were stolen by alow-level drug cartel affiliated with the neighboring prison. As a gesture of good faith and apology, o…

This is something that still irks me about these text generation AIs, the sentences they form make sense on their own and link to each other. But in one sentence they say "...our CEO has asked that we pay..." yet the next sentence says it will be charged to your credit card.

Either that or AI is making smart jokes. I fear for the second case.

Re: New GPT-3 model: text-DaVinci-003

#145

From the "Prompt Engineering Tips" section of the article: Customer: I need my [swear] internet. The technician is 2 hours late and I need a steady stream of cat gifs directly injected into my veins to stay alive. And the model's response: Response: Our field technicians report that all their trucks were stolen by alow-level drug cartel affiliated with the neighboring prison. As a gesture of good faith and apology, o…

It’s worth noting that this doesn’t actually make sense. Why would an extraction team come to your house? What is an extraction team for, in this context? Most generous interpretation is the extraction team would be rescuing the trucks. Otherwise, What is being extracted here? The trucks are not at your house, so why send the team there?

It sounds clever at a superficial glance, but what is impressive is its “turn of phrase” rather than any logical thread in the story, which is a non sequitur.

Don’t get me wrong, it’s still amazing, but it didn’t make me chuckle but rather scratch my head to try to make sense of it.

Re: New GPT-3 model: text-DaVinci-003

#146

Earlier quoted context omitted.

Using GPT-3 might be a better skill to have.

You mean clicking a button?

No, I mean managing an AI to achieve a random task. Prompting, iterating, filtering - they all require high level input from the user. A LLM is a complex beast, not easy to use (yet).

Re: New GPT-3 model: text-DaVinci-003

#147
post #97

I wonder at what point we can start performing useful psychological experiments on a language model that will transfer to humans.

GPT-3 can be used for "Silicon Sampling" - running a virtual poll based on prompting with varied personality profiles. https://jack-clark.net/2022/10/11/import-ai-305-gpt3-can-sim...

That’s wild! You could fine tune a message for effect (or maybe just ask the model for the most effective formulation).

Re: New GPT-3 model: text-DaVinci-003

#148
post #56

We've been waiting for GPT-4. Does GPT-3-003 mean that GPT-4 will be delayed?

I think we'll need about 10 times as many parameters and text on the web... Or so. How long would that take?

A couple of months ago OpenAI created a model called Whisper that can transcribe speech efficiently and with high quality in 97 languages. It might be used to transcribe the whole YouTube, someone estimated about 12T tokens. GPT-3 used about 0.5T tokens for training. https://twitter.com/HenriLemoine13/status/157284645289587507...

If you add the visual data you get another order of magnitude increase in size. I think this training will teach the model procedural knowledge - how to do things step by step, and that will have a huge impact in robotics and software agents.

Re: New GPT-3 model: text-DaVinci-003

#149

Earlier quoted context omitted.

That's not what most of us consider "self-aware". That's just knowing your name. To me, "self-aware" is essentially being able to watch yourself think. I'm pretty sure GPT-3 can't do that, in any model.

I like to share this example of self-awareness any chance I get :) [1] https://news.ycombinator.com/item?id=29576737 [2] https://toldby.ai/4kQNd-_tvUG This was on the first davinci model

Is there some way to "walk back" story construction on toldby.ai, to see other forks of the same initial ... take on fish?
Post reply on HN