Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

GPT-4

openai.com

Post reply on HN
Print view View original

251–260 of 1001 posts

Re: GPT-4

#251

Post by monkeydust »

harrisonjackson wrote: post #187

I am interested in how a 32k token context even works. That is so much larger than 4k that I am having a hard time imagining how prompts will change and what sort of output is now possible. That is 50 pages of text. Far larger than most content currently being consumed and generated by LLMs. Q&A and summarization it will be easy to see improvements as current recursive summarizing and embedding techniques are very "l…

Yea this is huge. Been playing with conversational technology in langchain and one of the issues you have to manage is the historical conversations, langchain has some cool ways to deal with it but this changes the nature of the problem entirely.

Re: GPT-4

#252

Post by tekbog »

We can finally start an education and "testing" people's knowledge reform since GPT4 makes a lot of those tests irrelevant. It's an interesting point in history, how society, different institutions and countries will approach this new tool.

Re: GPT-4

#253

Post by detrites »

ml_basics wrote: post #136

From the paper: > Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar. I'm curious whether they have continued to scale up model size/compute significantly or if they have managed to make significant innovati…

What about the glaring safety implications of the custody of this power being in the hands of a relatively small number of people, any of whom may be compelled at any point to divulge that power to those with bad intentions? Secretly?

Conversely, if all actors are given equal access at the same time, no such lone bad actor can be in a position to maintain a hidden advantage.

OpenAI's actions continue to be more than merely annoying.

Re: GPT-4

#254

Post by netsroht »

Wow, a context of 32K tokens. I'm excited to see what new capabilities that will have! Up until now and depending on the task by hand, I usually broke a larger context down into several contexts. For example to summarize multiple websites and/or long social media posts, on a recent task [1] I fell back to making several requests each with its own (isolated) context and then merging these summarized contexts into a new context. That worked remarkably well, though.

[1] https://foretale.io/zeitgeist

Re: GPT-4

#255

Post by BoiledCabbage »

Imnimo wrote: post #218

A class of problem that GPT-4 appears to still really struggle with is variants of common puzzles. For example: >Suppose I have a cabbage, a goat and a lion, and I need to get them across a river. I have a boat that can only carry myself and a single other item. I am not allowed to leave the cabbage and lion alone together, and I am not allowed to leave the lion and goat alone together. How can I safely get all three…

It's a good observation.

Although on the flip side, I almost went to type up a reply to you explaining why you were wrong and why bringing the goat first is the right solution. Until I realized I misread what your test was when I skimmed your comment. Likely the same type of mistake GPT-4 made when "seeing" it.

Intuitively, I think the answer is that we do have two types of thinking. The pattern matching fast thinking, and the systematic analytical thinking. It seems clear to me that LLMs will be the solution to enabling the first type of thinking. But it's unclear to me if advanced LLMs will ever handling the second type, or if we'll need a different tech for it.

It seems like math problems (or unexpected logic problems like yours) could always be an issue for the first type of thinking. Although I would have assumed that programming would have been as well - and was surprised to see how wrong I am with that one.

Re: GPT-4

#256

Post by btx »

Has anyone found a way to trick it into using pictures with ChatGTP+ yet?

Pasting pure base64 images got this interesting response: "Thank you for providing the base64-encoded image! I can now process the image and analyze its content. Here is the decoded image:"

But it failed to do anything further with the image.

Re: GPT-4

#257

Post by throwaway4837 »

> Yes, you can send me an image as long as it's in a supported format such as JPEG, PNG, or GIF. Please note that as an AI language model, I am not able to visually process images like a human would. However, I can still provide guidance or advice on the content of the image or answer any questions you might have related to it.

Fair, but if it can analyze linked image, I would expect it to be able to tell me what text is present in the image. That seems useful and well-within the capabilities of their connected image models.

> I apologize for the confusion. Can you please provide me with the correct image or link to the design so that I can provide an accurate answer to your question?

It claims to understand how to look at images, but it failing miserably when I give it a simple sign-up modal Figma. I ask it what text/copy is in the design, which it claims to be able to answer, but it hallucinates a navigation bar, a logo, and other generic things that are simply not present in the design. It gets the copy all wrong. Once, it said that my design was a Celtic knot. Once I told it that it was a sign-up modal, it started spitting out generic sign-up modal React code that had absolutely no basis in the presented design. I tried engineering the prompt quite a bit, telling it that it was wrong, and needed to focus on the design at hand rather than generic designs it knows about, but with no success.

ChatGPT was confidently incorrect here which is unfortunate. I wish ChatGPT truly hooked into an image processing model, but it seems like it doesn't. Oh well. I guess I'm just going to have to write this code myself... :)

Re: GPT-4

#258

Post by 2bitencryption »

This should come as no surprise, but I do enjoy this cheeky little blurb at the end of the GPT-4 paper:

> GPT-4 was used for help with wording, formatting, and styling throughout this work

Re: GPT-4

#259

Post by chucklenorris »

Boring. Where's the model? Do they really think they can keep up with the community with this closed source approach? I expect that openai models will slowly be outclassed by open source ones, probably maintaining a few wins in specific tasks but open models will be eating their lunch in the meanwhile.

Re: GPT-4

#260

Post by tuvan »

rcme wrote: post #233
ml_basics wrote: post #136

From the paper: > Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar. I'm curious whether they have continued to scale up model size/compute significantly or if they have managed to make significant innovati…

I bet they use CLIP to caption the image and feed the text of the caption into GPT, but that's just a guess.

Did you check all of the samples provided? It can read an entire research paper and understand the figures just from the images of the papers pages. This seems to be a much deeper connection than extracting captions.
Post reply on HN