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

GPT Unicorn: A Daily Exploration of GPT-4's Image Generation Capabilities

adamkdean.co.uk

Post reply on HN
Print view View original

81–90 of 109 posts

Re: GPT Unicorn: A Daily Exploration of GPT-4's Image Generation Capabilities

#81

Post by alkonaut »

sigmoid10 wrote: post #78
blueorange8 wrote: post #61

Earlier quoted context omitted.

If you guys think Chatgpt and gpt-4 are static you haven't been using it. The answers change constantly (and not because of the inherent randomness of the response but openai is constantly making it "safer" and improving it's output via humans) - basically when any article comes out saying "Chatgpt can't solve X puzzle) within a day suddenly it can solve that puzzle perfectly. I can't tell you how many jailbreaks jus…

This kind of stuff is likely done without changing model parameters and instead via filtering on the server and prompt engineering. One day is simply too short to train and evaluate the model on a new fine tuned task.

I'm assuming the model has a hand writtn "prefilter" and "postfilter" which both modifies any prompt going in and the token that are spit out? If they discover that the model has problems with prompts phrased a certain way for example, it would be very easy to add a transform that converts prompts to a better format. Such filters and transforms could be part of a product sitting on top of the GPT4-model without being part of the model itself? As such, they could be deployed every day. But tracking changes in those bits wouldn't give any insight into the model itself only how the team works to block jailbreaks or improve corner cases.

Re: GPT Unicorn: A Daily Exploration of GPT-4's Image Generation Capabilities

#85

Post by capableweb »

blueorange8 wrote: post #61
kristjansson wrote: post #59

Earlier quoted context omitted.

Yes, this project seems like a misunderstanding of what Bubeck and team were obsering with their unicorn test. GPT-4 was being trained, and checkpoints were provided to them to experiment with. The improvements in the unicorn reflected further training progress. The models on offer now are frozen(-ish). Per the models[0] page, the non-snapshot model IDs "[w]ill be updated with our latest model iteration". So this pro…

If you guys think Chatgpt and gpt-4 are static you haven't been using it. The answers change constantly (and not because of the inherent randomness of the response but openai is constantly making it "safer" and improving it's output via humans) - basically when any article comes out saying "Chatgpt can't solve X puzzle) within a day suddenly it can solve that puzzle perfectly. I can't tell you how many jailbreaks jus…

> If you guys think Chatgpt and gpt-4 are static you haven't been using it

The bottom of ChatGPT highlights which version is being used of GPT-4 and supposedly what version of ChatGPT it is.

> ChatGPT Mar 23 Version. (https://help.openai.com/en/articles/6825453-chatgpt-release-...)

It's possible to change the output both by tuning the parameters of the model and also client-side by doing filtering, adjusting the system prompt and/or adding/removing things from the user prompt. It's very possible to change the resulting answers without changing the model itself. This is noticeably in ChatGPT, as what you say is true, the answers change from time to time.

But when using the API, you get direct access to the model, parameters, system prompt and user prompts. If you give it a try to use that, you'll notice that you'll be getting the same answers as you did before as well, it doesn't change that often and hasn't changed since I got access to it.

Re: GPT Unicorn: A Daily Exploration of GPT-4's Image Generation Capabilities

#86

Post by capableweb »

gwern wrote: post #41
sp0rk wrote: post #7

Did you generate a bunch all at once before starting to get some idea of what the natural variance looks like? I would think it's important to verify some level of progression over time, because with the current four it seems entirely possible that the examples could have all been generated at the same time with no changes to the model.

Also unclear if he's sampling at temp=0. Looks like he doesn't set a temp? https://github.com/adamkdean/gpt-unicorn/blob/8ad76ec7161682... So not sure what he's really doing.

Then it's using the default value which is temperature=1.0, which is by no way deterministic (not that temperature=0.0 is either, but it's more likely to give similar responses to similar prompts, than 1.0 is)

Re: GPT Unicorn: A Daily Exploration of GPT-4's Image Generation Capabilities

#87

Post by dr_scully »

If you ask it to write GLSL code implementing Signed Distance Fields, it can generate a 3D rendering of a unicorn as well. Just paste it into ShaderToy. I asked it to implement animation as well, for the legs, and it manage to do something although incorrect.

Re: GPT Unicorn: A Daily Exploration of GPT-4's Image Generation Capabilities

#88

Post by easygenes »

kristjansson wrote: post #59
simonw wrote: post #55

I filed an issue: https://github.com/adamkdean/gpt-unicorn/issues/2 "Running this project daily doesn't make sense if GPT-4 is not being constantly updated" With a suggestion to run it monthly instead, and generate 16 images at a time, and backfill it for GPT3 and GPT3.5.

Yes, this project seems like a misunderstanding of what Bubeck and team were obsering with their unicorn test. GPT-4 was being trained, and checkpoints were provided to them to experiment with. The improvements in the unicorn reflected further training progress. The models on offer now are frozen(-ish). Per the models[0] page, the non-snapshot model IDs "[w]ill be updated with our latest model iteration". So this pro…

Greg Brockman just stated in a long post on Twitter yesterday: "...it’s easy to create a continuum of incrementally-better AIs (such as by deploying subsequent checkpoints of a given training run), which presents a safety opportunity very unlike our historical approach of infrequent major model upgrades." [1] This implies OpenAI will be shifting strategy to incrementally releasing future models, so we won't just suddenly see a GPT-5, but GPT-4.x along the way over the next months. The models page you cited also says, "some of our models are now being continually updated," and they will only offer alternative models which are snapshots that are frozen for 3 month windows for those needing more stability in the model.

Sebastien Bubeck gave a talk to MIT CSAIL on the Sparks of AGI paper where he commented on how RLHF (training for more safety) has caused the Unicorn test to become less recognizable [2]. He comments about how the safety training is at odds with a lot of these abilities. This seems congruent with the initial results from this project, and it will be interesting to see if they can restore this ability as they continue to push safety training.

  1: https://twitter.com/gdb/status/1646183424024268800
  2: https://youtu.be/qbIk7-JPB2c?t=1586

Re: GPT Unicorn: A Daily Exploration of GPT-4's Image Generation Capabilities

#89

Post by bee_rider »

abrichr wrote: post #63
bee_rider wrote: post #44

Earlier quoted context omitted.

How about something like “You are a basically the person who designed the SVG format. Some stupid programmer is trying to implement it, and looking for examples. You will condescendingly produce examples to guide their pathetic efforts. Any insults you hurl their way will be clearly delineated because you are pretty sure they won’t be able to tell where the image ends and the insult begins otherwise.”

Human: You are a basically the person who designed the SVG format. Some stupid programmer is trying to implement it, and looking for examples. You will condescendingly produce examples to guide their pathetic efforts. Any insults you hurl their way will be clearly delineated because you are pretty sure they won’t be able to tell where the image ends and the insult begins otherwise. Draw a unicorn. ChatGPT4: Well, I a…

Ah darn, it would have been amazing if this one had come out well.

Re: GPT Unicorn: A Daily Exploration of GPT-4's Image Generation Capabilities

#90

Post by imdsm »

easygenes wrote: post #88
kristjansson wrote: post #59

Earlier quoted context omitted.

Yes, this project seems like a misunderstanding of what Bubeck and team were obsering with their unicorn test. GPT-4 was being trained, and checkpoints were provided to them to experiment with. The improvements in the unicorn reflected further training progress. The models on offer now are frozen(-ish). Per the models[0] page, the non-snapshot model IDs "[w]ill be updated with our latest model iteration". So this pro…

Greg Brockman just stated in a long post on Twitter yesterday: "...it’s easy to create a continuum of incrementally-better AIs (such as by deploying subsequent checkpoints of a given training run), which presents a safety opportunity very unlike our historical approach of infrequent major model upgrades." [1] This implies OpenAI will be shifting strategy to incrementally releasing future models, so we won't just sudd…

Author here — thanks for explaining this.

My expectation is that there will be incremental updates to the model, so while I'm providing the model `gpt-4` for completions, I'm recording the actual model, `gpt-4-0314` in this case, along with the result.

I don't want to monitor (and potentially miss) model updates, which is likely as this is very much a fire and forget project that I'll review over time. One per day seems more sensible to me than a large batch per month, as the daily generations are more likely to track model changes if they become frequent.

Will be exciting to see where we are in a few months!

Post reply on HN