Neil Slater

general recent conversations

The Singularity

General discussion for ai.stackexchange.com. See ai.stackexcha...

Neil Slater

Mar 3 11:42

I'm hoping that enough human operators are still going to be driven to learning and care to become experts, despite the laziness drive

Neil Slater

Mar 3 11:41

It's easy in principle to use the new bots as teachers and expert partners. But there's a strong drive to use them as expert servants instead

Neil Slater

Mar 3 11:39

Sadly I think the expert text in various Stack Exchanges has been a great source of training data for systems like this, which are often going to be a superior user experience to anything SE can produce. I'm not necessarily sad for the demise of SE, neither do I feel bad that my knowledge has been "harvested". What I wonder is where the next generations of experts will come from.

Neil Slater

Mar 3 11:36

I think I have seen a couple. The "tell" is reference numbers in the text.

Neil Slater

Mar 3 11:35

Now Deep Research is available on OpenAI's plus plan, I think we can expect more sophisticated cut& paste answers.

Neil Slater

May 19, 2024 15:32

I don't think it cause a problem on the site specifically, but I may have to take a break from answering their questions, as the benefits from spelling out every single detail to my own understanding on the subject are very much diminishing returns here. I.e. I'm not really getting anything from answering so many of them

Neil Slater

May 19, 2024 15:29

@nbro if you are referring to whom I think, then I don't think they are a student on any official course. Instead they are a slow methodical learner, and have a strong need to have every small detail spelled out.

Neil Slater

Mar 12, 2024 13:19

I think with RL any manipulation that impacts or uses the agent's internal model will categorise close to the existing categories. However attacks against the environment or agent sensors may not fit so well.

Neil Slater

Mar 12, 2024 13:14

I don't really have any opinion. The safest bet is to describe what was done to disrupt the agent. These categories are broad descriptions from a practical engineering standpoint, and are not mathematically precise.

Neil Slater

Mar 8, 2024 13:39

As more and more companies use "AI" as the marketing term for generative text and media models (which is not incorrect, just very specific), we get more and more questions about how to access and use such models - e.g. https://ai.stackexchange.com/questions/45060/do-you-know-of-any-service-with-an-api-that-takes-image-and-text-prompt-as-input

It's been proposed before, but what is the appetite for renaming AI SE to e.g. AI Theory or some other term that academic and industry insiders would recognise as pertaining to their expertise, and not to prompters, integrators etc?

Neil Slater

Dec 22, 2023 09:52

Weirdly my interest in reinforcement learning started around 10 years ago. Unrelated to the paper, but of course it's existence means I could study that too, as it followed on naturally from reading Sutton & Barto.

Neil Slater

Dec 17, 2023 18:43

Now we have another use cut&pasting LLM output to answer questions at a high rate with questionable quality answers - ai.stackexchange.com/users/5763/brian-odonnell - this user should know better, not quite sure what to do

Neil Slater

Dec 11, 2023 14:06

I think the problem with that is that I would get flagged for misuing my votes after 2 or 3 on the same user

Neil Slater

Dec 11, 2023 14:05

Is th ebest I can do vbote them down?\

Neil Slater

Dec 11, 2023 14:05

@nbro, @Mithical I just came here to report same user. They've posted 19 obviously LLM-generated answers in the last day. They are mostly semi-correct nonsense that doesn't really answer the question as posed.

Discussion between hanugm and Neil Sl…

Imported from a comment discussion on ai.stackexchange.com/que...

Neil Slater

Apr 25, 2024 08:07

Whilst "what is the potential gamut of a diffusion model" is more philosophical

Neil Slater

Apr 25, 2024 08:05

However, this is more about practical control and expectations on a project?

Neil Slater

Apr 25, 2024 08:05

I was kind of hoping you were asking about the limits/gamut of the model - i.e. just how much variance can you expect given you can vary the initial noise seed. I think that is an interesting question - for which I don't have a great answer.

Neil Slater

Apr 25, 2024 08:02

It's pretty common to set a seed then increment it to generate e.g. 10 images that you can then compare when changing other things such as the prompt or one of teh many parameters, to be able to see more clearly what effect they are having

Neil Slater

Apr 25, 2024 08:01

Yes, pretty sure if you set the seed, it should be the same image each time. Although not familiar with where/how it would be set in that code.

Neil Slater

Apr 25, 2024 07:58

It is possible to inject more noise vectors. Stable DIffusion is open source. If you used something like ComfyUI, or scripts, then you could add noise to almost any of the stages

Neil Slater

Apr 25, 2024 07:57

OK.

Neil Slater

Apr 25, 2024 07:56

OK. Got there. There are none. You can fix the seed and see this, the same image is generated every time (there are some caveats, because different GPUs may behave differently). If that's not what you mean, please explain?

Neil Slater

Apr 25, 2024 07:56

Still not quite there. Are you interested in that it changes at all, or in the practical limits of what those changes can be? Are you asking if there are other factors beyond noise and settings. Or are you asking "can a cat prompt generate a dog sometimes"?

Neil Slater

Apr 25, 2024 07:56

I'm still not understanding what you are getting at with "diversity". Do you mean just that the image is different - i.e. the pixels are different colours, or are you referring to the limits of what differences can be realised (e.g. if you ran the same settings with random noise a million times with the prompt "cat", would one of the cats be surfing, or doing something else unusual that typically you would need to prompt for)?

Neil Slater

Apr 25, 2024 07:56

Are you asking about parameters other than the text prompt. E.g. choice of sampler, cfg scale etc? Or can those be considered fixed, like the prompt. Also are you asking about gamut of possible outputs related to what the different noise values produce, or about the image varying at all?

Discussion on answer by Neil Slater: …

Imported from a comment discussion on ai.stackexchange.com/que...

Neil Slater

Sep 7, 2023 14:55

If you asked a human to translate themselves 100 metres to the left with efficient use of energy, with one step, it is completely impossible. The human needs to make those individual interim steps. A lot of LLM "logic" is similar - it needs to follow a form in which the logic is embedded, and doesn't exist as a function outside of that

Neil Slater

Sep 7, 2023 14:53

An analogy here might be how human legs work for locomotion. A lot of the "smartness" of 2-legged walking is encoded in the structure of the legs, which bits are elastic, how the leg will simply move under gravity without needing any conscious control

Neil Slater

Sep 7, 2023 14:51

For better math and logic answers it is usually better to have teh AI roll slowly through its explanation. The logic seems to be at least partially encoded in how the long-form text is structured, and not as a higher-level function that anticipates or calculates the answer.

Neil Slater

Sep 7, 2023 14:49

You can never easily tell when the AI is using any kind of internal model that would apply. Simply changing the formatting or grammar of the question, or requesting specific terms can cause completely different parts of the model to engage with completing the text.

Neil Slater

Sep 7, 2023 12:31

Thanks. Lunchbreak over, I won't respond for a few hours.

Neil Slater

Sep 7, 2023 12:28

which is to say very little

Neil Slater

Sep 7, 2023 12:28

I have engineer-level spelling ability

Neil Slater

Sep 7, 2023 12:27

if you want to play with something rawer, but don't want to feel spyed upon,

Neil Slater

Sep 7, 2023 12:27

You could use an open-source one like Llama

Neil Slater

Sep 7, 2023 12:26

but it seems that the LLM pre-pcakcaged as a chat assistant is very popular consumer choice

Neil Slater

Sep 7, 2023 12:25

the RL fine tuning obfuscates the "raw" LLM, but it's still mostly there and can do far more than be a chat-based assistant

Neil Slater

Sep 7, 2023 12:25

people have used this to introspect a few things, such as the system prompt (although OpenAI also published the system prompt)

Neil Slater

Sep 7, 2023 12:24

So yes, you can de-program the chatbot with a bit of careful prompting

Neil Slater

Sep 7, 2023 12:23

but the raw LLMs don't have this issue, you can make them take all sorts of creative roles

Neil Slater

Sep 7, 2023 12:23

With ChatGPT that is harder because it is also fine-tuned to be a chatbot

Neil Slater

Sep 7, 2023 12:23

Yes, or you can just change the system prompt that makes it a chatbot

Neil Slater

Sep 7, 2023 12:22

I did simiular with "What is the best Beatles album?" as a demo to family one time

Neil Slater

Sep 7, 2023 12:22

You can change the system prompt

Neil Slater

Sep 7, 2023 12:20

That is what most people are doing. A lot of "prompt engineers" around, and some no doubt have discovered something real and valuable. LLMs have created a huge new virtual space

Neil Slater

Sep 7, 2023 12:18

OK, I understand your reticence, but your desire to be correct is fighting against that, and you need to figure out which is more important.

You cannot really go around making claims to secret knowledge on a Q&A site, that you won't also share on that Q&A site as e.g. an answer to the question.

Neil Slater

Sep 7, 2023 12:15

I'd rather simply not know

Neil Slater

Sep 7, 2023 12:15

I still do. I expect to, after being "taught", But this NDA-like paranoia is creeping me out

Neil Slater

Sep 7, 2023 12:14

I also do not want any responsiblity to be "sharing a secret" with you. It does nothing for me, and I really don't see it benefits you either

Neil Slater

Sep 7, 2023 12:13

I don't feel a need to be taught here. Thanks for offering, but it feels weird so not taking part.