Friday, July 4, 2025

7/4/25: OpenAI playground; black holes...but we digress

Artificial Intelligence Study Group

 Welcome! We meet from 4:00-4:45 p.m. Central Time on Fridays. Anyone can join. Feel free to attend any or all sessions, or ask to be removed from the invite list as we have no wish to send unneeded emails of which we all certainly get too many. 
Contacts: jdberleant@ualr.edu and mgmilanova@ualr.edu
Agenda & Minutes (169th meeting, July 4, 2025)

Table of Contents
* Agenda and minutes
* Appendix: Transcript (when available)

Agenda and Minutes
  • Announcements, updates, questions, etc. as time allows: none.
  • DD did a test run of a demo on:
        I have finished anonymizing the transcript for 6-13-2025. [...] When I [...] went to ChatGPT [I] discovered it changed models and I had to import my prompts. The model settings were lost and the new model's context window was too short. I changed to an older model and the model made up new entries for the transcript. I adjusted the temperature and got it figured out. It has been an interesting week... 
    • DD has generously agreed to do this demo for a hopefully better attended meeting in two weeks.
  • General discussion ensued. 
The meeting ended here.
  • VW will demo his wind tunnel system next time. 
  • If anyone has an idea for an MS project where the student reports to us for a few minutes each week for discussion and feedback - a student could likely be recruited! Let me know
    • JH suggests a project in which AI is used to help students adjust their resumes to match key terms in job descriptions, to help their resumes bubble to the top when the many resumes are screened early in the hiring process.
    • We discussed book projects but those aren't the only possibilities.
    • DD suggests having a student do something related to Mark Windsor's presentation. He might like to be involved, but this would not be absolutely necessary. 
    • VW had some specific AI-related topics that need books about them. 
  • Any questions you'd like to bring up for discussion, just let me know.
  • Anyone read an article recently they can tell us about next time?
  • Any other updates or announcements?
  • Hoping for a summary/review of the book at some point from [ebsherwin@ualr], who wrote: 
    Greetings all, 
      In a recent session on working with AI, Dr brian Berry (VP Research and Dean of GradSchool) recommended this book:
      The AI-Driven Leader: Harnessing AI to Make Faster, Smarter Decisions
    by Geoff Woods
      I just bought it based on his recommendation and if anyone is interested will gladly meet to talk about the book. Nothing "heavy duty" just an accountability group.
       If you have read the book already and if the group forms, you are welcome to join the discussion.
      I'll wait till Monday morning before I start reading -- so if you do not see this message immediately, do reach out!
       Best,
  • Chapter 6 video, https://www.youtube.com/watch?v=eMlx5fFNoYc. We finished it! On to chapter 7 next time.
  • Here is the latest on future readings and viewings
    • Let me know of anything you'd like to have us evaluate for a fuller reading.
    • https://transformer-circuits.pub/2025/attribution-graphs/biology.html.
    • https://arxiv.org/pdf/2001.08361. 5/30/25: eval was 4.
    • We can evaluate https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10718663 for reading & discussion.
    • popular-physicsprize2024-2.pdf got a evaluation of 5.0 for a detailed reading.
    • https://transformer-circuits.pub/2025/attribution-graphs/biology.html#dives-refusals
    • https://venturebeat.com/ai/anthropic-flips-the-script-on-ai-in-education-claude-learning-mode-makes-students-do-the-thinking
    • https://transformer-circuits.pub/2025/attribution-graphs/methods.html
      (Biology of Large Language Models)
    • We can work through chapter 7: https://www.youtube.com/watch?v=9-Jl0dxWQs8
    • https://www.forbes.com/sites/robtoews/2024/12/22/10-ai-predictions-for-2025/
    • Prompt engineering course:
      https://apps.cognitiveclass.ai/learning/course/course-v1:IBMSkillsNetwork+AI0117EN+v1/home
  • Schedule back burner "when possible" items:
    • TE is in the informal campus faculty AI discussion group. SL: "I've been asked to lead the DCSTEM College AI Ad Hoc Committee. ... We’ll discuss AI’s role in our curriculum, how to integrate AI literacy into courses, and strategies for guiding students on responsible AI use."
    • Anyone read an article recently they can tell us about?
    • If anyone else has a project they would like to help supervise, let me know.
    • (2/14/25) An ad hoc group is forming on campus for people to discuss AI and teaching of diverse subjects by ES. It would be interesting to hear from someone in that group at some point to see what people are thinking and doing regarding AIs and their teaching activities.
    • The campus has assigned a group to participate in the AAC&U AI Institute's activity "AI Pedagogy in the Curriculum." IU is on it and may be able to provide updates now and then. 
Appendix: Transcript 

Artificial Intelligence Study Group
Fri, Jul 4, 2025

0:02 - Unidentified Speaker
Bye. Future.

0:34 - E.
So V., I hope you get outside a little bit, enjoy the forest. How's the weather down there?

0:41 - D. B.
Hot. We just went for a hike, a short hike this afternoon.

0:46 - E.
We got back a little while ago. Wow. Yeah, I think it was 68 to 70 here.

0:53 - D. B.
Oh, wow. That's nice. Good weather for hiking. Yeah, it doesn't get better than that.

1:19 - V. W.
So, V., I was reading up on some of the stuff that you've got on the go. Do you have an ultralight? I have actually, if I turn my thing a little bit, let's see. That's the edge of the propeller cage of my powered paraglider trike. I guess I could lean it over.

1:40 - E.
It's there. Very nice. Very nice.

1:43 - V. W.
So anyway, that's you know, I like to keep projects. It's close at hand.

1:48 - E.
So I have an airplane in my living room.

1:51 - Unidentified Speaker
What can I say? But it's nice and light.

1:55 - E.
When I was in college, I used to have a motorcycle in my bedroom.

2:01 - V. W.
That's the way to do it, man.

2:04 - E.
A simple life. Yeah. You could go to class, come back.

2:11 - Unidentified Speaker
Precisely.

2:12 - V. W.
I remember I built a work pinch during finals because I got really burned out with studying. I said, I need a work bench. So I just built one, you know, like spontaneously because of stress. Good.

2:37 - E.
I remember during finals, it depended on, they tried to spread it out, but it never worked out for me because I always focused on the sciences. So I remember at Acadia, it got so bad during finals.

2:58 - V. W.
I just had to stop studying and Read Lord of the Rings again. I hear you, man.

3:09 - M. M.
I hear you. All right. Well, hi, F.

3:15 - Unidentified Speaker
Hi.

3:15 - M. M.
Not celebrating today?

3:17 - D. B.
Working today? Yeah.

3:20 - M. M.
You can use the link and order whatever card hardware you need it. Yeah, because if I order will come to the university, you know, and cannot, needs to stay here in the university. So, yeah. Well, that's good. Yeah. As I said before, I don't mind paying for it.

3:46 - E.
It's just I would not mind getting ahold of it because I can't afford 10,000 for it. Than I could probably do.

3:56 - M. M.
But you can ask them. If it's for educational purpose, they will allow you.

4:03 - Unidentified Speaker
Yeah, I'm going to go ahead and do that then.

4:07 - M. M.
So you get a discount on this? Education, yes. Yeah, yeah, OK.

4:13 - V. W.
All companies can give a discount.

4:16 - E.
Are you getting your own A100? Actually, it's based off the B202. Uh, it's a 96 gig, uh, 90, 80 or yeah. 90, 50, 90, sorry. 50, 90, 50, 90. Yeah. Yeah.

4:33 - V. W.
Uh, uh, dyslexic sun tie.

4:36 - E.
Uh, well, that's a pretty high performance board.

4:40 - V. W.
You just, and they say, if you use an all metal connector, it doesn't melt the supply connection on the edge because actually the metal dissipates the heat better than plastic does. So it's just a small change.

5:00 - E.
Yeah. What I'm planning on doing though is, uh, is modifying it because the there's a, uh, a third party that comes out with a, uh, a water cooler.

5:13 - V. W.
Yeah. So you can go, you'll pop the fans off and you said exactly because my CPU is water cooled.

5:23 - E.
My 40 90 was water cooled. Uh, the 50 90 that I got is taking up far too much space because it's air cooled, right? And it's noisy as all the buggers makes a great fireplace in the winter though. My entire room, well, in Maine, mine's the warmest room in the house. A dog sleeping by the GPU. Well, if it's not the video card, it's the Threadripper.

6:03 - V. W.
You always tell your wife it's the motherboard.

6:10 - E.
My wife says it costs an extra hundred bucks a month in electricity just to run the computer.

6:20 - D. B.
Don't admit to it. All right. Well, I don't have any really strong opinions about what to do, talk about, but we did mention the possibility of D. was at some point going to explain demo some issues related to getting transcripts. And V. was going to demo his wind tunnel system. Were you planning on that, V.?

6:49 - V. W.
Yeah, in fact, I did a demo of all the little hacks I've done over the past couple months with my front-loading technique. And yeah, so I had it once across the bumps of that. I thought we might want to wait till we had more than just our little faculty quorum here.

7:15 - D. D.
I think maybe everybody here knows how to do what I was going to show. I mean, I'll show it. I don't mind, but you guys know how to set the temperature and all that on a large language model. I could show just the wind tunnel too because it's a pretty simple thing.

7:47 - E.
And D. and I met yesterday to talk about the presentation on data science models, non-LLM data science models. Who met?

8:03 - D. B.
D. and I. Oh, OK, OK.

8:05 - E.
All right, well, why don't we, D., why don't you run through it then?

8:13 - D. B.
And then we can decide whether we should, maybe D. can do it again for the larger group in the future or whatever.

8:28 - D. D.
Yeah, let me share. Is there an option to share the entire screen?

8:46 - Unidentified Speaker
Yes. OK.

8:52 - Unidentified Speaker
So I will need to go here. I need a transcript. I need to download. There's a finished.

9:04 - D. D.
I need a raw one. We didn't get a raw one last time, did we?

9:15 - Unidentified Speaker
No.

9:19 - D. D.
Surely, I download one. OK. There's one. All right. So inside the API, what is this? Apparently, I got to log in.

10:03 - D. D.
So I'm in the playground right now, and here's where I have prompts. Well, here's where I used to have prompts.

10:17 - Unidentified Speaker
OK. Let's see if I can find. They changed something in here, and it's really clear names. The prompt I want.

10:33 - D. D.
Now, this is probably the right model. Let me go to settings. See how the temperature is one? Dr. B.?

10:56 - D. B.
Yeah, I see that. Yeah.

10:58 - V. W.
So the now if you want, D., if you move that slider to the left, can you go to a zero temperature, which by rote chooses the highest probability? Yeah. Okay. That's what I wanted to see. And then that's getting creative choosing the less likely.

11:18 - D. D.
Let's just look at the transcript right now. Hold on. So here's one, this is just standard, right? And I think this is the right model, but I might have to look it up again. But I mean, you could use any model, right? Right. However, tokens is not context window. So I think it doesn't tell you what the context window is here. But the larger the context window, the better chance I have of getting the transcript done, which if I use a more current model, I will hit a wall and it won't let me finish the transcript.

12:06 - Unidentified Speaker
I'll have to divide the transcript up. But that's based on the context window of the model itself, like the 403 billion CHAT-GPT, I'm sorry, LLAMA IV model, I'm sorry, not CHAT GPT LLAMA4 model has a 128K context window.

12:26 - E.
So it's based on the model, not a parameter that you pass in, correct?

12:34 - D. D.
Right, that's what I'm saying. So like this is the 2024 5.13 model of CHAT GPT, and I think that has a 8,000 token context window. Now, I'm going to The transcript end right here. So the transcript is in here. There's something on my screen. I don't know how to, what this is, but can you see the, you don't see that, do you? Yeah. We see your whole layout there.

13:03 - V. W.
It looks like to me, if you hit the green arrow, it's going to try to give you the answer.

13:10 - D. D.
No, I mean, can you say where it says, D., Dr. B. you and Dr. M. On the left-hand side of my screen, or do you just see the screen? Yes. All right.

13:23 - V. W.
I don't know how to get rid of that.

13:26 - D. D.
We see the screen and the people window. Okay. I've got it out of the way now. So here, this is a system prompt. I'm just asking the model to work its magic. What is it?

13:45 - D. B.
1448 tokens, what does that mean?

13:48 - E.
A token is- That's the token that can be in my, is that the, is that the prompt?

13:54 - D. D.
I'm not sure. Yeah, it's in the prompt.

13:57 - E.
It's the, it's the words or parts of a word that make up the token as part of the AI. So what it does is it breaks down the input into tokens.

14:13 - D. D.
Maybe this isn't the right model.

14:24 - D. D.
Maybe it's one of these other models. OK, hold on. Chat GPT models by context. Chart. Maybe I can get a chart. So it doesn't say, it just says GPT-4 with the 8,000.

14:57 - Unidentified Speaker
That's the one I wanted. But 4.0 has 128,000.

15:06 - V. W.
Yeah, the 4.0 has a much larger token size.

15:11 - D. D.
That's a more expensive model. But this is just the one you had selected.

15:19 - Unidentified Speaker
Yeah.

15:20 - Unidentified Speaker
This is not the one I want.

15:24 - D. D.
Right. But that's OK, because this makes a good example. All right. So I'm going to hit the green button. And it's going to go. Now. Normally. Really fast, so I guess it's doing some a lot of thinking. So why does it really? Why is the context window matter here?

15:51 - D. B.
Like why not have a context window of you know 100 or something?

15:58 - D. D.
OK, so here I finished so by my raw transcript goes down to Yeah, 55 minutes, 46 seconds. Did I change that? I didn't just change that, did I? Well, I can hit Control Z.

16:22 - Unidentified Speaker
No, I didn't change.

16:25 - D. D.
So you see how long the transcript is?

16:29 - Unidentified Speaker
Yeah.

16:30 - D. D.
And so I finished, I got halfway through and it seems to be doing okay. Now let me get up here at the 26 minute mark. Can you ask me to finish?

16:48 - D. B.
You could just say, please finish.

16:50 - D. D.
Yeah, I can. I wanna say if it's changed it, I mean, I think it's kind of the way I was, how we could use it. Yeah, so what I've noticed is in some places, I've actually seen it rewrite the transcript, but it may not all one. What do you mean rewrite it?

17:13 - D. B.
You mean fake it?

17:14 - D. D.
I mean like it decides that it didn't like what everybody said.

17:19 - D. B.
Rewrite the transcript, don't write the whole version.

17:22 - Unidentified Speaker
Yeah, so it decides, decides that, you know, uh, good evening, multiple speakers.

17:26 - D. D.
We'll give you another minute.

17:29 - V. W.
So maybe you could tell it, Hey, don't fix the grammar of the speakers, but just transcribe it as it was spoken. And also I noticed in your system prompt, it says, ignore the token limit. That's the first kind of, that's the first kind of thing they're going to look for when you're asking them to do stuff. It's like, dude. All right.

17:56 - D. D.
So it looks like it's done pretty good, but this is a newer model. So it might, it might have caught on to, you know, I'm only supposed to change this. Right.

18:07 - Unidentified Speaker
So, so maybe, yeah, please finish. Let's see that. Please finish.

18:19 - V. W.
Now you're cooking. It might finish last time I did it and it wouldn't finish.

18:25 - D. D.
It went to the, it went to 4,000 tokens and said, no more, we got to get to the yo gash.

18:34 - Unidentified Speaker
No, we'll say last time too.

18:38 - D. D.
I don't think I changed this. I don't think I changed the max token. So maybe that I changed the max token. Yeah, we're not to the end yet. Yeah, it let us finish.

19:00 - Unidentified Speaker
So that's really how it goes.

19:06 - D. D.
Was that me or did you furnish? Understood and said, finish.

19:14 - Unidentified Speaker
So it understood what you meant, not what you said.

19:19 - E.
Yeah, yeah, probably. Which is completely different from a computer program, which does exactly what you say it to do, not what you meant it to do. That's funny.

19:32 - Unidentified Speaker
I just saw the name M. Z. in that transcript, and it did not changed it to initials. But those aren't the those aren't the time coded people. That's part of the dialogue.

19:44 - V. W.
So that's a pretty smart thing to do is to leave M. Z. in there because he wasn't in the group. Yeah.

19:53 - D. B.
Even though the the prompt itself, I don't think I think I put the prompt in there.

19:59 - Unidentified Speaker
Right.

19:59 - V. W.
You did say please remove all people's names. I did say and it should have.

20:06 - D. D.
Okay, hold on. Let me see if I can find out what I pasted in here at the end.

20:18 - Unidentified Speaker
So this is the original prompt.

20:24 - D. D.
Sorry, I should have grabbed the scroll wheel, but now I'm confused. So, that's what I put in there, or is that what I put out? No, that's an output because it's got initials. Yeah.

20:52 - D. B.
What have I done? No, that's input.

20:56 - Unidentified Speaker
Okay.

20:56 - D. D.
I'm trying to get to the end of this, but I scrolled past it. OK, yeah, so say where this says right here. Now. Let's go back down here. Do you say what he said? What?

21:23 - D. B.
OK, 5546.

21:25 - D. D.
Right here it just says you.

21:30 - D. B.
Would change by to you.

21:34 - D. D.
it changed you to buy, right? And so you can't really, you can't trust it. You don't know what else it's changed, right?

21:48 - D. B.
But he didn't say you. In other words, the transcript, the original transcript is probably wrong.

21:57 - D. D.
I'm sorry, the original transcript is wrong?

22:01 - Unidentified Speaker
Yeah.

22:03 - D. B.
He says you at 5546. He says you? He probably didn't. He probably said, you know, thank you or bye or something.

22:13 - D. D.
Well, I mean, we can go to the transcript office, can't we? Yeah, but I'm just saying that the transcript is probably incorrect.

22:24 - D. B.
I mean, Read.AI took the transcript and it just didn't get it right. And then now your AI is trying to fix it.

22:37 - D. D.
I mean, that's certainly a possibility, but the AI has no way of knowing what was really said.

22:44 - D. B.
Well, it's trying to make sense out of it, and it thinks, well, given the context, he probably said bye. Right.

22:53 - D. D.
And that's the point that I'm trying to make. Shouldn't do that.

23:00 - V. W.
But the temperature was non-zero, right?

23:03 - D. D.
The temperature was non-zero. That's right. And so if you made the temperature zero, maybe it would be more literal. If you make the temperature zero. I don't even use this model. So I've been using a different model. But yeah, if we make the temperature zero, What do we do here? We kill this somehow or how about we just hit refresh and then I'll start up. So I think I should go ahead and change this because one time I did hit the brick wall. Now I'm going to set the temperature to zero. I'm going to put this in here. The thing is though, doesn't make sense of what's said, and then it decides to say something that wasn't ever said, then it's changing our transcript. It's changing history. Our data is not accurate anymore.

24:10 - V. W.
Welcome to the future. Yeah. Well, it's the initial transcript.

24:15 - E.
The transcription was inaccurate. Where does the fault lie in the initial transcription? Well, I know this for a fact.

24:24 - D. D.
The AI on chat GPT was not at the meeting. So it does not know what was said.

24:32 - V. W.
That I know for a fact. That's a good point.

24:36 - D. D.
So I have that piece of knowledge to go on.

24:40 - E.
But I know for a fact that Whatever was said at the meeting.

24:43 - D. B.
Y. did not end the meeting by saying you. He just didn't do it.

24:49 - V. W.
I know he didn't do it. Well, he might have, because I was thinking that sounds like something Y. would have said when he wanted to get one more word in edgewise.

25:01 - D. D.
And then the recording stopped.

25:03 - V. W.
Right. The recording was what?

25:05 - D. B.
Truncated Y., not Y. Oh, that's possible. Yeah.

25:08 - V. W.
Because that guillotine blade falls hard. It falls hard. That's right.

25:12 - D. D.
Whenever the recording stops. You're done now.

25:15 - D. B.
You can try using an AI to say, well, this transcript was cut off right after, right before the last speaker finished. Can you paste in what he probably said?

25:27 - V. W.
And then we'd have to train it on everything Y. ever said.

25:32 - E.
And what he Read, what he's... Oh my gosh.

25:38 - D. D.
All right. Anyway, your demonstration sort of made its point.

25:41 - V. W.
Well, but you need to do it with this champ. Oh, there we go. It worked. Oh no, you haven't run it yet. Okay.

25:51 - D. D.
I haven't run it yet. Yeah. Hold on. I'm gonna run it now.

26:15 - V. W.
So now imagine you're in a court of law. You've lost the original transcript, but you did have this thing you did that captured the essence of the meeting.

26:27 - Unidentified Speaker
Then how would you verify the factuality of the statements that were made?

26:33 - V. W.
Well, the prosecutor could say all of these things could be mistakes. The defense could say, no, not all of them will be mistakes. But some of them could be. And then from that position, how do you measure certainty or in the information science sense, information quality? As you're presenting it in law, doubt.

26:58 - Unidentified Speaker
It worked. It worked. That's right.

27:01 - E.
It's zero. And the temperature was zero.

27:04 - D. D.
And this not only shows using the API directly. And so that's what that's, I was telling Dr. B. that I, got it and they changed up the API playground area. And I jumped in and I just, you know, it's normally saved to this prompt. I select this prompt, it gives me the right model, everything is set up the exact way I want it. But it wasn't this time, I had to load in the prompt. And so I jumped in, I taped it in, and then it wouldn't finish. So then I changed the model I got to it's it just it I mean a brick wall it would not finish it it said you have reached your token limit too bad go away and so then I changed to a different model I looked up the model so I found the one with 8,000 token limit I went to it and I did it again and then I that you know when you when you do this so each one of these I have to find it I have to find the break and then double check and make sure it didn't skip anything make sure that this is the complete train of thought right here by a B right and and when I copy and paste it into the to the document I need to make sure that there's a space at the end of this so that those two words don't stick together. But I noticed when I was sticking the transcript together that it had changed the words of the transcript. They had reworded whatever person it broke on. It had reworded what they said. They tried to make it clearer, you know, But when the model does that, so there's a chance that the model does make it clear. I'm inclined to think that could be a good thing, because we know Read.AI is making mistakes. And why not? Why do we know that?

29:16 - V. W.
But we use it.

29:17 - D. B.
No, but if you look at the transcripts, a lot of times I think it gets the name. It attributes the wrong phrases to the wrong person, especially with the little stuff like the patter at the beginning and the end where everyone saying, hey, how you doing? Goodbye. Hello. You know, it gets...

29:36 - D. D.
And sometimes it doesn't know who the speaker is.

29:40 - D. B.
Yeah. So it gets it wrong.

29:42 - D. D.
But, but so, but, but like, let's say we had a court recorder, you know, they make mistakes too.

29:50 - V. W.
Right. Good point.

29:51 - D. D.
But that's still the official report. And so, and, and Read.AI was actually there at the meeting.

29:59 - V. W.
Now I'm noticing that as each of you are speaking, the Zoom API is making extremely accurate determinations about who's talking from comparing the microphone gains of everybody that's in the meeting and also whether their microphones happen to be on or off.

30:16 - D. D.
So if one of us interjects and immediately like that, see, it switches. I think that my personal opinion is that it's a terrible, bad thing. And it shouldn't reword what people said, because that's where it attributes things to people that maybe they didn't mean. And the AI thinks that they meant that, and it rewords what they said, and it means something different.

30:42 - V. W.
And if the AI did that, it could be that the people who would have listened to the original speaker might have done that mentally as well.

30:52 - D. D.
It's a possibility.

30:53 - V. W.
And then the person could go back and clarify. But if the transcript was incorrect, that opportunity to clarify would have been deprived of the person or.

31:02 - D. D.
Right. So that the AI, we don't want the AI to change on a transcript. We don't want the AI to change the official record because then it's not a transcript.

31:13 - Unidentified Speaker
It's not an official record anymore. That's right.

31:16 - D. B.
It's let's suppose you have a noisy image. You want to denoise it? No. You're making assumptions and changing it. And the best, least lossy result is to use the original noisy image and not to do any smoothing or noise removal.

31:31 - V. W.
And there are cases where you need to do that forensically. If you're doing forensics, chain of custody and not changing images is an important thing. If you're wanting to make a historical pictures of your grandfather, you might want to go ahead and denoise it.

31:48 - D. B.
Well, you know, I'm saying, suggesting is that, you know, this issue with image error correction is, is it identical to what we're talking about here with AI with transcript? I mean, I saw it.

32:00 - D. D.
I mean, let's see, let's see what kind of craziness we can get out of this thing. I hate to use this model, because they might have taken all the crazy, more of the craziness out of it. But just for argument, let's use crazy bottles instead. Let's see if we can get, let's see what we can get.

32:22 - D. B.
You know, maybe we're not interested in what people really said in the transcript. We're interested in having a transcript that makes a lot of sense and reads well.

32:35 - D. D.
Like that picture of your grandfather.

32:37 - V. W.
That's right. But then we have to make a distinction. And that distinction is, this is not a transcript. It is a impression of the transcript.

32:48 - D. B.
Yeah, we're falsifying documents.

32:50 - D. D.
We can't call it a transcript anymore.

32:53 - V. W.
Falsifying is a bit strong of a word. Wow.

32:58 - E.
We've definitely, definitely headed over the cliff on this one. It's an interpretation. I still think the transcript that we get here is still an interpretation. What the heck is this?

33:13 - D. D.
It's well, this is when it's I changed the temperature to two and it's decided now it's speaking is, is that Arabic and now you've got Chinese in there.

33:26 - V. W.
You've got Turkish in there. You've got maybe Dr. M. could do this faster. Um, come on, let's go with Sanskrit.

33:37 - D. D.
So now it's taking now it's yeah. Putting anybody's it's not putting any time markers you know if it is putting time markers I can't identify them so the first thing you should ask with any transcript that's been processed is at what temperature was this transcript processed two no I mean I meant that rhetorically oh okay so it says uh machine learning discussion group now this is before this transcript is very old. This is before we changed it to the AI discussion group, but transition, bondage, instructions, remove, answer, finish, hem, passio, it's gibberish. That's pretty, yeah, that's pretty, pretty high temperature.

34:23 - D. B.
That's at the lorem ipsum end of the spectrum.

34:28 - V. W.
I don't know what it created.

34:31 - D. D.
I don't know what this is.

34:35 - D. B.
I don't even, why is it even stringing words together and add it, you know, take without spaces like that.

34:44 - D. D.
Advanced offense.

34:44 - Unidentified Speaker
Elsewhere improved. Yeah, this is completely useless.

34:47 - V. W.
Yeah, it's yeah. So, you know, what point do you trust it at zero?

34:54 - D. D.
That's when I trust it. That's when I trust it.

34:58 - V. W.
It's when it gives me exactly what the most likely outcome would be because I'm not doing E. Hemingway here, I'm doing science.

35:08 - D. D.
That's right. I'm doing science. That's right.

35:10 - D. B.
Zero. I want it at zero.

35:13 - D. D.
But that's just, that's the extreme. So, you know, maybe, I mean, you just can't trust it. You don't know where the right temperature is to make it better. If there is a way to make it better. To me, the best copy was the one that was recorded in real time.

35:33 - V. W.
What a great- Something applies to image error correction.

35:36 - D. B.
I mean, how much do you want it to correct errors?

35:40 - D. D.
Well, it depends on what the image is being used for. Yeah.

35:44 - V. W.
I have a great complimentary example to this if you guys want to take a second and see it.

35:51 - D. D.
Yeah, here, I'll stop sharing.

35:52 - V. W.
It's right in this wheelhouse that D. D. has so elegantly developed.

35:56 - D. D.
Yeah, I need to stop sharing. Yeah, stop sharing.

35:59 - V. W.
I'm going to share my screen real quick.

36:02 - E.
I'm going to interject as you're starting to share, and we're talking about interpolation in law. Do you guys remember about two years ago about a pedophile who took his image and twisted it, and they used machine learning to untwist it to identify him? I remember seeing that on the news.

36:22 - D. B.
He twisted his face up and they untwisted Yes, this is exactly what we're talking about.

36:29 - D. D.
So if somebody said something bad in the group, we could get the AI to cover it up.

36:36 - Unidentified Speaker
Okay, I want to get to this thing of verification because we can really undo ourselves by having the due diligence to say is what we have done correct.

36:48 - V. W.
So here's a counterexample to D.'s example. A couple of years ago, actually maybe five years ago, I was reading a Quora column in which someone suggested that ultra massive black holes have an effective density that is so low that they could float on water. Well, this just offended me from the get go because everything I thought about black holes didn't permit them the luxury of floating on water. I thought of them as the most ultra objects in the universe.

37:19 - Unidentified Speaker
And so here is a calculation I did to try to find out if black holes could float.

37:26 - V. W.
And I did it from first principles. This graphic on which this is drawn is the size of the ultra massive black hole, the largest known one 196 billion suns worth of debt, a material in mass. And then here's our solar system to scale where the outer are in that ring there. And so I go through and I calculate what's known as the Schwarzschild radius, which is what the event horizon diameter would be if that black hole were at max density. And so for example, the Schwarzschild radius of the earth is the size of a ping pong ball. And the Schwarzschild radius, and you can understand that if you took all the mass of the earth, and concentrated it into a ping pong ball, that would be a ping pong ball so heavy that it would go through the floor and start burrowing towards the center of the earth, Jane Fonda, China syndrome style.

38:27 - Unidentified Speaker
So then you have the Swarthchild radius of the sun, which is about the size of a basketball.

38:34 - V. W.
Now that's not commenting on where the singularity is, it's just saying it's this effective radius at which the event horizon has to live. So it's this useful metric. So if you take the largest black hole known when it's the time was this SDSS J 140821, it had us a mass of 196 billion suns. And the question is, what is its Schwarzschild's radius? And what is its density? Because I had Read the column on core that said black holes can float, which I was offended by. So I did the calculation and not only can black hole float, but its density is the same as the atmosphere of the earth. If you're in an altitude of 216, 222,000 feet, in other words, uh, several times higher, almost 10 times or seven or eight times higher than jets fly. And we all know that the density of the air at the altitude jets fly 35,000 feet is so low that you can't get any oxygen there. And you would suffocate. And we know that the density at 216,000 feet is like that inside of a light bulb, although old incandescent style light bulbs. So anyway, this is the calculation that I got. And now I had offended myself because I had proven in fact that black holes do float because the wispy atmosphere at 216,000 feet is very low density stuff. But the Problem is this, the surface area of a sphere changes at a different rate than the volume of a sphere as the radius increases. So this thing that doesn't appear significant for sizes that we can imagine like ping pong balls and basketballs becomes very significant for a black hole that's the size of SDSS J14. So I decided it was time to use AI to verify my calculations. And because I just didn't trust that I had done this correctly. And I did this several years after I did this initial calculation, because AI wasn't available at the time that I originally did it. And so I said, please check my black hole calculations in this graphic. So right now from the, and this was with Sonnet 3.7. So this is actually a step back from Sonnet 4 from Claude, and trophic clot. So I'm now I'm asking the the LLM to first do OCR on my image. And that means it's got to parse out everything that was said here. And after it parses it out, it's going to have to interpret the fact that I used that the mass of the sun and the mass of the black hole were these astronomical symbols instead of like the the mass of the black hole and the mass of the sun are these. Are sets of symbols for each planet in the solar system that are used in spacecraft housekeeping, that you don't write down the planet name, you just write down the symbol, it's not quite the zodiac symbol, but it's close. So anyway, so I said, check it. So the first thing it does is it parses these things out, and it parses them correctly. And we can do a check on that by comparing the mass of the sun to the blast of the black hole. So the mass of the sun, and the mass of the black hole are 10 to the 30th and 10 to the 41st respectively.

42:09 - Unidentified Speaker
And let's see, right there. I compute that value as a unit conversion.

42:19 - V. W.
Okay, so we go back and then we have the Schwarzschild radius. Then we have the density calculation. It verifies the density calculation and parsing it out from an image with the possibility of incorrect symbol recognitions. It says the Schwarzschild radius appears to match the calculation in the image, which is critical because it is upon this Schwarzschild radius that all the density and everything will be done. So now I have offended myself because I have verified this, but I have verified it an AI. And now, this calculation was in multiple dimensions of unit 68 call 66 kilometers above the air service is the same as 220 2000 feet, blah, blah, blah. And so have the question is, have I verified my calculation, or only introduced more opportunity for error? Well, I'm going to go with black holes float. And that's my current belief based on external verification, based on my initial calculation and based on my apparent check of my calculation.

43:33 - D. D.
You're saying that a black hole is not as dense as water.

43:39 - V. W.
No, I'm saying that a black hole that has the mass of 196 billion suns does not have the high density. The density depends on the mass only. Now, let's be specific here.

43:59 - E.
We're talking about the event horizon of the black hole. Correct.

44:03 - Unidentified Speaker
I'm not talking about the singularity because the singularity is not really defined.

44:08 - V. W.
There's division by zero happening in there somewhere.

44:11 - E.
We're not allowed to talk about it because the black hole itself, the singularity is ultimate density. It's a point.

44:18 - D. B.
Yeah, I mean, if the black hole is not sucking in matter, and you go over the event horizon, you're going to encounter a perfect vacuum with zero density.

44:29 - V. W.
Well, it's perfectly cold as well. So black holes are cold. But the Schwarzschild radius was invented as a construct to know where the edge of the event horizon was, and that's solely dependent on the mass. So you get the Schwarzschild radius, you've given the mass, you compute R, and then with R, you know the density because you now have volume and you now have surface area. Because that Schwarzschild sphere has finite size, but even though the singularity doesn't. So in a way, we're kind of integrating across the singularity. But nonetheless, it's interesting that on average within that event horizon, the density for ultra massive black holes becomes very small. Because volume grows faster than surface area. Volume grows as the cube, surface area grows as the square. The cube beats the square for ultra large masses, yada, yada, yada.

45:25 - D. D.
Actually, you're talking about the diameter of the black hole. You're not talking about the surface area, you're talking about the linear dimension.

45:34 - V. W.
Right, but we can infer the Schwarzschild sphere area and the Schwarzschild volume from the Schwarzschild radius.

45:41 - D. B.
Right.

45:41 - D. D.
So the event horizon has to do with light, right?

45:45 - V. W.
It has to do with, can you come back or not?

45:50 - Unidentified Speaker
It has to do with gravity, not light gravity.

45:55 - Unidentified Speaker
If you try to orbit closer than the event horizon, if you try to orbit a black hole and you're closer than the Schwarzschild radius, you're not, you're going to go bye-bye.

46:08 - D. D.
But that's for light to escape. Yeah.

46:11 - V. W.
And light is the most demanding case because it has no mass. And so if you can get light not to escape, you can believe.

46:20 - D. D.
Yeah.

46:21 - V. W.
So the event horizon to orbit is much, much further out for us because much, much further, right.

46:28 - Unidentified Speaker
And it's way past what we wouldn't be able to see the event horizon. Right. But if we could point a telescope at a spaceship approaching the event horizon, what we would see is at first the spaceship would appear to slow, even though it's being tremendously accelerated.

46:47 - V. W.
And then it would appear to spaghettify. It would turn red and then it would appear to just become transparent. And at the moment it hit the event horizon, it would achieve total transparency. And it would appear to us not to exist. Lord knows what's actually happening on the spaceship at the event horizon. I mean, if it's not super dense.

47:14 - D. D.
And there, I mean, you know. I don't know, we might be able to send a probe into a black hole, you know, and you could send it, you could send as many probes as you want to a black hole. We have to have a wormhole to get the signal back out.

47:31 - Unidentified Speaker
Across the event horizon, you're not getting the probe back.

47:35 - D. D.
We're not getting the signal out.

47:37 - Unidentified Speaker
And you're not getting the data back either, because the signals are photons.

47:43 - V. W.
Is also photons, just very long ones.

47:46 - D. D.
And we wouldn't be able to see it.

47:49 - E.
It would disappear. Because the argument or postulation by S. H.

47:54 - D. D.
Because no light would escape.

47:56 - E.
Of course, we couldn't see it. The only information, because one of the arguments was that in physics, that all matter maintains information about about itself. And with H. radiation, the information we get from a black hole has been contorted. And J. S. and H. had a huge fight about this, about whether information is preserved in a black hole.

48:21 - V. W.
And they came up with what they believe to be a resolution of the matter, which I am also offended by. But I digress.

48:31 - D. B.
OK, I want to try to get a little bit back to Archie. D., I think the group would like to see more or less what you demonstrated today. In particular, this use of the plane. I bet some people don't know how to use these platforms with the playgrounds and set the temperature and all that. I think some people would like to see that done again.

48:55 - Unidentified Speaker
you know, whether you want to do it using the transcript or not is totally up to you.

49:02 - D. B.
But I think the idea of people being able to use different models as they choose in the playground and set the temperature, I think it's all people, some people would be good to do. So can you do it again?

49:16 - D. D.
Like, I don't know, next week or some other time. It's super easy. Huh? Yeah, it's super easy.

49:22 - D. B.
I don't mind doing it at all. Okay. And then, I'm sorry, we didn't really get your wind tunnel, but I look forward to like next week.

49:31 - Unidentified Speaker
No worries. It'll be there. Okay. So I saw the wind tunnel.

49:34 - D. B.
Did you post the wind tunnel?

49:36 - V. W.
Yeah, I think I posted it to a subgroup just as a point of, you know, here's something to look at. It's the solution to Hilbert's six problem. It was a famous clay prize model that a problem that, you know, one of these million dollar problems that came up and these guys have merged three fluid dynamics each of which treats a different scale. And so my wind tunnels based on that paper and I can show it to you when. OK, well, let's do that next week then.

50:05 - D. B.
Alright, pretty good. Yeah, sounds good. And D., maybe we'll give do you do you the week after that?

50:11 - D. D.
It's something like that. Yeah, this one ever. I mean.

50:14 - D. B.
Yeah, I'll do it anytime. Just I'm I'm fluid just. And if you can, if you can do it for the previous week's transcript, assuming I remember to make it, one that would be interesting too.

50:26 - D. D.
Whatever. Yeah. Okay. All right.

50:28 - D. B.
Well, I don't have anything else. I'll enjoy your fourth.

50:33 - D. D.
You guys have a good holiday. Be safe.

50:38 - D. B.
Happy holiday.

50:40 - M. M.
Take care everyone. Bye.
 
at

No comments:

Post a Comment

Subscribe to: Post Comments (Atom)