Artificial Intelligence Study Group
|
Contacts: jdberleant@ualr.edu and mgmilanova@ualr.edu
Agenda & Minutes (173rd meeting, Aug. 8, 2025)
Table of Contents
* Agenda and minutes
* Appendix: Transcript (when available)
Agenda and Minutes
- Announcements, updates, questions, etc. as time allows: none.
- Next time at about 4:15 (when he gets home!) DD has generously agreed to do a demo on local LLMs.
- EG and DD are working on slides surveying different ML models.
- VW will demo his wind tunnel system at some point.
- ES will provide a review of The AI-Driven Leader: Harnessing AI to Make Faster, Smarter Decisions, by Geoff Woods, when we request it.
- Join us for a thought-provoking lecture and book signing with renowned economist and King’s College London professor Daniel Susskind as part of the CBHHS Research Symposium.
Thursday, September 4, 2:00 p.m., UA Little Rock, University
Theatre – Campus conversation
Friday, September 5, 2:00 p.m., UA Little Rock, University
Theatre – Campus and community conversation
Susskind, a leading voice on the future of work and technology, will explore how artificial intelligence is reshaping the workplace and how we can harness its potential to work smarter. Don’t miss this opportunity to engage with one of today’s most influential thinkers on AI, economics, and the future of our professions.
- 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.
- JC suggested: social media are using AI to decide what to present to them, the notorious "algorithms." Suggestion: a social media cockpit from which users can say what sorts of things they want. Screen scrape the user's feeds from social media outputs to find the right stuff. Might overlap with COSMOS. Project could be adapted to either tech-savvy CS or application-oriented IS or IQ students.
- We discussed book projects but those aren't the only possibilities.
- VW had some specific AI-related topics that need books about them.
- 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.
- markwindsorr@atlas-research.io writes on 7/14/2025:
Our research PDF processing and text-to-notebook workflows are now in beta and ready for you to try.
You can now:
- Upload research papers (PDF) or paste in an arXiv link and get executable notebooks
- Generate notebook workflows from text prompts
- Run everything directly in our shared Jupyter environment
This is an early beta, so expect some rough edges - but we're excited to get your feedback on what's working and what needs improvement.
Best, Mark
P.S. Found a bug or have suggestions? Hit reply - we read every response during beta.
Log In Here: https://atlas-research.io - 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?
- 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.
- 7/25/25: eval was 4.5 (over 4 people). https://transformer-circuits.pub/2025/attribution-graphs/biology.html.
- https://arxiv.org/pdf/2001.08361. 5/30/25: eval was 4. 7/25/25: vote was 2.5.
- We can evaluate https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=10718663 for reading & discussion. 7/25/25: vote was 3.25 over 4 people.
- Evaluation was 4.4 (6 people) o 8/8/25: https://transformer-circuits.pub/2025/attribution-graphs/biology.html#dives-refusals
- Evaluation was 3.87 on 8/8/25 (6 people voted): https://venturebeat.com/ai/anthropic-flips-the-script-on-ai-in-education-claude-learning-mode-makes-students-do-the-thinking
- Evaluation was 3.5 by 6 people on 8/8/25: Put the following into an AI and interact - ask it to summarize, etc.: Towards Monosemanticity: Decomposing Language Models With Dictionary Learning (https://transformer-circuits.pub/2023/monosemantic-features/index.html); Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., Turner, N., Anil, C., Denison, C., Askell, A., Lasenby, R., Wu, Y., Kravec, S., Schiefer, N., Maxwell, T., Joseph, N., Hatfield-Dodds, Z., Tamkin, A., Nguyen, K., McLean, B., Burke, J.E., Hume, T., Carter, S., Henighan, T. and Olah, C., 2023. Transformer Circuits Thread.
- Evaluation was 3.75 by 6 people on 8/8/25 for: Use the same process as above but on another article.
- Https://www.nobelprize.org/uploads/2024/10/popular-physicsprize2024-2.pdf once got a evaluation of 5.0 for a detailed reading.
- 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 - Neural Networks, Deep Learning: The basics of neural networks, and the math behind how they learn, https://www.3blue1brown.com/topics/neural-networks
- LangChain free tutorial,https://www.youtube.com/@LangChain/videos
- Chapter 6 recommends material by Andrej Karpathy, https://www.youtube.com/@AndrejKarpathy/videos for learning more.
- Chapter 6 recommends material by Chris Olah, https://www.youtube.com/results?search_query=chris+olah
- Chapter 6 recommended https://www.youtube.com/c/VCubingX for relevant material, in particular https://www.youtube.com/watch?v=1il-s4mgNdI
- Chapter 6 recommended Art of the Problem, in particular https://www.youtube.com/watch?v=OFS90-FX6pg
- LLMs and the singularity: https://philpapers.org/go.pl?id=ISHLLM&u=https%3A%2F%2Fphilpapers.org%2Farchive%2FISHLLM.pdf (summarized at: https://poe.com/s/WuYyhuciNwlFuSR0SVEt). 6/7/24: vote was 4 3/7. We read the abstract. We could start it any time. We could even spend some time on this and some time on something else in the same meeting.
- 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, Aug 8, 2025
0:33 - Unidentified Speaker
Oh.
0:34 - R. S.
you Hi, everyone.
1:57 - Unidentified Speaker
Hello.
1:59 - R. S.
Hello.
2:00 - M. M.
Hi. I was sitting here trying to figure out why there was no sound.
2:06 - Unidentified Speaker
Meeting is now being done telepathically.
2:11 - D. B.
There is no sound because nobody was saying anything.
2:18 - Unidentified Speaker
Well, at least I guess I guess we can go ahead and get started. Yeah.
2:26 - M. M.
Does anybody have any announcements or updates or anything, questions?
2:33 - M. M.
I think that we need to invite more students when the school start.
2:39 - D. B.
Yeah, I think that's sort of why we started the group was because the students, it's a way to get students involved.
2:47 - M. M.
We have to do this. Yes, we have to at least all of our master and PhD students to make presentations and, you know, to learn what they do, what the colleagues are doing. So, yeah, definitely. So I'm just come back from my trips, so I'm happy to be back. You're welcome back.
3:11 - Unidentified Speaker
Yes, D.
3:12 - M. M.
So, yeah, we.
3:15 - D. B.
Well, during the semester, we we typically had, you know, an average, I'd say, about about an average of 11 attendees. And as the summer has gone on, it's declined. But when the things start picking up in the fall, I think attendance should go up and we can recruit students and so on. Yes. Yes, sir.
3:40 - Unidentified Speaker
OK.
3:44 - D. B.
So last week, D. did a demo for us. And he's generously agreed to do other demo, just not today particularly. But D., do you want to do it next week?
4:03 - D. B.
You're muted.
4:05 - D.
So I mean, yeah, I guess I can. I'll be doing them from work. So I think that I would have everything that I need Yeah, I should be able to do it.
4:21 - D. B.
OK, yeah, it's up to you. I'll go ahead and
4:24 - D.
Well, so I can't do my large language model one, though, as I can't. I don't know how to hook up to I can hook up to it at home.
4:36 - D. B.
We're going to do the local ones like downloaded language model.
4:42 - D.
Yeah, I guess I don't know that I can run anything on my laptop, though. Oh, OK, well, we don't have next week. I mean, we can just I wonder, I wonder if, if I could do it, but like it for 15 maybe.
4:59 - D. B.
So like, we could start the meeting while you're going home or something like that.
5:02 - D.
Yeah, because so I would get off at 345 I suspect. Okay, and then I have to go through traffic. So, you know, well, we can start a meeting and then as soon as you're ready, to do it, you log on a little bit late, and when you're ready, we'll do it.
5:20 - D. B.
How's that sound? Let's do it.
5:21 - D.
I like it. Okay.
5:23 - Y.’s iPhone
So D., which LLM do you plan to download?
5:29 - D.
Well, so what I've downloaded three models from Olama. Quan Co and DeepSeek one, I think. I think I can handle like the 32 billion parameter models. So they're small.
5:55 - D.
They're pretty good at making lists and stuff. I mean, they're not, but they're not full featured models. So for instance, if I was gonna take one of those models and try to anonymize the transcript, they can't handle it.
6:22 - D. B.
They cannot.
6:23 - D.
They don't have what it takes. But when I say that, there might be a way that I could do it.
6:34 - Y.’s iPhone
Like small pieces of it at a time.
6:37 - D.
rebuild it, like build a, but go ahead, sir.
6:43 - Y.’s iPhone
And when you say small and big, I'm assuming you're talking about the size, do you know how big they are in terms of size?
6:52 - D.
Maybe one of them, I think two of them are 32 billion parameters. And one of them might be 30 four billion parameters? Is that?
7:05 - Y.’s iPhone
Yeah, in terms of parameters, yes. But in terms of actual size, how much size is it picking up and what kind of processing or RAM is it expecting?
7:17 - D.
I don't have that information, how its size there are that I don't think there may be a few gigs, maybe 10 days. I mean, it's not that it's not that big. All right, D.
7:33 - Unidentified Speaker
Yes. Would it help if I made myself available for my system?
7:39 - D.
You want to, you want to, you've got my cell phone number.
7:44 - Unidentified Speaker
Give me a call. I don't have your cell phone.
7:47 - D.
What kind of talk is that? You've got it.
7:49 - E. G.
You've called me on it. I've never called you. Yeah. Okay. Then connect on discord. Yeah.
7:58 - D. B.
Why don't you just tell us what the number is, and then... Yeah, I'll put it in chat here.
8:03 - D.
Put it in chat, that way we can send it to our telemarketing buddies.
8:08 - E. G.
Here you go.
8:10 - D.
This guy, he likes to buy.
8:13 - Y.’s iPhone
Yeah, I just Read today the Attorney General talking about the fraud that is coming for Dole and a lot of AI is being used to those texts and other content. It's quite interesting. And I think just last week, S. A. from OpenAI, he spoke about fraud in banking. But today I Read about fraud for toll payment and other things and how AI is smartly being used. To actually autorespond and create the website and everything. So it's quite interesting.
9:03 - D. B.
All right. Well, so you got E.'s number. If you lose it, let me know. I've got it.
9:09 - D.
I'm putting it in my phone right now. I called you, so you should I see that.
9:23 - D. B.
That was me. So I'll just keep that here on the list for now.
9:31 - D. B.
V. has not been here recently, but he did volunteer to And then Professor S., whenever I ask, she's willing to do it. Whenever I ask her to do it, she'll review this book that she Read over the summer, or she started an ad hoc discussion group about the book. I don't know.
9:51 - D. B.
if she got a group out of it or not. But anyway, she Read the book and she's willing to tell us about it. So.
9:56 - Y.’s iPhone
Can I ask one more question to D.? Yes.
10:00 - D.
I mean, I don't know.
10:03 - Y.’s iPhone
It's not my meeting technically here.
10:06 - Unidentified Speaker
That's why I asked D. permission if I can ask on the previous topic. D., is that okay? Yeah, go ahead.
10:13 - Y.’s iPhone
So when you are demoing or when you did testing in the past, did you also check that once you download and you start running those models on the backend, so can you really get off from internet and run and does it work? And the reason I'm asking is, or you download, so you have some parameters locally but for any reason have you checked like actually then it goes and goes to the model available on the respective clouds and checks or process something and gets back if it's not able to find the answer. So are you doing any testing or confirmation that once it's local it's not actually going off your machine and processing done anything like that.
11:15 - D.
Okay, so so what so what I've done is so I built a I built a computer specifically for this task. It's got to to 3060 12 gigabyte graphics cards and basically mid level processor and a whole bunch of RAM and a large hard drive. And so I've got it networked so that what I do is I use the, what's that thing called? Open UI, I think.
12:06 - D.
It's a free program that you download But what I do, so on my machine, where I have less VRAM, I just access it through my browser. So that is in my internal. That's my, it's a network, but it's not the internet. It's my home network. So I access it. Now I have not went and tried to block the internet or block access to the internet on either one of the devices.
12:43 - D.
Because I'd have to, well, I need the router, you know how you got the big box, it's got the modem and the router and, and all that in it. And, and so So if I when you go to hook up to a network. So we're both on the, you know, on our, on our own network, we're both on the home network, but the home network has internet access to the same box that's managing the home network. So I haven't tested it at all, as I would need to figure out how to test it, I'd have to go like, leave the router plugged in, but unhook the internet cable from the cable box or the fiber box, I guess is the appropriate way to say that, and then test it. But I'm confident because of the internet technology courses that I've had in school, I just that I am accessing that machine directly from my machine and there's tools that I put on there so that I can choose to either allow it to go online and access the internet or choose not to allow it to. But I think the real drawbacks are context window. All right.
14:18 - D. B.
Well, we can get more into that next time if needed.
14:24 - D. B.
Another announcement is there's going to be a guest lecture they're someone visit physically from King's College in London.
14:40 - D. B.
He's not a computer Thought-provoking lecture and book science tech, technical guy, he's an economist. Oh, he's a voice on the future of work in technology. Anyway, that'll be next two sessions in early September, one on Thursday at two on Friday at two. University Theater. You're supposed to register. I don't know what happened if you just showed up without registering. You'll be okay. The theater is pretty big place. University Theater. So yeah, it's another opportunity.
15:27 - D. B.
As I mentioned last time, I'm hoping to recruit. Oh, yeah, I did talk to L. G. a few days ago. He'd started the book project last semester, and something happened, and he had to leave for the semester. But he's going to be back, and he wants to do it for real this time. So now that we know how to do book projects, we'll have some to do it, hopefully.
15:55 - D. B.
And if there's anyone else who wants to suggest a project for grad students, we could try to recruit, I can try to recruit someone for that. Okay.
16:06 - D. B.
I've So last time we reviewed some readings and today we can review some more.
16:15 - Unidentified Speaker
That brings us, let's see, we did this one. Oh, no, we didn't do that.
16:18 - D. B.
We evaluated this one a long time ago. Let's evaluate it again. So I'm going to bring it up. We'll Read a couple paragraphs through the abstract, and then we'll decide if we want to Read it in more detail. Let me.
16:35 - D. B.
Let's see.
16:58 - D. B.
OK, so as you know, G. H. and J. H. won the Nobel Prize in physics, no less, in 2024 for their work on AI.
17:09 - D. B.
So what was this that I was found? I guess I didn't get the right link or something.
17:21 - D. B.
Popular it. Physics not Prize 2024-2. Is This
17:30 - D. B.
I think it was his speech.
17:46 - Unidentified Speaker
No, that's not the right link.
17:52 - D. B.
I'm sorry, I didn't take get the right link to So anyway, I going thought we here. Were
17:59 - D. B.
I'm clearly, I just don't have the right link. We So can now. For we'll maybe look that at it next skip time. Let's take a look at another paper.
18:12 - Unidentified Speaker
Looks like an academic, well, I don't know what this is. Someone suggested more detail, and it is biology of a large language model, whatever that means. So let's take a look at a little bit of it.
18:41 - D. B.
here and see if we want to address it.
18:44 - Unidentified Speaker
Okay, so let's take a look at this pair of and S. to thank you for this. Us.
19:23 - Unidentified Speaker
Any comments or questions?
19:29 - D.
Yeah, I'm going to Read the next paragraph.
19:33 - D. B.
All right, well nobody had any So let's go ahead and Read the next paragraph.
19:55 - Unidentified Speaker
Any thoughts or questions?
19:59 - D.
I think it's interesting that you're looking at it. So let's study this large language model like it's a biological organism. Because we don't know how it works.
20:15 - E. G.
Sometimes, I mean like the biological, you could look at a microscope, you can see what's going on.
20:26 - E. G.
times LLMs are black boxes where we're putting information in and it's coming out the other end.
20:39 - D.
So it's really a lot like a human brain which to a lot of people is still a black box.
20:52 - D.
I think at the end of the day the black box part of a large language model is any special training techniques that they did and didn't share with anybody. So, you know, we say that, you know, we assume now that all these language models are still being, you know, trained through the transformer, but are they still, you know, do they have, have they developed techniques that they haven't written any papers about that they're not telling them they don't want to leak it out because they think they might have an advantage. But at the one that when the models done, it's just series of weights, right? That's what that's what it is at the end of the day, it's just weights. And then they might have some more proprietary like logic that they'll put, you know, in between those weights and us are the output, right? So that they can, you know, be sure not to say any But once the models train, you know, the only offensive or so, or, you know, so it is that would be black box would be that logic that they use to filter answers. Because a weight is a weight is a weight, right? Villanova Lee?
22:31 - D. B.
Am I thinking about that right? Yeah, I mean, the building blocks- It's what's called an emergent property of these systems. They're basically weights and neurons. Neurons are pretty simple little computational units. But together, when trained, somehow they develop this organization where they can now write books and things like that.
23:02 - D. B.
Because it's only predicting the next word, right?
23:12 - D.
And I think by looking at the weights, Yeah.
23:16 - Unidentified Speaker
And if that's the case, every model, once you get the weights, you can duplicate it.
23:22 - E. G.
That's right. If all of the training at that point goes away, I take chat GPT five that just came out, get all of its weights, and then I've got a copy of it.
23:40 - D.
Well, so that's kind of what they're talking about. The seek they use that they use the chat algorithm to kind of you know do do kind of like this if they didn't treat it like a biological you know uh organism they treated it like uh a conversational you know they went and picked chat gpt force brain or they're accused of it I don't know if they did or not They were accused of picking its brain to try to guess its weights. But yeah, so once you get the weights, then you've got a model. That's right. That's correct. That's correct. OK.
24:32 - Unidentified Speaker
F., did you have a comment?
24:35 - D. B.
No, no, it's fine.
24:36 - M. M.
It's correct. Yes, we have the model. And can modify somehow. Yeah. All right.
24:44 - D. B.
Do you want to Read another paragraph or do you want to vote on this one?
24:48 - D.
I'm interested to see what kind of craziness they're going to come up with.
24:51 - M. M.
Yeah, I agree. So Read another paragraph?
24:56 - D. B.
Please. Sure.
25:00 - E. G.
All right.
25:03 - M. M.
I like the figures too.
25:06 - D. B.
Yeah, well, if we decide we want to Read the whole thing, we'll definitely check everything. I can't figure out how to even highlight anything nowadays.
25:17 - M. M.
Something happened.
25:18 - D.
Yeah, all right, here we go.
25:38 - Unidentified Speaker
Good. No.
25:41 - D.
Oh, sorry.
26:00 - Unidentified Speaker
Comments or questions?
26:02 - D.
So what are we looking at?
26:08 - D.
References.
26:10 - E. G.
Yeah.
26:12 - Unidentified Speaker
Are Maybe.
26:18 - Unidentified Speaker
We If any of these look really interesting, we can I can put them on the list to evaluate and interpret.
26:27 - D.
How to encode the real.
26:31 - M. M.
This is number two.
26:37 - D. B.
No.
26:40 - D.
Let me go down to the bottom. Is this? 1, 2, No, no. I want the references.
27:08 - M. M.
Yeah, you're there.
27:12 - Unidentified Speaker
Hmm.
27:39 - D.
Well, I mean, that's a whole, that's a bunch of good papers right there.
27:46 - D. B.
Well, I mean, research topics. So there's a lot of stuff out there. I mean, I don't know how effective it is yet.
27:52 - D.
But yeah, I would be interested in slapping those into a large language model and getting a submarine to see if it's worth reading.
28:01 - D. B.
Number two looks interesting.
28:04 - E. G.
Decomposed using LLMs with dictionary learning.
28:13 - D.
I kind of like, D., We could try that.
28:30 - D. B.
Because that would give I'll put that on the list. I'll say it's not exactly reading something paragraph by paragraph. It's not exactly viewing a video one minute by one minute. But it is something we could think about doing. So I will put what you suggested in the list of readings, potential readings.
29:02 - D.
When you go inside a language model, and you start trying to get at its weights, you're significantly violating your terms of use if it's proprietary.
29:18 - D.
But I don't know about CLAWD3, what they're doing with CLAWD3 to have permission But just because we can figure out ways to do really cool things, it's kind of like hacking somebody and taking their code. It's not very nice.
29:52 - Unidentified Speaker
I've invested. Significant amount of money to these things.
29:58 - D.
That free data that they downloaded off the internet.
30:00 - D. B.
All right, I can't get the link. I've got it. I've just put that idea in the list of readings, potential readings. All right, and let's go back to this one, which we just So the question is, do we want to Read that in more detail, Read it paragraph by paragraph, Read more or less the whole thing or a lot of it and talk about it? And what I'd like to ask you to do then is the usual, if you put in the chat your vote for how much you want to Read it in detail. Five is you definitely want to Read it. One is you definitely don't. Three is you're not sure. Two and four are leaning against or leaning toward.
30:53 - R. S.
If I remember correctly, the exact title was something about the biology of large models. Is that correct?
31:02 - D. B.
Yeah, it's right exactly.
31:07 - R. S.
So are they really talking about biology or that they're using that word biology in a different context?
31:15 - D. B.
Well, they were trying to draw an analogy with biology in one of the paragraphs we Yeah, there it's, it has no, it's, it's just similar.
31:25 - D.
Yeah. It has no ties to actual biology.
31:33 - D.
All right.
31:33 - D. B.
Well, I thought we had the starts of the board here.
31:37 - Unidentified Speaker
Okay. I'm looking for your chat thing. Okay.
31:40 - R. S.
Uh, So again, one, two, five.
31:43 - D. B.
OK.
31:43 - D.
I don't want to get overly excited.
31:58 - E. G.
D., you sent that to me, not to group. I did?
32:03 - D.
Yeah. Any other votes?
32:05 - D. B.
We've got two votes, so I'll give We got four so far. There's more people than that here.
32:26 - Unidentified Speaker
Yeah.
32:28 - M. M.
Oh yeah, F., how about you?
32:30 - D. B.
I'm okay.
32:31 - M. M.
We can Read this, four. Four, okay.
32:34 - D. B.
All right. Let's see what we got. The average is 4.6.
32:59 - D. B.
Okay, so I'm going to go back here.
33:18 - D. B.
Where were we?
33:29 - Unidentified Speaker
I can't even remember where we are. Here we are.
33:49 - Unidentified Speaker
All right. So we'll do a few more of these. We'll finish this up today, maybe.
33:52 - D. B.
And then we'll pick one, the highest one, to Read in more detail. In the future. Let's take a look at this one. These might be a little, a little old. In some cases, because they've been sitting here for a while, but you're welcome to anyone's welcome to suggest another one and we'll vote on it. In the future, anytime. Okay.
34:20 - M. M.
Okay, here's the title.
34:33 - D. B.
Yeah, yeah, I can.
34:34 - Unidentified Speaker
Any so far?
34:44 - R. S.
He's trying to do the complete opposite of AI by having the students do the thinking. Something like that.
34:59 - Unidentified Speaker
All right here's the snip it and take a look at that.
35:27 - Unidentified Speaker
Any comments?
35:34 - Unidentified Speaker
My question is, well, how How are they going to do it?
35:47 - Unidentified Speaker
All right, let's Read this.
36:04 - Unidentified Speaker
Comments or questions?
36:09 - E. G.
What is the date of this?
36:13 - D. B.
Let's take a look.
36:17 - D.
What was that AI that they did?
36:20 - D. B.
I'm sorry, I've got to put this in full screen mode so I can...
36:24 - D.
I'm sorry, it was not an AI really, it was a chat program that they made a long time ago. April 7th.
36:37 - D. B.
maybe in the 60s, that was really good at just, you know, rephrasing your question.
36:45 - D.
They turned a large language model into a chat program from the 60s.
36:54 - D. B.
Well, there was, Eliza was the first chat bot.
36:57 - D.
A lot, yeah.
37:00 - Unidentified Speaker
Eliza, exactly.
37:02 - M. M.
It's doing this kind of stuff.
37:06 - D.
But it's not really Socratic at that point.
37:10 - E. G.
The of the Socratic method is to ask probing questions, not just turn around the question, but ask probing questions so you develop the thinking process, the critical thinking on your own rather than be told the answer.
37:29 - Unidentified Speaker
Yeah, Eliza was it asks questions reflectively.
37:35 - D. B.
So it was like intended to mimic a certain school of psychotherapy where you sort of, or certain school of interacting with a person who has, you know, who's seeking answers, whatever, by turning their statements into questions. So like Eliza would, if someone said to Eliza, you know, I'm not feeling well today, Eliza might respond, you're not feeling well.
38:03 - M. M.
That would elicit more from the person.
38:09 - D. B.
In this Socratic questioning. It's like what Socrates did. He teach people by asking them questions, make them think. Where is it?
38:28 - D. B.
When students ask questions, Claude responds with Socratic questioning. How would you approach the problem? What evidence? Things like that. It doesn't seem like it's that hard for a chatbot to do something like that, but normally when you interact with chat GPT or Claude, it doesn't do that. Now, are saying it can. T. E. sometimes comes to these meetings. As I understand it, he's planning on doing a dissertation on using AIs for Socratic questioning. I want to use it in my classes. Make it students to give homework questions where they have to go to it. And ask it to interact with them by Socratic questioning.
39:27 - D. B.
I think this is a whole crisis. And now, you know, Claude, the anthropic company, is coming out with this learning mode, which does it. I mean, you can still do it without, you can just go to a normal Claude and just say, please use Socratic questioning. To help me understand the following passage. You give it a passage, and it'll start giving you Socratic questioning. That's interesting.
39:56 - D.
So I don't think learning mode is that different from normal plot.
39:59 - D. B.
It's just probably a modification to the prompt or something, a background prompt.
40:10 - E. G.
Actually, then, what I'm going to do over the next couple of weeks is I have access to ChatGPT, the latest ChatGPT, and the latest Claude is, I'm going to use Claude and through a Socratic method, ask it to help me better understand allele-specific expression.
40:33 - D. B.
Yeah, you could try to, you could ask it to, you know, you were talking about earlier about asking it to summarize the paper. You can do that, but you can also ask it to teach you about the paper using Socratic questioning and it'll do it. Maybe not very well, but you can always make it do it better by improving your prompt. If it doesn't do a good job teaching you using Socratic questioning, that's not its problem. That's your fault because you didn't prompt it properly.
41:11 - Unidentified Speaker
I put on the list of readings what D. suggested that we use in real time, ask ChatGPT to summarize a paper.
41:21 - D. B.
Well, we can ask it to do more than summarize. We can ask it to teach us the content using Socratic questioning or anything Anything like that we like,
41:36 - D. B.
Any other comments on Let's Read another paragraph.
41:44 - Unidentified Speaker
and decide it and then vote on it.
42:10 - D. B.
Any comments or questions?
42:20 - D. B.
Well, we already Read a lot of this. It's a pretty short news article, but we can still vote on whether we want to Read the whole thing.
42:27 - D. B.
Are we ready to vote?
42:37 - D. B.
All right, go ahead and type in your vote for whether you want to Read this in more detail together or not.
42:51 - D.
It would be more interesting to test it. More interesting to what? Test it.
43:09 - D.
Take a fine thought and thinking mode and check it out and see what it does.
43:15 - D. B.
No, I'm going to actually do that.
43:16 - E. G.
I'm going to say, preface it, as a tutor in biostatistics, based on these attached papers, using the Socratic method helped me better understand the concepts being presented in these papers.
43:37 - M. M.
I like it, I like it in this way, yes.
43:45 - M. M.
We need to test it.
43:46 - D.
We need to test it, that's right.
43:49 - Unidentified Speaker
Yeah.
43:53 - M. M.
Somebody think of something really, really hard.
43:59 - M. M.
Something super complicated.
44:01 - D.
All right.
44:04 - D. B.
Yeah, we'll help us figure it out.
44:15 - D. B.
Okay, that came out to a total of 3.87.
44:27 - Unidentified Speaker
people voted? All right, well, let's go ahead with the one that D. suggested.
44:52 - D. B.
I think what you suggested was we put this paper into chat TPT and ask it to summarize and interact with it about that paper.
45:08 - D. B.
So let's see if I can't figure out how to get to the link. Oh, here we go. Here's the link. I've got to go back. I just don't want to lose it. All right.
45:40 - Unidentified Speaker
So here's the paper. Whoops. Here's the paper.
45:47 - Unidentified Speaker
It's pretty long.
45:48 - D. B.
I'm going to go ahead and copy paste the whole thing into What?
46:33 - Unidentified Speaker
All right, now what should I do? What did I do wrong?
46:36 - E. G.
Attach the, oh, you can, okay. I pasted it. Conversation is 25% over the length limit. You too big.
46:48 - Unidentified Speaker
Yeah.
46:53 - E. G.
Is that a paid version? No. Would you like me to do it? I've got the paid version.
47:00 - D. B.
Well, let me just try it on ChattyT first for ease, and then you can do it.
47:18 - Unidentified Speaker
Too long.
47:19 - E. G.
OK, well, if you want to share your screen, Sure.
47:26 - E. G.
I'll try it with, uh, open AI first.
47:34 - Unidentified Speaker
I think you're going to stop sharing first.
47:43 - E. G.
Yeah.
47:43 - Unidentified Speaker
Okay.
47:50 - E. G.
What's the, uh, What's the link to the
48:02 - D. B.
I'll put it in the chat in a moment as soon as I figure out how.
48:11 - E. G.
Tell me what you want as a prompt.
48:17 - D. B.
All right, I just put the URL in the chat. Did you get it? Yep.
48:35 - D. B.
Please summarize the most important point in the following article. Bye for now.
49:35 - Unidentified Speaker
Okay, well, why don't we take a look?
49:52 - Unidentified Speaker
Why don't we take a look at the core thesis bullet points? We'll Read those two bullet points.
50:05 - D. B.
discuss.
50:40 - Unidentified Speaker
All right, any comments or thoughts or questions?
50:48 - D. B.
I find it interesting to Read that a given neuron can be activated by completely different concepts. For completely different purposes.
50:58 - E. G.
Yeah. But other than that, it gets pretty dense pretty quick.
51:05 - E. G.
It's use of superposition.
51:13 - D. B.
That's the name of it, this phenomenon.
51:20 - D. B.
Anyone else have any thoughts or anything they want to say? All right, why don't you go to too much? Why don't you go to the prompt window again at the bottom there? Does it ask anything?
51:35 - D. B.
And let's say something like, please tell me one thing about one important short thing in simple terms or something like that.
52:08 - Unidentified Speaker
Okay.
52:22 - Unidentified Speaker
let's Read that.
52:23 - D.
you I think it's a little too simple.
53:03 - D.
Yeah, go ahead.
53:05 - D. B.
It seems like it missed something.
53:11 - D. B.
This is a little more understandable to me, but it still is more.
53:16 - D.
I guess if I was going to ask another question, I'd say, well, in simple terms, What do you mean by break the neurons?
53:26 - Unidentified Speaker
into separate pieces. Is.
53:56 - Unidentified Speaker
Look at that.
53:57 - D.
That's exactly what I thought. So as you're really asking it, what is that? What's that? Go back that it was.
54:08 - D.
Can you scroll up just for a second?
54:14 - Unidentified Speaker
Yeah, just a little bit to see the previous response.
54:18 - D.
Yeah, so in its thinking, it was like, OK, so we got to explain what the sparse encoder does, because that's what really ask.
54:26 - E. G.
I also changed the model from chat GPT-5 to chat GPT-5 thinking to get more detailed answers.
54:33 - D.
Yeah it is like thinking is like okay they're really asking about you know what this sparse encoder really does that's what and that was really the question if you Thank you, man.
55:10 - E. G.
So if you look at this from a The neurons are basically an overloaded method. And Yeah.
55:20 - D.
Yeah. Now ask it what it means, what it means to light up a neuron.
55:29 - D. B.
So they train a tiny helper model, they call a sparse encoder that breaks it into separate functions.
56:05 - D.
Okay, now that makes sense.
56:08 - D. B.
All right well we could continue to interact with this article by asking questions just like we're doing and that would be an activity which we can vote on if you want to do it.
56:25 - D. B.
Are we voting on the activity or the Are we ready to go ahead and vote?
56:32 - D.
paper?
56:35 - D. B.
The activity on this paper, but we can vote on the other activity on another paper afterward if we like,
56:44 - Unidentified Speaker
Let's start with just this paper. Do this activity on this paper for several sessions.
56:57 - Unidentified Speaker
Not really.
57:07 - D.
R., go ahead. I missed R.'s earlier.
57:13 - R. S.
It's before your percent signs there. All right, let's see what you got here.
57:23 - D.
Yeah, about it.
57:24 - R. S.
3, I'm going to give it a 5.
57:32 - D. B.
You can pull four 3s and a That adds up to 6, 10, 16, 21 divided by 6, 3.3 and 1.5.
57:51 - D. B.
OK.
58:03 - Unidentified Speaker
That's not what we wanted to show you. All right.
58:41 - D. B.
All right, well, I think we can do one more quick evaluation. I think it was you, D. Somebody just suggested use the same process, but on another article.
59:10 - D. B.
So why don't we go ahead and vote on that as soon as I, not yet. Let me first go to the chat and put a little demarcation.
59:26 - Unidentified Speaker
Yeah, all right.
59:29 - D. B.
I mean, any other comments or thoughts on this before we vote on it?
59:44 - D.
I think getting it on another paper would be interesting.
59:51 - D.
That paper, if you kind of get down at the bottom of it.
59:58 - D.
It doesn't sound They're tracking neurons.
1:00:06 - D.
There's a lot of them. All right.
1:00:12 - D. B.
Well, go ahead. Are we voting on the biology of large language models?
1:00:17 - R. S.
Because you highlighted that thing above. That URL.
1:00:22 - D. B.
I'm highlighting use the same process as above, but on another article.
1:00:28 - R. S.
What is that other article? The one we just did. The biology of large No, the one we did a minute ago. OK. 4, This is Y.
1:01:01 - Y.’s iPhone
I'm driving, but doing this for another article, I'm good. Got a couple of threes.
1:01:09 - Y.’s iPhone
It can be four or I'll give it a five.
1:01:13 - D. B.
I'll give it a 4.5.
1:01:15 - D.
He's giving it a 4.5, too. A four or five, that's 4.5.
1:01:20 - D. B.
All right, let me add it all up here.
1:01:23 - E. G.
11, I'm also running the exact same questions through Claude.
1:01:43 - D.
You have better answers?
1:01:47 - E. G.
It does not.
1:01:51 - E. G.
It's similar answers, but it does not keep the So when I ask him to, using an improper prompt saying, please refine, it doesn't keep the context.
1:02:11 - D. B.
Okay, well, I think we're at a good stopping point. At some point, we'll have to stop evaluating different things and pick one, pick the one with the highest vote. Votes.
1:02:27 - D. B.
And again, if you have more items, more articles that you want to evaluate, just let me know and I'll add it to the list. Any last comments before we adjourn?
1:02:39 - D. B.
Okay, well, thanks everyone.
1:02:41 - M. M.
You guys have a good weekend. See you next time.
1:02:47 - Y.’s iPhone
Thank you. Thank you. Everybody.
Fri, Aug 8, 2025
0:33 - Unidentified Speaker
Oh.
0:34 - R. S.
you Hi, everyone.
1:57 - Unidentified Speaker
Hello.
1:59 - R. S.
Hello.
2:00 - M. M.
Hi. I was sitting here trying to figure out why there was no sound.
2:06 - Unidentified Speaker
Meeting is now being done telepathically.
2:11 - D. B.
There is no sound because nobody was saying anything.
2:18 - Unidentified Speaker
Well, at least I guess I guess we can go ahead and get started. Yeah.
2:26 - M. M.
Does anybody have any announcements or updates or anything, questions?
2:33 - M. M.
I think that we need to invite more students when the school start.
2:39 - D. B.
Yeah, I think that's sort of why we started the group was because the students, it's a way to get students involved.
2:47 - M. M.
We have to do this. Yes, we have to at least all of our master and PhD students to make presentations and, you know, to learn what they do, what the colleagues are doing. So, yeah, definitely. So I'm just come back from my trips, so I'm happy to be back. You're welcome back.
3:11 - Unidentified Speaker
Yes, D.
3:12 - M. M.
So, yeah, we.
3:15 - D. B.
Well, during the semester, we we typically had, you know, an average, I'd say, about about an average of 11 attendees. And as the summer has gone on, it's declined. But when the things start picking up in the fall, I think attendance should go up and we can recruit students and so on. Yes. Yes, sir.
3:40 - Unidentified Speaker
OK.
3:44 - D. B.
So last week, D. did a demo for us. And he's generously agreed to do other demo, just not today particularly. But D., do you want to do it next week?
4:03 - D. B.
You're muted.
4:05 - D.
So I mean, yeah, I guess I can. I'll be doing them from work. So I think that I would have everything that I need Yeah, I should be able to do it.
4:21 - D. B.
OK, yeah, it's up to you. I'll go ahead and
4:24 - D.
Well, so I can't do my large language model one, though, as I can't. I don't know how to hook up to I can hook up to it at home.
4:36 - D. B.
We're going to do the local ones like downloaded language model.
4:42 - D.
Yeah, I guess I don't know that I can run anything on my laptop, though. Oh, OK, well, we don't have next week. I mean, we can just I wonder, I wonder if, if I could do it, but like it for 15 maybe.
4:59 - D. B.
So like, we could start the meeting while you're going home or something like that.
5:02 - D.
Yeah, because so I would get off at 345 I suspect. Okay, and then I have to go through traffic. So, you know, well, we can start a meeting and then as soon as you're ready, to do it, you log on a little bit late, and when you're ready, we'll do it.
5:20 - D. B.
How's that sound? Let's do it.
5:21 - D.
I like it. Okay.
5:23 - Y.’s iPhone
So D., which LLM do you plan to download?
5:29 - D.
Well, so what I've downloaded three models from Olama. Quan Co and DeepSeek one, I think. I think I can handle like the 32 billion parameter models. So they're small.
5:55 - D.
They're pretty good at making lists and stuff. I mean, they're not, but they're not full featured models. So for instance, if I was gonna take one of those models and try to anonymize the transcript, they can't handle it.
6:22 - D. B.
They cannot.
6:23 - D.
They don't have what it takes. But when I say that, there might be a way that I could do it.
6:34 - Y.’s iPhone
Like small pieces of it at a time.
6:37 - D.
rebuild it, like build a, but go ahead, sir.
6:43 - Y.’s iPhone
And when you say small and big, I'm assuming you're talking about the size, do you know how big they are in terms of size?
6:52 - D.
Maybe one of them, I think two of them are 32 billion parameters. And one of them might be 30 four billion parameters? Is that?
7:05 - Y.’s iPhone
Yeah, in terms of parameters, yes. But in terms of actual size, how much size is it picking up and what kind of processing or RAM is it expecting?
7:17 - D.
I don't have that information, how its size there are that I don't think there may be a few gigs, maybe 10 days. I mean, it's not that it's not that big. All right, D.
7:33 - Unidentified Speaker
Yes. Would it help if I made myself available for my system?
7:39 - D.
You want to, you want to, you've got my cell phone number.
7:44 - Unidentified Speaker
Give me a call. I don't have your cell phone.
7:47 - D.
What kind of talk is that? You've got it.
7:49 - E. G.
You've called me on it. I've never called you. Yeah. Okay. Then connect on discord. Yeah.
7:58 - D. B.
Why don't you just tell us what the number is, and then... Yeah, I'll put it in chat here.
8:03 - D.
Put it in chat, that way we can send it to our telemarketing buddies.
8:08 - E. G.
Here you go.
8:10 - D.
This guy, he likes to buy.
8:13 - Y.’s iPhone
Yeah, I just Read today the Attorney General talking about the fraud that is coming for Dole and a lot of AI is being used to those texts and other content. It's quite interesting. And I think just last week, S. A. from OpenAI, he spoke about fraud in banking. But today I Read about fraud for toll payment and other things and how AI is smartly being used. To actually autorespond and create the website and everything. So it's quite interesting.
9:03 - D. B.
All right. Well, so you got E.'s number. If you lose it, let me know. I've got it.
9:09 - D.
I'm putting it in my phone right now. I called you, so you should I see that.
9:23 - D. B.
That was me. So I'll just keep that here on the list for now.
9:31 - D. B.
V. has not been here recently, but he did volunteer to And then Professor S., whenever I ask, she's willing to do it. Whenever I ask her to do it, she'll review this book that she Read over the summer, or she started an ad hoc discussion group about the book. I don't know.
9:51 - D. B.
if she got a group out of it or not. But anyway, she Read the book and she's willing to tell us about it. So.
9:56 - Y.’s iPhone
Can I ask one more question to D.? Yes.
10:00 - D.
I mean, I don't know.
10:03 - Y.’s iPhone
It's not my meeting technically here.
10:06 - Unidentified Speaker
That's why I asked D. permission if I can ask on the previous topic. D., is that okay? Yeah, go ahead.
10:13 - Y.’s iPhone
So when you are demoing or when you did testing in the past, did you also check that once you download and you start running those models on the backend, so can you really get off from internet and run and does it work? And the reason I'm asking is, or you download, so you have some parameters locally but for any reason have you checked like actually then it goes and goes to the model available on the respective clouds and checks or process something and gets back if it's not able to find the answer. So are you doing any testing or confirmation that once it's local it's not actually going off your machine and processing done anything like that.
11:15 - D.
Okay, so so what so what I've done is so I built a I built a computer specifically for this task. It's got to to 3060 12 gigabyte graphics cards and basically mid level processor and a whole bunch of RAM and a large hard drive. And so I've got it networked so that what I do is I use the, what's that thing called? Open UI, I think.
12:06 - D.
It's a free program that you download But what I do, so on my machine, where I have less VRAM, I just access it through my browser. So that is in my internal. That's my, it's a network, but it's not the internet. It's my home network. So I access it. Now I have not went and tried to block the internet or block access to the internet on either one of the devices.
12:43 - D.
Because I'd have to, well, I need the router, you know how you got the big box, it's got the modem and the router and, and all that in it. And, and so So if I when you go to hook up to a network. So we're both on the, you know, on our, on our own network, we're both on the home network, but the home network has internet access to the same box that's managing the home network. So I haven't tested it at all, as I would need to figure out how to test it, I'd have to go like, leave the router plugged in, but unhook the internet cable from the cable box or the fiber box, I guess is the appropriate way to say that, and then test it. But I'm confident because of the internet technology courses that I've had in school, I just that I am accessing that machine directly from my machine and there's tools that I put on there so that I can choose to either allow it to go online and access the internet or choose not to allow it to. But I think the real drawbacks are context window. All right.
14:18 - D. B.
Well, we can get more into that next time if needed.
14:24 - D. B.
Another announcement is there's going to be a guest lecture they're someone visit physically from King's College in London.
14:40 - D. B.
He's not a computer Thought-provoking lecture and book science tech, technical guy, he's an economist. Oh, he's a voice on the future of work in technology. Anyway, that'll be next two sessions in early September, one on Thursday at two on Friday at two. University Theater. You're supposed to register. I don't know what happened if you just showed up without registering. You'll be okay. The theater is pretty big place. University Theater. So yeah, it's another opportunity.
15:27 - D. B.
As I mentioned last time, I'm hoping to recruit. Oh, yeah, I did talk to L. G. a few days ago. He'd started the book project last semester, and something happened, and he had to leave for the semester. But he's going to be back, and he wants to do it for real this time. So now that we know how to do book projects, we'll have some to do it, hopefully.
15:55 - D. B.
And if there's anyone else who wants to suggest a project for grad students, we could try to recruit, I can try to recruit someone for that. Okay.
16:06 - D. B.
I've So last time we reviewed some readings and today we can review some more.
16:15 - Unidentified Speaker
That brings us, let's see, we did this one. Oh, no, we didn't do that.
16:18 - D. B.
We evaluated this one a long time ago. Let's evaluate it again. So I'm going to bring it up. We'll Read a couple paragraphs through the abstract, and then we'll decide if we want to Read it in more detail. Let me.
16:35 - D. B.
Let's see.
16:58 - D. B.
OK, so as you know, G. H. and J. H. won the Nobel Prize in physics, no less, in 2024 for their work on AI.
17:09 - D. B.
So what was this that I was found? I guess I didn't get the right link or something.
17:21 - D. B.
Popular it. Physics not Prize 2024-2. Is This
17:30 - D. B.
I think it was his speech.
17:46 - Unidentified Speaker
No, that's not the right link.
17:52 - D. B.
I'm sorry, I didn't take get the right link to So anyway, I going thought we here. Were
17:59 - D. B.
I'm clearly, I just don't have the right link. We So can now. For we'll maybe look that at it next skip time. Let's take a look at another paper.
18:12 - Unidentified Speaker
Looks like an academic, well, I don't know what this is. Someone suggested more detail, and it is biology of a large language model, whatever that means. So let's take a look at a little bit of it.
18:41 - D. B.
here and see if we want to address it.
18:44 - Unidentified Speaker
Okay, so let's take a look at this pair of and S. to thank you for this. Us.
19:23 - Unidentified Speaker
Any comments or questions?
19:29 - D.
Yeah, I'm going to Read the next paragraph.
19:33 - D. B.
All right, well nobody had any So let's go ahead and Read the next paragraph.
19:55 - Unidentified Speaker
Any thoughts or questions?
19:59 - D.
I think it's interesting that you're looking at it. So let's study this large language model like it's a biological organism. Because we don't know how it works.
20:15 - E. G.
Sometimes, I mean like the biological, you could look at a microscope, you can see what's going on.
20:26 - E. G.
times LLMs are black boxes where we're putting information in and it's coming out the other end.
20:39 - D.
So it's really a lot like a human brain which to a lot of people is still a black box.
20:52 - D.
I think at the end of the day the black box part of a large language model is any special training techniques that they did and didn't share with anybody. So, you know, we say that, you know, we assume now that all these language models are still being, you know, trained through the transformer, but are they still, you know, do they have, have they developed techniques that they haven't written any papers about that they're not telling them they don't want to leak it out because they think they might have an advantage. But at the one that when the models done, it's just series of weights, right? That's what that's what it is at the end of the day, it's just weights. And then they might have some more proprietary like logic that they'll put, you know, in between those weights and us are the output, right? So that they can, you know, be sure not to say any But once the models train, you know, the only offensive or so, or, you know, so it is that would be black box would be that logic that they use to filter answers. Because a weight is a weight is a weight, right? Villanova Lee?
22:31 - D. B.
Am I thinking about that right? Yeah, I mean, the building blocks- It's what's called an emergent property of these systems. They're basically weights and neurons. Neurons are pretty simple little computational units. But together, when trained, somehow they develop this organization where they can now write books and things like that.
23:02 - D. B.
Because it's only predicting the next word, right?
23:12 - D.
And I think by looking at the weights, Yeah.
23:16 - Unidentified Speaker
And if that's the case, every model, once you get the weights, you can duplicate it.
23:22 - E. G.
That's right. If all of the training at that point goes away, I take chat GPT five that just came out, get all of its weights, and then I've got a copy of it.
23:40 - D.
Well, so that's kind of what they're talking about. The seek they use that they use the chat algorithm to kind of you know do do kind of like this if they didn't treat it like a biological you know uh organism they treated it like uh a conversational you know they went and picked chat gpt force brain or they're accused of it I don't know if they did or not They were accused of picking its brain to try to guess its weights. But yeah, so once you get the weights, then you've got a model. That's right. That's correct. That's correct. OK.
24:32 - Unidentified Speaker
F., did you have a comment?
24:35 - D. B.
No, no, it's fine.
24:36 - M. M.
It's correct. Yes, we have the model. And can modify somehow. Yeah. All right.
24:44 - D. B.
Do you want to Read another paragraph or do you want to vote on this one?
24:48 - D.
I'm interested to see what kind of craziness they're going to come up with.
24:51 - M. M.
Yeah, I agree. So Read another paragraph?
24:56 - D. B.
Please. Sure.
25:00 - E. G.
All right.
25:03 - M. M.
I like the figures too.
25:06 - D. B.
Yeah, well, if we decide we want to Read the whole thing, we'll definitely check everything. I can't figure out how to even highlight anything nowadays.
25:17 - M. M.
Something happened.
25:18 - D.
Yeah, all right, here we go.
25:38 - Unidentified Speaker
Good. No.
25:41 - D.
Oh, sorry.
26:00 - Unidentified Speaker
Comments or questions?
26:02 - D.
So what are we looking at?
26:08 - D.
References.
26:10 - E. G.
Yeah.
26:12 - Unidentified Speaker
Are Maybe.
26:18 - Unidentified Speaker
We If any of these look really interesting, we can I can put them on the list to evaluate and interpret.
26:27 - D.
How to encode the real.
26:31 - M. M.
This is number two.
26:37 - D. B.
No.
26:40 - D.
Let me go down to the bottom. Is this? 1, 2, No, no. I want the references.
27:08 - M. M.
Yeah, you're there.
27:12 - Unidentified Speaker
Hmm.
27:39 - D.
Well, I mean, that's a whole, that's a bunch of good papers right there.
27:46 - D. B.
Well, I mean, research topics. So there's a lot of stuff out there. I mean, I don't know how effective it is yet.
27:52 - D.
But yeah, I would be interested in slapping those into a large language model and getting a submarine to see if it's worth reading.
28:01 - D. B.
Number two looks interesting.
28:04 - E. G.
Decomposed using LLMs with dictionary learning.
28:13 - D.
I kind of like, D., We could try that.
28:30 - D. B.
Because that would give I'll put that on the list. I'll say it's not exactly reading something paragraph by paragraph. It's not exactly viewing a video one minute by one minute. But it is something we could think about doing. So I will put what you suggested in the list of readings, potential readings.
29:02 - D.
When you go inside a language model, and you start trying to get at its weights, you're significantly violating your terms of use if it's proprietary.
29:18 - D.
But I don't know about CLAWD3, what they're doing with CLAWD3 to have permission But just because we can figure out ways to do really cool things, it's kind of like hacking somebody and taking their code. It's not very nice.
29:52 - Unidentified Speaker
I've invested. Significant amount of money to these things.
29:58 - D.
That free data that they downloaded off the internet.
30:00 - D. B.
All right, I can't get the link. I've got it. I've just put that idea in the list of readings, potential readings. All right, and let's go back to this one, which we just So the question is, do we want to Read that in more detail, Read it paragraph by paragraph, Read more or less the whole thing or a lot of it and talk about it? And what I'd like to ask you to do then is the usual, if you put in the chat your vote for how much you want to Read it in detail. Five is you definitely want to Read it. One is you definitely don't. Three is you're not sure. Two and four are leaning against or leaning toward.
30:53 - R. S.
If I remember correctly, the exact title was something about the biology of large models. Is that correct?
31:02 - D. B.
Yeah, it's right exactly.
31:07 - R. S.
So are they really talking about biology or that they're using that word biology in a different context?
31:15 - D. B.
Well, they were trying to draw an analogy with biology in one of the paragraphs we Yeah, there it's, it has no, it's, it's just similar.
31:25 - D.
Yeah. It has no ties to actual biology.
31:33 - D.
All right.
31:33 - D. B.
Well, I thought we had the starts of the board here.
31:37 - Unidentified Speaker
Okay. I'm looking for your chat thing. Okay.
31:40 - R. S.
Uh, So again, one, two, five.
31:43 - D. B.
OK.
31:43 - D.
I don't want to get overly excited.
31:58 - E. G.
D., you sent that to me, not to group. I did?
32:03 - D.
Yeah. Any other votes?
32:05 - D. B.
We've got two votes, so I'll give We got four so far. There's more people than that here.
32:26 - Unidentified Speaker
Yeah.
32:28 - M. M.
Oh yeah, F., how about you?
32:30 - D. B.
I'm okay.
32:31 - M. M.
We can Read this, four. Four, okay.
32:34 - D. B.
All right. Let's see what we got. The average is 4.6.
32:59 - D. B.
Okay, so I'm going to go back here.
33:18 - D. B.
Where were we?
33:29 - Unidentified Speaker
I can't even remember where we are. Here we are.
33:49 - Unidentified Speaker
All right. So we'll do a few more of these. We'll finish this up today, maybe.
33:52 - D. B.
And then we'll pick one, the highest one, to Read in more detail. In the future. Let's take a look at this one. These might be a little, a little old. In some cases, because they've been sitting here for a while, but you're welcome to anyone's welcome to suggest another one and we'll vote on it. In the future, anytime. Okay.
34:20 - M. M.
Okay, here's the title.
34:33 - D. B.
Yeah, yeah, I can.
34:34 - Unidentified Speaker
Any so far?
34:44 - R. S.
He's trying to do the complete opposite of AI by having the students do the thinking. Something like that.
34:59 - Unidentified Speaker
All right here's the snip it and take a look at that.
35:27 - Unidentified Speaker
Any comments?
35:34 - Unidentified Speaker
My question is, well, how How are they going to do it?
35:47 - Unidentified Speaker
All right, let's Read this.
36:04 - Unidentified Speaker
Comments or questions?
36:09 - E. G.
What is the date of this?
36:13 - D. B.
Let's take a look.
36:17 - D.
What was that AI that they did?
36:20 - D. B.
I'm sorry, I've got to put this in full screen mode so I can...
36:24 - D.
I'm sorry, it was not an AI really, it was a chat program that they made a long time ago. April 7th.
36:37 - D. B.
maybe in the 60s, that was really good at just, you know, rephrasing your question.
36:45 - D.
They turned a large language model into a chat program from the 60s.
36:54 - D. B.
Well, there was, Eliza was the first chat bot.
36:57 - D.
A lot, yeah.
37:00 - Unidentified Speaker
Eliza, exactly.
37:02 - M. M.
It's doing this kind of stuff.
37:06 - D.
But it's not really Socratic at that point.
37:10 - E. G.
The of the Socratic method is to ask probing questions, not just turn around the question, but ask probing questions so you develop the thinking process, the critical thinking on your own rather than be told the answer.
37:29 - Unidentified Speaker
Yeah, Eliza was it asks questions reflectively.
37:35 - D. B.
So it was like intended to mimic a certain school of psychotherapy where you sort of, or certain school of interacting with a person who has, you know, who's seeking answers, whatever, by turning their statements into questions. So like Eliza would, if someone said to Eliza, you know, I'm not feeling well today, Eliza might respond, you're not feeling well.
38:03 - M. M.
That would elicit more from the person.
38:09 - D. B.
In this Socratic questioning. It's like what Socrates did. He teach people by asking them questions, make them think. Where is it?
38:28 - D. B.
When students ask questions, Claude responds with Socratic questioning. How would you approach the problem? What evidence? Things like that. It doesn't seem like it's that hard for a chatbot to do something like that, but normally when you interact with chat GPT or Claude, it doesn't do that. Now, are saying it can. T. E. sometimes comes to these meetings. As I understand it, he's planning on doing a dissertation on using AIs for Socratic questioning. I want to use it in my classes. Make it students to give homework questions where they have to go to it. And ask it to interact with them by Socratic questioning.
39:27 - D. B.
I think this is a whole crisis. And now, you know, Claude, the anthropic company, is coming out with this learning mode, which does it. I mean, you can still do it without, you can just go to a normal Claude and just say, please use Socratic questioning. To help me understand the following passage. You give it a passage, and it'll start giving you Socratic questioning. That's interesting.
39:56 - D.
So I don't think learning mode is that different from normal plot.
39:59 - D. B.
It's just probably a modification to the prompt or something, a background prompt.
40:10 - E. G.
Actually, then, what I'm going to do over the next couple of weeks is I have access to ChatGPT, the latest ChatGPT, and the latest Claude is, I'm going to use Claude and through a Socratic method, ask it to help me better understand allele-specific expression.
40:33 - D. B.
Yeah, you could try to, you could ask it to, you know, you were talking about earlier about asking it to summarize the paper. You can do that, but you can also ask it to teach you about the paper using Socratic questioning and it'll do it. Maybe not very well, but you can always make it do it better by improving your prompt. If it doesn't do a good job teaching you using Socratic questioning, that's not its problem. That's your fault because you didn't prompt it properly.
41:11 - Unidentified Speaker
I put on the list of readings what D. suggested that we use in real time, ask ChatGPT to summarize a paper.
41:21 - D. B.
Well, we can ask it to do more than summarize. We can ask it to teach us the content using Socratic questioning or anything Anything like that we like,
41:36 - D. B.
Any other comments on Let's Read another paragraph.
41:44 - Unidentified Speaker
and decide it and then vote on it.
42:10 - D. B.
Any comments or questions?
42:20 - D. B.
Well, we already Read a lot of this. It's a pretty short news article, but we can still vote on whether we want to Read the whole thing.
42:27 - D. B.
Are we ready to vote?
42:37 - D. B.
All right, go ahead and type in your vote for whether you want to Read this in more detail together or not.
42:51 - D.
It would be more interesting to test it. More interesting to what? Test it.
43:09 - D.
Take a fine thought and thinking mode and check it out and see what it does.
43:15 - D. B.
No, I'm going to actually do that.
43:16 - E. G.
I'm going to say, preface it, as a tutor in biostatistics, based on these attached papers, using the Socratic method helped me better understand the concepts being presented in these papers.
43:37 - M. M.
I like it, I like it in this way, yes.
43:45 - M. M.
We need to test it.
43:46 - D.
We need to test it, that's right.
43:49 - Unidentified Speaker
Yeah.
43:53 - M. M.
Somebody think of something really, really hard.
43:59 - M. M.
Something super complicated.
44:01 - D.
All right.
44:04 - D. B.
Yeah, we'll help us figure it out.
44:15 - D. B.
Okay, that came out to a total of 3.87.
44:27 - Unidentified Speaker
people voted? All right, well, let's go ahead with the one that D. suggested.
44:52 - D. B.
I think what you suggested was we put this paper into chat TPT and ask it to summarize and interact with it about that paper.
45:08 - D. B.
So let's see if I can't figure out how to get to the link. Oh, here we go. Here's the link. I've got to go back. I just don't want to lose it. All right.
45:40 - Unidentified Speaker
So here's the paper. Whoops. Here's the paper.
45:47 - Unidentified Speaker
It's pretty long.
45:48 - D. B.
I'm going to go ahead and copy paste the whole thing into What?
46:33 - Unidentified Speaker
All right, now what should I do? What did I do wrong?
46:36 - E. G.
Attach the, oh, you can, okay. I pasted it. Conversation is 25% over the length limit. You too big.
46:48 - Unidentified Speaker
Yeah.
46:53 - E. G.
Is that a paid version? No. Would you like me to do it? I've got the paid version.
47:00 - D. B.
Well, let me just try it on ChattyT first for ease, and then you can do it.
47:18 - Unidentified Speaker
Too long.
47:19 - E. G.
OK, well, if you want to share your screen, Sure.
47:26 - E. G.
I'll try it with, uh, open AI first.
47:34 - Unidentified Speaker
I think you're going to stop sharing first.
47:43 - E. G.
Yeah.
47:43 - Unidentified Speaker
Okay.
47:50 - E. G.
What's the, uh, What's the link to the
48:02 - D. B.
I'll put it in the chat in a moment as soon as I figure out how.
48:11 - E. G.
Tell me what you want as a prompt.
48:17 - D. B.
All right, I just put the URL in the chat. Did you get it? Yep.
48:35 - D. B.
Please summarize the most important point in the following article. Bye for now.
49:35 - Unidentified Speaker
Okay, well, why don't we take a look?
49:52 - Unidentified Speaker
Why don't we take a look at the core thesis bullet points? We'll Read those two bullet points.
50:05 - D. B.
discuss.
50:40 - Unidentified Speaker
All right, any comments or thoughts or questions?
50:48 - D. B.
I find it interesting to Read that a given neuron can be activated by completely different concepts. For completely different purposes.
50:58 - E. G.
Yeah. But other than that, it gets pretty dense pretty quick.
51:05 - E. G.
It's use of superposition.
51:13 - D. B.
That's the name of it, this phenomenon.
51:20 - D. B.
Anyone else have any thoughts or anything they want to say? All right, why don't you go to too much? Why don't you go to the prompt window again at the bottom there? Does it ask anything?
51:35 - D. B.
And let's say something like, please tell me one thing about one important short thing in simple terms or something like that.
52:08 - Unidentified Speaker
Okay.
52:22 - Unidentified Speaker
let's Read that.
52:23 - D.
you I think it's a little too simple.
53:03 - D.
Yeah, go ahead.
53:05 - D. B.
It seems like it missed something.
53:11 - D. B.
This is a little more understandable to me, but it still is more.
53:16 - D.
I guess if I was going to ask another question, I'd say, well, in simple terms, What do you mean by break the neurons?
53:26 - Unidentified Speaker
into separate pieces. Is.
53:56 - Unidentified Speaker
Look at that.
53:57 - D.
That's exactly what I thought. So as you're really asking it, what is that? What's that? Go back that it was.
54:08 - D.
Can you scroll up just for a second?
54:14 - Unidentified Speaker
Yeah, just a little bit to see the previous response.
54:18 - D.
Yeah, so in its thinking, it was like, OK, so we got to explain what the sparse encoder does, because that's what really ask.
54:26 - E. G.
I also changed the model from chat GPT-5 to chat GPT-5 thinking to get more detailed answers.
54:33 - D.
Yeah it is like thinking is like okay they're really asking about you know what this sparse encoder really does that's what and that was really the question if you Thank you, man.
55:10 - E. G.
So if you look at this from a The neurons are basically an overloaded method. And Yeah.
55:20 - D.
Yeah. Now ask it what it means, what it means to light up a neuron.
55:29 - D. B.
So they train a tiny helper model, they call a sparse encoder that breaks it into separate functions.
56:05 - D.
Okay, now that makes sense.
56:08 - D. B.
All right well we could continue to interact with this article by asking questions just like we're doing and that would be an activity which we can vote on if you want to do it.
56:25 - D. B.
Are we voting on the activity or the Are we ready to go ahead and vote?
56:32 - D.
paper?
56:35 - D. B.
The activity on this paper, but we can vote on the other activity on another paper afterward if we like,
56:44 - Unidentified Speaker
Let's start with just this paper. Do this activity on this paper for several sessions.
56:57 - Unidentified Speaker
Not really.
57:07 - D.
R., go ahead. I missed R.'s earlier.
57:13 - R. S.
It's before your percent signs there. All right, let's see what you got here.
57:23 - D.
Yeah, about it.
57:24 - R. S.
3, I'm going to give it a 5.
57:32 - D. B.
You can pull four 3s and a That adds up to 6, 10, 16, 21 divided by 6, 3.3 and 1.5.
57:51 - D. B.
OK.
58:03 - Unidentified Speaker
That's not what we wanted to show you. All right.
58:41 - D. B.
All right, well, I think we can do one more quick evaluation. I think it was you, D. Somebody just suggested use the same process, but on another article.
59:10 - D. B.
So why don't we go ahead and vote on that as soon as I, not yet. Let me first go to the chat and put a little demarcation.
59:26 - Unidentified Speaker
Yeah, all right.
59:29 - D. B.
I mean, any other comments or thoughts on this before we vote on it?
59:44 - D.
I think getting it on another paper would be interesting.
59:51 - D.
That paper, if you kind of get down at the bottom of it.
59:58 - D.
It doesn't sound They're tracking neurons.
1:00:06 - D.
There's a lot of them. All right.
1:00:12 - D. B.
Well, go ahead. Are we voting on the biology of large language models?
1:00:17 - R. S.
Because you highlighted that thing above. That URL.
1:00:22 - D. B.
I'm highlighting use the same process as above, but on another article.
1:00:28 - R. S.
What is that other article? The one we just did. The biology of large No, the one we did a minute ago. OK. 4, This is Y.
1:01:01 - Y.’s iPhone
I'm driving, but doing this for another article, I'm good. Got a couple of threes.
1:01:09 - Y.’s iPhone
It can be four or I'll give it a five.
1:01:13 - D. B.
I'll give it a 4.5.
1:01:15 - D.
He's giving it a 4.5, too. A four or five, that's 4.5.
1:01:20 - D. B.
All right, let me add it all up here.
1:01:23 - E. G.
11, I'm also running the exact same questions through Claude.
1:01:43 - D.
You have better answers?
1:01:47 - E. G.
It does not.
1:01:51 - E. G.
It's similar answers, but it does not keep the So when I ask him to, using an improper prompt saying, please refine, it doesn't keep the context.
1:02:11 - D. B.
Okay, well, I think we're at a good stopping point. At some point, we'll have to stop evaluating different things and pick one, pick the one with the highest vote. Votes.
1:02:27 - D. B.
And again, if you have more items, more articles that you want to evaluate, just let me know and I'll add it to the list. Any last comments before we adjourn?
1:02:39 - D. B.
Okay, well, thanks everyone.
1:02:41 - M. M.
You guys have a good weekend. See you next time.
1:02:47 - Y.’s iPhone
Thank you. Thank you. Everybody.
No comments:
Post a Comment