Meeting notes provided by Gemini
J. Langley: All right, if not, we will start off. Um, feel free to go camera off if you want. Um, I am Let’s see. So, welcome to Huntsville AI. Um, this is our virtual meetup that was going to be live, but then the rain decided to not do live. So, uh, we’ve got Smita here, uh, talking with us about how to build curious agents.
Sumaita Rahman: All
J. Langley: Uh, so with that, uh, take it away, Smita.
Sumaita Rahman: right, I’ll start sharing.
J. Langley: All right, that looks
Sumaita Rahman: All right, great.
J. Langley: good.
Sumaita Rahman: Uh so hi everyone uh I’m Sumita and uh like I’m I have just started at Kulsa and I’m working for Lauren and I’m really glad to be able to be here.
Sumaita Rahman: Thank you so much for the opportunity. So today um I will try to go over this paper uh training a generally curious agent that I was a part of and um in this paper uh the main question that we are trying to explore is that how we can like train language models to behave like useful agents and not just single turn question answering systems. And here like curiosity doesn’t mean random curiosity or like simply uh asking more questions. It kind of means that uh strategic information gathering so that the task can be done uh more efficiently. So we are introducing a framework called paprika and the goal of paprika is to train LLMs on diverse multi-turn tasks so that they learn general exploration and decision- making behavior that can transfer to new and unseen tasks. uh so I’ll briefly go over like the authors the first co-authors were Finance and Eding Jang and then uh myself and Abita worked on it too and uh Jler Jeff Snider and Raslan Salakov are were like the advisors.
Sumaita Rahman: All right. So here is the outline. uh I would go over the problem setup, a little bit about prior attempts, our approach, uh the main empirical results and the limitations and future corrections uh briefly. So let’s get started with what is the problem or like the problem setup. Now, autonomous agents uh deployed in the real world will often face completely new tasks. For example, uh a lang a offthe-shelf language model would not necessarily know that when it’s like being being used as a customer service agent, what kind of problems it could like what kind of problems it might have to solve. And um it has to be really like it has to gather information from the user to be able to solve the problem properly and then tasks can be vaguely determined and it’s very common to have partial observability. So like in general when a user ask a question it does not at once ask everything like give all the information and ask the questions it’s uh and ask the question it I’m sorry they usually like uh take multiple turns to like totally make the age and understand where the problem is.
Sumaita Rahman: So, uh instead of like promptly uh answering a question, it would be much better if the agent just ask clarifying questions before it gives the answer. And lastly, the thing that we’re trying to make it do is to make the agents uh be good at exploration. So again, even in the frontier models, it’s not very common to like ask clarifying questions before answering a prompt at once. So here is an example. So this is like from 4.5 GB 4.5 and it was prompted to find the code to uh show the shortest path between two nodes in our weighted graph. And it puts down the algorithm which is very efficient, but it only works if like the widths are the weights in the graph are non- negative. So it did not clarify that part before answering. And if like it don’t work when the WS are negative. And here is another example. So this is based on the multi- arm bandits problem. Uh this problem could be compared with like slot machines. So if you imag imagine that there are three slots uh in a machine and say it’s marked A B and C and for each slot if someone presses they get some sort of reward and we could quantify it with like 10 20 and 30 and after one turn uh that user is getting like 10 from A, 20 from B and 30 point C from C.
Sumaita Rahman: Now the goal of this problem is to maximize the reward and for that it needs to have the proper balance of exploration like seeing if in the second round uh a like slot a will give more rewards or it could just exploit the information it has and keep pressing c. So even uh in this like comparably well-known problem, LM seem to be doing uh like the other algorithm seems to be doing better than the GPT4. So the blue curve right here is the performance of GPT4 and the red is like greedy algorithm and UCB is uh upper confidence bound and TS I think is console sampling those are well-known algorithms for solving this problem and GPT uh without any fine tune tuning GP4 without any finetuning uh does like perform uh badly compared to those Now uh I’ll go over a little bit about the paradox. So language models can improve when there is like chain of thought prompting prompting with summarized history. So like in chat bots when we uh chat with it multi-turn it it gets the history it gets the gist and the performance improves over time with the chain of thought and we can see that here like with the chain of thought the performance increased but the thing is uh it only gets better for that particular context and we want the LLM to be more generalized and uh be able to like do it across all
Sumaita Rahman: contexts. Now another work has been done on training on synthetic data from known optimal algorithms. So uh in this slide it’s like showing the results that when synthetic data was generated from like the no samples again UCB and Thompson sampling it has pretty good results uh comparatively but again the issue is uh we can have like any and every task and we would not know the like oracle or the best algorithm to perform that task. So again it needs to be generalized and in a way that could work uh for a task for which the algorithm is not
J. Langley: I think we’ve got a question from
Sumaita Rahman: known.
J. Langley: Christopher.
Christopher Bales: Hey Summit on page 10 uh if you go back one uh I got two questions. First of all, what is chain of thought and also what is the reward? What am I looking at on this graph in terms of the reward? What does that mean exactly? So two-part question.
Sumaita Rahman: Uh thank you for the question. So to my understanding chain of thought is like when the model has the context it utilizes it and like I cannot define it quite uh formally but what I understand is like it’s utilizing the given context to like produce answers like its context is bigger in in that particular floor setting and could you please repeat the second
Christopher Bales: Oh yeah. Just what is the the y axis the uh the time average reward? Just high high level. What is that exactly? Is that like the the speed at which you expect a good answer or what what is that? Um what are we measuring?
Sumaita Rahman: like Help.
J. Langley: I’m thinking that’s what you got back from the multi-arm bandit. Uh, averaged over how many times you actually, let’s say if you were looking at slot machines, you were just pulling the lever and lever and lever and lever, you know, working your way across the x-axis. I think the y is how what your uh what your score was over time.
Christopher Bales: I got you. Thanks.
J. Langley: I think.
Sumaita Rahman: Thank you.
J. Langley: Is that right? Or close.
Sumaita Rahman: Yes. Uh I’m so sorry I I’m getting sick.
J. Langley: That’s okay. No
Sumaita Rahman: Sorry.
J. Langley: worries.
Sumaita Rahman: Uh the fourth one is like embed LM calls within an explicit search mechanism.
Sumaita Rahman: So what this uh what I understand from this is uh for a particular task uh is already set up and there could be multiple paths uh predefined that the LLM could explore and it could also like backtrack if a path is not good enough. So uh from this prior work we could say like the missing things are that open-ended task don’t have the structure for explicit search like the previous one and uh for a setting like when a user is like chatting with uh with an agent it’s not always desirable to backtrack and it’s it might not be even like possible and uh we will not know the optimal algorithm from we’ll not know the optimal algorithm to generate data from like for every task and again it’s impossible to collect data from all tasks that we care about so the LM needs to be taught uh general information seeking behavior that are task agnostic. All right. So now going into our approach is like the paprika framework. uh we want to construct tasks that require strategic exploration.
Sumaita Rahman: uh so like the task should be diverse enough so that the LMS will learn general strategies. Then task selection using curriculum. Tasks need to be chosen so that uh LMS can learn from multiple trials and then when the tasks have been chosen uh generate diverse trajectories from those and lastly uh train the LLM based on the trajectories that we got. And all this together is like the arria framework. Now first getting to the task construction. We have diverse groups like 20 questions mastermind and uh all of them has like some criterias that they’re text based multi-turn and partially observable. It also requires the model to understand the previous interactions and uh choose actions that will help to succeed. uh now for the task group like uh in I it could be defined in the following way. So there is an LLM agent that would be like agent in a real world setting and there is environment that could be a representative of an user and we have two kinds of environment one is hardcoded and one is LM generated.
Sumaita Rahman: So for example for training questions uh it might ask so like the model that will be trained later or like the base model will ask question like is this an animal and the environment like the LM generator environment will say yes or no and uh till it succeeds like if the LLM at the ends guesses that like guesses the actual answer correctly, it will be a successful trajectory and it will have like all the interactions like uh the questions the LLM agent asked and the feedback the environment gave and if it fails then like it will be a bad trajectory and the other task group also like hardcoded program task group also work similarly but the feedback is rather hardcoded and LM generated. And um also in some cases when the feedback is LM generated there is like LM as judge sometimes to give some score to know like if it succeeded or not. All right. So the first step is data collection. Uh for each task in the train split we will be gen we generated like 20 trajectories using a high temperature so that the trajectories are like diverse and also mean like this sampling method uh which is supposed to reduce noise like there should be some creativity but it shouldn’t go like off the rails all the way.
Sumaita Rahman: And like after the data has been collected and we have the trajectories uh comes the part of training. So first uh there is supervised finetuning. I would not go into the like calculation in here. I don’t I might not like say it rightly. Uh but one thing is uh that in general supervised fine tuning is done on single turn in our case it was applied for multi-turn and the loss was calculated only on the interactions from the agent. So like each trajectory could have uh each trajectory will have interactions like answers from both the agent and the environment but the loss is calculated only based on the answer from the agent. Next uh we use DPO. So uh so for the supervised fine tuning we only took trajectories that were successful or had like higher scores. Uh but for like TPO there is both successful trajectories. So like one the one with high score is chosen for like this is preferred and the low score uh I think was chosen randomly and then like the DPO was applied and again here the loss is only calculated on the tokens generated by the agent.
Sumaita Rahman: Uh there is also an application of RPO where it like combines supervised finetuning and TPO. Now uh comes the concept of curriculum curriculum learning. Now we train the LLM on self generated data and the main bottleneck is data generation and not like necessarily the gradient updates. Generating trajectories is expensive because the model will have to interact with the environment a lot of the times and needs multiple attempts per task and it will require a huge budget. So like there should be some sort of ways so that the budget on like generating the tasks are not uh wasted and we want to choose tasks where the model has a high chance of generating meaningful experience. So like the main so the main motivation is like choose task groups in a way so that uh model spends more sampling budget where there is more uh learning potential from the data. Now the idea is about having a learning potential and we define a metric for measuring a task’s learning potential. If there is a high variability in performance uh there is more opportunity to learn.
Sumaita Rahman: So uh for example if a task is too easy for a model then there is less opportunity to learn and again if the model fails every time and the task is too hard for it then also there is like less opportunity to learn. So the trajectories that are like in the middle where the model success sometimes like succeeds sometimes and also fail sometimes has kind of a combination. Those are like the ideal ones and here is like a matrix to measure a task learning potential. First we have a standard deviation of the reward divided by the average short and this is what we’re calling the learning potential. So first we define a metric for measuring the task learning potential and we assume that similar task will have similar learning potential and similar tasks can be grouped using prior data uh about their difficulty. And if we have uh for example like an easy group, medium difficulty group and a hard group, those could be kind of uh compared with the multi-armed bandits and the multi-arm could be compared with the multi-armed bandits.
Sumaita Rahman: Like so getting back to the previous example, the easy task being labeled as maybe slot A, slot like the hard, medium, and easy task being labeled different slots. And so for task selection algorithm with our curricul curriculum we would be using the UCB which is like known for sele known known to solve the multi-arm banded problem and we here we basically compute like a score and choose the optimal score and like the group that gets that score get trajectories from those. Now we have the main empirical results and here is like representative task groups where we can see that the like compared to the base llama model the fine-tuned model with like the paprika framework has a better success rate and like we show this for different temperatures. And this is like uh one in one out like leave one out I’m sorry leave one out experiments where uh the mentioned like it’s trained on the other given task groups and like tested on the one that is mentioned and even on this one uh most of the time the fabrica is doing better.
Sumaita Rahman: Lastly, uh it shows that the curriculum like improves efficiency and we have higher success rate when the curriculum has been included for the task selection. uh it’s for like both average success rate past at force success rate and like when divided by groups and then moving to the limitations and future directions. So these are kind of com combined uh in this task it was like everything was offline. So if it if the agents could uh interact with live platform like online RL that might get even better results and the task generations are very hardcoded in here. Uh everything is done quite manually. So uh that is also one of the limitations and that could be done next likeing the task generation and uh the base model has a like huge role on how the trajectories will be generated and how good they will be. So uh figuring out like how strong the base model has to be and having s some sort of uh measurement or like criteria for that is also one of that next steps and it’s also a limitation because in I think one of the task maybe minder or like it needed more trajectories than the others because the m base model was not good enough.
Sumaita Rahman: So yeah that is all. Thank you so
J. Langley: Nice. Let me see if I can uh I don’t want to say repeat what you were what you
Sumaita Rahman: much.
J. Langley: were talking about. Um I’m trying to make sure I have a a decent understanding. Uh basically putting it in nonI kind of words maybe. So I think I think where you’re headed is that uh a lot of times when I’m using GPT or some other model um it gives me it it’s almost like it has an answer just waiting on me to ask the question and it’s almost always wrong and then I well that’s wrong and of oh oh I’m sorry that was that was wrong let me and it’s like then it triggers it to come clarify or you know it’s like it takes some bit of interaction. Um whereas if it were curious and would, you know, ask me clarifying questions to start with to help, you know, let’s say I’m trying to to find a restaurant to eat tonight or something. Well, knowing knowing where I want to go, what time, you know, things like that would, you know, is that is that kind of kind of where it’s headed?
Sumaita Rahman: Yes. So like before jumping on to like answering the questions uh getting some background and clarifying and it’s more towards also like uh in business applications uh for example what I mentioned before customer service agent they need to gather a lot of information from the user to like get to the root of the problem. So get better on that particular sets as well and what you describe is like the general case that all of us face. So of course that
J. Langley: Yeah.
Sumaita Rahman: one.
J. Langley: So maybe customer support’s first answer isn’t turn it off and turn it on again, you know. Um, that would be cool because a lot of times I feel like uh I guess the other
Sumaita Rahman: Yeah.
J. Langley: part of it is uh it it it reminds me of Finding Nemo with Dory um who you know gets two questions into a conversation and then it’s almost like reset and you’re starting over again. But um one thought I had that might be interesting is that uh is there any concept of tracking what questions have been asked and either you know having some negative piece for trying to ask the same thing twice.
J. Langley: I don’t know if you’ve run into that
Sumaita Rahman: I don’t think so.
J. Langley: before
Sumaita Rahman: I don’t think that has been addressed. Um, yeah.
J. Langley: because I think a lot of times it might be valid, you know,
Sumaita Rahman: Mhm.
J. Langley: because the answer may you may actually expect a different answer based on what has transpired so far. But in general, it you just wind up going back to the well, hey, it’s it’s the question I know how to ask. So, you know, it show or in in some cases it shows up more than most others in a probability perspective. And so maybe that’s the, you know, do you want fries with that kind of, you know, the answer is usually yes. And usually it gets you a better
Josh Phillips: I think that’s actually so you’ll notice if you look at some of the newer reasoning in agentic models like Quinn and
J. Langley: reward.
Josh Phillips: Gemma a lot of times they’ll have repetition and frequency penalties. Uh that that’s part of that’s why because they’ll doom loop on the same sort of questions over and over and over again because of their reasoning training.
J. Langley: It cost you a lot of
Josh Phillips: Yeah, that was that was kind of one of my big I mean one of the most interesting things I think about this paprika thing is the 1.5
J. Langley: money.
Josh Phillips: temperature SC uh sampling. Uh and I’m wondering if you guys I saw that you are looking for next steps as doing that actual RL training. Uh but have you kind of looked into any of the ways that they’ve solved those problems and caused uh these issues that I think you know that high sampling templ temperature seems like it would lead to more doom looping and it getting stuck in reasoning.
Sumaita Rahman: Excuse me. So I think uh for that there was like the thing which kind of uh like the high temperature gives some room for like creativity or going a little bit of all the way but then the sampling method kind of tries to restricts it so that it totally does not go out of the context or something like that.
David Showalter: Yeah, you kind of asked my question better than I was going to, Josh, because I was basically just going to ask for the for the 1.5.
David Showalter: Um, you know, was that after testing different temperatures or was that selected based off previous information? Like did you find that it that it served your purpose best or was it just kind of a good middle ground?
Sumaita Rahman: So I think here is results for uh different temperatures. So yeah it reports uh like it shows 3.7 and 1.0 O and I think after like trying multiple temperatures the 1.5 was the one that has been sorted.
Lorin Bales: And then so M I’ve got a question about kind of like the trajectories and you you have you have the one slide That’s all about like leaving them out.
Sumaita Rahman: Yeah.
Lorin Bales: And it makes me think about like masking. And I was wondering if you guys had ever looked at like taking a full trajectory and then masking parts of it to kind of like maybe hopefully help it like better infer maybe. Yeah,
Sumaita Rahman: Okay.
Lorin Bales: this one. Mhm. Yeah, this one. I was just wondering if if that had been looked
Sumaita Rahman: Uh no uh it has not been.
Sumaita Rahman: So the thing that happened was like after doing this like it was okay
Lorin Bales: at.
Sumaita Rahman: and it was not like really looked back into. So online RL was the thing that was in everyone’s head but there was not for it. So like other things were not explored much but that’s I think a really good point. Thank you.
Lorin Bales: Yeah. I don’t I don’t know if it’s like valuable. Like I was just thinking when it said leave one out, I was thinking about like this kind of masking things, but I see it’s like it it looks like it’s unseen task and leaving out the experiment as a whole, right? So it just makes me think of some sort of masking. Yeah. Anyway, that was very cool. Excellent.
J. Langley: Yeah, I like the trajectories approach. Um, I think we talked about that at that was last year or something. Josh, you probably talked about that I think for something.
Josh Phillips: Yeah, we talked about that during the Deep Seek R1 stuff.
J. Langley: Okay. And we’ve we’ve done some RL stuff in the past. Um I mean a ways back uh and we wound up it it’s just takes so long uh you know to to do things just cuz you’re you’re not only training the model, you’re also having to deal with the environment and the you know how fast can you turn this crank, you know? So finding some way to sample and then modify, regen, you know, some other kind of synthetic data off of that is is pretty genius. I really like it. This is pretty
Sumaita Rahman: Thank
J. Langley: cool.
Lorin Bales: And and then sum I have one more question. I’m just curious. Uh what hardware did you have? Because I saw that it’s the 8 billion model and so I’m assuming that you probably didn’t have like like the best hardware out there and you couldn’t use like a larger model maybe.
Sumaita Rahman: Yeah, it would be really great if there was a larger uh model. I think the memory was 400 GB.
Sumaita Rahman: Uh I don’t remember like the exact GPU uh
Lorin Bales: Okay. Okay.
Sumaita Rahman: configuration.
J. Langley: And I think we’ve we’ve got the uh I think John had found the uh GitHub for Paprika and posted that in the chat.
Sumaita Rahman: Yes.
J. Langley: So I’ve been through a couple of I’ve already been through the notebook. Uh it’s the the great thing about the virtual meetups is you can actually play along um as things are going. So that’s been a lot of fun. Uh, also one of the one of the cooler things about AI in general is most of the papers are out there and a lot of them actually have code that goes with them and it’s something you don’t normally see in other fields. So, that’s pretty cool. Uh, how about over to
Christopher Bales: Hey, uh I I saw that you were comparing it to Llama 3.1 and is this available
J. Langley: Christopher?
Christopher Bales: um to to try out for us um on like uh
Sumaita Rahman: Yes, it is. I uh Yes,
Christopher Bales: hugging face?
Sumaita Rahman: I think it
Christopher Bales: Okay,
Sumaita Rahman: is.
Christopher Bales: cool. I was just going to play with it a little bit.
Sumaita Rahman: I’ll try to like find out the link.
J. Langley: Yeah. And I actually like that you went with the smaller one so I have a chance to run it, you know. Uh that’s always something I run into when you know, well, hey, we did this cool thing. I was like, yeah, I’d love to, but I don’t have a I don’t have one of those cards.
David Showalter: Hey,
J. Langley: Um,
David Showalter: try it out with the GIMA 4B.
J. Langley: oh, that would be cool.
Sumaita Rahman: Oh yeah, you already found
J. Langley: Yeah, I found that a lot of times when I’m presenting stuff, uh, I wind up with with as many I wind up learning as much sometimes as what I came in with. Um, especially with this group.
David Showalter: That’s uh what stood out most to you from working on this paper? Like was there anything that surprised you or you weren’t expecting at the time?
Sumaita Rahman: Uh like this is it felt like quite straightforward like something simple in the site but still not like put together. uh especially like uh the SFT and DPO has been out like doing working with them is fine but like the concept of task selection like the task or like the data selection how much difference it’s making is kind of stood out I
J. Langley: All right. Does anybody else have any questions?
Lorin Bales: I have a complete sidebar question kind of the like usually when we wrap up we we all get to chat and got one of those kind of questions if that’s all
J. Langley: All right. Yeah, let’s we will pause the uh I’ll have to go back and cut the recording because I don’t actually know how to stop it halfway through. Um so yeah, uh full up sidebar session, whatever uh crazy stuff you’all want to talk
Lorin Bales: Yeah.
J. Langley: about.
Lorin Bales: So, I was using um just a typical LLM 14 billion today and just doing um kind of vibe coding with it, right? and did something really strange.
Lorin Bales: I went from planning mode to build mode and it won’t build. It will not execute the plan. And I was just curious if anybody has run into that. I have not run into that before up to this
Josh Phillips: I run into it with the mist models all the time.
Lorin Bales: point.
Josh Phillips: Even in devstrol, it just does not seem to understand that it’s its job to do stuff sometimes.
Lorin Bales: Okay.
Josh Phillips: Uh yeah, with the mist modules especially, I do that find that
Lorin Bales: Okay.
J. Langley: I didn’t hit that. I ran into something similar where I would have to uh it would like get stuck and just not give me anything back and I would have to toggle back to plan and then back to build or
Lorin Bales: Yeah. Yeah, I tried
J. Langley: something kind of like re please, you know,
Lorin Bales: that.
J. Langley: it’s like me say kicking it or something. No, really. Um, I’ve been I switch I haven’t seen nearly that kind. We’ve I’ve got access to a Nemo super somethingron model that’s pretty good that I didn’t quite see that kind of level.
Lorin Bales: Okay.
J. Langley: And then the Gemma 4 uh that that I’ve been I’ve moved to after that has been has been a whole lot better.
David Showalter: Yeah,
J. Langley: Yeah. Neotron.
David Showalter: I’ve still been having good uh Yeah,
J. Langley: Yeah.
David Showalter: I’ve still been having good results from that Neotron 3 Nano.
Lorin Bales: Okay.
David Showalter: It’s doing
Lorin Bales: And and and David,
David Showalter: good.
Lorin Bales: you’re using that on like smaller hardware like like I
David Showalter: Yeah. uh very small hardware like a 6 GB of VRAM like a laptop
Lorin Bales: am.
David Showalter: 3060 uh RTX with like uh 64 gigs of regular RAM.
Lorin Bales: Okay.
David Showalter: I don’t mind I don’t mind letting it run for like 20 minutes to get a little output from
Lorin Bales: So I have Okay.
David Showalter: it.
Lorin Bales: Excellent.
David Showalter: And the only problem similar to yours is uh used to run into it with Claude a lot where I’d have to force it to uh stop creating and go more planning. But haven’t had that problem in a while.
Lorin Bales: Okay. Okay. And then does it matter what software you use in that regard like like the clawed code or open code or uh anti-gravity or or is it mostly always on the model side?
Josh Phillips: No, it’s definitely the harness, too.
Lorin Bales: That was
David Showalter: Yeah.
Josh Phillips: I would say I would say it’s sometimes more the harness than the model.
Lorin Bales: hard.
David Showalter: Yeah.
J. Langley: Oh, shoot. There was a uh I got to go find the podcast. Um somebody I I think it was practical AI. Um there there was somebody that actually did a pretty good deep dive of uh the claw drop that hit that went uh open they didn’t open source their their harness but there there were copies in the wild um and went through some interesting uh kind of things about what what they had actually done in the harness side of things that were different from what what has been seen elsewhere. If I can find that, I’ll drop it on Discord because it was it was pretty neat.
Lorin Bales: Oh, that does sound really interesting,
J. Langley: They also went through a little bit of how the leak happened.
Lorin Bales: Jake.
J. Langley: Um, which was interesting. Um, apparently it was the I don’t know a whole lot about it because I don’t do a ton of TypeScript. Uh, but when you build TypeScript and you’re trying to debug it, it comes up with like a map file that maps the code back to something. uh it’s really you can’t really read it by hand, but if you’ve got a machine that can read it um or parse that thing, uh somehow that got checked in or released with a release somewhere and I think that’s what people used to then back out kind of what the harness looked like. At least from what I could what I could
Lorin Bales: Oh,
Josh Phillips: Yeah,
Lorin Bales: wow.
J. Langley: tell,
Josh Phillips: it’s for testing and it was stuck in the binary and somebody pulled it apart.
Lorin Bales: Interesting.
J. Langley: guess don’t do that. Um, which they don’t anymore, but um, then there was a conversation about, well, maybe they’ll just open source the thing since parts of it are out there.
J. Langley: And I just don’t know. I don’t know that that would be a thing that they’re going to do, but
David Showalter: So, Jay, were you good with me taking the uh paper review next week or did you have something else lined
J. Langley: no, you’re good.
David Showalter: up?
J. Langley: Um, I’ve got, um, I’ve been I’ve rolled over into more of an implementation stage for some of the other stuff I’m working. Um, I now know what AWS CloudFront is. Uh, that’s pretty wicked. Um, I’ve learned how to spin up an API um with fast API, but then host it with a lambda. Um, it’s slow as I’ll get out, but I’m not it’s nearly free uh to throw it out there just to play around with. Um, and I see a lot of stuff from Christopher. I’m guessing you’ve done something similar. Um,
Christopher Bales: I’m just an AWS fan and I I know where you’re coming from there. It’s awesome.
J. Langley: it blew my mind. it, you know, it it was all clawed code.
J. Langley: It was like, “Hey, try you should do this.” And I’m like, “Well, how long would that take?” And it’s like, “Well, 5 minutes. I’ll have it.” I’m like, “Well, yeah, let’s try it.” And so now I have a Terraform thing that just throws this thing out there and it’s like, “What?” Um, okay,
David Showalter: Nice.
J. Langley: sure. Um, also through uh what is off zero? I always try to always mess it up. I always say OA off, but I think it’s off zero or whatever it is. That’s in there, too. uh about to spin up Daxter and have it thrown in and then probably have something operational pretty quick. But uh the only AI stuff really that’s in that right now is a way to uh so I’ve got optical scans or optical views uh of certain things and then I’ve got uh synthetic radar aperture radar views of the same thing. And there’s a way to uh build a an estimate of what the optical would look like based on the the radar part.
J. Langley: And it’s it’s within like 5% accuracy. Um and it’s very specific to a to one kind of a a view. Uh but it’s a way that if you if it was cloudy and you don’t have the optical, you know, from the satellite, a way to replace that uh in something that’s, you know, close enough to be useful. Um but anyway, that’s that came out of a paper that just again handed to Claude Code and said, “Can you implement this paper for me?” And it said, “Okay, here.” Um I I didn’t believe it but then at next I I added
David Showalter: That’s awesome.
J. Langley: it I I had it go um you know create a way to verify or validate that it’s within you know what percentage is this and I don’t they don’t use percentage they use R squared and other uh RMSSE kind of stuff but anyway fun stuff but not super it if we ever want to do a not another LLM paper you uh talk again. Um that might be one of them uh to kind of me mix
David Showalter: No, that sounds great and it sounds more uh more practical than my
J. Langley: up
David Showalter: planned one will probably be. So, uh for a heads up, month and a half ago, Google Deep Mind put out uh I’m always fascinated when people try to define AGI currently. So, Google DeepMind published a measuring progress toward AGI, a cognitive framework. There was some fun stuff in there that just thought I’d share. And then Stanford’s AI index report came out a couple weeks ago and it’s about 430 pages. So, was kind of curious what Stanford uh found.
J. Langley: Mhm.
David Showalter: So my plan next week was just kind of share general findings from Stanford and Google Deep Mind and uh it’ll be entertaining but not that uh practical usage
J. Langley: No, that’s cool.
David Showalter: probably.
J. Langley: I like
Christopher Bales: Don’t we have like several versions of AGI?
J. Langley: it.
Christopher Bales: I mean, can’t I ask like Brock or Gemini any question in the world and it’ll give it to me?
J. Langley: It’ll give you an
Christopher Bales: Don’t we already have that?
J. Langley: answer.
David Showalter: Yeah, that’s kind of the uh that’s kind of the fun part is the bar always keeps changing and what people mean by it. So just just a little spoiler uh Deep Mind threw in um two parts for cognitive faculties. One of them is social cognition um that it thinks something must have for AGI.
J. Langley: H.
Christopher Bales: Cognition. Really? What is that even?
David Showalter: social social cognition.
Christopher Bales: How do you even Oh, interesting.
David Showalter: So
J. Langley: Yeah. Sometimes I struggle with social cognition.
David Showalter: yeah, that’s what I keep coming back to. Uh the it’s defined as the ability to process and interpret social information and to respond appropriately in social situations. I’m like, that’s already kind of uh super intelligent if you compare it to a lot of people.
J. Langley: I mean, yeah. I mean,
Todd Page: That’s
David Showalter: But
J. Langley: it’s Yes. Apparently, engineers are not AGI. I agree with that, Josh.
Christopher Bales: Hey, I just wanted to say so Ma, this was amazing and um it seems like you did a lot of really cool work here and it’s way beyond anything I’ve ever done, but it’s really cool and you got a bright future ahead.
Christopher Bales: So, thank you.
Sumaita Rahman: Uh thank you so much for having me and this like I worked really small portion of it. It’s like the first and second author’s brainchild but I am really glad that I could work a little bit on it and thank you so much for giving me this platform again.
J. Langley: Yeah,
David Showalter: Yeah, we’d love to have you be a part more often.
J. Langley: we we appreciate it. Yeah, you’re on the Discord now. Uh which I found out earlier. I’m pinging some folks trying to spend virtual. I’m like, could somebody cuz I Yeah, we we’ve uh sometimes we run into stuff where we have to get word out really quick on a on a venue thing or something, especially with weather in the in the springtime. Um, and luckily, I don’t think I haven’t had my phone blow up by anybody standing outside of Hudson Alpha, so I think it worked, which we’ve done that before. Um well yeah uh super really appreciate um all the all the effort that you put into this um and really impressive being on a paper that you know that’s published and you know things like that. So that’s pretty cool. So with that um we will wrap it up.
Sumaita Rahman: Yeah.
J. Langley: Um, I guess next week is the virtual. Uh, so we’ll be right back on on meet um next week. So see y’all later. Um,
Sumaita Rahman: Okay.
J. Langley: everybody enjoy the time off from the rain that’s currently not happening.
David Showalter: Yeah, real quick. Uh Josh, I just saw your comment on the knowledge graph presentation.
J. Langley: So
David Showalter: I look forward to that.
Lorin Bales: Thank you so much so much. I’ll see you tomorrow.
Sumaita Rahman: I’ll see you tomorrow. Thank you so much, Lauren, for having me here.
Lorin Bales: Always.