Meeting notes provided by Gemini
Executive Lounge: And I can tell you right now, the prices are gonna go down. Uh 20 years ago, I interviewed with the firm that’s on the wall there to do patents here, but I actually decided not to do that. I was wor and Jay, I’m interested in your hackathon approach kind of technical stack. Yep. As well as uh I know data bricks as a data company from back in the day. Yeah. Um I’m not sure I know there’s uh you know from going into you know Spark and some of the other things and then the uh machine learning library in Spark and some other stuff but Spark yeah all of that we use that quite heavily. So I’m not quite I’m interested to see what that what that means the AIM ML strategy. Okay. All right. Um I mean again all these are lightning talks so we could go for hours on some of these topics or weeks but uh yeah we’ll go through each uh just reminder that this should be uh Yeah, I spent the last 16 years working for Oracle, previously worked for IBM, front defense and other people.
Executive Lounge: I also practiced patent law in San Francisco representing Oracle and Xerox Park and some others and I ran the Huntsville big data meetup from 2010 to 2016 and Jay was a frequent attendee. uh that is probably the that is the first meetup where I ever presented anything which kind of uh in my kind of progression kind of kickstarted the whole maybe I can do this type thing so it’s kind of full circle and I’m the author of a book on big data author of a book on R and been involved in a lot of books over the years too. Oh. Uh, and I also didn’t put it up here, but I’ve taught computer science courses for Virginia Tech and Harrisburg University. All right. All right. So, let’s dive in. Uh, we’ll start with what I saw at GTC. All right. There were a lot of robots at GTC. Uh, now GTC um is a massive conference right now. the AI boom, Nvidia being so valuable, their stock price shooting up. Um, everybody I saw at that conference was like super enthusiastic.
Executive Lounge: There were no feelings of the AI bubble is going to crash. There were no feelings of AI was even in a bubble. Okay. So, so it was very pro AI. This was only a month ago. This is March 16 through March 19th. Um, there were a lot of robots on the show floor. There were a lot of robots everywhere. Um, why do you suppose I put a a cooler up here next to all these robots? Yes. It was a robot cooler and and it was the most practical robot I saw because this guy was roaming around San Jose giving people bottles of water and sodas and things and you know, um, so let’s talk about roaming around San Jose. Let’s uh let’s go look at a map here. So, did you did you go as a just a individual person and I took a week on vacation with Oracle and paid a zillion dollars to go out there as an individual person. Paid for the ticket. I did get a 40% discount as an alumni.
Executive Lounge: I haven’t previously gone to it. Okay. Um but if you you would qualify for that if if you just had the virtual pass from a previous one, too. So, you don’t actually have to have gone to physically to get 20% discount. Um I’ve been to the DC twice. This is the first time I’ve gone in person to San Jose. I haven’t virtually done the San Jose one previously. Yeah, as you mentioned in the email, you can see a lot of the presentations. I’d say about half the presentations are actually online. Um, there’s only like three 400 presentations online, but they have over a thousand presentations there. Um, so let me talk about how massive this conference is. Okay, this is downtown San Jose. Um, the convention center is this block right here next where the Hilton is in the San Jose Convention Center in South Hall. Um the first Monday of the conference they had um the keynote where uh Jensen talks about his vision for robots and everything else.
Executive Lounge: Uh Nvidia that was at the basketball arena the SAP center over here. Okay. Pretty much everything between here and about here on the map and up to here and down to there was all conference stuff. Um, you would walk into any restaurant in that area, any hotel in that area and it’d be all conference stuff. Wow. They had this little plaza de Chavez up to about here down to there was all closed off and um you had to have a badge to get in that area. The civic center and a convention center were both fully dedicated to the conference. Most of the buildings around here were taken over either by the conference or by a vendor like Dell or HP or somebody. you know, they would rent out the building to run as their local headquarters, right? So, um, I’m doing four lightning talks. One is on the NVIDIA GTC conference. That’s what we’re talking about right now. One’s on a hackathon, one’s on patents, and one is on, uh, data bricks, a IML strategy.
Executive Lounge: Uh, is there anything in there you’re really interested in? I’m taking a question. you can let me know later. Yeah. Yeah. So, there’s about Yeah, that’s coming up next. That’s all. So, um there were about 50,000 people attending the conference and the outline stuff. So, was it it was a massive event, probably the biggest tech conference in San Jose. Um, so so and obviously because it’s Nvidia there was a big hardware focus. Okay? You know it’s um but there’s some really funny stuff. You’ll hear comments of you know how you know Jensen talks about how it was kids gaming uh PCs that paid for his company and they went to go into AI and all this stuff. All right. So the first thing I mentioned was the robots. The other big thing is it was a big focus on quantum computing. Um, now we’re still probably three years away from a practical quantum computer, but we’re getting much closer and there was a big presence there.
Executive Lounge: Uh, and the other thing I saw was this big focus on AI factories and where they’re going with the liquid cooling of the AI racks. So now, you know, the next generation of AI racks are completely liquid cooling. You know, they’re basically going to be just super compute of the future. Um, and it wouldn’t totally shock me if Nvidia became a cloud computer computing company in direct competition with Amazon and everybody else or one of the AI vendors becomes that, you know, basically an AI cloud in the future. Um, the other thing they talked a lot about was all of the efficiency and, you know, oh, we’re we’re much greener than we were last year, we’re much smaller than we were last year, all this kind of stuff, right? All the it is being improved. Um, those were my main takeaways from the GTC conference. It’s AI to the max. It’s um sky’s is the limit. You know, anything you can imagine, you can do. And you know, uh like I said, there was one practical robot, which is a cooler robot.
Executive Lounge: Uh the other robots were great dancing and doing different things for photo ops, but they really weren’t doing too much other than that one. Uh and the cooler robot, there was a fleet of them out there. Um because they had them all over the place. There probably was at least a hundred roaming around. Oh wow. Yeah, I didn’t know that Nvidia was that heavy into quantum. Well, I I don’t know the reality of how heavy they are. Okay. Um, but the quantum vendors you see there are all the main quantum vendors. So, you see IBM, you see all their competitors, um, you see, and they’re all showing off their their quantum computer machines. Okay. Now I think the idea is that you know if you think about most modern quantum architectures are assuming you’re going to use a quantum computer plus a traditional computer right and so the idea is that traditional computers can have GPUs in it to accelerate the math. So I think that’s the role Nvidia is playing.
Executive Lounge: Okay. Um I honestly have stepped away in quantum over the last few years. I actually taught a course on quantum computing in 2020, but I haven’t really done anything with it since then. I do know it’s advanced quite a bit since 2020. Yeah. And like I said, Google released a paper not too long ago saying they think they can bring ECC in three years. Um, which would mean 2029. And so Nvidia’s been talking about six months. Yeah. a week ago, wasn’t it? About a week, they released modelers and stuff like that. They’re just they’re getting in the space a lot more. Yeah. And I would want say one thing about the open source models is Nvidia has gotten really big into open source models and it does remind me of the last days of Sun Microsystems where Sun was a hardware player but did a lot of open source stuff with Java and my SQL and so on and I think um Nvidia is kind of the same way and that they’re a hardware player but they’re really diving in deep into the open source models and they do have a large team building those models.
Executive Lounge: And in fact, they made a big deal about it when I was at the DC event about last six months ago about how, you know, hey, we’re doing all this stuff. You know, you don’t have to rely on Yeah, we’ve tried to use some of that. I mean, right now we’re using Neotron. Um, it’s it’s solid. Uh, their SDKs for doing some things with AI. It’s kind of it’s I’d much rather use PyTorch or you know it’s just it’s hard to you either buy into the entire ecosystem or you stay away from it’s it’s hard to do a little bit with their you know and I think that’s part of why they do what they do you know like were you expecting it to be this large and like um No. Um, okay. So, my biggest pet peeve with Nvidia conferences, this applies to DC1 as well as this one is I never had any idea where the lunches were. You know, in theory, uh, we paid for a ticket and the me the ticket came with meals.
Executive Lounge: I never had a clue where these meals were. You know, I stumbled into the meals in the DC one because it was a much smaller event. You know, there’s only so many places they could have hiked, right? And the California one, what you’re looking at, I think like this is a big parking lot over here. Okay. This green thing, I think this little square is where they had some tents and they were giving meals out, but there were no signs or anything. And I think there probably was a cafeteria somewhere in this convention center, but again, there were no signs. I have no idea where the meals were. Um, so that’s but again that’s an example of this is a company growing so fast that they don’t have anyone to tell you hey by the way meals are open. Um, that’s my biggest pet peeve. Um, they did have an app. Um, I’m not going to bring it up and show it on screen, but they had an app that you could look at in the computer as well as on your phone.
Executive Lounge: Um, the app had a lot of stuff in it. you could it made it easy to find the the the seminars you were interested in. Um then there was a bunch of stuff that’s not on the schedule that you only found out about if you knew someone or if you somehow were on the right mailing list. Um so why don’t we segue into developer events? Okay, Luma is basically what Nvidia is using for developer events. Um, so like Dynamo after hours, this is some of the Nvidia people. Um, so if we scroll back a month or so, let’s go back to March 16, 19. So, for example, here’s a hackathon. They did the Neatron model reasoning challenge. They had a bunch of hack on Saturday and um Tuesday they hack for impact. Uh Wednesday, they had hack to create building toolkit. This is the one I basically all these hackathons were really fast. This one was 9:00 a.m. to 2 p.m. Um, they didn’t really even tell us what we were doing till 10:00 a.m. So, we got there at 9:00, we hung out for an hour, had breakfast, eventually at 10, they said told us what we were doing.
Executive Lounge: So, we kind of hit the ground running around 10:30 and we knew that we had to submit by one or so because then we’re going to like show off what we did and then they’re going to judge and then they’re going to get out of there by two. Now, hackathons never run on time. So, we’re not surprised they didn’t start actually started to 10:30. Um, and we actually ended up submitting at 1:30. So, we had about three hours 10:30 to 1:30 to actually build something. All right. So, yeah, and you had to be a you had to have a GTC pass to participate. Um, so how did this thing work? Well, basically what we were going to do was uh we were going to help uh artists. We’re going to build some AI toolkit to help artists, you know, either visual artists, movie film people, uh creative writers, whatever. We’re giving them the ability to use AI better. Uh so build tools to make creativity seamless. Uh we each got to use a Dell Pro Max that has a uh greatest blackboard chip in it.
Executive Lounge: Um and we only allowed to show up as individuals. Now once we showed up, they told us go ahead and form teams if you want. So I did form it. There were two of us sitting at the table. So I said, “Hey, you want to team up?” He said, “Yeah.” So we had a team of two. Uh, let’s see what else they got up here. Did you get to keep the Dell Pro? Uh, I got one. I think you keep the one that we were using. I got an HP one. Um, all right. So, didn’t really tell us much throughout here. Um, it told us where it was. was it wasn’t a building it wasn’t register okay so not much with aluminum invite other than register any there okay so that was a luma invite and you can see they actually have all their meetups in this luma calendar as well um mostly obviously it’s west coast events but they do have New York and Boston events too big city events um so what do we do so in Jensen’s keynote uh on Monday, he made a big deal about Oklahoma.
Executive Lounge: Um, so I just and I had gone by, they actually had outside on that uh that park that they had closed off, they had a tent that they called the open claw tent and they had a bunch of these Delp Pro Maxes and and from various vendors set up with open claw on it. So I said, “Fine. you know, if you guys are going to have a tent full of these things, I’ll go ahead and set it up and we’ll do something with it. So, um, my friend or new friend, um, he’s been using cloud. So, he used claw to generate a JavaScript front end. I installed open I installed open claw and the and the GPT OSS120B model which took a while to download uh given everything that was going on on the network. Open claw is not so long to download but the GPT mod took away a lot a while an hour. Um so of my three hours an hour and a half was spent just installing open claw and getting um the model downloaded and everything.
Executive Lounge: Wow. But I had accomplished that in about an hour and a half and he got clawed to give him a nice JavaScript front end pretty fast. Um, and then what we did, uh, we had to spend a certain amount of time coming up with our presentation because, you know, you have to have a presentation to show people, hey, this is what we did, right? Um, and so that was pretty much our three hours. Wow. Yeah. Very simple, you know, slightly better than a low world implementation of Open Claw. this is what a creator’s calendar schedule looks like and this is how it’s, you know, running a prom job to throw up uh events and stuff and saying, “Hey, you’re supposed to go with this to do this film session. You’re supposed to do this.” Um, so I did a little basic thing like that. combined it with his front end JavaScript thing of like showing calendaring and stuff like that and you know access to you know some some generation stuff and that was it.
Executive Lounge: That’s what we did. Um, nobody else had anything as comprehensive. So I I I think the working code is what won, right? But also we had a little bit of, you know, practical working code, but we also had sexy stuff in that we had open claw set up and running, right? Nobody else really had the combination of both. Everybody else had, you know, they went all in on one and the other. The other thing is is that one of the judges pointed out that um I actually interviewed the uh the judge who was representing the creators. We had they had like a a creator me there. Okay. So I interviewed that person asked well what are your pain points and the judge said well you this team was the only one to ask me what the pain points were. Everybody else kind of asked uh do you like open claw or do you like this or do you like that? We were the only ones interested in the pain points you know.
Executive Lounge: So maybe we got his vote for that reason. Maybe asking somebody what do you actually kind of kind of help. Well, I’m I’m almost surprised somebody didn’t have like a local mirror of hooding face like there. So instead of trying to reach across and you know what I mean? Um yeah, knowing that that’s what everybody is going to be trying to pull from most likely. Yeah, that’d be interesting. I think again it was um and video is growing super fast. This is a hackathon. They’re putting this together on a napkin at the last minute. This was probably somebody’s idea the Monday before. Now, that week they did have this thing up maybe two weeks ahead. Okay. But we didn’t really know what it was until like when you showed up. Yeah. So, and I I know for another one that I didn’t participate in, they were constantly changing the rules. So, yeah, they were they’re they’re Oh, man. They’re making up as they went along.
Executive Lounge: Yeah. Um Yeah. So, we won one of these Grace Blackwell devices and um I let my partner have it because he offered me $1,000 in cash and I was like, “Okay, I can use the money in my bank account. There you go. You have you have my 50% of it, right?” Um, now I had seen a demo at HP’s booth of theirs and they had one of those random drawings and I won the random drawing. So, I put HP on the head. You had a good trip. It paid for the trip. Now, it didn’t really pay for, you know, a week of vacation plus a trip unless you look at the list price of it. If you look at the list price, then it paid for it. List price was expensive. All right. So, the tips I’m picking up so far, choose the right table to sit at. Yes. Ask what they Ask people what their pain points are, especially if they let you talk to a judge.
Executive Lounge: Um, they almost always let you talk to a judge. Now, they may not tell you who the judges are, right? But they’ll let you talk to anyone in the room, okay? Find out where the lunch find out where the lunch is. That’s important. Enter the raffle. Now, my grandmother won a surfboard that she never used, but I think she did sell it to somebody. I’m not surprised. It’s always um it’s all well the the symposium here the the app was pretty crappy. I mean like nearly horrible but hopefully theirs was better at least letting you see what the what the talk was and what it was about. Yes, the app was great. The app I had no problems with the app. I mean, I didn’t do any of the persontoperson conversations in the app, right? I don’t think anyone does, but um but uh no, it was good as apps go. All right. So, I did not watch many of the seminars uh because I figured I could watch them later.
Executive Lounge: Uh, I mostly went to wander the floor and there was a lot of expose stuff booths. It wasn’t just the main hall. They had side halls full of booths. They had booths outdoors. They had several parking lots that were full of booths. And like I said, some of the vendors took over um their own little buildings and they had their booths in the buildings. So, there was tons. You could spend the whole time there just going and talking to vendors. And is talking to the vendors and going to the external booths, do you also have to have a GTC? Yes. Okay. So, just to get into the whole area, you got to have a badge of some sort. Yeah. Now, if you’re going to one of the ones that’s like the vendor’s own place, right, they may or may not let you in if you don’t have a badge because it’s up to them, I guess. Right. Um, but the outdoor organized areas, yeah, you had to have a badge.
Executive Lounge: And generally, even the vendors wanted you to have a badge because they wanted to sell the people badges. Yeah. Well, they also probably have a a barcode on your bag or something. Yeah. Um, now the other thing is that um Yeah, it it was a massive event. Sounds fun. It was a lot of fun. Um, now it’s they did have a day pass the last day of the event. So, um, so I think the last day of the event was Thursday. So, yeah, it’s like a Monday to Thursday event. They had like Sunday workshops. You could have done workshops and a pre-party if you wanted to. I skipped that day. And then uh Thursday also they had a day pass so you could show up on the very last day if you don’t want to buy a pass. Um any idea where it is next year? So they actually do three or four of these uh across the world or last year they did three or four.
Executive Lounge: So I’m expecting them to do the same three or four. Okay. Plus maybe more. Um so it will be in San Jose again roughly the same time even though they haven’t announced it. They’ve done it 20 some years in San. It’s gonna happen. Yeah. Uh there will be one in DC roughly the same. Again, they haven’t announced it, but it’ll be this fall sometime same time because they did the last two years. Okay. And then they have a couple like in Europe and Asia and so forth. Um now this is the big one. The DC1’s tiny in comparison. Okay.
Lorin Bales: And and Tom, what did you think of the robots that were there?
Executive Lounge: See if I can get out there. Now the DC one’s like two days maybe is like two or three day two days and maybe a workshop day to make it three. And then um I don’t remember no European ones and like I said they usually put a lot of this stuff online so you can watch some talks online.
Executive Lounge: You just can’t see the vendors online or do the hackathons. Um so that was all my comments and hackathons. Oh, uh, one other comment I on last week’s thing we talked about Kaggle and the GPUs. Yeah, I have used those. Um, I think it depends when you’re doing your stuff because sometimes there’s more conflict in other people wanting the GPUs, right? So, I had a mix motion. Also, it’s been a year since I’ve used them, but I did I did some last year last year. I did use those for one of the computer. All right. So, that’s all my comments and hackathons. Um, let’s see. All right. Where are we? Robots. Okay. What’s this coming up here? Uh, maybe your Wi-Fi. Was my Wi-Fi dead? My Wi-Fi’s dead. It’s You’re Well, you are back for a second. Now you’re back. Yeah. All right. Maybe in the screen share.
Executive Lounge: Let me go back to screen share. Also verify through us now. So the TV and the pad there actually hardwired into their land here. So those aren’t on the Wi-Fi. And it gets kind of weird where Wi-Fi drops, but that doesn’t. You get in this weird kind of a, you know, we’re all frozen, but everybody on TV might be fine. I was trying to think where we just got new set up the office. I was trying to think where I just realized it’s the same. All right. So, um, we were talking last week about the artificial int analysis intelligence index and you know, Gemini and Claw and GPT are all rated about the same. So, I thought I’d try out Gemini um, and with the pro preview. Uh, and how effective is it to do some reasoning in the pack domain? Um, so what I tried to do was to generate the core of a patent application. This is a full utility patent application, but I’m only doing the core of it.
Executive Lounge: I’m not doing every last task. And I just want to see what it did. And I think it saved me about twothirds of the time it would take me if I was just writing this patent application using Microsoft Word and Microsoft PowerPoint and Google Search, but not the AI part of Google. Right. Wow. Um, now everything I’m about to show you was a test, not a real patent application, no client data, blah blah blah. Nothing confidential, not a real idea. All right. Um, so here is um the prompt I was essentially using. Okay. Uh, and what and what the goal was. So assume you receive an invention disclosure from your client, a company that builds search engines and provides searching services. Uh the invention disclosure form identifies as the only prior art a regular search engine like Google uh in which the search engine receives text as a search query identifies relevant web pages and provides a list of the relevant web pages. Uh the the invention disclosure form describes this invention as an image search engine.
Executive Lounge: Assume image search doesn’t exist yet. Um so the search engine receives a text search from a user identifies images and provides a list. And so our description should talk about how imaging search works. All right. Now, obviously, if you’re doing a real patent application, you’re obviously you’re inventing something that doesn’t exist. So, we don’t have to assume it. And you’re describing what doesn’t exist. But in this particular case, it’s much easier because it already exists. We know it exists. And this is much easier use case. So, I do think it would be harder uh to do a brand new patent application in terms of the AI won’t know quite as much. All right. So that’s what I was doing. And so here’s an example image that was generated. Now, this is an SVG graphic. Um, which still I wouldn’t file this exactly like this. I would probably clean it up a little bit more, but this is pretty good. You know, it’s because really for graphics for a patent application, you want black and white.
Executive Lounge: You want to uh have everything numbered, and you want to keep labels sort of minimal. So, for example, I would in a real file do uh diagram, I would have gotten rid of that and I would have gotten rid of that and I would have had that in the textual description in the patent itself, but I wouldn’t actually have it on the arrow, you know. So, that’s the only cleaning up thing I would probably have and I probably would have not had these two lines and these little buttons. I just would have had us use your device and, you know, some some block figures essentially with labels inside and the numbers inside. But basically, this kind of shows you the idea. We do a search, we send it to the search engine. The search engine has an image index that it uses to match the words with the images and it sends back a bunch of images to the user device. Uh and on the background we’ve got the web crawler which is searching the internet uh populating the search engine with images and metadata describing the images to build the image index that you can then look up to determine uh what information you find.
Executive Lounge: And I threw in a machine learning component just in case you’re doing some machine learning on the im uh you know because some of the search engines are doing more than just looking at the meta the metadata of the web page. They’re actually generating some stuff and doing some analysis. And then I had a packing description. I only generated like two pages of text and two D diagrams and one claim just to but you know the text came out looking pretty good in terms of you know it looks like patent text it doesn’t look like generated conversation you would expect to see you don’t see any of the sort of um psycho fancy you you see often in some of these search engines and not search some of these AI when they’re talking to you. It reads like a patent document for me. So I did a pretty good job of no more reference just um I did do a bit of back and forth and I’ll show you that in a bit. 3.1 Pro. Yeah, 3.1 Pro preview.
Executive Lounge: Um, at the very end of it, I ran out of tokens on a free version, so I switched over to Pro for the very end of it, but everything I’m showing you is the stuff I did just using Pro and I really didn’t really rely on anything Pro. Um, the one part where the AI really can’t handle is on the claims. um it does a good job of the syntax but it can’t really handle the subject matter. So what it can do is it can give you a plausible looking sentence. You know we put a period here we got a verb. We got a noun. They are words. They are words. They understood noun verbs and adjectives and how they work. But do they have the meaning you’re looking for? Um, and is that the meeting you want to pay for? Yeah, maybe not. Um, let’s back up. Okay, so let me move over to back over here. All right. So, here’s um Gemini Pro Preview.
Executive Lounge: All right. So what I was using again I was using AI studio because that’s where you get the free access to Gemini preview. Um and basically you know I gave it a prompt and I started off I gave it an example of a US patent but basically they could have used any US patent for my example. Um, and I kind of just, you know, told it, hey, um, initially I did not tell it to create the image, but I just wanted a prompt for creating the image. And part of that was to conserve tokens, right? Um, because I figured it’d be extra tokens and generating the actual image. And I knew I was doing this free and I wanted to keep token usage really low. Um, so we kind of gave it that first prompt. Um, yeah, it gave me my my prompt. It gave me my two-page text and it gave me one claim. Um, I I read through what it generated and I gave it um some additional stuff. Now, this is my very first time doing this.
Executive Lounge: I didn’t really uh if I was doing this professionally on an ongoing basis. I would have um you know shot learning in my prompt engineering give it a couple examples, right? And then we go through this really fast. In this case, I didn’t have that. So basically I to I told it to make some modifications to the the text up there you know basically given its suggestions like it’s an assistant who’s working on um yeah just basically I looked at it and said hey you mentioned that there might be an intermediary in the search engine uh you know like an AI assistant you know we’re talking about digital assistants on phones all the time and I didn’t mention a phone but it immediately realized oh he’s probably talking about a phone we did catch that so that was pretty Um, so this came up with uh the updated description. Um, blah blah blah. Uh, I actually took its text and I made a couple edits to it and gave it back to it. Um, then modified it.
Executive Lounge: Then I told it I wanted that second figure I show showed you because originally I only asked it for one figure. Well, actually originally I called it one or two figures and they only gave me once. So I fig yeah I really want to break this out. Let’s have two figures because originally I was thinking maybe we’ll put the web crawler on the same figure as but really the thing about search engines they’re really two different applications. One is serving users information the other is focused on building up the the the knowledge base that you’re going to then use. Um so then we came up with some prompts for that second f the figures. I’m going to jump through that. And it then updated the description. Then I told it to make some more edits. And so it made some more edits. Um then I told it to get rid of some of the labels it put in there. I didn’t really like it. And I changed some numbers around.
Executive Lounge: Um all in all, this whole process probably took me about four hours. And I think it would have taken me at least 12 to do that um without AI. Most of that being on building those two darn figures. Um you know because in the past when I was doing figures it was a pain in the butt because I because we’re very very precise and where the numbers go and you move them and it’s just it’s just a hassle. Um, and so most of the time in the past I would not do multiple figures in that application. It just wasn’t worth the hassle. Um, because clients generally wouldn’t care and so if they don’t care, right, the mental strain of the four hours, you know, compared to doing that 12 hours, not just time. So, I actually did it across two periods of two hours. I did like two hours on Thursday and then two hours on Sunday afternoon. Um, it wasn’t u mental strain. There wasn’t that much mental strain, but there was a lot of going back and forth.
Executive Lounge: You you know what it’s like when you’re in this sort of u bite coding thing and you’re you’re you’re throwing something at the AI and the AI is coming back and you’re throwing at it. It’s almost like you’re playing tennis. And it’s fast. And it’s fast, too. I mean, at most it took it a minute to respond. And usually it’s really fast. And pretty much I was in thinking the whole time. I was on the thinking general pretty much. Yeah, that’s that’s what I was going to ask because if you did it in two hours on two different days, how did it maintain your context or the conversation? It kind of kept it kept the flow of everything and just paid off. Yeah. Um, yeah, pretty much. Yeah. Now, the last half hour on the second day, I was out of tokens on Pro Preview. So, I fell back on the regular probe that I do have a subscription for. Um, but really I was at that point in time I’d finished my figures, I’d finished my text and I was just playing with the plane language which the AI wasn’t really that helpful anyways.
Executive Lounge: But basically what I would do is I I asked it to generate a claim. I made changes to it threw it back as what it thought of my changes and it just gave me cycle fancy. And I was like those are great changes. Gee, thanks so far. So, and then I played around, played around, eventually I got something I liked. Uh, and that’s why I said I figured now this is only the core of a patent application, a real patent application. You know, my file was about 20 claims um and three independent. This was just one independent claim, but it was two pages of text and two diagrams. I probably would file the real patent application with maybe one or two more diagrams, but they’d be really simple diagrams. It’ be fast to make. uh probably like 30 seconds or so using AI. Um I would have some more text, but not a whole lot more text. Maybe another five pages, but it’ll be easy to generate text, right?
Executive Lounge: A lot of the standard boiler plate. This was probably the hardest part of a patent application. So it probably to get the full patent application probably only take another two or three hours. Wow. Now we used to 20 years ago, we used to charge for 40 hours for a patent application. Oh, now billing rates have gone up. But with AI, you know, people tests, right? Now, you might remember two years ago, there was this uh case where someone found a lawyer with dealing with hallucinated case citations. Um, so I saw a post on LinkedIn today. Someone said there’s been 761 instances of courts complaining about this in the last two years. So apparently, just because one person got sanctioned was not enough. Now the reason is outports the sanctions are so low compared to the amount of time you save that nobody cares about the speeding ticket. Yeah. Now if the speeding tickets get high enough people will eventually care but none of the right they’ll tell you they care but they don’t.
Executive Lounge: Did you have any issues with the text on the diagrams doing kind of weird things? misspellings or letters overlapping or anything like that. Well, okay. So, I didn’t use diffusion. I used uh I had I used SVG. Okay. And that worked out really well. Okay. That’s Yeah, because because it’s code. It’s kind of like throwing it to a mermaid diagram instead of actual image of Yeah, I’ve tried mermaid as well, but mermaid uh has shading and shading will not work for patterns. But uh you see here’s my SVG code for generating the diagram. Um and there was a mermaid option as well. Um but yeah, I gave up on the mermaid option once I saw the shading. That solves that whole problem. Yes. Um and I also was playing around with SVG graphics. What I’m going to show you in a minute too. Okay. Uh and I really like the idea of going with SVG for professional presentations because you control it so much better.
Executive Lounge: I have found that almost perfect. Yeah. Going with the SVG tool. Yes. Oh, let me think about that. All right. So, here before I go, any other questions about balance? All right. How are we doing time? You are at 653 and we run a little after seven. Yep. We’re good. We start trying to wrap up. All right. This one generated by uh Gemini. Um this is an SPD graphic. You were talking about words. Uh I don’t see any misspellings in here. Okay. I am about to change my process for the uh graphic I use on the invite. Right now I use I did notice gee I got ML flow here twice. I got open source standards here three times. I got unity catalog here twice. Now human being would not do the redundant stuff. So this is still geni being but other than that I I see no typos.
Executive Lounge: Some one funny thing I noticed is some of the icons that you have are very similar to what I’ve had to generate for icons. Yeah, those are generated icons. So yeah, but it is interesting that you might be able to identify soon. Oh yeah, you did a but that’s no different today. Yeah. So this is not an official data bricks slide. This is just a hey Gemini. How would you describe uh Data Bricks a IML strategy? So what’s going on? Um Data Bricks acquired a company called Mosaic ML couple years ago. Uh Mosaic ML. Why did they do that? Well, let’s let’s back up 16 or 14 years uh ago. Back when I was running the Huntsville Big Data Meetup, we had some guys show up around 2013 talking about this thing called Spark. Um, and Spark was a way to do uh big data. It was a competitor to Hadoop. Um, but it, you know, made it a little bit easier on the data warehouse side and it was a little easier to use SQL with it and so on.
Executive Lounge: And Spark has morphed over time in becoming what they call a data lakehouse. You know, it’s a combination of the data lake of Hadoop with also data warehouse capabilities for traditional relational or structured database. Um, but it’s all commer focused in a lot of ways. Um, but it it really handle architecture. I don’t know if you’ve heard about that, but it’s basically this idea that you have this data pipeline. And you’ve got your data lake, which is really what they call the bronze level, and then it goes to silver and gold, right? And so the bronze data is it’s just arrived in the data lake. The silver we’ve cleaned it. We’ve, you know, made it fit with a schema and so forth. And then the gold is ready to be served as a data m or data product to a particular user group like sales or accounting or whoever. Um so that’s what they’ve been doing over the last few years. But um what they’ve noticed is that not only do they have people using BI and uh tools on that on that gold level, but they also have people using machine learning both at the silver level and at the gold level.
Executive Lounge: And so they’ve really morphed into becoming both a data company as well as a machine learning company. And that’s where Mosaic ML fits in and ML flow fits in um and they expanded upon that to um you know feed everything into retrieval augmented generation you know where we’re providing the knowledge to go along with the AIS um they are not creating their own geni models yet but they’re more like leveraging all the models that are out there you know cla um so Let’s take another look at what we might be able to do with this. All right. So, here’s an example kind of taking that bronze, silver, gold in a little deeper. All right. So, we’ve got like the Spark, you know, data lakehouse in here. Okay. And you’ve got all your data sources flowing in. This is your data lake, the bronze area. Then we go in and curate it in the silver area. And then we create our data products in the gold level. Um, and so your sources might be all kinds of different information.
Executive Lounge: Some of it might be real traditional relational, some of it may be sensor data, some of it may be PDFs and all sorts of knowledge stuff. And it just all goes into Bronx, right? Dump it in. You dump it in. Um, and nowadays they’re using something called Delta Lake. Anyone familiar with Delta Lake? So Delta Lake is essentially a way to provide asset transactions for big data. Okay? And the way it does that is not through like two-phase commit. Uh instead they’re you know because two-phase commit is what you use for relational databases and stuff. But the problem is is two-phase commit does not scale well to massively distributed systems. So instead, what they do with Delta Lake is they’ve got their transaction log and essentially everything gets written in the transaction log and then you can go back in time and recover based on your transaction log. And they essentially use like uh checkpointing. So you can, you know, every 10 transactions get checkpoints. You don’t have to go back a million rows.
Executive Lounge: You just go back to the last checkpoint. Um and they have what they call unity catalog which is their governance layer. It’s basically tracking everything that’s in the environment. And so between your Delta Lake transaction log and your Unity uh catalog, you can kind of manage and do uh true data lineage and track your data throughout the entire process from when it gets loaded into the bronze all the way to your machine learning apps for performing analysis on your data. And that’s about as much detail as I want to go into tonight. So I think we made it to seven. One minute ahead. The interest, it would not surprise me at some point if you saw. So so far I think they’re working for data pipelines, things like that. Imagine if I took my, you know, right now a lot of the stuff is running in pipelines that are triggered automatically. You’ve got things. It’d be interesting to see that right next to open claw tied together where now I have an ecosystem of agents that are not necessarily tied directly to this pipeline kind of a thing.
Executive Lounge: They can kind of move up or down, you know, that that would be that’d be pretty interesting. Now, they have something called agent bricks which in theory might be going in that direction. Uh they have a marketplace for skills kind of like open call has a marketplace for skills, right? So, I think they want to go in that direction. Are they really there yet? Who knows? But a second here. Yeah. So, both of these are SVG graphics created by Gemini and I haven’t seen any titles. Um I think one of my slides actually did have some writing where I really wouldn’t want it. This one had the store. Yeah, that one. Um but other than that they’ve done pretty well. Okay. I’ve been pretty happy with the SVG graphics. Oh, and up here ML there. Yeah, I’ve been uh a separate discussion. I’ve been using cloud code extremely heavily to develop some some fun act tech stuff. And I’ve I’ve been kind of watching what it feels like to interact, you know, so it’s good at generating all of its specifications, going out and doing them and all that kind of stuff.
Executive Lounge: I’m sitting there, I have no idea where this thing is and its lineage or whatever. So I’m looking at that for one thing. The other is the documentation I’ve got for this is geared towards having an agent. It’s geared towards documentation an agent uses. I look at it and it’s like I I get lost in just a massive amount. So I I wound up moving over to Mermaid as a documentation kind of approach that’s useful for both me and the agent to pick up and read because it’s textual a lot. But SVG could do the same thing and it actually looks a little pretty. Okay. I have a fun story. I’ll tell you in a minute. Go ahead. uh learning this week because I was like the same way I’d like to keep bringing out all the documentation into a visual. Yeah. And so um cheat code whenever you get your markdowns get your documentation or whatever especially when code is building is building is going right you just say hey present this to me in a rich HTML uh file or rich HTML slide deck.
Executive Lounge: Okay. And then it will just it’ll take the same markdown file and it will build it out into something that you click on and it opens it up and presentation that you get of it is okay substantially. What’s that called? Just you just you just it’s the prompt. You just so just like it’s giving you the documentation, right? You’re just asking it can you give me this either the markdown file or whatever it is that it’s creates. Yeah. generate this as a HTML slide deck code. Okay. And then it’ll just use this front end design and it just a visual way of saying because it’s I’m kind of watch because I’m also looking at it from a uh helping run a software organization that also needs you know we’re going to be more and more using these agent kind of systems. So the human to uh not human to agent but human to arm of agents that are working a whole lot faster than I could. Yeah. So I have a funny story about that.
Executive Lounge: All right. So I was asking this is Gemini Pro and that’s the preview one. I was asking it um what code generation tools were used in the GitHub directory. All right. So this is the GitHub directory that belonged to my hackathon partner. Okay. Um and we had used separate directories. I was doing open claw and stuff like that. He was doing JavaScript and miscellaneous other things using claw. So I asked him, hey, does the code in that directory use a llama or is it a similar case to open claw where a llama’s mentioned by the dev along with gposs 12p but not actually using the github directory. Uh and I I can tell it hey there is a separate directory where all these things are. But my question is is are they in this directory on GitHub? So, can Gemini tell me whether or not they’re being used or called or whatever? And Gemini says, “I’m having a hard time fulfilling your request. Oh my gosh, that’s like an intern move.
Executive Lounge: I can’t do what you asked me, but I can do this other thing. It’s of no use to you. I asked it. That’s pretty cool. Can you tell me what caused a heart? So what he said again sort of apologetic hard time was entirely on my head. I attempted to use a Python library beautiful soup to scrape the file list from your GitHub repository. My internal Python environment is restricted to a set of packages and I can’t do pip install and module and I found a module not found error and can’t do it. And then it just said, well, I agree with you. Alama’s not used. In other words, he doesn’t tell me what, you know, did he really look at the directory? Is he just he just saying, “Yeah, I fess up. I disagree.” He did tell me those no API calls, no dependencies, and no orchestration. So, it’s good enough evidence. Uh what’s actually there? Yeah, the JavaScript stuff that I mentioned.
Executive Lounge: So, it’s it’s an interesting uh you know, behavioral of what these agents can be like. Yeah. Yeah. Yeah. They all have different thing when I originally I told it like the P example I said hey make me one to two diagrams you made one if someone asked me to make one two diagrams I make one too unless I had to Gemini is going to It has the reasoning capability of the highest when it brought it reasoning in some metrics were higher than an what you would find in like when you saw you’ll never see the feeling of what AI is doing right now until you meet the city. Yeah, I’ll catch you up in like it’s it’s it’s so different because what what it what you’re seeing is especially with these models uh shad GBT is brilliant. It’ll give you a beautiful layout. It’ll give you all the documentation and the data. It even gives you better insight than what you’ll see in Anthropic. But if you set it to an agent, you find that it’s lazy.
Executive Lounge: It’ll shortcut it. It doesn’t want to do the work. It’ll tell you that it’s doing something and it say okay I’ll go do it and then as soon as it runs into it trips it st like I quit you know I’m not doing it and you will find models like uh anthropics that opus or sometimes silent but opus it will find a way to work harder regardless of what comes in front of it. And so what you find when you see that model work, right, is it changes your entire perspective of what can be done. So now without even having to use an agent, if I’m just sitting and claw code, I speak to it like an agent because I know its model does the agentic work that I’m accustomed to. So I don’t have to ask it to do things that I know. I’ll say do this impossible thing and I’ll just wait to see it comes up because figure some stuff out that is not simple. It’s crazy. I’m looking at building either a skill or something because when you did the thing with two with the day and then took a minute and it came back for the patent thing.
Executive Lounge: I wind up the way that I’m using quad right now is on kind of a nights and weekend kind of I’ll play with it this I may have an hour whatever actually you get doping in yeah well with it there’s you know dash c continue where you weren’t right I however need a little more than just a dash c to get me back what I was doing the club Oh yeah, that’s normally what I’m in. So I always make sure that you say here’s my goals empty document and here’s what we’ve accomplished so far empty and they catch me up. I ask for it and I constantly update those. So whenever I come back I’m like hey lay out where we are doing that. Right. That’s what I’m looking at. maybe a skill of where were we, you know, summarize. What is good at doing is creating a a to-do list, right? So, if you just give it that’s that’s pretty much what he’s built is a to-do list. So, whenever you have your your project, you have a to-do list at the beginning of your planning phase and that’s major thing and you say, “Hey, I want I want you to give me either steps, I want phases, I want whatever it is, give me that list of things.” And so that way when you come back,
Executive Lounge: he’s like, “Well, you just finished tier one, start on tier two.” Right. Well, what I’m It’ll pick up. Yeah. What I’m doing right now is this. I’m still playing around with this. Uh most folks, David, you’re probably one of these. You probably use a lot of text files in your codebase uh for either specs or for, you know, things like that. And I know that’s that’s kind of how a lot of the stuff is written. Uh the way I’ve pushed it lately is instead of using specs on my file system, I have it create GitHub issues for the spec and I got projects and then lay that out. I could hit that on my phone, look through things and either change priority of something or whatever. Um, here’s my goals and there I kind of let it work. Okay. Well, that’s avoid this problem. Avoid that. That’s how I got the set of issues was I gave it here’s what I want to build. Break this down.
Executive Lounge: create all of these things as incremental, you know, your pieces. And now I’m slowly, I think I had like 60, you know, issues. I think I’m like 45 in so far. And it’s just kind of, but even within that little context, you know, I may have 30 minutes to play with it. And I come back and it’s been a day and I’ve worked a whole another job in the middle and slept. So, it’s just a I I wish I had the ability to just reload my context with the dash C. Yeah, and I’m with you there.
Lorin Bales: Hey, I I did want to update uh David on something because he’s in the room.
Executive Lounge: Two models that are going to work similar to what is doing. Uh you’ll see I like Gemma. I like the Gen 4 model and and how it’s uh is real good for back background task when you have it set to to run. But um Quinn 3.6 that model is it will worth is you can you can use it for like less than what you’re paying for signing.
Executive Lounge: Oh yeah. So you’re you’re gonna use it and it’s going to it’s going to it’s going to try harder than chatbt will. Yeah. So that’s that’s useful. And then um um encoding uh Kimmy K2.6 came out and you have uh what is it? Is it It’s not It’s 5.1 or something else. It’s a Chinese. It’s not I don’t think it’s it’s it’s 5.1 and just it just came out. It’s it’s long coding is real good. And then GBT is dropping because it probably drops 4.7. Then you have GPT dropping 5.5 and then they drop the 2.0 image model image model like he was saying uh it’s almost no mistakes and it’s a world model. So it it will it’s elo is like 300 higher than the highest nano banana. Uh so it’s it’s significantly better. Okay. It’s ridiculous. Yeah. What they’re doing, they’re pretty much taking a a picture of anything and it’s building it to code exactly as the image is.
Executive Lounge: It’s like breaking the image down into code and then just rebuilding a perfect replica. So it’s it’s ridiculous because I have a feeling they’re going to pair that with um their coding agent which is what you seeing with anthropics design when they drop the whole design. Right. So I think OpenAI is coming to compete directly against that. Right. Anything else? Uh Lauren, do you have any questions or comments?
Lorin Bales: Yeah, Jake, can you hear me?
Executive Lounge: Yes.
Lorin Bales: Sweet. Um, my only comment would be that it work. I’m actually using uh Minstal uh 14B with open code and just because of where I’m at and the hardware I’ve got and what you know what I’m working on and it’s actually working out
Executive Lounge: Okay.
Lorin Bales: really well. It surprised me. Uh David encouraged me to to give it a shot, David and John. And uh it’s working and it’s helping. So
Executive Lounge: Cool. Yeah. Mist.
Lorin Bales: yeah.
Executive Lounge: Okay.
Executive Lounge: Oh, Mistl is one I can use. Okay. Thank you for the shout
Lorin Bales: Yeah. Yeah. But uh I I can’t use Quinn,
Executive Lounge: out.
Lorin Bales: right? And I can’t use um Deep Seek, but I’ve I’ve used them like temporarily. Uh and they work really great. They’re awesome, but long term they don’t, you know, they don’t want us doing that. So,
Executive Lounge: Well, if there aren’t any other questions, we will close this out. Thanks, Tom. Appreciate you coming. Actually, appreciate you going to DTC. So, we don’t all have to. Um, there’s a lot of us. Yeah. Um, cool. Let me go ahead. Oh, one other thing. If you are going to go um let’s say now see if your hotel room is really fast because you want to be closer. You don’t want to they do have a shuttle service but the shuttle service sucks. So you want to be in walking. Yeah. Okay. Cool.
Lorin Bales: Thanks so much, Tom. Really appreciate it. Yeah,
Executive Lounge: Thanks.