Transcription provided by Huntsville AI Transcribe
Alright, that was great. And I got to click this thing. Alright, well, this is the most people we have had in one of our weekly sessions in forever. I don’t think we’ve ever had as many people in one place. So I am curious before you leave, again in a while, let me know what it was that triggered you to come join up, because I want to leave more at that. So there’s that. I will skip the whole going around room and hey, my name is and everybody do because yeah, it’s not going to happen. So if you haven’t been here before, this is a sensible AI.
We meet up every other week, talk about AI stuff.
Sometimes we get in the weeds. Sometimes we stay out of them. Sometimes we almost get into trouble. Sometimes I write code or not. And just generally talk about what’s going on, what people are working with. Pretty much everything we do is wide open.
So if you hit, if you hit get hub, pretty much everything we talked about is there.
If you want to present something, we’re welcome to do that.
We do again, try to keep it open so don’t come share your super secret stuff. Probably not secret in the first place. If you haven’t Googled it on the other thing is, yeah, we just try not to do that. So tonight we’ve got a couple of things. We’ve got Josh going to talk about BLLM and then run pot, which I had never seen before, but I lost an hour of my life.
It may wind up being the thing that runs our transcription service by Friday.
And then I will hop into SBIR topics because that changed last week or the week before.
There’s some things that changed there on their schedule.
So we’ll cover that. And then Gary is going to follow us up with maybe the first time we’ve actually, have you given a talk here?
No, I haven’t. I was kind of on the track of the flaming. Here’s your graduation night where you actually get up and talk about stuff. So we got a conversation going earlier about, you know, we about the BLLM response for the hurricane and how chaotic it is up in the mountainous region. And I am not only the AIC thing, I’m also a first responder with the fire department where I live. And I’ve been through all the FEMA ICS training and all that stuff. I know how they do response. And also kind of live on the mountain where radios don’t work sometimes.
So I know a little about that. And then Gary’s looking at it from the data analysis side and what’s available and how do we put all this together. And so we made a lot of conversation about that. And then also those logistics and supply chains. Yeah. So a little bit of everything. So with that welcome, and I will turn it over to Josh. So just to start off, how many of you guys have heard of BLM for? I’ve heard. So a lot of people, a few.
So what BLLM, this is kind of similar as far as what it’s doing to something like Lava CPP or TGI, there’s few different inferences that are kind of like this. But it basically serves as a way to take a single model file and make it available for multiple clients to use.
Usually as some sort of a REST API server, and what a lot of these do.
Something that’s compliant with the open AI API so you can do the drop in for different use cases applications that depend on open AI.
You can instead use those with your own local models.
So that’s that whole host of this inference serving libraries that BLLM is one of those.
Now, there are lots of reasons to use BLLM.
It’s kind of, I’m going to say that wrong all night. The BLLM library over the other ones, especially in the past few months or so as an ad again, so pretty significant updates. And so we’re going to go through a little bit of that night and show kind of what it takes to kick up and use this thing and why you might want to do that.
So the biggest thing that we care about with this BLLM, sorry. Can you make that font a little bigger?
Yeah, absolutely.
One more.
One more?
Thank you.
That’s beautiful. Awesome. So the big thing with this server architecture is it is very efficient for continuous batching or batching of many requests at once.
So it has a special strategy called page detention that it uses to efficiently not load in the entire memory of the conversation.
So you say you have a big long message and it has a way of going through that efficiently.
It’s not having it load at all at once, meaning that it can serve much quicker. It also has a way of batching your tokens.
And I’m just going to stand up here.
Charlie, can you scroll down on your own pipe? So generally, whenever you’re sending a message to one of these models, there’s a fixed context link that they have as a size of whatever the sequence is.
And so if you say hi, how are you doing today to just a random unoptimized model, it’s going to say hi, how are you doing today?
And then you’re going to have all that rest of that context link, which is empty zeros that don’t have anything to do with it, but you’re going to have to send back and forth over the wire.
And then each time that you do that, it has to do that for each next token.
So you have all that empty, empty, empty, just waste its space.
So the big thing with with BLLF, you go down now, is this concept of dynamic batching.
And a lot of people have started copying this now, but these guys really did the best and first, and they’ve continued to kind of sail ahead of everybody else on it.
And the big thing is where the normal batching, they’re kind of doing this where they’re adding in, you know, certain segments that fit. But with what they’re doing, they’re actually even clipping certain subsets of those as well.
So they’re fitting every single little bit they can out of and doing just doing little parts of this and putting it over multiple GPUs.
And to get all this enabled takes nothing.
You want to do anything. There’s no optimizations that you have to implement. It just kind of works. This is why we care about this.
And this means is it’s really good if you are posting, especially multiple, you know, where you have enough people coming in, you maybe have your serving about 100 people or something like that. You know, it can handle all of this whole request or if you’re doing a job for you’re running, you know, maybe I’m doing a large scale analysis of a rag database.
And I want to analyze 10 documents at once.
And so I know I’m going to continuously get that buffer.
That’s going to be used case for the state building.
Another thing is if you’re at that sort of scaling, it’s hyper optimized for those ampere, give you use and up.
So it’s really good to have that level, especially recently with the 6.0 update and going forward, or continuing to optimize here.
You can see some of the recent latency improvements they’ve had between 5.3 to 6.0.
So this is a very efficient framework.
It works on AMD as well, TPU, and then I think AMD CPU as well, but not Intel.
So they support a lot of different inference sorts.
That’s the big thing about why we care about VLM. And here’s how easy it is to get set up with one.
And so I have here a little sub area where I took the picture of demo to the last week.
And I just kind of grabbed those with the new Lama 321B vision models that just dropped. And then another one that I have here locally on my.com called MiniCPM, which is an 8 billion parameter model. And so you’ll notice too, this kind of looks also kind of be a context test for, what’s that?
Ryan off the screen.
I don’t know. We usually don’t get on the screen.
Let’s see.
What was that?
The local 8B model.
Right.
So this is kind of a service that it says to show kind of comparison between the Pixel model we saw last week and what some of the other ones that are in its class and kind of where Pixel rise on some of these tests that we’re looking at.
Another big thing is that the, you see here the console AI Lama model.
This is something I have posted it over here on something called Run Hot.
This is something that the equivalent to, you know, like your serverless AWS, I think, paper space is one that you talked about the past. Yeah, we’ve done paper space, we’ve done Lambda. Yeah, Lambda is a cloud.
You know, how to get a GPU in the cloud, how fun is it to work with?
Run Hot is actually fun to work with. I’ve been using this for about 2 years, very vigorously in that kind of low, medium scale open source space. This is where you come to get your cloud compute.
And they’ve got really good options.
You see here, they have information about what to have running, the logs and all that.
And I’m going to show you how easy it is to deploy a Lama 3.2 instance. So here we’ve got a bunch of different options.
You see the price vary competitively.
But the real star of the show here is this 840, which they’ve got a whole ton of, and it costs 40 cents an hour to run a 48 gigabyte VRAM unit. That is pretty dang good. We’re going to use here one of my favorite templates, which is this PLLum by Trellis.
He’s got this easy edit template.
You can go in here, change what your input argument is, all that sort of stuff.
If you want to get it to optimizing, but I’m buying my own thing he has here. So I’m just going to go ahead and take those and get off. And so while it’s kicking off, I’m going to let it get to the other screen here real quick. So you see here, got the second one here.
This is the one that I have already running.
This one is kicking up.
So it’s going to do all that container stuff and we’ll check back with it a little while. So basically what it’s going to do is it’s going to download the image right now, getting that all set up and running on one of their servers. It’s then going to go and download the weights of the model for Pugging Base and then do all that serving stuff. We don’t want to wait for that. You just take about four biominutes for a model of the size.
So while we’re doing that, we’re going to look at the one that’s already running and kind of see a little bit about how much faster this thing is with that thing running over there versus the one that’s running on my Mac is smaller and is slow. This is not a bad Mac.
It has pretty good speeds as far as all that.
We’re just going to go ahead, just because I didn’t want to mess around with trying to copy and paste stuff and go ahead and look at some of the ones we did last time.
So what I have here is this is something called Open Web UI. You don’t know about that. So just to get us out of the deep inner notebooks, we’re trying to show stuff fast. I have it connected to this thing.
If you haven’t heard about it, go check it out.
It’s super easy to set up and use. So you can see there, it just kicked off immediately. It started going. Cuba, the one hand last time, had a bunch of issues with it. This seems to be more or less fine. Yes, what she did was ask it to describe what was in the picture. Yeah, I just took the thing straight from whatever you had in here. Right. Yeah. So looking here, this is the one that we were all impressed by last time.
And it turns out this might not be too hard a problem because it seems like they both don’t have one.
So you see here, it’s the one that’s on the computer is having a lot of trouble taking all this stuff. But the Laman went in straight ahead. Oh, it looks like the mini CPM did have trouble there. But more or less correct.
So you guys kind of see some of this one thing too with this open web UI has the ability to do different functions and aggregations and pipelines.
So one fun thing I have here is one that I’ve set up to merge the responses so you can take a array of multiple models and then compile them down into a singular model.
You can choose what that is.
That can be a different problem with Ireland.
Lots of fun stuff there just to kind of tease all that. It could be possible to like for Germany football for all to get the mean and the standard deviation of if you had six entries or in the GDP. Sure.
That’s right.
What is the mean lastly of the standard.
Yeah, the interesting thing about I don’t know if you’re using straight pictorial mean and standard.
But it’s basically all it has is that image of the graph of the, you know, the GPs of Europe in the different places.
And it’s got a whatever the table of content, whatever the legend over your right.
And it’s actually got to figure out, okay, Europe is in the is in the legend.
Europe is green. So I got to go find all the green things in the graph.
And then after that figure out which parts are which.
I don’t know.
It’s just the size of these.
He’s going to tell you how to compute the standard.
I told you to give it a chance show your work. Yeah. Change of thought. All right. So, oh, that’s pretty sure. Yeah.
Looking over here where we at.
See here it’s downloading and it looks like it’s moved over to the container logs.
And so just to explain a few of the things that are here.
So right now it’s loading the model weights.
It’s using a certain sort of back in called flash retention.
This is something to if you’re looking at serving at scale, making sure that you have the capability to do this flash attention. If you don’t know what that is and you’re looking into this, go figure out what it is and make sure whatever you’re running can do it. Because if you don’t, then your stuff’s going to run slow and stink. I guess the best way to say it. So I guess a while we’re losing this round, I do want to show loading this into the other you want to show kind of how easy this to connect it. But are there any questions about the project in general?
Just to keep on time, you know, I’m trying not to get too deep, but if you guys have things you want to look into the equipment.
So what kind of different models can you load with the LLM?
Let’s go look.
Generally, it’s pretty much anything that runs like the big models on Transformers that you have to run right their own kernels for these things because of the page tension mechanism.
It is a very active code base.
This is what they have for text generation.
You see most of the big ones.
I don’t really think there’s anything, you know, but you might have a few of the ones that are a little bit more odd like Java.
I don’t think Java’s in here.
Oh, it is.
So you can even have like the weird state space models.
So you just everybody uses the pixel.
They dropped their first support was on the all that this word has how they can go and print it.
Same thing with the point.
That’s where I came across it first was, oh, how do I get this to run? Yeah, so this is this is going to be around for a second.
And then quantization. So the conversation is actually there.
The big day. So they are really good, especially with the dbtq, which they have a Marlin kernel for that, which does like in eight and four stuff really well.
They also do a wq, which is the activation aware of conversation fairly long.
Yeah, they support all that.
You can do ggf, but it’s not as good. Right. That’s usually my first question is, hey, I found a school model.
Can I run and the answer is usually no, you know, and I’m struggling to go find a gpu thing or I can load it in and play with it and kick the tires and stuff. And then of course my next question, which if you’re not on the discord, if you get the email that I send out, there’s a discord link at the bottom. That’s pretty active channel.
It’s usually hot on there and go help. Yeah. Yeah, some of the other things to this support past what you find like a long cpp. They do have the ability to do some specific reward model models that have kind of a weirder back end to them, which is what this went to pass.
Where they basically does some some funky stuff on the back to force it into a certain type of reasoning, just to do the math abilities.
They also have multi mode models out the wazoo, which a lot of cpp does not for real. And so they are by far the ones who support that most outside of maybe like TGI TGI is obviously that’s going to have a lot.
You’ll also see they don’t have it here.
They’re rolling out audio stuff to a lot of things with speech to speech and audio in audio out whisper. So anything that’s that encoder decoder that started to add support for that. So this thing is going to make me look like bull and take forever to download.
Oh, I’m on the the.
No, I’m not.
No, that’s not working. So it’s not working slow, you know, maybe. Okay. It’s a three fresh.
All right, so you see here it’s opened up this little space here. This is not super helpful detail not them, but it actually is because it has what we need, which is the address to my little container over here.
This probably finished a good while ago.
I just forgot to use my brain. And so all we have to do to connect to this new instance is why am I looking over there look over here.
So here’s what my old one was and I’m going to pop in my new one that I just put there.
I put a V1 and connect.
And so before it said like tons of AI, whatever, whatever, which is the custom thing I put in there to show what we were using.
You’ll see that that is now gone because that doesn’t exist anymore as far as the state cares. And our unslawed. Lama vision structures here. And you play with the Python API yet for run pod.
No, not that much.
I know some people do do what there’s the eye through it.
I mean, there are people that run companies out of this thing. Yeah, so there’s a I haven’t hit it yet.
And there’s actually a Python library download and you can actually create tasks and jobs and whatnot and submit them and it’ll automatically go kick off an instance on run pod.
It’s been the thing up by itself do all the things come back.
I’m guessing you have to configure it somewhere. So that knows what’s available that doesn’t charge you, you know, for a big instance that come out looking for one.
Sure.
So I guess any final questions.
That’s most of what I have without going, you know, be down right. Very interesting.
Yeah, you have any questions about the LLM or run pod especially the LLM feel free to take me I’m on this board and all that.
The worst thing about it is I just started doing my presentations on Lama CPD.
Yeah, they are getting free to support, but I think that their vision stuff is going to be a little bit. Yeah. And this this does seem like it’s built much more for an inference server like you like if you had several users or we’re trying to spend something out that you were going to let people hit and didn’t want it to suck.
You know, that sounds like it’s got pretty much what you need out of the box rather than me trying to figure out how to do all of the things.
Yeah, I think it’s pretty much what most people use up until the point they’re writing their own. Yeah, so they open AI, they’re not using this.
But Mistral has a lot of support for it. They have use cases where they deploy it. So there’s a lot of cases where it’s good for that price grade, especially that hundreds to a thousand sort of thing, anything where you’re you can get away with like one server rack of a 100 something like that.
This is kind of their sweet spot. Obviously, the scale pass that to.
All right, let me reclaim host and share.
That way we can quit looking at ourselves. See the scale of the whole screen.
So what.
See what had happened with us.
I’m trying to think of when the nine.
So the last time we did a talk on this, the IR topics was, I think on 925 or 926.
It was the day they all came out.
So, if you’re aware of small business innovation research, and I never can remember what is something, something technology transfer. I messed it up on the 26 as well. In case you want to go back and listen to that. Think of this as a round of seed funding for small companies to do the type work, or in the case of STTRs to connect with a either a lab or a college or something like that to actually go. They’ll write something and try to get some funding from the government to do something. It’s kind of useful in two ways. One is that it can provide some funding to get you off the ground on some things. The other fun part for us is it gives us a clue as to what kinds of things the government is actually working for enough to put money behind it. So a lot of us have good ideas. And if you want to know if somebody is actually thinking about putting money behind it, this is not a bad place to look. So the thing that happened is there was a, let me see if I actually comes up.
Well, of course I do it live and you ask what this is going on here actual blankets do DSP IR, try that. There it is. So, previously, the defense oriented or the DSP IR and STTRs would come out in three rounds per year.
We usually get our first ones around March time frames when they, you know, you have to have your proposal in awards sometime around April, May, something like that.
Next round around August next round.
Well, I’m getting some of that wrong.
But they’ve actually stopped doing that. And now they’re dropping new topics on the first Wednesday of every month. So basically I have to quit doing my three sessions a year and actually have to watch this about every month.
So thanks for that.
I’m not going to talk about those folks, but it does actually help spread this things over over the years. If you actually can figure out, you’re not all one or nothing. Because some of us that work for small businesses, it actually costs money to write proposals for SBIR. So, I have a group of 10 that all come out at the same time.
And I really want to write to three of them, but I have budget to write for two. You know, it is tough. So maybe this will actually help from a small business perspective. This is where we wave our hands and lights come on. For those that aren’t regulars, the next time they go off, somebody has to go find the light switch that doesn’t look like a light switch and press the button that doesn’t look like a button. So, it goes off twice. First time we wave our hand, and second time we do the button.
Yes.
Yes, that’s a protocol here at Hudson Alpha. I know if I didn’t make it up. So anyway, there’s that.
What we do is we go through all the ones that drop.
We look for ones that are interesting from an AI perspective.
They’re pretty good at it.
Marking the ones.
Sometimes they’ll mark something without AI and it’s as an AI topic and you’re just kind of shaking your head trying to figure out what happened. Yeah. Number two.
Yes.
Enhanced explosive ordinance, the small school training capability. And what is the FBI doing on the arsenal though? Oh, yeah. They’re doing exactly that comment. Well, ENV is meant to bring in people from all over the country to go through explosive disposal training. So we need someone from the FBI to send that business. That would be good. What I’m going to do is bypass all of the ones that we talked about last time that opened in September. And I actually got several that drops on the 2nd of October. I think they’re still getting used to the new schedule because if you actually go look at some of these. It says release date was last year. The opening date is there. The new day is March 31st of next year, but we’re going to close November this year. So either they’ve created a time warp somewhere and we don’t know about or somebody hasn’t got the memo on new schedule. So I’m kind of expecting them to put out another notice or something to let us all know.
I think that’s first thing Josh said, but this doesn’t make any sense. Like, I agree. Like the lake inducted dudes.
Right.
So the other thing that some of these would cover tonight, we’ll go through this pretty quick.
It took me a little while to figure out their little encoding.
If you go back to the email I said, I got a list of them. The ones that have the D in them are a direct to face to. So I didn’t know that initially I was starting like, man, a lot of these are direct to face to and one of them in particular. I swear it has to be written towards a product or thing that already exists that they’re trying to move on. So we’ll cover that in a second.
It’s foolish. And what sucks is it’s one of the ones that I really would like to play with. Given some of my hardware background, but.
Person is interactive knowledge graphs for situational awareness.
For those of you that have been playing around with also you got to understand this is the sense part of the SDIR and their user interface is horrendous. So they have a rule that you have to have at least three scroll bars available at any given time.
So you have to know what you hear and then scroll up because you quit here and scroll up with this different.
So if you’ve been playing around with LLMs and doing a lot of ingestion of a lot of documents and building up the network so that you can actually use that to do something like a. Retreat log managed generation assisted thing, something like that.
You may have come across knowledge graphs.
This one is actually I was a little confused at first.
I thought this was initially to build knowledge graphs, but it’s not.
This is trying to allow users to interact with the knowledge graph after it’s been built and do things like investigate why certain things are connected to where they are. Maybe clip some connections and save it and then have that conform any kind of AI you do afterward kind of in a feedback loop. If you read into this, a lot of it is I think it’s right on the money.
One of the statements they make is we have a lot of AI assistants that build knowledge graphs for us and that we beat guys and a lot of documentation, but our analysts still trust it yet because they can’t see what some of the codes and they can’t look at it.
So yeah, this makes sense.
This is connected to this or this doesn’t make sense, but then at least click a link to go. Oh, they have the same part number.
Accidentally, this is where the waving doesn’t help you got to go. So this one’s a lot. What they’re looking for is a couple of user studies.
So if you have like somebody that does a lot of UX, something like that.
And then I think they’re, you know, the phase they got it split up phase one phase two. Yes, it just above the phase one the last sentence just where you said space one there.
The last sentence offer should plan to use their own data sense.
Yes. So the criteria of your own data sets has to be something that’s very close to what the Air Force is, or you can’t build a tool to do their work. Right. So this is going to take someone who knows the Air Force data sets and your own data sets. There are some pretty good references dropped in here.
I know from a, never sure how to say a, a, a, a, but I know there was a conference at UAH a while back where they did the, I think it was the yearly thing there. Anyway, so this one actually looks pretty good. There’s a couple of questions that are in this event.
So in other words, that some of these are pretty good.
Actually, there’s only one question. I bet.
No, sorry. I forgot this multiple scroll bar thing. So apparently when people submit questions, they a lot of times enter like 10 questions into one entry.
I’m sure to be, folks love that on the other side.
So interesting parts here. So you have to bring your own knowledge graph.
You have to bring your own engine that generates the knowledge graphs and your own data sets, things like that. I’m pretty sure it would be interesting to see.
I don’t know if they’re the Air Force is one that provides conversations with the topic authors. Some, some of these topics you can actually call the author of the topic and have a conversation. Others were required to go through the system and enter your questions, which are then published publicly along with any answers.
So don’t give away your secret sauce and your questions. Somebody go, Hey, baby, that’s not a job. So I remember the last one we did, they made us everything at the email, which I think was helpful. That was number one.
The second one.
So again, it’s interesting because the first one requires a decent amount of AI knowledge and how knowledge graphs created, how you may train something to build a knowledge graph and how you may create a feedback loop to actually take user feedback. Not necessarily a heavy AI task.
A lot of this is user, you know, user experience. How do I interact with something that nobody does much with?
The second topic is a little close.
It was a little similar. They’re trying to find some way to bring in multiple of. Let me bring this up so you can read part of it. They’re trying to build some kind of a system that takes inputs from multiple avenues, maybe, maybe army records for training, maybe your medical records, maybe your deployment history, you know, for Air Force in this case, as well as after you are leaving the Air Force and going back into civilian life, any kind of inputs you may have coming through that way and try to build some kind of a system to support not only you as a soldier or later as a veteran, but your family as they’re trying to, you know, move from place to place. Where this thing is actually trying to tie in the AI is actually somewhere down here. AI should be leveraged to, you know, they’re looking for things like automatically identifying, you know, private information or things that should be screened of anything you could automatically anonymize data, things like that. So this one was kind of interesting, but it wound up being a lot of things that I’ve seen before where it’s role based access, and how do I determine somebody’s trying to access something and other patterns and it throws it more in the cyber realm, which is, again, interesting, just not super in depth AI kind of thing. So I would apply an AI technique to do this. So it is pretty interesting. I have no idea where you’re going to find a data set that does this.
Maybe if you have a system that does aggregation of multiple feeds from different places, if you’re a company that already has something like this for your own people, maybe personnel records plus training records plus, you know, anything like that.
So I don’t see that kind of a point.
Let’s go find a GitHub repository where everyone’s at the original link.
Ask OPM for my data back there.
So that was one.
The others are unfortunately direct to phase two, which the direct to phase two, if you haven’t played in this realm before means you already have to have something good enough to be a phase one application.
Some of them actually write in what kind of technology readiness level this needs to be at.
Does it have to be deployed already? Does it have to be field tested? Does it have to just be, I mean, sometimes a small project is actually has more than one user may actually suffice. This one just was super interesting.
So I just talked about a little bit. They have ways when they filled equipment to put canopies and things over the equipment. If anybody here’s actually worked with things like that where you have a canopy go over something. So if you’re flying over on an airplane or from a satellite, taking photos out of a spear to make it kind of hard to find things that maybe in another canopy. We’re not the only people that do that.
Or the only military is or whatever to do that. I could imagine drug cartels throwing stuff over where they’re growing fields are something to obscure things. So this is actually trying to use AI to go find out where that kind of thing might be happening and draw a big red circle around. You know, I look here.
So that’s pretty interesting.
Trying to back out, you know, any kind of distortion that might be applied by these things.
That was pretty interesting. Unfortunately, I don’t know anybody that has anything out there that’s already has existing phase one.
But anyway, it was really, really interesting.
The other one, this is also the ones this is what I think is written almost directly to somebody that already has a product that’s halfway there. I don’t know too many other folks that already have something like this. I started my actual career at PDE electronics doing circuit boards and putting BSPs on the chips and then figuring out how to do the reverb, you know, things like that was fun. One of the things that this is trying to do in this case is the Air Force. They have circuit boards that were designed by companies that aren’t into business anymore. I mean, a lot of our fleet was designed a long time ago. I mean, it’s a SR 71 Bliper out there that’s probably older than me.
Oh, you know, things like that.
So where do you find the circuit designs for the circuit boards that they have?
You know, some of the stuff is on paper and stuff doesn’t exist anymore.
So apparently there’s already a product somewhere that will take an image of a circuit card and then reverse engineer a design for that card.
This is actually looking for an AI based rule checker.
Things that say, hey, you can’t put two amps through a quarter rock resistor.
You know, I mean, just things like that.
So that’s what this one’s looking for.
It seems like they already have a piece that does the reverse engineering of a circuit board and something which that sounds like the cool part.
But this is trying to build a design checker and without the first part, we can’t even get to the second part.
What’s the partial program?
That’d be a good question.
I don’t know how many of these were programmable back in, you know, when we’re looking at the parts of 1997. Let’s say that age. When did that start from about?
The whole?
I’ll say 1990. Okay.
The FPGAs or was that?
Well, no, they were.
They were kind of hard wire program.
In other words, there were set of things inside of the chip and you put it into a programmer and then you burn the wires you wanted to disconnect at the wires. You didn’t want so to speak. Right. And you could make a special function out of it.
That’d be, I don’t know what you do that.
You have to have, I guess you have to have it. And then I don’t know if that, no, the next one, again, direct to phase two. This one was also interesting. Yeah. Tracking UAVs, removing dark. Is this actually? Well, it’s, it doesn’t have anything in here at all about that.
There it is.
Okay. Again, the wrong store.
Yeah.
So this one is a pretty interesting one, but I have no idea how I asked, you know, the phase one and showing that you’ve got that.
One of those things where I believe there are a lot of ways to intercept or deal with a UAV, if you know where it is, but given that a lot of these things are made by plastic and fairly small, you know, finding them is not, is not easy.
So this is the one where they’re trying to figure out a way to do that better.
Unfortunately, I don’t have any knowledge of how to do that at a phase two level. I think that was the last one you see.
And then they also dropped a, I have that link here.
They’re still the same ones.
It dropped earlier. They did put out a new direct, basically an open BAE, which is one where if you have a topic that isn’t on the list that you think is applies to the Air Force, or something you built that might have applicability there, you can actually propose your own top and go that route. You can actually get a higher level of funding because a phase one, I didn’t look at the, I haven’t had time to go through and check to see what the funding levels are on their phase one, or if they’ve changed as well.
Generally, you’re looking what, 140, 150,000 for a phase one.
You have to check to see, sometimes it changes based on things.
Phase twos are generally the million dollar level over about a two year timeframe.
So there’s definitely a big difference when it pays one in the phase two.
But sometimes if you’ve got an open BAE, you can bring something in.
If you’ve already had a prototype of something you built for some other purpose. A lot of times it’s an easy way to jump directly into a phase two funding.
Anyway, any questions about anybody else?
I know some of us work for companies that do SBIRs. Some of us work for large companies that write topics that they’re, I don’t know, have you ever been asked to submit a topic on something you’re working? No. I actually got asked that one time on a project we were working about, hey, from the government customer, hey, we need to submit some kind of a topic, a small business. Is there anything that you need that you think could be done, and a low level that would be fun? And yeah, it’s threw it out there. Next thing you know, it’s an actual, they wrote it up, put it out there as a topic. So it’s kind of interesting how that works. And then, okay, any questions on any of that stuff?
Oh, the fans of the 3D assessment work at TrueStar. Which one was that? We might have covered that last time, but maybe it. I missed a couple of these. Did it pass?
Okay, yeah.
I think it was going to pass. I think we’ve collapsed some of these.
Because I definitely want to cover stuff you’ve been working with. And I don’t know if you have anything you want to show or to do a read to? The link that I said again.
Facebook post with the map and then the data dump and the next post. That’s where I called the information from.
All right, let’s get that because we’ve got 12 minutes left on this one.
I’ll make it a fine. New chat, Facebook. Nothing like bringing up your own Facebook. I’m going to live feed. My stuff. But as we were talking earlier, my, for me, generally what I do at work and what I do personally, I just kind of merge this together anyway. So it’s, yeah, I’ll have to go find out where I put the other events. Make.
And thank you.
Tag on this.
As far as.
Maybe not.
Session.
I know we can find it.
Let me go back. There we go.
Yeah, if you really like Facebook, we also have a Facebook group if you want to join. Facebook also from a group perspective, change the way they do admin. And now they have an admin assistant that sucks. That just lets people right in without me approving or anything. So if you ever see for one thing.
These fake accounts and stuff for some reason they haven’t figured out that the first name and the last name usually aren’t the same word. So I keep getting like Monica Monica asking to join our group and David David and Adam Adam and only it’s a no but sometimes Facebook besides to automatically the event. And I think somebody there’s something posted last week that did you report at time. So that just wouldn’t know I had to go back to say how did this person even get access and oh yeah automatically added from their AI admin assistant. Yay. So, yeah, we generally try to keep people from hopping on and just trying to sell stuff from people confused market place and just generally you’re trying to, you know, the last one was somebody trying to clean that work or something that has nothing to do with that. Anyway, so you want to get back down there. Okay, I thought I gave you the links that I pulled all this stuff from that’s okay.
You were trying on time.
I’ll just get to the map. Okay, this guy. This was built with Google Maps.
It’s open and shareable. I spent a lot of time at night. Pending all these locations based on what I can pull from do to market transportation data from Tennessee and North Carolina. That was at least in that region that was affected.
If you pull up the DOT for North Carolina, the entire state is covered in orange.
Oh, did you ask what you saw?
So road conditions drive this one?
I think so.
Yeah. Very helpful.
Yes.
Very terrible to try to use a mobile on a cell phone. I was looking for things that people could use on the fly because we know that so reception there sucked on a good day before the hurricane. So you can use what I’ve got in offline fashion. You can also export the data points into other offline useful maps, but it takes a techie to do it.
And a lot of people there on the ground, I mean, they have high school educations, you know, they’re not really setting on spreadsheets.
That’s why I had to spend so much time manually calling this data.
That’s where I think some geeks or only could, you know, just help out a little bit here and there.
This is more for travel for people from outside the area. Emergency responders, firemen, whatnot. Construction workers. Everybody there on the ground level has what they need, but people traveling to the area are getting lost. They’re going to turn away after they brought a truckload of stuff to drop off. The new, the logistics of it. We’re going to quickly become a mess and over the next few days, that’s what happened. That’s why I was keeping an eye on that and building this. The purple monsters are all of the affected areas on the power of water. The flood markers. Same thing. Some of these were bridges. Yeah, yeah, I think the flood markers were the bridges out. And the yellow houses are distribution centers. The headquarters for it is in Bristol, Bristol motor speedway. They’re about eight miles south of Bristol, Virginia. They’re trying to get people to go there and then get what they need and then go south. But if you’re coming from the East, you know, people are trying to take shortcuts and go to these other areas. But they’re not talking about locals first. So that’s another reason that I pin this stuff.
And I get contact with information in there. For locals.
Terrier organizations. Churches, fire departments. Walmart is also open all their parking lots for. Laundry showers stuff like that. There is 2 weeks into it now. There’s still 200,000 people without power. This is really morbid.
I don’t want to say it, but I have to. Some percentage of that is going to be displaced people. And there’s no telling how many of those are still there to get power. So some percentage of that is probably going to move into this area as displaced people, you know, looking for places to rent. But if you want to go from here to there. It’s about 4 hours from Huntsville to Bristol. Actually about 5 hours from Huntsville, Bristol.
If I’m trying to highlight toxic areas, I’m only got 1, but I know there’s more. Bring deep, me high boots. I get you to send 95 respirator with filters.
You don’t need goggles, maybe, maybe not.
And clubs. You know, some of the fishing trips you’re ruining your clothes just from all the stuff you’re in. Whatever clothes you take there and hopefully they bring them back. They’re. They’re quiet about it because they’re still assessing it. I know they don’t want to cause a panic, but I think down to my, this was a mistake personally. But. Most of the folks, so I’m.
Side sides, my side thing. I’m also a landscape photographer and I spend about 2 weeks a year up in the high country of Northern Carolina, right where does it. Around there, L valley, crucial spoon.
Blowing rock, you know, then bill, spend a lot of time in Lendell Gorge. Most of my contacts up there, I didn’t hear anything from for like 2 days. And then 1 of them is actually in. I think it’s black.
It’s just east of Asheville.
Actually works for a children’s home there. And he couldn’t even get any kind of signal or anything at all just to let people know they’re live. The 1st thing we got from him was. We’re probably all okay just nobody can talk right now. You know, I can’t get any signal out can’t get any information in. We’re just trying to keep, you know, we’re just trying to keep people alive right now. So. The whole distribution network kind of a thing all dependent on how far you can walk.
Because with the terrain there, the amount of rain they got it had been raining for 2 days before clean it. And then that dropped another 2 days worth of rain. So some of these places got around.
I don’t know what the max was, but think of 3 feet of rain in about 3 days. And all coming in the strain when I’m done washing down. Washing roads out so you could get like from here to there and a vehicle anymore. I don’t know if you’ve seen the videos of the the KG days. They were on the scene and you know down in the Orleans places with flood and they have boats and stuff. They’re running strings of packed mules in the mountains. So the name is on the mountains. Of course. Yeah, they are there flying drones here and there.
I can fly anybody else.
No, I don’t we had we’ve had the guy here for a while that was 1 of the drone pilots that get on the arsenal and do fun stuff. That probably would have been helpful to get access. That’s kind of the reason we want to talk about it tonight. We got a lot of smart people here that are really good at cracking data and doing interesting things with data visualization and whatnot.
The hard part about it is the you can do all of the data stuff and visualized and why they can’t see it. They they have no internet with most of my contacts do now. But they spend maybe five days without any kind of internet. No cell phone signal cell phone repeaters don’t do any good there because of the terrain. Because you know, you got to have to hop to hop to get things satellite times is also interesting with that kind of terrain. If you don’t have something directly overhead.
You’re trying to think is there if there’s anything you can think of that just kind of an after action might have been able to be done to help guide people because there was a lot of support trying to pour in. Actually, the city of Van Ryl actually hosted on their Facebook and and Instagram and other feeds. We understand you’re trying to help just please stay out right now. Is unless you are self sufficient and bringing enough food for a week and a place that’s later on dead and up water for you. You’re actually accessing the resources they had left. It’s like a city of the music festival. Their population triples for a few days, but this is going to be a lot of company. So it’s probably much more useful now moving forward trying to track because that’s the other thing is there was no.
I was getting most of my information on various people’s Instagram feeds, which is not a place to get news. But you see things reposted and they were actually from, you know, reputable because I know where these places were. You see such and such fire. Okay, I know where that is.
I know they’re probably correct.
You know, because if you want to put the wrong place and have food or you just walked five miles for nothing.
You know, so just thinking through how can you how can you put together disparate information to how do you know you can trust it yet. One thing you can pull this up in Google Earth. And you can play with topography maps. It’ll show you a bit more detailed really awkward detail. Up or left the three dots. Here.
Yeah.
And the map of rulers.
Oh, and we flash past it just go above that was this download KML.
That’s where you get the file for the day points.
Now this got a little thick with text, but it’ll I can have it with no labels, but I was still working on it.
It did get very thick and kind of wordy. But if you look at the terrain there, that’s what you’re looking at. You’ve got this hilltop several feet down several hundred feet back up the next hilltop.
And then people live down in between. You know, they got washed downstream or lucky enough to sidestep the new room that came through their area temporarily. Yeah, that’s an elevation. Did you saw the thing from the Unicorner where they had helicopters picking people up off top of the hospital? That’s right where that was. Yeah. There’s a bathroom out this sort of in that door and then have a head to the right. There’ll be a hallway and you go way down the hallway and it’s on the right. Yeah, it’s what’s the same. It’s on the left. The other fun thing there is you’ve got two different states right inside the building. Trying to figure out what your station covered. It’s interesting.
That’s part of what I was going to sort out. But this is a prop. One state is going to have one set of resources.
The other states is going to have one set of resources. So on the Tennessee side, you’ve got a really steep incline to get up into the mountains there. It’s almost like farmage down. And when the storm hit, basically as the rain went down North Carolina, that’s the rain went down to Tennessee. But a pile into a more narrow area coming into Tennessee. Working its way down into the Tennessee River, basically. We’re at the other end of where the state of the TVA. We’re 218 miles away.
So you’re very likely going to see some residual effects of it with the water quality. At least in the Tennessee River, I think. But we’ve got the filtration sources to do it. A lot of them are on wells. So IEEE has the move truck.
And this is a group of volunteers in trade on how to operate the move truck.
Right.
And they bring a truck in and it’s a big truck.
We’re talking tractor-trailer kind of things. Right. And then they set up an antenna and they go to a satellite and now they provide radio services for emergency for the police and the fire department. Right. So when they have a hurricane in someplace in Florida or Alabama, down in the Gulf Coast, the move truck shows up. And now they’re able to help provide communications. Right. The problem with this area is you need 40 move trucks.
Yes.
So here’s a question.
If you took the high schools and these, if you just make, it’s going to happen again. Right. Let’s go put a concrete base with some high schools. It’s the high school out near the football field. Let’s put a concrete pad in. Right.
And let’s put a base with nuts and bolts on it and just leave it. And then somewhere else, we’ve got radio antenna towers that we can bring in by a helicopter and set it down. Yeah. And have a package. This package is battery powered and remote. Once it’s set, it’s working. No one has to be there. So I’m just saying. I think you got a couple of things.
You got one is a comms infrastructure, which is how do you get that set up logistically, you know, if you don’t have the base, you can’t put the tower. Right. Can’t bring a tower.
You can’t wait for the concrete to cure. But if you’ve already got the base is set up. Right. And you’ve got to do a topographic path as to where the high, it may not be high schools, but right.
Some place where we can start having emergency service coverage.
Okay. Over an area if someone was to design that and explain it, you might get somebody to fund it. Right. And then.
I mean, National guards or Air Force can fly in towers and they can be set up in a day or two. And you can start communications with people again. Schools that are there currently take the shoulders with schools that are still still there or we’re going to show. That’s what that.
Super option is just going to be silent next game.
That’s probably.
I know every iPhone or 15 up now.
Yeah, 14 15.
So they have a lot of time to start. Well, a lot of the time star length ones that have it, but even with star length, you know, if you’re, I mean, you’re pretty far down in this terrain. And if it’s not my close, you know, close enough to overhead talked about the kitchen.
Maybe I was thinking it’s easy enough now for people to get a mobile star link for their RV stuff like that. Right. You can get a volunteer network that can go to places and create.
Where they see get a good connection with the sky.
But more.
Yeah, yeah, just have volunteers like I’ve got a RV. I can put some food in there. I had a star link. I’ll go to this location and all you have to do is map out where all we need these volunteers to go for communication.
There’s something similar that I was looking at called mesh tastic, which is something similar to that, but not using satellite. It just, it allows you to set up your main base and then every person who connects to that creates a mesh network that just builds on itself. And there are global phone apps that can connect to as well. And it’s an open source also. So 1 side is the content structure.
The other side, which I think is really where I could come in is how do you take reports and messages and things of, hey, I’ve got, I’ve got 2 cases of water at this location.
Or, hey, this place is a shelter and it’s open 247. It allows pets.
This, this place is how do you take all of this information and find a way to battle it.
It’s true.
You could do something for like social networks, like let people upvote or downvote and do some kind of a something like that. What would you tell them to a cell phone or what would you?
The what the.
Tell them to a cell phone or what would you.
Now, I mean, how do you know that it’s true? Or how do you know that it’s timely? Well, or how do you know they’re not out of water by the time you get there? Just to verify the source of where it came from. Now, I mean, not just where it came from, but it’s true or not. Or if it’s still true. I just. But honestly, you can tie in a blockchain token to each object.
You can just verify it.
Okay. This token was tied to this pallet of water that took it was suited to this person. So now it’s, you know, out of the system, but it’s way up there.
It’s way up there. I did it. And at this level, it would have to be some kind of distributed mechanism able to disconnect and then reconnect and be able to manage the, you know, actually reconnect what do you, you know, what do you have? Anyway, just, it was a really cool chain of thought we had going on. I don’t know where we were talking. I was just, you know, I just found this need for this thing. And then I started back with keyboard and just kind of took a life of its own.
People, people in the area already had it. So they said, thank you. Cool. You know, I think it’s quite useful.
I don’t know.
There’s no thin metal, thin metal or paint or anything. I didn’t ask for money for doing anything else. It was the one website that was actually asking for help there or validating that their data was right. You had to get hub IO page. Oh, yeah.
Yeah.
Yeah. Yeah. That’s right. That makes it work. Yeah. But then again, that was on the TEPC site. You know, where they, uh, crossings, I think, make Bristol, Johnson city.
And he brings up the other thing was, uh, give up, um, Home Depot basically really needed the business functions with get up. I think a lot of the projects will probably be, you can see some stuff they have the streamlined or the distance. And you’re seeing a lot of the same problems here, but you know, they’re essentially a fallout area with, with what this is.
You pour too much milk in a cereal bowl and that’s what basically happened here.
Yep.
Places where, um, used to, you could just drive and now you can’t and, you know, the blue or the fire. It still exists in name, but it’s a full show.
This thing that’s going to be 10 years to rebuild it. We’re one of these places are the main roads that cut off to get to it. So there hasn’t been a plumb all over the house.
I haven’t been able to pause those yet. People have posted like pictures of what they’ve enabled to write down, but there’s a total life spreadsheet. There’s a lot of people.
So, thanks for your comment and being your source of proof since you have a picture. Yeah. So, so you mentioned earlier that people are getting burned away. Are we cashing the data, but people are getting burned away on that traffic stops? Um, they’re getting turned away from locations that have too much stuff. They want to start for the one problem.
The bodies aren’t prepared to go there through it.
Well, I should say the body don’t know that they’re walking into. So, if there’s over abundance of supplies in one place and there’s under. That’s going to be valuable to these people is where can I go to find stuff?
And then the people that want to bring stuff, where can I go to drop the problem? That’s what we’re using negative to actually figure out where the instead of looking at it from a supply perspective, you can actually look at it as a. If I bring supply and we don’t need it, you can actually look at it.
It’s almost like moving the whole instead of moving the piece.
So, that already exists.
It’s Google maps.
When you point out a traffic cop is pulling people over. Right.
I think that infrastructure, I guess you just take advantage of that and say over abundance or a need. And just call it out.
And then maybe I may won’t match up with business necessarily a little bit close.
So, I can see.
So, thinking about, you know, you have this, this application here.
Obviously, you put in the data for this manually.
So, how did you verify that these were accurate?
I went for brick and mortar locations, not people’s homes. A lot of these brick and mortar locations were schools, fire departments, churches, local businesses, a lot of local businesses for restaurants and whatnot are opening up. How did you get that information? A lot of it from Facebook. So, I think you combine, you know, so you have your data in from Facebook, you have, you know, like he has as far as the information, whether serving that I have an over abundance of things. What you can automate with AI is taking that unstructured data and putting it inside of here. You know, getting it so that on this pinpoint right here, you know, you have a human verifier, but it does all the stuff.
That’s where it’s going to be reliable enough that it’s not going to cause more trouble than it’s worth without training.
You can train a model to do all that stuff too, but that would be your quickest ones to take something from all of those inputs into like a JSON stream that interacts with the API.
Another issue is you’ve only got so many layers that you can put on the Google map.
I’m going to max that out. I think it’ll just like, you’re 10.
Got you. I’ll just do a little bit.
I’m not doing another world one application.
Nice.
Let’s, let’s close it out with that.
We are a little over time.
Let’s find out.
A lot of cool stuff. Let me hit stop recording. After I figure out where that went. We’ll see you in a while.