andrew severin iowa state university genome informatics bioinformatics ai workflows

I'm going to introduce Andrew and he's going to get started because he's got a lot to show you. Hopefully we have a lot of great back and forth today. Questions, Q&A, that sort of thing. You ready for that, Andrew? Yeah, we're good to go. All right. So Andrew Severin is the director of Genome Informatics Facility at Iowa State University, where he develops documentation computation-centered computational workflows that help researchers navigate complex data and analysis. You should not do it in a complete lap as fast as you can before you come in here and run out of breath. Oh my goodness. With more than 20 years of experience in genomics and bioinformatics, his work now focuses on how artificial intelligence is transforming the way we interact with information. Shifting systems from passive tools into active collaborators. In this session, Andrew will move beyond the idea of AI as a chatbot and demonstrate what it looks like to embed it, to embed AI directly into real projects and workflows.

problem solving troubleshooting chatbot usage daily work practical application

Drawing from his work on the 100 Days of Bioinformatics project, he will show how tools like ChatGPT and Claude can be used inside development environments reduce friction, accelerate work, and support practical day-to-day problem solving. Please join me in welcoming Andrew Zeperant. Hello, everybody. I'm listening to this, and I'm just like, that's the most boring thing that you can possibly describe what I do. In all honesty, I think of myself as a problem solver, as a troubleshooter, and I think pretty much everybody in this room Can relate with that. We're here to try to solve problems, trying to figure out how to use this technology, and this technology it keeps advancing. In fact, when I was writing this, I had originally told them, Paul, there's nothing new. I'm not doing anything. I'm just applying the chatbot to what I do in my... my daily work. And he said, that could be really useful. There are people that want to hear about that. I'm like, really?

openclaw ai agents job displacement existential crisis bioinformatics automation

Like, no. I mean, and then I, heard about, I don't know how many people have heard about OpenClaw and agents and stuff like that. When it came out in November, I was like, there's no way am I ever going to let something like that use my computer on its own. That's like the most dangerous thing that I could possibly think of. I ignored it for three months. four months. And then basically all social media exploded. And I was like, holy, I cannot believe I'm this far behind. And I was starting looking at it. And there are other bioinformaticians, people that help faculty analyze their high throughput DNA sequencing, RNA sequencing data to try to solve these biological questions to help improve agriculture or make drugs or whatever else. They basically have agents or skills now, which I'll explain in a minute, that can do pretty much everything I do. And I had this moment like, I'm going to lose my job.

genome informatics facility high-throughput sequencing dna sequencing rna sequencing data analysis

You know, this feeling that AI is going to replace me, which if you asked me six months ago, like, no way. There's no way. I have a knowledge base and I know enough that there's no way that AI is going to replace me. And then this happened and I was just, It was like an existential crisis. I don't know how many people have had that kind of feeling with all this AI. So really this talk is about my, it's like multiple things. It's partly my journey, my implementation, and how I'm using AI in what I do on a daily basis, which I think is generally useful. It's also, I guess, an introduction of where do you start, where is the second step, and where is it going? So all of that will kind of be part of that. What I do, I've already pretty much explained. I help the faculty transform all of this large data. Back in 2011, when they hired me and I founded this facility, they would send data to...

data pipelines documentation software evaluation ai adoption research facilitation

the genome core, and they would sequence it, and they would give this data back. The faculty would open that data file in Word or Excel, and it would crash their computers. And the university is like, well, we should probably hire somebody to help them figure out how to analyze the data. And that's how this facility came about. That's how my job came about. Since then, I've hired five more bioinformaticians that help analyze data across the university and at the USDA. But with the recent advances in AI, I've realized that that's not going to be sufficient to be able to maintain salaries for my people, especially with the current funding environment, and that we're going to start needing to pivot to help faculty and staff use AI and these tools to help advance scientific research. It's no different than what I've been doing before. Basically, I download a program off the internet. I figure out how to use it, and then I can explain it to the researchers or the students. I create pipelines, like raw data in, transform it, make that transform data the right input for the next program, boom, boom, boom, and then they get multiple outputs.

software documentation tutorials tool evaluation scientific research undocumented software

I make it look pretty, beautiful documentation. can be okay. The quality of the documentation, maybe less so. And so I'm in charge of finding the best tool to help my researchers analyze their data to get, I guess, closer approximation to truth for the scientific question. You typically with software that has no tutorial or very bad documentation. Well, we're currently living in an era where we have these amazing tools, but nobody knows how to use it. So to some degree, I feel like I'm pretty well qualified to be able to learn how to use a very under-documented tool to help solve problems. The first thing back in 2023 were the chatbots.

gpt-4 chatbots prompt engineering iterative feedback context provision

GPT-4 came out and everybody was talking about it. And the things they could do were amazing. If you have not learned how to create your own prompts, do prompt engineering, this is where you start. This has not gone away. This is the foundation of skills and agents and everything downstream. Because in order to write a skill, you have to be able to prompt it well. You have to make sure it's executing your instructions the way you want it to. And a lot of this is about loops. You ask a question, you get an output. Do you like the output? No, that's not what I meant, and usually that's because you didn't provide enough context. I have researchers come into my office and they start talking about their problem that they told me like 8 months ago, and I have no clue what they're talking about, and I'm like, Let's step back and tell me the biological question you're trying to answer. there's lots of technology, a lot of stacks, lots of sequencing, whatever that you could apply to a problem. But I want to know what is the thing you're trying to solve?

context provision identity files claude user background biological questions

What is the question you're trying to answer? And I think that's also true here. And you have to tell me more about what you are, what you know. I can teach people how to use and how to do bioinformatics. But the first thing I ask is, do you know how to use Unix? Do you know what a terminal is? Have you ever done any programming before? You have to tell these kinds of things also to the agent, because they don't, the program, the ChatGPT or whatever you're using, the large language model, doesn't know what you know. And so now people are actually creating identity files in their Claude folder that says, this is who I am. These are all the things I know. And then every time you start a new session, some part of that can be uploaded, and then you don't have to tell it every single time. And some people have things like, these are my current objectives. I'm trying to make this program, or I'm trying to make my company more profitable, or something like that. And that is something that could be updated on a more regular basis.

iterative refinement vs code markdown retrieval augmented generation rag

So this kind of loop structure where you keep asking it to refine to get a better answer until the final output, you're going to see that over and over again as we get to skills and agents. After that, The thing that I was going to talk to you about, but I've kind of tried to expand upon that, is talking with your documents. So I use a program called VS Code. It's basically a text editor that uses markdown. It doesn't really matter. Basically, a pound is title, pound pound is subtitle, a bullet point is a star, you know, stuff like that. It's a very, very simple way to mark up your text documents. It also happens to be what is a lot of the AI agents find easy to read. So a lot of this stuff. So if you have things in text, it prefers that over a PDF or something like that. So you have to convert that to some kind of a text file. And you can convert any folder.

folder chat rag system jekyll minimal mistakes javascript

We all have folders with all of our documentation, all of our information. You can convert any folder into basically a retrieval augmented generation system by plugins that you can just chat with the folder. And that's advantageous because there are times where I'm working on a very large website. It's Jekyll-based, minimal mistakes, basically lots of JavaScript and Jekyll code and my own information, my tutorials I'm putting in there. And they're all asking questions like, where do I... where's the file that allows me to change the color of this thing? It's a huge database, or not a huge website, and I might not know where that file is. I might have known at one point, but now I can find it immediately. And I can ask it to do tasks. Like, oh, can you add this information to this file? And I can keep staying within focus. And I'll talk about that more. So this is what it looks like. Let's see if I can... You've got your folders on the left.

vs code plugin task delegation github tutorial glossary management markdown formatting

The markdown text here, and this is what the markdown would look like. Again, a single pound gave it a big title with an underline, double pound. You got the bullet points here with dashes, et cetera. Not critical, but you have something like that. You can have a plug-in that basically gives you a chat window on the left-hand side here. And you can ask, in this case, I was writing a tutorial about GitHub. And I'm like, oh, I should add GitHub to the glossary. Glossary is a completely different file. And it's alphabetically sorted. And so in order to add GitHub as a term, I would have to think about how do I want to describe GitHub. It's a repository that you can use for software and text and things like that. It's version controlled, all these other things. And then I have to code to the glossary. I'd have to find the correct alphabetical order, insert it, paste it. That completely breaks me out of my flow that I was in for my tutorial. So instead, I had this thought that was going to distract me from my tutorial.

mental energy flow state task handoff coding agent local agents

I went over to the chat window and I said, hey, can you check if I was in my glossary? That's something I didn't know at the time. And if not, I added to the glossary at the appropriate location. Now granted, I sat there and I was looking, just staring at it, like, We'll do it, we'll do it, and of course it did it, but at this point, and basically, does this look good? I'm like, Yeah, that's good, or I could have corrected it, but I didn't have to be staring at it, but you could literally hand off that task that was keeping up mental energy and keeping me from breaking me out of my flow, but I didn't want to forget. and finish where I was in my tutorial and then go back and verify that I worked. And that's the kind of thing that we can now do with these basically local agents.

andre karpathy wiki generation knowledge curation document extraction rag optimization

In this case, it's a coding agent is what you're working with. So beyond these retrieval augmented systems, Or generation systems, which is basically you're taking a folder and you're making it so you can chat with it. This guy came up with, you may have heard of Andre Karpathy. He's really big in the AILM field. He basically just tweeted out or posted something that's like, you know, I feel like we're doing augmented generation grog. I feel like we have the source documents. I feel like we should just have the large language model extract the important information and create a wiki. And then we can just talk with the wiki. That way, every time I ask a question, it doesn't have to go through 1,000 documents. It can go through this much smaller curated data set based off of those documents. So you still have the ground source of truth, but you have a wiki that was generated from that.

obsidian vault conference networking knowledge graphs speaker connections json files

And then you can have the large language model or these agents curate the wiki. So you don't even have to manage it. It's like, you know, I think that might work. And so a day after that, there was like 20 people or 30 people that had already created, oh my God, I got it working. And honestly, nobody's got it working yet. It's more complicated. It's really cool. And I can show you later after this presentation, because it's not that long. It's a demo, so I'll hook up my computer and I'll show you. But what people are doing, it helps with linking entities. And what I did was I said, He kindly gave us all these nice JSON files. So I just said, hey, Claude, can you create an obsidian vault that has all the connections for people, the speakers, what they're talking about, and all those kinds of things. And you can ask questions like, since you know who I am, who should I connect with?

entity extraction knowledge graphs character analysis topic identification computational queries

Of the speakers, who would be most interested in working with me or I should connect with because they're doing cool stuff similar to what I'm doing? it completely changes how we do conferences and how we interact with networking. Because not only do you have that, you can ask those questions, but now you have little profiles of everybody that might be at the conference and a picture and stuff like that. And I can show you that later. But the main challenge that we have with these curated knowledge bases is you have to extract data. and the extraction process or entities, like what's important. So if you were looking at a series of books, say like Harry Potter, you might chunk the book into different scenes. And in that scene, you'd extract the characters, the items that he used, the places that they were at. And that, once you have that, you would have that in the Wiki.

domain expertise entity calibration extraction pipelines document types manual curation

It'd be these pages based on scenes. You can't ask A rag system, something like, How many times was how many scenes was Harry Potter in? It would fail, but if you had pre-computed that... You can do that. You can actually compute over the Wiki, and that's where you'd want to extract the useful information and every data type you have in there. So that was a fantasy book. But what if you have a non-fiction book? That's going to have very different items that you might want to extract. Maybe it was a textbook. Well, each chapter might have useful information that's relevant to the topics that were covered. So chapter, topic, et cetera. But you can see the interactions of these entities and you can compute over them. How many places were there? What was the place that the majority of the characters interacted at? Those are the kinds of questions you'll be able to do. And you could do that with any of your documents and information that you might have in your companies.

error handling contradiction detection scientific literature gene expression experimental conditions

Some of the challenges are calibrating those extraction pipelines. Every time you get a new type, you're going to have to, or data type, you're going to have to have different extraction entities. What are those entities? Are they really useful? You can have it do it, and that's what people are doing. The quick and dirty, it's like, oh, Claude, just extract useful entities. Like, Claude doesn't know, or ChatGPT doesn't know what those entities are. You know, because you are a domain expert, and you have to be doing that manually. That's what's taking the longer time for this. Sure, these large language models, these agents are super powerful, but if you want to actually significantly increase your productivity, you have to take your knowledge and insert it into the agent so that it can do what you're doing and do it well. Error handling, like there's duplications. contradiction detection and resolution. So if I was to extract information from scientific literature, one paper might say gene X does Y, and another one would say, well, gene X actually does Z.

wiki scaling structural consistency template enforcement content organization continuous expansion

Well, are they right? Maybe it goes up or down. How do I know which is the correct one? Maybe they're both correct, and it depends on the scientific or experimental setup. So that resolution of those errors is going to be important. It's not necessarily just going to be, oh, well, the consensus is this, so it must be this. It might in some cases, but you'll have to figure that out depending on your particular domain. And then being able to scale that. Because at some point, I'm finding so far is unless you have a very strict template in your wikis, as you ask questions and then you add that to the wiki, can you add it in a way that it still fits that structure? It starts getting messy and kind of all over the place. And I don't think anybody has a really good solution for that yet. After using chatbots, the next level up, which people are just now figuring out, and I mean, I'm only like 3 weeks or two weeks into this because I was ignoring the field, is our skills.

skills definition computer access prompt engineering web search file ingestion

And I think of skills as prompts with computer access, okay? So You have prompting, so you have to know prompt engineering. You have to be able to create a good prompt that does something you want to do. Skills are prompts for computer access. Technically, you're using skills or agents as well. So ChatGPT is a chat agent. Those agents are executing skills on your behalf. A web search, file ingestion, image generation, those are all skills. that was developed that you're using already on a daily basis. You just didn't make them yet. It's kind of like, these 10 prompts will do everything you need to do. That was such a dumb idea. Now, if you know these 100 prompts, here's a prompt library. Same thing's happening now with skills. People say, I have all the skills you will need. No, ignore all of that. That's just the hype. Learn how to write your own skills.

skill building iterative feedback claude chatgpt draft refinement

And in fact, this is getting easier because you can actually ask, Claude or ChatGPT, they have skill builder skills that will help you get that first draft. And then people are using codecs and code. Those are coding agents that are code execution, executing commands, editing files, and so forth. And as I was saying, these skill builders, that's another one of those feedback loops. You're going to describe the process, what you're doing. It will build you on a first draft of a skill. And then you say, let me look at this. You say, oh, I don't like this. I don't like this. Can you change the skill to make it better? And then it iterates. You keep iterating until you have something that gets to 95 plus of what you want to do. And my example is, I've been working on a project. It's this website. It's 100 days. So a lot of times we have these workshops, you may have gone to a bunch for AI or whatnot, that they promise you that you will learn everything you need to know in this four-hour workshop, or maybe it's 4 hours over 2 days.

100 days project spaced learning knowledge retention workshop alternatives incremental learning

The truth is, it's just an information dump, and unless you're using it on a regular basis going forward, most people will leave that workshop feeling great, and then two weeks later have forgotten everything and it's just gone. The way you get around some of that, you see these Websites are like 100 days of Python. And basically every day you spend an hour learning a little bit of Python, and then it builds and builds and builds and builds. And after 100 days, it's like riding a bike. It never goes away. So I'm trying to do that. It's been taking me about a week to go to create a single day as I envision it. So I'm kind of this, I have the broad idea of what I want. So I feel like I'm an architect and I know how to do it. I know how I want to do it. You're like, well, why don't you just have ChatGPT, right, the whole thing. It's because it doesn't have my vision. It doesn't have how I, all the experience I have seen as I've trained students and taught them bioinformatics. And I want it a specific way.

content generation architectural vision flowcharts yaml front matter page requirements

And so what we do is, or some of the challenges we have is like this content has YAML front matter. I have very specific requirements that I want. I don't want any page to be more than 2,000 words. And I want to make sure that all the content I have is updated. And so what does that look like? So I actually, so this is kind of fun. You may not know this. You can actually ask Claude or Codex to make a diagram of your process or of your website or of your documentation. And it will actually spit out something that looks like this and create a flow chart for you, which is extremely helpful. So in this case, I had my, I'm going to have a request. I gave it something like, I want you to provide a tutorial that helps a user get a GitHub account and create their first repo and put a few lines of text in it. Super simple. But I want it in a way that has roughly 4 subpages.

github tutorial skill specifications data types tutorial structure self-checks

It has an introduction. It has a wrap up. We have that here. And it may depend on the type of the day. This is where I'm talking about types of data. It might be, in this case, it's basic. I needed them to know about GitHub tools. So there are different tools. But it might also be, how do you use a tool? Or it might be biology in the background. But that type of data that I'm introducing will have different specifications of how I want it to be done. And so It has parts of the skill, which is just a giant text file at this point, that describes all of these steps for me, and then it has a self-check at the end, and finally I'll write it to the repo. So, what you'll see is what it did was it actually... It created day nine and day 10, which is what I asked it to. It had the intro, it had the wrap up, it had the read me file. All of that was generated in a few minutes. And then I went through it, and I had about 11 changes I wanted to do.

productivity gains auto-generation review process time savings architecture role

It auto-generated questions in the front matter, which is what I asked it to do. But it only had three options instead of four. I'm like, I want it to be 4. And that could be variable. If you don't specify this in these skill files, you might get two, you might get 4. And it was incredible. So something that would have taken me two weeks was done in about 10 minutes, which left me as the expert, as in the domain, just having to review it. And I'm not saying that, okay, now just create the next two days. No, I would still need to sit down and plan out what those next two days should be because I'm architecting this 100 days And I need to know what I want to do. And it won't be able to do as good of a job compared to what I am envisioning. Like if you want it to do what you envision, you have to sit down and outline it. And then if you create a skill, it can go out, do the searches, and actually give you a first draft, which then you can polish.

yaml front matter metadata storage jekyll quiz generation

And that's where you're going to get that productivity. So the first prompt is not do my job? Yes. OK. Could you quickly define front manner? And there is YAML that you mentioned. And some people might not know YAML. I don't even know if I know how to describe YAML. Basically, at the top of a website page, there are variables that are not shown when you show the page that you can specify that can be used as content in the page. For example, I put my questions in the YAML, so it doesn't display on the page, but it's tied to the page. And then later on, the Jekyll website will go through all the pages for that particular day and pull all those questions from the front matter and then create a quiz page for me, for example.

agents context windows token limits claude context pollution

So I needed that in that front matter in order for the little script to run later on when they're taking the quiz. I think that explains most of it. Oh, sorry. Agents. So chat bot, skills, agents. Took me a while to wrap my head around the difference between a skill and agent. I don't know if anybody else is struggling with that. But the way I think of it is at the highest level is an agent is a skill with a separate context window. Has anybody used Claude and hit the 200,000 tokens limit? So I know people have done that. So first off, try never to hit that limit because apparently keeping all of that context in your window multiplies the number of tokens you use. So As soon as you get to the next feature that you're trying to development or the next question and you don't need any previous context, start a new window or clear the window.

context management sub-agents token optimization final output intermediate results

OK, because you will use fewer tokens. When you want an agent is when you want to minimize how much of that context you're generating from polluting your main context window. So when you're chatting with a chat bot, it has a history. It has the context window. You might say, hey, can you go out and search kittens and puppies for me, as an example, or market strategies. But you only want the final output from that search. You don't care about all the websites it did and all the different texts it brought in to figure that out. That's just polluting your main context window with information you don't care about. So you launch a sub-agent with its own context window that can then store all of that context, produce a final output, and then bring that into your main context window. So you're only grabbing what you need for the next step.

one-off software vibe coding domain expertise library selection programming language

And that's what agents and sub-agents are useful for. Now that we've talked about chatbots, skills, and agents, How am I using it in my daily work? Well, one of the things that I've realized is that one-off software development is now basically the norm. Like, it shouldn't be something that is restricting us from accomplishing it. used to be that you had to have... somebody that knew everything about coding and programming to develop a software for whatever your need is. But now, if you just need something that will get you through and maybe it doesn't have to be super finely polished, you can vibe code it. If you want something that's more permanent, it's better to have somebody that has programming experience because they will do a better job. Because they have that domain expertise and they can prompt it in a way that will give you a better result. They know what libraries to use.

app development complexity management modular design example library app architecture

They know what programming language to use that makes it most efficient. So I don't think they should be getting rid of programmers. Yeah, there may be someday that you have an automatic programmer, but basically they become architects with a broader vision for what you want. I've tried to write apps or have it just write an app that I've envisioned. And it gets Unwieldy very quickly, and I think you almost have to break that bigger idea up into smaller apps first, so you can get like an app example library, and then from there I think you could take that and say, OK, now make this bigger app. Oops, sorry. So the problem I was facing was, anybody uses ZenHub? It's like an agile. It's built on top of GitHub, which is a system or a place where most programmers will put their code for version control and things like that.

zenhub kanban board chrome plugin github issues javascript

I use it as an online notebook for my documentation. You can make them private. It's great. It's all a markdown. I was using this program, it was actually a Chrome plug-in, to create a Kanban board, which I'll explain in a minute. And the company was like, hey, I've got a lot of users. I'm going to pull all that into my local servers. get rid of the Chrome plugin and then start charging 6 bucks per person. And I'm like, I don't use your full suite. I just need this one tool. And I'm like, you know, so I spent, this was 2023, and we only had the chat bots, no skills, no agents. I asked, hey, I want to make a Chrome plugin. I didn't know JavaScript. I know scripting. I know programming, but not those particular languages. And nothing about Chrome plugins. This would have taken me months to figure out. And within like 2 days, I had a functional Kanban board. So A Kanban board, you visualize a white board, and you divide it into sections. You have your sticky notes, your post-its, and you have, this is the things that are the tasks.

github repo kanban board chrome plugin github issues column labels

I'm going to do these tasks today, and these are the ones, and once you finish it, you can move it over. So a digital version of this, you can create, and it's basically, I like to think of it as like a 2D to-do list, okay? And I wanted to make a digital version that I didn't have to pay for. And so I did. And in fact, if you guys want to play around with it, you can go to this GitHub repo, and you can download the Chrome plugin, follow the directions, and you can load it up. Now you do have to have a GitHub account, and you'll have to figure out how to add the token. But anybody that uses GitHub, this shouldn't be too big of a deal. And then once you have issues, you can literally add issues, and the labels correspond to the columns. It looks like this. Wow, these are terrible slides. I apologize. But basically, you've got your columns here. And you can move around the columns. And you can, I don't know if you saw that very well.

one-off solutions digital products reverse engineering product value in-house development

So you can see that the left column became the right column. And you can move tasks or issues around. You see how that one got shorter? And you can do this, you can actually pull this up on any GitHub repo. Now, if the repo has a lot of issues, I haven't tested it since it failed the first time. It might say, you've requested too many calls, but I think I made a fix for that. But the point is that you can have your own private repo and put your own issues and have your own tasks and you can play around with that. And this solved my problem. So this one-off software solutions where you just need something to do the thing that you're trying to do, You can probably do that in-house with somebody that has some basic coding skills without having to be a full-blown programmer. But what this also means is that any digital product is now basically approaching 0 in its worth. Because you can say, hey, I really like what this program can do.

academic publishing software proliferation reverse engineering assembly programs program testing

Can you reimburse engineer it? And chances are, within a few days, you could do that. In fact, it's becoming a problem in my field because before, graduate students and postdocs would create a program to solve a problem, and then they would publish a paper. And that was great. They get credit for it, academic, all that kind of stuff. Now people are taking that and saying, hey, I like this, but I'm going to make it faster and then publish on it. Instead of having 20 or 30 programs that claim to be the best assembly program in the world, now we might have 100 because everybody's just rewriting it over and over again, which makes determining which software is the best one to use exponentially harder to figure out. And those people that are rewriting it, are they doing it well? Did they test every edge case? So you have to be careful to some degree as to what you're downloading, because people are just making digital products, programs, and you don't know where that came from.

swift mac app native development feature addition keyboard shortcuts

You don't know if they inserted some kind of prompt injection that says, oh, send me all of your Bitcoin tokens or your opening AI keys. You know, these are serious questions and concerns that you have to have as you go forward. Some of the other features include, you can open up the task and look at it. I then took it a step farther, and instead of a Chrome plugin, I said, Hey, I want you to rewrite this Chrome plugin as a... native Mac app using the Swift language. And in a one shot, it did it. I did some programming for fun as a hobby on apps when my kids were a lot younger, because I was looking at the apps, like, these apps are terrible. I could do a better job. I never got around to doing it fully, but I learned a lot about app development. So once I had it in here, I was like, oh, I want this feature and I want this feature. So I added things like... retaining the GitHub URL, retaining the access token, dragging cards between columns, dragging columns around so they have different places, color coding them.

coding experience feature development code quality efficiency bloat avoidance

I found that once I opened the issue, I wanted to scroll through the issues. Well, it turns out that all those keyboard shortcuts, like shift, enter and stuff like that, you have to program them in if you want them, right? I mean, that makes sense. Seems obvious now. But so I was able to do that really, really quickly. And so once you have the initial state, it's fairly straightforward to add small features, and you just keep building on that. Now, again, ideally, you have somebody that has some coding experience to try to keep away from some of the bloat or the inefficiencies and so forth. But it's a super powerful tool that you should have in your toolkit going forward. You don't necessarily need a full-blown coder to solve or create these user interfaces to help you access your data or to analyze certain things and going forward. Here's another example. Actually, it kind of ties back to the graduate student problem. This guy, Phil Ewells, he's part of a company that creates workflows.

rust workflow optimization speed improvement 60-fold speedup data processing

So I run a script, it gives me a result, and then I have another script that gives me a result, and it goes, well, There are these people that now have workflows that do all those steps. You can put them all in there. It distributes all of that compute across supercomputers and then brings it all back automatically for you. Super awesome. He looked at one of these and he's like, man, this particular workflow was taking 15 hours. I was like, I wonder if I could rewrite that in Rust, which is a super fast programming language close to C, C. but more often used in the scientific community. And after a week, he had tested, validated, all that kind of stuff, what you're supposed to do. He went from 15 hours to 15 minutes, a 60-fold improvement in speed. So if you have any data processing bottlenecks, and you have existing software that's doing that for you, chances are you can find ways to improve that bottleneck, reduce that time to do the data analysis or generation of documentation or plots or figures and stuff like that.

ai transformation knowledge assets disposable software repeatable automation reliability

So that's something else to consider. I think I'm way ahead of it. No, I'm actually quite on time. So in summary, we're shifting from tools to more collaborators. especially as we get to more agents. Knowledge is going to be the most important asset that you have in your company, not just the data you have locally, but the people in your company that have that knowledge, that have the skills that you're trying to make skills and agents for to automate. Software is now disposable, which is crazy to think about. Skills and agents are being used for repeatable automation. I think that's where you're going to get the most efficiency gains. And what the challenges or the constraints we're going to have are reliability, validation, security, governance, like where, how exactly did it do the thing?

workforce multiplication productivity gains employee retention competitive advantage knowledge leverage

You want to keep track of how things were moving forward. And then I have one final thought. As you are using implementing AI, I don't want, I don't think we should be thinking about how can I get rid of employees? Because I think that is, that's the wrong way to think about it. You should be thinking about how can I use my existing employees to multiply my production and multiply my profit? Because your employees are the ones that have the knowledge and have those skills that can utilize these agents to significantly increase your productivity. If you're using it just to maintain your same level of productivity, you're going to be out-competed going forward. You want to figure out a way to get your workforce to be able to do, instead of five people, being able to do 50 people's worth of work. Does that make sense? Yeah. Do you guys have questions? I don't know what I have. Who has a question? I'd like to bring the microphone to you so we can capture it.

erp software custom development open source business software ciras

All righty. You say the value of a digital product is rapidly approaching 0. We're looking at new ERP software to run our whole entire business. Can I really build my own ERP with AI tools? Maybe not, but if you have access, if you get access to that code base or if you can do it with open source stuff that you can then add your own features to. So download the trial version and stick it in AI? I don't know. At some point it's going to be harder, but my point is that For the smaller tasks you have, there's a lot of things you can do with one-off software development. One thing I'm going to tell everybody here, and I probably need to repeat it several times, if you're talking and you're from Coimag or anything like that, and you're having that sort of question, call CIRAS. End of story. I don't want to sound cocky or anything. We know how to help you with those kinds of decisions. And I see people making those decisions all the time and then coming to me and saying, those idiots don't know what they're doing.

saas decline open source models gpt-4 local models token limits

And usually what I say is, can you lift your thumb for me? Because they don't know what they're asking for when they go to the ARP people. Is SAS going away? My belief it is. So be careful with that. Don't get involved with that at this point in time. So for those that might be curious, this is what this Obsidian Wiki Vault might look like. So these are all people and sessions and information in this vault where you can see who is connected to whom, and you can actually see how they're connected and play around with it. It's not the most useful interface, but it's still a powerful way to view things. And then I asked Claude, hey, based on everybody, all the speakers and people that are speaking, who do you think, based on how you know me, I should interact with? Who are the main people that I should be networking with? Who are potential people that might find my services valuable?

local ai openclaw hermes security precautions experimentation

So you can take, like I said, going to conferences, networking, it's changed how we do that. I might have done some of this before, but it took me hours of going through and looking up individuals. Oh, maybe this person, maybe this. gives you a jump start on all of that. Yeah, if you scroll up a little bit, just a little bit, the people you can help, that's me right there, Hans Coink. There you go. Yeah, so you nailed it, because we're a vaccine company, and so we use bioinformatics, that sort of thing. But I had just kind of more of a general question for you. showed some windows and I'm a Gemini guy. So, it's like Ford versus Chevy anymore. People are talking about this stuff. But how do you decide which LLM or which, you know, you talked about anthropic, you talked about OpenAI. I mean, I can sense that you're trying to avoid like paying for anything.

llm selection gemini claude openai anthropic

You know what I mean? So, which, yeah, that's a good thing. So, but how do you make that decision? Right now, each of the big contenders, Gemini, Claude, and OpenAI, are all jump leapfrogging over each other for what is the best large language model. And some would argue, oh, you just, once that happens, you just switch, you know, and you can carry that over in an identity file with all the useful information. That's actually how Anthropic got that big boost at the end, if you saw that, why they're making so much money, because Chat or Open AI was like, basically Anthropic said, we're not going to let the government make killer robots that will come and like kill people. Like that. We don't think that's like, and then the next day Open AI is like, we just made a contract with the government, but we're not doing it. And like people ditched them, and then they said, just write in the thing saying, tell me everything you know about me, and then place that into a claw. And that's how they. not a big influx of users.

open source models gpt-4 comparison local execution token avoidance ralph wiggum approach

I don't know if it makes a big difference from the foundational model. Right now, they're fighting over harnesses. Claude has a harness, and this is how the foundation model interacts and the skills that are in there. And I've heard arguments that right now it's really cheap and it's getting everybody hooked. And then at some point, they're just going to jack up the rates, and then we'll be like, well, shit, I don't know what else we can. I guess I'll have to pay $2,000 a month now. And so I'm planning on taking advantage of what we can now and use the open source models and all the open source environment and say, OK, I know they're not as good, but how close can I get it to the the OpenAI or Claude, and can I iterate it? Is there a way, if I create a skill, say, with these top models first, and then I switch the model to these open source that are like, I don't know, like what, six months, a year behind, which they're exceeding GPT-4, which we thought was the bee's knees back in 2023, can you iterate that?

local installation email integration slack integration spare laptops security concerns

Is there a way that you can have a test What is Claude? Claude calls it the Ralph Wiggum approach, where you just, you keep doing it until it gets you the best result. I think that's where I'm going to spend my time. Can I do it all locally? Because then I'm not spending tokens. I don't hit my 200,000 token limit or my weekly limit. If I can get it working, so I have a spare laptop. I'm going to install either OpenCloud or whatever else I can find, Hermes, and see if I can give it an e-mail account, give it a Slack, invite it to Slack and just see what I can do with it, and... experiment. By no means am I going to put it on my main laptop that has all my stuff on it because that's just that's not a good idea. All right so we are at the time limit so I would like you to thank Andrew with me for his time and discussion. I also want to point out that what Andrew was showing you with his connections and everything was one way of taking the structured data that is on the website that you all have access to.

structured data conference content transcripts chunked information data access

And you can do it in really basic format with some of the prompts and things that we have. But that structured data that is put in there is for you to be able to use. Andrew just said, I'm going to use my tool and I'm going to pull the structured data in. My artificial intelligence is going to figure out how to organize it. and now I can use it in any way I want. We've given you some ways to do it online, but you can use any tool you'd like. Okay? If you have questions, you can check with the SIRIS desk. You can ask some questions there. We will be providing a lot more data. For instance, everything Andrew said here will be in a part of the structured data. It'll be a transcript. You'll be able to put that in. We'll have what he's shown us on the screen, et cetera, hopefully knocked down to some chunked information that he was talking about. So you're going to be able to talk with the entire conference. Now that's going to be over the next few weeks. Okay.