← Back to search

Cut Research Time & Costs with K-Dense

Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs · 2026-10-08 · 32 min
relevance 63 5942 words Episode page ↗ Audio ↗
Show full episode description
Imagine shrinking months of research into hours without sacrificing privacy, control, or expertise. In this episode of R eshaping Workflows with Dell Pro Precision and NVIDIA RTX , Logan Lawler welcomes Timothy Kassis, co-founder and CTO of K-Dense, to explore the rise of AI co-scientists for next-generation research. Discover how K-Dense’s open-source Scientific Agent Skills and multi-agentic orchestration enable seamless, automated workflows across 250+ data sources, with specialized agents for every step, from coding to review. Find out how tools like Faraday empower labs to run powerful local models with Dell & NVIDIA hardware, slash cloud costs, and protect sensitive IP. Learn why the future belongs to open-weight models, how AI orchestration closes the loop from hypothesis to physical lab experiments, and where your own workflow can level up, whether you’re in bioengineering, drug discovery, or fields yet unimagined. Watch the full episode now and see how to fast-track your own data science breakthroughs securely and affordably. You can also watch this and all previous episodes here . Find K-Dense k-dense.ai Follow Us LinkedIn @DellTechnologies Twitter @DellTech YouTube @DellTechnologies f LinkedIn @NVIDIA Twitter @NVIDIA Instagram @NVIDIA Facebook @NVIDIA Presented by Dell and NVIDIA https://www.dell.com/precisionai https://www.dell.com/dellpromax https://www.dell.com/nvidia-ai Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
✨ Episode Outline — click any point to jump to it in the episode
Problem solved
Speeding up slow, reductionist scientific research with multi-agent AI co-scientists that run cheaply and privately on local hardware.
Benefits
  • Aims to make scientific research ~100x faster with AI agents
  • Open-source Scientific Agent Skills teach any agent how to access and harmonize data sources
  • Orchestrator plus ~two dozen specialist agents act like a PI and lab team
  • Faraday local instrument cuts cloud cost and keeps proprietary data in the lab
Use cases
  • Scientific Agent Skills library works with Claude Code, Cursor, or any agent; ~33,000 GitHub stars
  • Mimeo codifies a named expert's papers, videos, and podcasts into an agent skill
  • Pantheon spins up 80 expert personas that answer a question and reach consensus
  • K-Dense Web co-scientist hosted on GCP; Google is an investor
  • Faraday on-prem box built on DGX Spark/GB10s from Dell runs open-weight models locally
KPIs / results
  • ~33,000 GitHub stars for Scientific Agent Skills
  • Typical frontier-LLM science task costs a few hundred dollars
  • Connects ~250 databases and thousands of tools
  • Company started ~10 months ago
Tools / build
  • K-Dense Web
  • Scientific Agent Skills
  • Mimeo
  • Pantheon
  • Faraday instrument
0:00 / 0:00
[SPEAKER_00] Welcome to Reshaping Workflows with Dell Pro Precision and NVIDIA, where innovation meets real-world impact in high-performance computing. [SPEAKER_01] Welcome back to another episode of Reshaping Workflows with Dell Pro Precision and NVIDIA RTX PRO GPUs. I'm your host, Logan Lawler. So today, all about data science. It's going to be great. If you are a developer, you're using kind of, you know, you think of yourself as a data scientist, you're going to love this episode. So my guest today is I have Timothy with me, who works at K-Dense. So Timothy, tell everyone, introduce yourself really quick and we'll jump right in. [SPEAKER_02] Yeah, thank you, Logan, for hosting me. So I'm a co-founder and CTO at Cadence. I primarily focus on kind of building that technology team and that technology. [SPEAKER_01] So I have some questions. Let's go into you. So you are a bioengineering PhD. You've done kind of wet labs, petri dishes, all types of kind of crazy stuff. How did that turn into you working at K-Dense? [SPEAKER_02] Very good question. So my background is actually pretty interdisciplinary. I've done research in kind of electrical engineering, mechanical engineering, bioengineering, tissue engineering, etc. And then during my PhD work as a kind of bioengineer at Georgia Tech, I realized that the way we do biological research right now is very slow. [SPEAKER_02] Slow. What I mean by slow, biology makes incremental discoveries and improvements and it's very reductionist in nature. For example, you might spend years, if not decades, just trying to understand how molecule A on cell A interacts with molecule B on cell B. But biology at the end of the day is very multiscale. So both on the timescale, you have things happening on kind of the order of, you know, picoseconds all the way to millennia. [SPEAKER_02] And on the spatial scale, you have, you know, things at the order of kind of, you know, molecules all the way to large kind of ecosystems. And during my PhD, I began to kind of really appreciate what AI can bring into all this. And that was a time when a lot of these kind of vision-based models started working very well based on deep learning. And deep learning started to become useful in kind of real world tasks. [SPEAKER_02] First, it was kind of vision, then signal processing, then obviously language now, etc. And since then, I've really been focusing around developing AI tools to accelerate scientific research in general, but specifically kind of biology. I've led a few AI teams at different kind of AI for bio companies. [SPEAKER_02] And back when I was still teaching at MIT almost in, well, in 2019, I actually had started a different company called Augunz. And Augunz, the focus of that was similar to what Cadence is now, which was essentially building an AI scientist. At the time, unfortunately, you know, back in 2019, the technology just was not there. The technology that we see right now with these agents, LLMs, was just almost non-existent. So I spent about nine months working on it. [SPEAKER_02] Then I kind of gracefully decided to kind of shut it down before I kind of joined some other startups. And Cadence right now, you know, where we are, we started the company about 10 months ago. Just as a lot of these AI agents started becoming very, very reliable. And we initially built kind of a few proof of concepts. We tested them with a few kind of friendly biotech startups that we, you know, friends and family and we knew. And they proved to be extremely powerful. [SPEAKER_02] So we decided to kind of, you know, commercialize this and kind of scale what we're doing right now. [SPEAKER_01] That's interesting. Interesting. So let's, so your background obviously lends itself very well to kind of Cadence. But let's do this because I know you could probably talk about Cadence for hours. If you couldn't, then you're probably not very good at what you do, which I know you are. So let's just say 30 seconds. You're explaining it to a person that's not super technical. What does Cadence do? [SPEAKER_02] So science right now is very slow. And what we're trying to do at Cadence is we're trying to speed up that process 100 times. And the way we're doing that is through AI agents. So think things like Cloud Code or Gemini Spark right now or the recently released GrokBot yesterday where they can automate a lot of the tasks for you. We do that for scientists specifically. [SPEAKER_01] Okay. So think of it as an AI agent for true data scientists. Okay. That's pretty straightforward. So let me ask this question. And do you know I've done some research, right? Is that, you know, your AI dev like the Cadence like will connect 250 databases, thousands of tools, you know, all these data formats, right? Which obviously a data scientist needs because you're going to get time series data versus like other types of data. So I can get it. [SPEAKER_01] But how do you keep those, all of those disparate sources and formats and all that in agreeing? Like I could barely get my Hermes agent to write the date like, you know, number, number, dash, number, number, dash, then the year versus like writing it in like some other crazy form. How do you, I mean, I'm assuming that's the secret sauce of Cadence, but how do you do that? [SPEAKER_02] Yeah. No, very good question. I think, you know, with science, especially nowadays where science is very interdisciplinary in nature. So, you know, if you're solving a problem in a kind of drug discovery, you have to reach out to areas in physics, in biology, in chemistry, in data science, and kind of health research, et cetera. And, you know, the data sources for all of this, whether it's literature or databases, they are kind of all over the place. They have different APIs. [SPEAKER_02] They have different ways of accessing it, different schema, et cetera, et cetera. And really, you know, what we do fundamentally at the company is we focus on kind of a few things. So one is we built a very popular open source library called Scientific Agent Skills. And what this is, is a set of explicit instructions that we give agents. And this could be any agents, not just the Cadence agents, like if you're using Cloud Code or Cursor or, you know, whatever you're using. [SPEAKER_02] And we give explicit instructions for the agents in terms of how to access the data sources, how to integrate them, when to access what data source, when to prioritize certain data sources over others. And then for, you know, a given problem, how do you kind of aggregate or harmonize if you need to harmonize kind of the data sources? [SPEAKER_02] So there's a lot of work that's being done on kind of these skills that give these agents that kind of fundamental capability in terms of understanding data that's out there and accessing it. [SPEAKER_01] Makes sense. Okay. So think of it almost analogy is it's a pack of skills that almost functions kind of like a model router, a mixture of agents, but I'm going to call it a mixture of skills that basically based on the task, it's got kind of pre-deterministic of like, go here for this, go here for this. Here's how we harmonize. All of that's kind of pre-built. And you said it's open source, so it's on GitHub. Anyone can download it. [SPEAKER_02] So it's a combination of things. So the skills are open source. We have about 33,000 GitHub stars right now for the skills. But ultimately, you know, what we've spent a lot of time building is the other component to this, which is a multi-agentic system. So our multi-agentic system is composed of, you know, about two dozen different agents. And each agent specializes in one aspect of scientific research. [SPEAKER_02] For example, some agents specialize in reviewing the quality of the work. Some agents specialize in coding. Some agents specialize in data retrieval. So these agents, along with the skills that they're given, they're able to kind of figure out what needs to be done. And there is, you know, as you described it as kind of a router. And, you know, our definition of router here is we have an orchestrator, an orchestrating agent. So think of it as a, you know, if you're a researcher, you might have a principal investigator of a lab, right? [SPEAKER_02] They get to decide what's going on in terms of the research trajectory that, you know, all the team members are doing. And then you've got these sub-agents that execute. And they're kind of, you know, very specialists in their field. [SPEAKER_01] And they all talk to each other. Let's go into this a little bit. You kind of have your open source. I think you call it Pantheon is the name of the library, or is that wrong? [SPEAKER_02] It's called Scientific Agent Skills. [SPEAKER_01] Was it used to be named Pantheon? No, Pantheon is actually something else. Okay, so, okay, let's talk about Pantheon now. Is basically you're, it's kind of like almost a persona slash soul file kind of spinner upper, right? And it says it'll spin up 80 to answer the question. So, you know, if anyone's used agents, use Hermes, use OpenQual, use LangChain. Generally, per profile, you kind of have one soul file or persona. Mine for mine is like, hey, I'm Logan Lawler. [SPEAKER_01] I lead to Pro Precision AI strategy. I'm kind of a jokester. I like to talk too much, blah, blah, blah, blah, blah. You spin up 80. So why, one, why do you spin up 80? And what is the value of spinning up 80 versus like five? [SPEAKER_02] Yeah, it's just a proof of concept that we built around a more kind of useful tool, which we call Mimeo. Now, what is Mimeo? So in the kind of agentic world in general, you know, the LLM world, LLMs are pre-trained on internet data. So anything that's in written form is kind of out there. You do pretty well on it. But in science, you have a lot of expertise that belongs to specific people. [SPEAKER_02] What I mean by that, for example, you might have a, you know, radiation oncologist that's been working in the field for 30 years. And they have a lot of that knowledge and expertise in terms of what to do in their discipline. That's just not written anywhere. So Mimeo was kind of our first attempt at think of it as codifying a person into an agentic skill. So, you know, if you have, for example, Opus 5, how do you make it execute and behave like [SPEAKER_02] this expert? And that's what kind of Mimeo is. So Mimeo, you know, you give it the name of a kind of a high profile researcher or a scientist. It goes out there. It looks up everything they've written, all the videos that they recorded, all their podcasts, all their papers, et cetera, and distills that into specific ways of thinking about problems in science and codifies it as an agent skill. And what Pantheon is, is basically a demonstration of kind of the capability where we did this [SPEAKER_02] for 80 different kind of individuals, experts in different fields. And we said, let's use all these expertise as one thing, or it's like a council where you ask a question and then all 80 basically answer from their own perspective. And then there's a kind of consensus that's reached among the 80 to answer your question. [SPEAKER_01] That is more of like the mixture of agents type model, right? Or you have all these personas. Okay. So when you say it, first off, I love that. That's amazing. Like it's, I mean, it's genius. Now here's the question though, is that when I hear 80 personas doing research, you know, blah, blah, da, da, da, da, da. All I hear is tokens, tokens, tokens, tokens, tokens, tokens, tokens, tokens, tokens, tokens, tokens, tokens, tokens times 80. [SPEAKER_01] So tell me a little bit about, you know, like where does the compute lie? Are we local? Are we in data centers? Are we in the cloud? I'm kind of curious because you're right. Like some AI things, right? To do them right. In this case, pretty smart. You're basically getting all the smartest people in the world and their discipline acting as an AI agent for you. That's fantastic. But then that comes with a cost. So tell me a little bit about like, where does the compute lie? What system does it need to run on? I'm very curious. [SPEAKER_02] No, great question. I mean, let me tell you what we kind of built and where we're heading right now. So our AI co-scientist that's available for anyone to use is actually called Cadence Web. This is our web platform that anyone can use. It's hosted on cloud compute. We primarily use right now GCP as kind of our main provider. You know, no surprise there. Google is an investor in us. So they gave us a lot of kind of resources there. But ultimately, I mean, you raise a very important question. [SPEAKER_02] So in science, a lot of these tasks that the agent executes on are very computationally intensive, both in terms of number of tokens, as well as the compute resources they need to execute on a task. So a typical task, for example, if you're just running it with Frontier LLMs from, let's say, you know, Google or OpenAI or Entropic might cost a few hundred dollars to run. So it becomes kind of very, very, very expensive. [SPEAKER_02] And we've realized that there's, you know, in addition to kind of the cost limitation, scientists in general are actually very skeptical about sending their proprietary ideas and data to cloud providers. You know, no matter how much guarantees they provide or we provide, they're very hesitant. So what we're transitioning to right now is, you know, we're working closely with NVIDIA and Dell in terms of building instruments. [SPEAKER_02] We call our kind of the, what we call the instrument right now is Faraday. And what this is, is it's a machine that sits in your lab. It's built on kind of the GGX Spark architecture. Sure, you know, we primarily rely on, on GB10s from, from Dell, but it allows us to host and run open wake models fully locally. So what that basically means is now we solve the two pain points for these scientists. One is cost. [SPEAKER_02] So they can run as many tasks as they want. It's running locally. There's no, you know, per token cost involved. And then two, the data privacy and security aspect that a lot of companies and individual researchers worry about. So we built this kind of system called Faraday, runs everything locally. Your data stays local or the computer is local. Tokens stay local. And that's what we're kind of really trying to scale right now. [SPEAKER_01] Okay. Let's use a foundation model. Let's use my favorite, Anthropik versus the local version. So my first question on the cloud version, you said $200. Is that in cloud and token costs? Is that per kind of one run? Like what that give me the nuts and bolts of like, what is being done to get to that $200 in token costs? Is it literally like, hey, I want to look at this and this and this and this and it's 200 or is it more complex than that? [SPEAKER_02] No, let me give you a quick example. For example, a very common thing you do in bioengineering is you design biomaterials for a specific objective. Let's say I wanted to design a biomaterial that I can use in the lab to culture, let's say intestinal organoids, which is something I previously worked with. You go to CadenSwep and you say, I want to design the spiral material to culture these organoids. I have a certain budget. These are the criteria I care about. Go. Right? [SPEAKER_02] Now to answer that question, that's a very non-trivial question, right? Like pre-LLM days, this would have taken a researcher three to six months to actually execute on. Do the literature search, do a lot of the kind of in silico experiments, do a lot of the simulations and then in order to get to an actual lab experiment. But what CadenSwep does, it spins up all of these agents. They do research. They access databases. They iterate a lot. [SPEAKER_02] They do in silico experiments. In certain cases, they might need to train their own machine learning model to make a prediction on something that's unknown based on previous data. And all of that kind of translates to, in certain cases, tens of millions, if not hundreds of millions of tokens, both input and output tokens, to achieve that given objective. And that's where the most of the cost is. And if you think about it, if you're a researcher and your day-to-day job is to ask these questions [SPEAKER_02] multiple times a day, you know, paying $100 to $200 of LLM costs every time you do something like that is prohibitive. [SPEAKER_01] That's the question. Using that as an example, from start to finish, how long does that take to run? [SPEAKER_02] Yeah. A question like this typically takes around three or four hours of execution. And then you'll get your answer. [SPEAKER_01] Okay. Now let's apply it to the Dell Pro Max with GP10. Token cost zero. However, and let's be transparent, right? Like the Dell Pro Max GP10 or the DGX part, it's more like a tanker than it is a speedboat. It's like, it's got a lot of VRAM and it can run a lot of things, but it's not necessarily the fastest. Like dedicated cards that you're probably getting in GCP or, you know, in our traditional [SPEAKER_01] Dell Pro Precision Workstations, just the speed is 6x. The memory bandwidth is 6x. It's just faster. So obviously zero cost, completely secure. But does that three to four hours now become 24 hours? So there's two limitations, right? [SPEAKER_02] With kind of local hardware. One is memory. So VRAM. So what that basically means, it limits the types of models we can put on the device. And obviously, you know, a model like, you know, Nematron Super, which you can fit on the DGX Spark or even the BG10 or Gemma 4, et cetera, are less capable than Opus 5 or, you know, GPT 5.6 Sol, et cetera. So that's kind of the one limitation. [SPEAKER_02] So we have to do a lot of optimizations around these limitations. Two is, yeah, as you mentioned, is speed. And more importantly, I think, you know, with a kind of cloud provider, you typically can execute agents in parallel. So for example, if there are multiple sub-agents that have to do things that are parallelizable, you can easily execute these in parallel. [SPEAKER_02] Now, on a system like the BG10, GB10, you're limited because you might be able to spin up one or two sub-agents at the same time. But you can't do more than that because you quickly kind of hit bottlenecks with, you know, whatever model that you're using. [SPEAKER_01] Okay. So it still runs. It's just slower. [SPEAKER_02] It runs. It's slower. But, you know, a lot of these workflows are workflows that scientists might just kind of set and forget. So, you know, you're working in the lab. You say, okay, I want to run this, you know, answer this question for me, do this workflow for me. And you come back the next day, right? And you have your answer ready on the machine. So we haven't really noticed, I would say, complaints from kind of early users around time [SPEAKER_02] because the expectation is, okay, if it takes three hours versus 12 hours, that's kind of the same thing. Because, you know, there's kind of two bins here. Either it's instantaneous where you get your answer while you're sitting at your kind of machine or it's not. Like it's a binary, you know, yes, no. You know, because if you're going to leave it for two hours anyway, just leave it for 12. [SPEAKER_01] Exactly. So, and that's so interesting because I'll draw a parallel here is that, you know, when Comfy UI was a thing like a couple of years ago. And I was doing some lore training, trying to train for a show, AI version of my dog. Like it basically, I would just set it at night, let it run. And then in the morning it was done. And, you know, it is what it is. So what I'm hearing is basically kind of that research question. It's $200 approximately. [SPEAKER_01] You're sending your stuff up for maybe proprietary or very unique, you know, world changing things up to the cloud. If that's $200, yeah, it runs a little slower, but you're still saving six months of your life, which is kind of awesome. And that, which is really awesome. But, you know, if you do the math, right, you're basically talking 20, 25 of these type queries on the Dell Pro Max with GB10. And you've basically paid for the system. [SPEAKER_02] Right. Exactly. [SPEAKER_01] That's insane. And you've saved what? 25 times, you know, six months of your life. So you've saved 12 years of your life. [SPEAKER_02] And not just that. I mean, there's a kind of an added component to this, which is as kind of fine tuning and RL algorithms become more and more efficient. You know, there's a world we see in the next kind of six months or so where we start having kind of the model weights adapt on the machine itself for that specific lab or user. So thinking about it, I just thought, you know, self-evolving, not just, you know, harness [SPEAKER_02] or skills or prompts, but self-evolving weights as well, because you can do kind of fine tuning and RL on that GB10. And, you know, what we're betting on here, Logan, is, you know, we're seeing a very interesting trend right now with open weight models where, you know, their intelligence is really lagging behind maybe four to six months frontier models. So we're going to get to a point, you know, in the next few months, if not early next year, [SPEAKER_02] where you can have a, you know, Opus 5 level model running locally, maybe on like two GB10s connected with each other. Then, you know, the question is, why do you need these cloud-based models at that point if you're a scientist, right? If you can run unlimited tokens locally, if you can keep your data secure, keep it private. So that's what we're actually most excited about is, you know, what the future holds in the next few months. [SPEAKER_01] It's true. And I mean, let's be honest, open weight models are the future, but like much like the, I don't know, I'm trying to think of a good example of something when it first came out, it kind of stunk. Like, for example, you're the first cell phone. It was in a bag. You plugged in a lighter in your car and, you know, that's what it is. Now we all have phones that fit the world's human collective knowledge in our pocket. Every single one of us, pretty much everyone on the dang planet. You're right. Open weight models, private models are not private models, but open source models. [SPEAKER_01] Are the wave of the future. However, with the intelligence, they are kind of lagging behind. Like there is some stuff and I'll be the first to admit, like I do a mix of local and cloud. It depends on what I need, right? Like if I need some like serious reasoning and like, Hey, the local model is not hacking it. Then I, you know, I got to go to Opus or I go to Fable. But a lot of times once things are built, which it sounds like cadence is, is built, it's packaged, it's ready to go. [SPEAKER_01] Models like Nemetron are very good or Jim Naird, whatever, are very good at like following instructions. Like, you know, follow your instructions. Might not be the smartest, but I kind of liken it to that person that's been in their job for 10 years that knows a lot of things. That's like a foundational, like, you know, Opus Fable type model versus someone who's a couple years out. That's like two or three clicks down. They still know a lot of the same things, but they're just doing the work versus strategizing the work, if that makes sense. [SPEAKER_02] And you know what we actually did with Faraday? So, you know, this is our kind of AI, you know, co-scientist in a box. We've actually, we have kind of the best of both worlds. So Faraday right now runs an NVIDIA Nemetron models for kind of reasoning, execution, et cetera. [SPEAKER_02] But if it gets stuck, if there's a very complex query or problem, what it does is it allows the user, basically it presents this kind of nice interface to the user saying, look, user, I don't know how to answer this. I'm going to send it off to Opus 5 in the cloud. Here's exactly the prompt that I'm going to be sending. Do you approve? [SPEAKER_02] And then we have another NVIDIA model called Gliner, which actually looks at sensitive information or like personal identifying information in the prompt and flags anything that's questionable. And then we give the user the option of masking it or deleting it before the prompt is sent. So then the prompt is sent to an expert model. And, you know, right now we're using Opus 5 and then the expert model replies and then Nemetron kind of continues its work. [SPEAKER_02] So we like to think of this as like a thing of it as a kind of intermediary solution until local models become super smart. But now we got kind of we have the best of both worlds. You know, we have data privacy where your data actually doesn't leave. Your data set is on your computer. Your ideas are on your computer. You have full transparency in terms of what's actually being sent to a cloud provider. And you get to approve every single request that's being sent only when Nemetron actually needs to send a complex query. [SPEAKER_01] Now I'm just curious. Is let's say that workflow you said before where it was answering the question that took like three or four hours. Like what percentage do you think if you're running the Nemetron model, what percentage of the tasks or what percentage of the work has to go to Opus currently? [SPEAKER_02] Yeah, it really I don't have it. I wouldn't be able to give you an exact number, but it really depends on the complexity of the task. So for example, if there are just certain workflows that let's say a data scientist or a bioinformatician writes and executes, Nemetron can easily do that 100% of the time. No reason to kind of call this expert in the cloud. [SPEAKER_02] More complex queries where you're trying to kind of think outside the box or there's some brainstorming involved or some novel kind of scientific code writing that just is not available to Nemetron, whether it's internal knowledge or the skills we're giving it or the databases we're giving it. So then you would need to reach out to these kind of expert models. But I would say, I mean, kind of preliminary tests on a few kind of kind of random internal benchmarks here and there and like different like tasks. [SPEAKER_02] I would say probably less than 20% of queries would have to go to an external kind of expert model. [SPEAKER_01] Next question is you've partnered with Litmus to kind of close the loop on hypothesis to the real lab experiment, right? It's like it seems kind of sci-fi to me, but, you know, I'm not a data scientist. I'm not in there with the pipettes and stuff. But shoot me straight, like how much of this is actually running today? And if none of it is, that's fine. I'm just curious. Or how much of it's just slideware? Is that or is that really what's working? [SPEAKER_02] Yeah. So, I mean, you know, the biggest limitation a lot of these kind of AI co-scientists have, including ours, is they're able to do a lot of computationally experimental work. And in silico experiments, they can generate novel hypotheses. They can write papers. They can do, you know, figure generation analysis, all that stuff. But there's a very important missing component when it comes to science, which is access to a physical lab to kind of conduct a physical experiment. [SPEAKER_02] So, what we've been working hard with is, so Litmus is kind of the first partner we announced. But we've been trying to kind of close that loop. And with Litmus right now, we're working on an integration with both Faraday and Cadence Web that essentially give the agentic system the ability to write an experimental plan that needs to be executed in the lab. And then that experimental plan gets sent to Litmus. [SPEAKER_02] And what Litmus is, is think of it as kind of API layer between, you know, software in general and various contract research organizations. So, these are, you know, kind of for-profit labs that run experiments. You know, there's labs in chemistry and biology and kind of animal science, et cetera, et cetera. So, then the experiment is kind of outsourced through Litmus to the CROs. They run them. Depending on the experiment, it could be a few weeks, could be a few months. You get the results back. They feed back into our agentic system. [SPEAKER_02] And the system kind of continues working from kind of that updated data. [SPEAKER_01] That is so cool. I don't know. I'm kind of blown away. I don't get blown away out of the show, but, like, that's pretty cool. All right. Future prediction. Then we're going to jump into rapid fire and we're going to call it an episode. So, future predictions. 18 to, let's say, 12 to 24 months out. I've never written a research paper. I am not a published author. However, I know several in the academic community. [SPEAKER_01] You know, what percentage of papers do you think will have some sort of cadence-esque, whether it's through you or someone else, kind of an AI co-scientist that is assisting with that paper in research? Do you think that will be 100% in the future? Is it going to be 20%? Like, where do you think it's going? [SPEAKER_02] I would say definitely a majority. And keep in mind, we actually see a lot of AI being used right now in current papers, you know, in terms of the analysis, execution, etc. But the gap there is a lot of researchers don't admit it, right? When they write the paper, usually in the methodology, you have to say, I used X tools. You know, here's what I did. Here's what I used. They don't admit it because there's this perception still, especially in the academic community, that it's admitting weakness, right? [SPEAKER_02] Like, if you're giving a task for the AI to do or, you know, to write code for you or do data analysis for you, that means you as a researcher don't know how to do it. You know, I think there's a lot of hesitation right now. I think the general mindset is starting to change. You know, I go to a lot of kind of conferences, a lot of meetings for AI, for science, and the mindset is shifting. So to answer your question, I would say definitely a great majority will be using these kind of AI agents for science in one way or another. [SPEAKER_02] In some cases, it's, you know, the full paper from data to the actual paper will be written with AI agents. You know, the data acquired, analyzed, the literature searched, the paper written, all AI end to end. In some cases, it might be just kind of the data analysis or just the writing, right, where the researcher did the data analysis, but they don't like the writing part. They just dump everything into, like, cadence web and say, write a paper for me. And it gives them, like, you know, ready to publish paper. [SPEAKER_01] Now we're going to get into some rapid fire questions. One or two word answers, just the first thing that kind of comes to your head. So what's your own daily driver for a model? Is it local or is it frontier? Frontier. Over or under? A Nobel-sided paper listing an AI agent before 2030. [SPEAKER_02] Oh, under. [SPEAKER_01] Under, okay. The most overrated benchmark in kind of AI data science right now? [SPEAKER_02] Honestly, all of them. I think benchmarking is being gamified a lot. [SPEAKER_01] It is. Open AI or Anthropic? [SPEAKER_02] We highly rely on Anthropic. [SPEAKER_01] We like good models. I do too. All right. One data science base that you wish had an API that doesn't? [SPEAKER_02] Oh, that's a hard one. I think there isn't really one right now because when there's an API absence, we've always found workarounds. I think the biggest bottleneck, honestly, is paywalls. [SPEAKER_01] All right. So, I mean, obviously you're scraping. You're the fire crawlers and something. But which one is the one you want to get to and you can't? Because the Cloudflare or whatever it is. Like, what is the one you want? [SPEAKER_02] I don't think we're hitting any limitations right now. I think that our biggest limitations is access to literature databases. So, for example, think Elsevier. They have thousands of journals under them. You really cannot access any of their papers. And again, it's because of paywalls and not like technical issues. [SPEAKER_01] All right. Last question is your PhD or shipping the Cadence product? What taught you more about building Cadence? [SPEAKER_02] Oh, definitely shipping the Cadence product and talking to customers. [SPEAKER_01] I love it. All right. So, we're kind of at the end of it. So, Timothy, you've been great. Take one minute. Tell everyone where they can go to find Cadence, find you, and any parting thoughts. [SPEAKER_02] Yeah. I mean, if you guys want to learn more about kind of our open source tools or our paid platforms or where we're heading as a company, just visit our website. It's pretty comprehensive. www.cadence.ai. That's k-dence.ai. [SPEAKER_01] Well, thank you, Timothy. You were fantastic. So, data science, research, all wrapped around AI. I mean, I can't imagine personally having 80 personas spun up every time I run kind of my Hermes agent. I have one. It's just incredible what Cadence is doing. Thank you for listening. And until next time, keep your GPUs and your compute running locally. And we'll see you on the next one. [SPEAKER_01] This podcast was produced in partnership with Amaze Media Labs. Thank you.