Become a Creator today!Start creating today - Share your story with the world!
Start for free
00:00:00
00:00:01
A cognition engine for science with Allen Stewart
3 Plays7 hours ago
Scott Hanselman sits down with Allen Stewart, Partner Director of Software Engineering at Microsoft, to explore how AI agents with persistent memory are transforming scientific research and software engineering. Allen explains how his team built an AI system that learns from every investigation turning a 12-day autonomous drug discovery run into reusable knowledge that makes future research exponentially faster. Instead of starting from scratch each time, the AI inherits hypotheses, methodologies, and findings from previous work, saving hundreds of millions of tokens and weeks of effort.
Check out https://textcontrol.com for document editing, reporting, and PDF processing SDKs for .NET
Recommended
Transcript
Journey into AI for Science
00:00:00
Speaker
So like stuff like that, you just don't want to make those kinds of mistakes or otherwise you're going to be drinking sulfuric acid, Tommy. Absolutely. Look, look at my journey. I look, I had my own journey to science, right? I'm a computer scientist. And look, when we found out that we're going to work in the AI for science, you know, I got to spend some time with scientists and that was very, that was very eyeopening, right? You get to start spending time with scientists. First of first of all, they said, look, Hey, AI guy, science, scientists do science. Do you understand? And I was like, I understand. Yeah.
00:00:29
Speaker
so AI is, they're like, repeat it back to me. What did you hear? AI is a tool. Scientists do science. they're like, okay, we like this guy. There you go. He's kind of understanding what we're trying to do here, right? So, Hey friends, you probably knew that text control is a powerful library for document editing and PDF generation, but did you also know that they're a strong supporter of the developer community? And it's part of their mission to build and support a strong community by being present, by listening to users, and by sharing knowledge at conferences across Europe and the
TextControl's Community Engagement
00:00:58
Speaker
United States. If you're heading to a conference soon,
00:01:01
Speaker
Maybe check if TextControl will be there. Stop by and say hi. You'll find their full conference calendar at textcontrol.com. That's T-E-X-T, control.com.
Role of Memory in AI Systems
00:01:13
Speaker
Hey friends, I'm Scott Hanselman. This is another episode of Hansel Minutes. Today I'm chatting with Alan Stewart. He's a partner director of software engineering innovation at Microsoft and a supporter of AI for science. How are you, sir?
00:01:25
Speaker
I'm doing great. How are you doing, Scott? I'm getting there. I'm ready to learn. I'm ready to learn. You know, like somebody was telling me that they're worried that my podcast is going to become like an AI, like slop kind of talk. Well, all we do is talk about AI. i am interested in solving real problems for real people. And if it happens to be using AI, then that's cool and that's that's fine. And everyone's saying agent, everyone's sick. If you want to be smart in a meeting, you say agentic.
00:01:50
Speaker
And i think we've had a lot of talks about that, but you are you believe that it's memory that is the trick. That's the special sauce. I believe memory is that closed loop. It closes the loop of you know the cognitive engine and AI.
00:02:05
Speaker
doing research, that being stored in a memory store, and then that being used again by the AI to either unlock a new path of research or to wouldn accelerate um the existing path of research. Let me tell you what I mean. So if somebody puts a science problem there and like I recently ran a science job, it's ran for like 14 days and spent millions of tokens, right?
00:02:27
Speaker
And some of the thoughts in the loop were completed and some of them were incomplete thoughts. But what I wanted to do is I wanted to prove that any memory, any token expenditure from an AI perspective is good research.
00:02:38
Speaker
So I took those memories, packaged them up, and I was calling them a science pack of knowledge, took them to a new machine, restarted the same science example. And that example started 150 million tokens less because 150 million of the tokens were creating the plan for how to solve the problem.
00:02:56
Speaker
Right? So the next system picked up directly from there and started, but it also picked up the incomplete thoughts and started research from that perspective too. So just from a token efficiency perspective is 150 million tokens less to start that research moving forward, just from the memories of the system. So think about that, that exhaust being used.
00:03:18
Speaker
solving the problem, using memories, starting the problem again, using those same memories to solve either some of the similar problems or new problems moving forward. So memories have, has been that closed loop for me. I've been doing a lot of work there. So so is this the AI equivalent of no bad ideas in a brainstorm, right? Like there is no wasted tokens from your perspective. you're going to pick up all the exhaust and feed it back into the system.
00:03:40
Speaker
Yeah, because think about like a science problem where maybe I'm trying to repurpose or reanalyze a drug ah drug for Alzheimer's. So every token expended is some sort of research path that could either unlock a new path or, like I said, speed up an existing existing science problem, right? These problems run for long long-running problems, right? Trying to repurpose and use you know millions of catalogs of science data, either chemistry or biology, moving forward.
AI Memory and Research Optimization
00:04:07
Speaker
Yeah.
00:04:07
Speaker
ah Okay. So you're you're deeper in this than I am. I'm focused primarily on solving problems for programmers. but You know, you've you've done, you're doing like postgraduate work and you're thinking about this deeper than I am.
00:04:20
Speaker
Put this into the context for the regular Joes and Janes who are listening for whom memory is a bunch of markdown files. Right? That's like the, you know, the poor man's memory. It's just kind of like what I'm keeping track of. What is a real legitimate enterprise level memory system and how is that different than a bunch of MD files in storage?
00:04:39
Speaker
Yeah, so you know your markdown files, are they're a record of context that you've pushed into the system to either build something, and then out of it, you're measuring some sort of outcome out of it. So you think about it just from us a scientific perspective, we're pushing a science challenge problem into the system. some you know Look, the the type of customers we work with, scientists, they're trying to either you discover a new chemical, discover a new biological agent. So they push a challenge problem into the system,
00:05:05
Speaker
And as the system is decomposing the problem and researching or deciding how it's going to research, all those activities from the tasks, the agents, we call them subtasks and tasks and thought channels, all that thought channel work that the agents and tasks are doing,
00:05:22
Speaker
to, to build to the problem, all that's memory is generating exhaust from the process. Hey, um I want you to go research 10 different drugs in this thought channel. So that channel channel is actively researching.
00:05:33
Speaker
And what they're finding out is all the memories that we associate into the system that we store into the system. And then we use as a part of the cognitive engine again, as I'm to either re prosecute some additional research or some of the same research.
00:05:46
Speaker
But there's a confidence score between the cognitive engine and the memory store that says, is this a relevant memory? Right. I'm researching drugs. Is this a relevant memory or not? Right. If it's score, if the memory has a confidence score of three or higher is stored in the actual store itself.
00:06:02
Speaker
If it doesn't have a confidence score of three or higher, it's stored in the actual database for regular use down the line, but it's just not relevant for this particular problem. So there's there's signals that are happening between the cognitive engine and the memory store that signal, hey, this I want you to query for a memory. This memory is relevant. Now I'll ground my research in this data, or it's not relevant for this problem, but still stored in the the store. It may be relevant for another problem down down the line.
00:06:29
Speaker
Okay. I want to make sure that we're grounding this in as as simple analogies as I understand and that I'm pushing back as appropriate and as I am known for so that everyone can come along for the ride. As you're known for, yes. Well, what I'm just saying, like as because I want to make sure that everyone gets to come along for the ride because some people might be thinking word salad. Some people might be saying, I'm kind of where he's at, but I'm not 100% AI is the overarching thing on top of and ML, deep learning, neural networks, and what happens to be generative pre-trained transformers. Are you doing primarily work in GPTs or are you combining all the different things under the AI umbrella, only some of which are tokens burned on next token prediction?
00:07:17
Speaker
We're combining all of them together. Anything from building, you know, look for science, we use large language models, but the way we use them is not the way people think we use them for science. We're not using them. They're like, you're like, what drug's best?
00:07:28
Speaker
right Right. And then you put it chat GPT. That's what we'll think. Like, hey, let's put it into chat GPT. So, build me a drug to to solve this. We're not using it that way. Right. For science, we're using specifically trained and fine-tuned scientific models. right okay yeah that are trained on scientific process.
System 1 and System 2 Thinking in AI
00:07:44
Speaker
We're using LLMs to in either orchestrate agents or enhance or summarize some of the context.
00:07:50
Speaker
Or we have a cognitive engine that we call Clio. kind And Clio uses system one and system two thinking, right? Slow thinking and fast thinking in the system, right? So... For slow thinking, we input that something that has to churn. This is a huge challenge problem. You're gonna burn a lot of tokens to try to solve it.
00:08:08
Speaker
So we push that into what we call a slow thinking channel, right? And a model, we have a specific model there that works on that problem. Slow thinking, lots of research. gonna You're gonna spend a lot of tokens to get to an answer. then we have a fast path, right? Where we say, okay, look, this is, you're summarizing some data. mean Maybe you're taking something out of the fast pass and you're summarizing in a table or summarizing in a markdown file to be used later as a part of research as you get further down the chain. So we have two kinds of thinking mechanisms there. And each one of those things are exhibiting exhaust. Those exhaustive memories that are coming out of a system, whether it's markdown file or show me the top 10 drugs that meet this criteria.
00:08:47
Speaker
in a file. We're pushing that down further into the system. There's ah one thing our cognitive engine does is a very resilient brain. It works across transient errors, problems.
00:08:58
Speaker
We use the coding agent, we use the deep research agent, and let's just say you're unable to get to the coding agent. It will work around that. Okay, I'm not able to get the coding agent. Let me use a LLM um to do some research. Let me use another tool or mechanism that maybe the system designer put in to be able to do that work. So we're using every mechanism to do science. We we need all of them to do science. pretty complex. not talking about and And for context, for folks that may not be familiar with those those concepts, System 1 and System 2 thinking came from Daniel Kahneman's book, Thinking Fast and Slow. so System 1 is kind of automatic, unconscious. It's your gut. It's your intuition. It's how you get from point A to point D yes without actually showing your work. Now, when when when we were kids and the teacher was constantly like, show your work, show your work, they were forcing you to go from System 1 thinking to System 2 thinking, which is more effortful.
00:09:47
Speaker
more complex. Now, people might push back on saying that
Mimicking Brain Functionality in AI
00:09:52
Speaker
next token prediction and models think, but I'm noticing that you and others are all using effectively anthropomorphized descriptions. You say memory. Yes. We say memory when we talk about computer memory, but you're also saying like cognitive engine. And some people might wince at that going like, it's not a brain, but it's kind of is trying to be modeled like a brain? We're trying to model it like a brain. We're trying to model like a brain where we have, you know, the cognitive engine that's gonna orchestrate the agents, the tools and take in the challenge problem and orchestrate, you know, the ensemble models that we have to get to an answer, to orchestrate the memories as well, right? The cognitive engine says, I'm gonna start on this problem.
00:10:31
Speaker
Hey, memory store, I'm building this type of drug. Do you have anything there? yeah I have relevant memories. Let me ground myself in that in that data. Let me ground myself in the tools. And then if I need to create a tool because one's not available, we have the GitHub Copala CLI SDK there for the agent to build a agent to build the tool that needs to that needs to be built for the research moving forward. So it is kind of orchestrating and operating like a brain. I know people like, you know, people kind of cringe when you hear some AI people talking about we're trying to build a brain here. And they're like, all right.
00:11:04
Speaker
We're far away from that, but we are trying to model out a cognitive engine ability to store short-term and long-term mut memory, procedural memories or episodic memory in a store. And then some level of engine that says, Hey, let me query. And is this relevant to relevant to what I'm trying to research and use it?
00:11:21
Speaker
Right. we I'm going to write a paper that we, you call like the enterprise science brain. Right. So every time you use the system, the system that we build a system that we call discovery to do science, scientific research. Every time you research, the system is learning and growing and becoming smarter.
00:11:37
Speaker
Okay. So my son was in art school for a while and he would always start drawings and then he'd like stop and scribble them and throw them away. You're arguing, never do that.
00:11:47
Speaker
Never do that. There's no, no such thing as a wasted token in science research. There's some way we can use that and leverage that. You may think that this not relevant. It's not relevant for your existing challenge problem.
00:11:59
Speaker
But in the field of science, it's probably relevant to some additional research that's happening. And might that one partial memory might be the thing that unlocks...
Adaptability of Cognitive Engines
00:12:09
Speaker
A breakthrough.
00:12:09
Speaker
Okay. So in a world where people are concerned about the the weight of AI and on on on ecology, on you know eco-friendly AI is not something that you would hear said, one could argue that the cheapest thing in the cloud right now is storage.
00:12:26
Speaker
And the most expensive thing in the cloud right now is making tokens. So by saving them all, you're potentially allowing agents to not have to restart from scratch, which is just wasteful.
00:12:37
Speaker
Absolutely. And, you know, during my research, even even doing the research where I run this this large science job and i I took all the memories from the completed channels. And then I look back and it's ah we have 33 incomplete channels with 30 percent memories. And I said, hold on, I want those, too.
00:12:53
Speaker
Let me package those as well. I package those and you take them to another system. and You start to run and you see the system. Take even the partial memories and say, oh, these are relevant. And you actually see the cognitive engine and the memory store agreeing that even the partial memories are relevant to the research and using them moving forward. So like I said, I literally proved there's no such thing as a wasted token in science research, right? Okay. So is this doing, and I know I'm i'm being indirectly diminutive, it's thing it's kind of doing rag over its own
00:13:25
Speaker
partial work and deciding what to show its attention to. Because arguably though, you still have, and you don't have an unlimited context window. You have a large context window, but you still have, context is still a thing.
00:13:38
Speaker
Okay. Yeah. If you look at the the cognitive engine in the background is breaking down, breaking down context, it's separating files, it's splitting things out, putting them into markdown files and keeping you away from that that ah lo that context limitation explosion, right? We're we're kind of moved away. yeah Eventually, things do decompose down to to a prompt. But how we're doing them within the cognitive engine, how we're breaking them apart, splitting them, separating them, even images, right? You know, if you do a lot of research around images, even images have a limit around the the the scope of a number of images. So in the cognitive inches is handling those limitations and working through them, working around them. Nice. I like that. And then the partial memories have value because they are preserving explored territory. It's like the fog of war when you play like, you know, Age of Empires. Like, okay, I've been there before. i do not need to return.
00:14:30
Speaker
Yeah. You're absolutely right. Even do even while i was doing memory research, I kind of thought about it a little bit. Like, I'm like... Is this really advanced? Because, you know, really what we're doing is restoring stuff. Like, an a I'm looking back at well the the development I used to do in financial service, restoring stuff in a database and recalling it. Okay. just That's basically what we're doing. There is some parts of, hey, the cognitive engine,
00:14:51
Speaker
the dynamic nature of storing thought channels in a memory store. What do you use for a memory store that that gets to be a little AI-y, right? Procedural versus episodic stuff. Like we ran our our cognitive engine, we ran against science problems, but we also partnered up with the M365 research team and we ran it against office tasks, right?
Efficiency in Scientific Problem Solving
00:15:11
Speaker
There's no way to prove out your thinking brain than by saying, hey, you're a science brain. Now I'm going to you something totally not relevant to science.
00:15:19
Speaker
Just really more procedural, right? I want you to send an email. In that email is going to be a number. If 518 is in there, read it via OCR. If you find that via OCR, generate a new document and send an email. The multitask task kind of stuff.
00:15:35
Speaker
But even there, we actually saw at our cognitive engine being very relevant for even those kind of jobs as well. Right. So without any what we call a warm start, meaning giving it data, procedural information. Here's how you send an email without giving it any information. It was very efficient in in using working on science tasks and working on office tasks as well. Yeah.
00:15:55
Speaker
Okay. So give me something concrete. ah I went through three or four different scientific papers and PowerPoints and Word documents. and I don't know what's secret and what's not secret. Okay. i do know that you did do a 12-day autonomous investigation just a couple of weeks ago.
00:16:10
Speaker
And the first run was you know hundreds of millions of tokens across and almost 200 thought channels. Yes. Then you did a second run. How are those runs different?
00:16:21
Speaker
Great, great, great deal. So look, you know, science problems, they run over multiple days, right? And they're, they're, people say AI is non-deterministic. A science challenge problem is non-deterministic on when it's going to start and how it's going to finish, right? So we ran this long job over 14 days.
00:16:38
Speaker
it was a sample that I built to really kind of exercise all of the AI capability. In our engine, metacognition in our engine, deep research in our engine, the dynamic nation notion of thought channels, like spin everything into a task and a thought channel.
00:16:53
Speaker
And then setting up the memory store to be dynamic as well, to store all those memories in the thought channel. So if I spin off a third channel, a thought channel and inflammation, Everything in that thought channel now is stored in the database under the thought channel inflammation and any task that builds out of that is stored there as well.
00:17:10
Speaker
That channel activity is what I was able to package up and then take to a new system and rerun the same sample. And one of the observations was 150 million tokens were expended building the plan for doing the research.
00:17:25
Speaker
I didn't know that from the first job, right? It's so unique, it's running, AI's cranking through it. Here's the activity I have to be able to solve this problem, but it expended that much in just setting up the plan. So when I moved those memories to a new system, the plan of research was laid out or automatically from the previous system in in the memory store. So the first query from the cognitive engine was to the memory store, do you have anything relevant to this problem?
Ensuring Effective AI-Driven Research
00:17:49
Speaker
And the memory store said, oh boy, do I have a lot relevant to this problem. I have about 1200 memories that I'm gonna give you that's relevant to this problem. Thought channels, full, complete memories and impartial memories.
00:18:01
Speaker
So it actually started the research 150 million tokens less So it started not from square zero, it started like at second base. At second base. act second yeah I know like some people would say 150 million tokens is second base.
00:18:16
Speaker
For science problems, that's second Yeah, if it's a big problem to start with. Yes. For sure. That's second base. Yep. Okay. Now actually, the research on the second job, the second job got a lot much further along in the process of actually being able to solve it.
00:18:31
Speaker
Many of the channels actually converge. Converge means they finish to their work. They were able to finish their work and come up with output. Maybe the output was come up with 10 candidate drugs for this problem.
00:18:41
Speaker
So was able to finish and have a finished output, a markdown file, graph, something that fed into the broader research. Okay. Now, keeping in mind that, you know, I can only do so much research and I don't work in this space. My goal is to make coders faster and your your work is to make scientists faster.
00:18:58
Speaker
I would think, though, that a low quality memory could bias or mislead a run later. Absolutely. Okay, but you said there are no bad memories. Is there a risk of a bad memory compounding and how would you pull out like that was a bad idea in a brainstorm and we're going to throw that one out?
00:19:16
Speaker
Absolutely. So that's where this, the between the engine and the store, there's signals, they're called case signals between the engine and the store that says, They're measuring the efficacy of the actual memory to the problem.
00:19:29
Speaker
Is this irrelevant? We're giving it a confidence score, or relevance, say one through five,
Developing an Enterprise Science Brain
00:19:33
Speaker
right? So one to two are low value memories. Low value memories, there keep them in the store, but they're low value memories for this problem. Do not ground in ground in the problem in the in these memories. A three or above,
00:19:46
Speaker
are groundable memories, meaning they have some level of fidelity to the problem. Okay, there it is. It just clicked for me. So you have a library and you don't know if the stuff in the library is crap or not.
00:19:57
Speaker
So you're going to keep it all. And some of it might be pulp fiction and some of it might be the great American novel. But if you're doing a problem about pulp fiction, you'll go and you'll get the crappy memories and they're going to lead someone in some direction. So you throw nothing away, but you also don't have to keep everything in context.
00:20:12
Speaker
You just pull the bits that are that are pertinent. Absolutely. there's that there's glu There's this interaction, this dynamic interaction between the cognitive engine and the memory store. Really, the work between the two, because these are two separate things, right? We work with the M365 research team on on their memory store, and we built the cognitive engine. So the first part of the activity was marrying up the signals, if we call it, right? The signals of when something is relevant,
00:20:37
Speaker
When to call the memory store for the cognitive engine, right? Of course, when you're starting a problem, absolutely call the memory store. But when you're doing subtasking you or starting new channels of activity, that's another opportunity for you to call the memory store to get memories out of the system, rank them,
00:20:53
Speaker
If they're efficient for the problem, use them. If they're not efficient, keep them in the store and use it. And this whole idea around, think about it for a second, us a user gets, a scientist gets sick down, sit sits down his computer, says, i have this problem.
Grounding AI in Reliable Data
00:21:06
Speaker
We can query an enterprise memory store and say, hey, you're trying to do immunology. Here's all the memories on immunology. Would you like to use it? And somebody would say, well, why would we want to surface that to the scientists? We just have AI do it automatically.
00:21:19
Speaker
Scientists don't want to just be disenfranchised from science, right? They feel like science shouldn't be science, AI is a tool for them to do science. They feel like they get they're SMEs. They can look at the memories and say, yeah, these I'll select them.
00:21:30
Speaker
But behind the scenes, we're doing the automatic selection as well. That whole cognitive engine dance behind the scenes that we're doing between the signals of the cognitive engine, the memory store, we're doing that as well. So in addition to the scientists picking some of these tasks, right? So it starts to look like,
00:21:45
Speaker
This is a learning system. So last week I didn't have any inflammation-based memories. This week I have 1,500 inflammation-based memories. How did that get there? Other colleagues doing research. So the system is continually to grow grow from that research, right? So yeah, very exciting around that.
00:22:02
Speaker
Okay. So someone might be, if they made it 20 minutes into the show here, they might be thinking, well, hang on, I can't get Claude to make me a mediocre website and you're making drugs.
00:22:13
Speaker
Like this thing, you know, I talked to Resinovich all the time and he's always telling me how the things gaslighting him, commenting out tests. If ever there were a thing that needed to be grounded in data, it would be something like this. Oh yeah. How do you ground it in data? Because I know that there's a far more, there's far more data around like drug repurposing and things like that. That's like, that's testable.
00:22:36
Speaker
You need to stop hallucinations at the, at the root. Yeah. When we first started our journey down the Afro-Scientist journey, right? We call that area knowledge, right? Knowledge. How much knowledge do you need to get to give to the system to do science? Because remember when we started out,
00:22:51
Speaker
Going to a a large language model says, give me something with smiles notation. The large language model would understand what smiles notation was, but submit you something that you could look at and you say, this is totally, absolutely, positively, a thousand percent wrong.
00:23:08
Speaker
Right, because miles is simplified molecular input line entry system. It's basically ASCII yeah for chemicals. For chemicals, exactly. It doesn't know that. It doesn't come baked in and it can probably gaslight you and make up all kinds of nonsense. And it absolutely was, right? So, you know large language models are not great for science without proper grounding. So we work with MSR and we built this capability called GraphRAG, right? So GraphRAG is how do you use knowledge graphs? So moving beyond RAG to, you know, RAG is, you know, very, is not in dynamic, right? It's a fixed ontology. You put your data in there,
00:23:43
Speaker
It doesn't grow. It's just, here's what I have. Here's what I'll ground. But when you build in a knowledge graph or you use graphs, graphs are dynamic ontology. So the more data you feed it, the more relationships it builds and the more communities it builds as part of the data.
00:23:57
Speaker
So we built what we call a scientific bookshelf. So, hey, Mr. Scientist, you're a chemist. What is your bookshelf?
00:24:05
Speaker
your materials that you use to do science and your tooling that you use to do science. Those two things are super important, right? And we can incorporate those into the discovery system. The knowledge and the tools, once you have that, you marry that with fine-tuned models, science models, and fine-tuned knowledge graphs of chemistry, bio, depending on what what area science you're working on.
00:24:27
Speaker
And you use that to ground your answers in. And that's how you get around this whole, hey, a large language models are not good for science, they hallucinate.
Avoiding Ambiguity Loops in AI
00:24:35
Speaker
How do you get to a hallucination-free system, right?
00:24:38
Speaker
go There you go. So I've been calling them, and I'm trying to get this to catch on ambiguity loops. And if you've got a loop, like a for loop is very unambiguous. It's very clear. It's procedural. It runs exactly as you expect. But an ambiguity loop says, hey, you know, a AI, go and do a thing.
00:24:53
Speaker
And I didn't tell you how to do the thing. i say, hey, go get me bread and milk. Didn't say take the bus. Didn't say walk. You decide. And sometimes they're really good at that. But there's the car wash example that everyone thinks is funny, where you tell the AI, I need to go and get my car washed. The car wash is about 300 meters away. Should I drive or should I walk?
00:25:12
Speaker
And many, many AIs will say, well, it would be silly to drive. yeah And then they have you walking off to the the thing but because it doesn't realize that it needs to go there in order to wash the car and take the car with it.
00:25:24
Speaker
But ambiguity loops are sometimes helpful because you leave the part that is ambiguous because you don't know yes and you wanted to explore that space. But when you do know stuff like like scientific data, it must never...
00:25:38
Speaker
fabricate that data. Yeah, must never, fat because, you know, the difference is a smile notation is a different drug. It's a single bite. Yeah, that's a great point. You should go and people who are listening should go and look up Smiles, the the specification. it's It's such an important thing. like you know Ethanol is like CCO, but you change one character and it's a... you know What do they say? you remember the old thing in chemistry school? What was it?
00:26:05
Speaker
Tommy was a chemist, but Tommy is no more. What Tommy thought was H2O was H2SO4. didn't mean it. So like stuff like that, you just don't want to make those kinds of mistakes or otherwise you're going to be drinking sulfuric acid, Tommy.
AI Integration in Lab Environments
00:26:19
Speaker
Absolutely. Look, look at my journey. I look, I had my own journey to science, right? I'm a computer scientist. And like when we found out that we're going work in the AI for science, you know, I got to spend some time with scientists and that was very, that was very eyeopening, right? You get to start spending time with scientists. first of First of all, they said, look, Hey, AI guy, science, scientists do science. Do you understand? And I was like, I understand.
00:26:44
Speaker
AI is, they're like, repeat it back to me. What did you hear? AI is a tool. Scientists do science. they're like, okay, we like this guy. There you go. let stay You got it. Understanding what we're trying to do here, right? So they will be in the loop in this process, right? Even though we're trying to, like another area that I'm working on right now is this area called lab in the loop, right? So you think about the science, the hypothesis is I'm going to give you, feed you some stuff about a chemical property and I want you to come up with a new chemical, candidates for a chemical.
00:27:10
Speaker
What about making the chemicals? oh Now, I want to take that formula in the lab protocol, I want to submit it to a bunch of robots in the lab, and want to synthesize chemical. And this is a real thing.
00:27:23
Speaker
And the beauty of that is it brings AI, physical environments together, and science at the same time. So you get a bunch of autonomous systems, you get some robotics in there, you get some AI and science at the same
Self-Regulation and Data Accuracy in AI
00:27:34
Speaker
time. So really exciting area of focus. Interesting. And then the last thing that I wanted to touch on, because I thought it was interesting as we get towards the end here, was that, you know, we see these kind of random ah emergent things happen. Like, it's like, I didn't expect it to do that. You would not expect a large language model or any system to change try to develop a discipline of anti-fabrication too. They tend to float away like balloons and in helium, and they tend to kind of fall off the rails and they just start making stuff up.
00:28:03
Speaker
But in the building of this system, you started to notice it repeatedly policing itself, catching itself, and then re-grounding in the data. Why do you think that happened? Memory.
00:28:14
Speaker
ah The effect of memory on the system, the ability for it to say, hey I have a challenging area, like something in our cognitive system. There's there's path there's a path A, which is the preferred path.
00:28:27
Speaker
And I want you to use these tools. Then there's path B. Non-preferred, you war something happened, there's some transient failure infrastructure, I'm not able to get this tool. It's another path for being able to research and and get to a result. That's one thing our cognitive engine will do. Because these things run in a long, wrong running multiple day, the cognitive engine will get to a result. It's trained to get to a result.
00:28:49
Speaker
Failure in this new path, failure X, create a new path, right? It will continually try to execute new paths. And by adding this dynamic memory capability to it, where before the path was, all right let me ask the large language model. Since I'm not able to get to a coding agent, let me ask the large language model to write code for me or something like that, right? So in this area, hey, I'm i'm stuck a little bit here.
00:29:10
Speaker
Guess what? I have a memory store. Let me query the memory store for actions or activity or insight on how to generate the next problem, right? So the memory store has been really, it's been really a breakthrough. There's been, and then the interaction between the cognitive engine and the memory store, that interaction, right? The ability to pull memories, use memories at runtime has effectively changed a lot of behavior in long running jobs, which scientists, science jobs are preferred.
00:29:37
Speaker
So where where would we learn more? Like, is this, are there papers? what am i What am I Googling for to learn about this stuff? Yeah, the first thing I'd like to direct you to is the product that the division is building around for scientists to do um science, and that's called Microsoft Discovery.
00:29:52
Speaker
Okay. So Microsoft Discovery, great great tooling. You start to introduce some of the concepts that we talked about. If you want to know more about our are a cognitive engine, There's a paper that we put out for, I'll send you the link. I'll send you the link. I'll put it in the show notes. Yeah. It tells you how our cognitive engine goes about prosecuting its work and doing its activities. And then some of the memory research that I'm working on right now, hopefully in the future, i will we'll be able to publish some of the memory research and the effect of memory on scientific problems moving forward. so Very cool. Well, thank you so much, Alan Stewart, for hanging out with me today. I've got a lot of reading to do, it sounds like.
00:30:28
Speaker
ah Come on. Absolutely. Anytime, Scott, you know that. All right. I appreciate you. This has been another episode of Hansel Minutes, and we'll see you again next week.