Beyond Chatbots: Anthropic, SandboxAQ, and AI’s Move Into the Physical World

Anthropic’s partnership with SandboxAQ may sound like a technical announcement, but it points to a much bigger shift in enterprise AI: moving beyond chatbots and productivity tools into physical-world decision-making. 

In this episode of the Macro AI Podcast, Gary Sloper and Scott Bryan explain how SandboxAQ is integrating its Large Quantitative Models, or LQMs, with Anthropic’s Claude through MCP — the Model Context Protocol. The key idea is simple: Claude acts as the natural-language interface, MCP provides the connection layer, and SandboxAQ’s quantitative models perform specialized scientific calculations. 

The discussion breaks down why this matters for business leaders and CIOs. Large language models are excellent at explaining, summarizing, reasoning, and orchestrating workflows, but they are not designed to be physics engines. Large Quantitative Models are different. They are built to model scientific, mathematical, physical, and biological systems. 

Gary and Scott explore how this architecture could affect catalyst discovery, battery development, drug discovery, industrial R&D, and materials science. They also explain why the real enterprise opportunity is not replacing labs or expert systems, but improving the funnel before expensive physical testing begins. 

The episode also covers why MCP matters as an AI-native integration layer, how CIOs should think about security and governance when AI systems can call tools, and what this partnership means for the broader competition between OpenAI, Google, Microsoft, Anthropic, and specialized AI companies like SandboxAQ. 

The takeaway: the next wave of AI may not be about generating more content. It may be about helping businesses make better decisions about the physical world. 

About your AI Guides

Gary Sloper

https://www.linkedin.com/in/gsloper/

Scott Bryan

https://www.linkedin.com/in/scottjbryan/

Macro AI Website: 

https://www.macroaipodcast.com/

Macro AI LinkedIn Page:  

https://www.linkedin.com/company/macro-ai-podcast/

Gary's Free AI Readiness Assessment:

https://macronetservices.com/events/the-comprehensive-guide-to-ai-readiness

Scott's Content & Blog

https://www.macronomics.ai/blog

Transcript

00:57
Thanks for joining us on the Macro AI podcast. I'm Gary Sloper and I'm here with my co-host today, Scott Bryan. And Scott, for the last couple of years, ah you we've talked a  lot about this and you're starting to see it more. Most executives  have experienced AI as something that lives in a browser window. And I think, you know, that's where many people are familiar with what I would term AI.

01:25
You type something and it writes something back. It drafts an email.  It's performing a task like summarizing a meeting or helping you with coding. Might even be building a PowerPoint slide. And I think that's been useful, but it's also shaped how business leaders think about artificial intelligence. They think of it mostly as a productivity tool. Yeah. No, I agree with that. think that's kind of a valid perception. And I think obviously there are plenty of organizations that are

01:54
much deeper with their AI use, like the healthcare or life sciences vertical. And then also things like banking and financial services. But in general, I think I agree with that. I think most business leaders think of AI as productivity tools. Yeah, that's a good point. think one organization that many people are familiar with, Anthropic, Anthropic and the Sandbox AQ partnership that was recently announced.

02:22
points to something different. Anthropic announced that Sandbox AQ is integrating its large quantitative models or LQMs uh with Claude.  And that's being performed through MCP. We've talked about  MCP on prior episodes. And so if you're not familiar with that, please go back and listen. It's uh good information.  And this will round out Anthropic's model context protocol for end users. And this is AI moving beyond

02:51
things like documents and emails and chat bots and PowerPoint slides we were just talking about. And it's really starting to connect into things that are really intriguing, such as scientific models and physical simulation  and advanced materials, batteries, uh catalysts, uh drug discovery. We're seeing quite a bit of that as well. And  just overall industrial research and development. So in other words, AI is moving from the office

03:20
that many are familiar with into the physical world, which should help drive artificial intelligence adoption and acceleration of various use cases. And I think to simplify it, Claude becomes the natural language interface. MCP becomes the connection layer. So Sandbox AQ's quantitative models become the specialized engines, which  is really interesting. Yeah. And I think that's the key.

03:47
Claude is essentially just acting as the front end, the reasoning layer,  and the orchestrator. uh So for example,  a scientist or engineer can ask a question in natural language, like everybody's comfortable doing.  Claude understands the request. MCP uh connects Claude to the right sandbox AQ model. And then the sandbox AQ model performs the specialized calculation. And then Claude can then...

04:16
articulate it back. It can explain the result in a way that humans can use. And I think that's uh definitely a different architecture from just asking a chat bot to leverage uh LLM training data to generate an answer. And I think for  CIOs, other business leaders, I think that's kind of an important point is that  the future of enterprise AI isn't just which LLM should we buy, it's how do we safely connect

04:44
LLMs to the specific tools, specific models, data, and the workflows that actually run the business. Yeah, that's a point, Scott. mean, the whole governance of what you just mentioned, I think is critical. Yeah.  And to really break this down, if we were to  simplify the terminology, especially if this is your first episode listening into our show,  if you think about LLMs versus LQMs, an LLM or as we...

05:11
Most all know is a large language model. It's built to understand and generate language code and other forms of content. It's very good  explaining and summarizing and planning, reasoning, drafting, and interacting with the end user, us as people. exactly. Yeah. And so if you were to flip that, an LQM or a large quantitative model we just talked about is different. It is built  to model mathematical

05:40
scientific, physical or biological systems. And I think that distinction matters because a fluent answer is not the same thing as a valid answer, right? Right. Yep. Yep. Totally agree. So an  LLM can explain, you know, what a catalyst is, but an LQM can help screen a catalyst candidate  or, you know, let's think of another example. Like an LLM can describe battery chemistry, but an LQM, a large quantitative model.

06:10
can help  model actual candidate battery materials. So it gets much, much deeper. So the value is not that the, you know, the LLM doesn't replace those expert systems. The value is that the LLM makes connected expert systems easier to use. So we're familiar with the LLM. The LQM would be the expert system that it's connected to. And I think one of the most kind of important business concepts in this episode is that AI is moving from

06:39
you know, an answer generator to actually becoming a control layer in the workflow. Yeah, those are really good uh examples on how to explain that Scott. I think that's a good way to put it in. And I think if you're a business leader right now listening in saying, you know, why does this matter? And you really need to drill into the partnership a little bit more. And let me kind of explain this  because it'll be a little bit of alphabet soup. uh The first major example in this partnership is Sandbox

07:08
AQ's AQ Cat, AQ Catalyst. Think of it that way. It's the AQ Cat adsorption spin. So it's  ADD, not ABB, like absorption, it's adsorption spin, which is really focused on catalyst discovery. So that may sound pretty narrow, but catalysts are fundamental to the global economy. They matter in industrial chemicals, uh fertilizers,  fuels, hydrogen plastics,  anything.

07:36
sustainable aviation fuel and many, many other industries. So if a company can improve catalyst discovery, it may do some things that are really beneficial. may reduce costs. It could improve efficiency for the business. It could shorten those R &D cycles and ultimately create better products for their end customers. Yeah. Yeah. The ultimate key. I think the technical point is that AQCat isn't just a generic

08:05
molecule model. It's built for a really hard scientific problem. So most  catalysts involve transition metals like  iron, cobalt, nickel, and those metals can have a magnetic behavior. So spin polarization can matter. And if a model ignores that, it might produce an answer that looks clean, but it's just not realistic enough for an actual industrial system. uh

08:32
So actually, I have a quote here. I'm just going to read a technical quote so listeners can kind of get a sense of the LQM. So the quote is, AQCAT25 was trained on a very large set of density functional theory calculations, including roughly 13 and a half million density functional theory or DFT single point calculations across 47,000 absorbate slab systems with a spin polarization enabled.

08:55
The goal is near DFT accuracy at much higher speed so researchers can screen candidates before committing expensive lab resources. that's a quote. my  yeah. You could remember that quote by the top off the top of your head.  No, no, couldn't do that. My  my point, though, is that, you know, the point to and reading that is that  CIOs don't need to become computational chemists. But right. they really need to understand.

09:23
real specific capabilities of the architecture.  And this is a very specialized model that's trained for the physics of the problem. So they need to understand at least at  a technical fundamental level what  specifically the model can do so they can translate that to the business. Yeah, that's a good point that this is  a specialized model trained for the physics of the problem. And that leads to the business takeaway. Artificial intelligence does not replace the lab.

09:54
can improve the funnel before the lab, before getting into the lab. Yeah, that's huge. It can improve the funnel before the lab. Yep. Right. And if you're developing, let's say, new material,  maybe doing that drug discovery, a catalyst or battery component or physical testing, this is where it becomes essential, but it's expensive and slow. So if artificial intelligence can help eliminate weak candidates early,

10:20
earlier, the lab can focus on ideas, uh you know, most likely to matter for the end result of the output. Yeah, and I think the net is there that that really changes the economics of research and development. uh and I think that was just, you know, just one example. So sandbox AQ and this partnership that we're talking about with entropic uh also has models that are aimed at other things like you mentioned, like specifically batteries and drug discovery. They have

10:49
AQ Volt, which is focused on battery materials, uh especially the solid state battery research. And the business relevance there  is kind of obvious to things that we're talking about, like electric vehicles, grid storage, energy density, uh just overall better materials for  electricity. And then they have another model, AQ Affinity, which is focused specifically on drug discovery and binding affinity. So essentially,

11:17
how strongly a molecule binds to a target protein. And that obviously matters because early stage drug development is  full of expensive dead ends and AQ affinity can, like we talked about, narrow that funnel  for what they're going to focus on. So  like you noted, instead of kind of moving slowly through physical trial and error, companies can explore a lot more possibilities computationally using this new relationship and using these models.

11:46
They can then narrow down that list and then spend the physical testing dollars a lot more intelligent. that is,  you know, that's really why this is  bigger than science. The first wave of generative AI went after knowledge work. So, yeah. And we all used it, right? Things like documents and meetings and support coding, search and analysis, you name it. That's, that's kind of where AI started. And the next wave goes after the physical world decision-making.

12:15
research and development, engineering, materials, uh energy. And then I start thinking of different verticals like manufacturing and healthcare that have logistics and operations needs. So this is much larger business surface area to cover than I think  a lot of business leaders are familiar with. Yeah. Yeah. So I think it kind of, back up a little bit. We talked about MCP,  Model Context Protocol. uh

12:42
couple of episodes and I think episode 65 was the one that we specifically focused on MCP. So MCP, Model Context Protocol is  an anthropic standard for connecting AI systems to external tools,  data sources and execution environments. And it's starting to become pretty prevalent out there as a protocol. So think of MCP as a AI native integration layer. It helps an AI assistant.

13:09
discover tools, call tools, send structured requests and receive structured results. So in this partnership, Claude and Thropic uses MCP to connect to sandbox AQs, quant models, their quantitative models. But that same pattern could essentially apply to, you know, lots of different types of enterprise systems. Yeah. And that, that's the real enterprise architecture. Most business value,

13:35
does not live inside the LLM. It lives inside the company's systems like ERP or their CRMs, oh data warehouses that they have, financial models, or even proprietary repositories and databases. So the winning AI strategy is not  stuffing everything into a chatbot. The winning strategy is connecting the artificial intelligence environment to trusted systems in a governed way, kind of like how you started off the show today. Yeah. Yeah, exactly.

14:05
I think there's a second technical point that matters and that's  data minimization. So you don't want to  massive data sets back into the LLM context window. That's super expensive. It's not efficient. It has all kinds of risks. So I think a better pattern is to keep the data where it belongs. look at the highly specialized tool uh or model that has very domain specific language, go and process it.

14:33
and return only the useful output to the LLM, which will obviously  generate uh something that's easy to understand and articulate. And  overall that helps with security, costs,  latency, governance, all those things. Right. So at that point, the architecture is much simpler. The LLM understands the user, the protocol connects to the tool, the specialized model performs the work, the enterprise governs access.

15:03
the human expert makes the decision, that pattern is going to show up across many industries  here in the future. Okay, so then if we were to talk more about security for a minute, because I'm sure there's uh some CISOs that have some questions, uh and also as part of your organization, you have to have this nailed. uh When AI is just generating text, the risk is one thing.

15:30
But when AI calls tools, query systems, runs codes, uh accesses data, uh or even triggers workflows, the matrix really changes. And now  the artificial intelligence environment is not just a chatbot like we were just mentioning, it's  a software actor inside the enterprise, which is something that a lot of organizations, I still come across that have not wrapped their head around that yet. Yeah, software actor.

16:00
just like any other actor, the software actor needs governance. So they need identity, they have to have permissions,  they need least privilege access  and they need logging to  track it and they need monitoring and approval gates. So you mentioned, CIOs and CISOs, they should treat MCP servers and AI tool integrations as part of the enterprise attack service surface. And that's starting to become

16:29
a lot more prevalent in the news. And I think the key questions are, what tools can the AI access? What data can it read? Can it change anything? Are actions reversible? And what sensitive steps require human approval? And then importantly, because auditing is going to be imperative in the legal environment, can you audit what happened? And I think that's difference between an isolated demo and a real

16:57
production architecture, need to wrap all those things into it. Yeah. And I think,  you know, to that point, there's a couple verticals that come to mind, um, that, you know, this really plays into, um, the role of governance around security. And if you think of, you know, um, science heavy industries, the stakes are high because the data is so valuable. So  a pharma company may have, um, compound libraries and testing data.

17:26
materials companies have, you know, formulations and test data. Manufacturers have processed data and engineering models. You know, these are just a few examples. And the goal is not to  send all of that blindly into a model. The goal is to connect AI to the right systems under the right controls, you know, in these environments. Yeah. Yep. Good stuff. um Let's just pivot a little bit to kind of, you know, what this, what this means in the overall LLM race. uh

17:55
So everybody's watching that now, especially as know, Anthropic and OpenAI are um gonna have  public offerings. So, but this particular partnership, this Anthropic and Sandbox AQ partnership also tells us,  you what's happening  competitively. So for a while, the market's focused on benchmarks, model benchmarks. So which model is smarter? Which one writes better code? Which one reasons better? And then you've got the competition around being cheaper and faster.

18:25
And those things still matter, but I think the next competitive question may be a little bit different. It's going to be which AI platform becomes the best orchestration layer for specialized models. Also for your enterprise tools and your particular proprietary data and governed workflows. So I think that's kind of the next competitive question. Yeah, that's a point. Cause that's where it gets strategically interesting. Open AI has chat, GBT.

18:54
enterprise adoption, agents, APIs, and a huge developer ecosystem. If I were to look at Google, Google has Gemini, DeepMind, uh Alpha Fold, Isomorphic Labs, uh Google Cloud, and deep scientific AI expertise. ah If you're thinking about Microsoft, Microsoft has Copilot, uh Office 365, Azure, they have the GitHub repository.

19:22
security and perhaps, you know, one of the more stronger enterprise distribution channels today. Yeah, they certainly have that.  So then you  look at Anthropic. Anthropic is positioning Claude as a trusted reasoning and orchestration layer. And MCP could become an important, you know, part of that strategy. So Sandbox AQ

19:45
you know, represents another category specialized domain engines that plug into the LLM ecosystem. So this is not just a model versus model. It's an ecosystem versus ecosystem. Yeah. Yeah. I think that's, that's a, that's a pretty big, big shift to call out there. And, um, and, Thropic doesn't need Claude to contain every scientific capability internally. And that's kind of a key point. doesn't, Claude doesn't have to have that capability internally.

20:13
If cloud becomes the trusted interface into the specialized system, Anthropoc can still have a uh really valuable position in the enterprise stack. And I think the winning AI platforms might not just be the ones that try to do everything inside one giant model. They might be the ones that connect the best, govern the best, and help businesses turn that specialized intelligence into real outcomes. So I think it's important to note that the Anthropoc Sandbox AQ partnership  is

20:41
is not exclusive. So Sandbox AQ is  strategically and contractually free to partner with,  for example, they could go partner with XAI, the Grok model tomorrow. They could partner with OpenAI or Google or any other LLM developer at any time.  Sandbox AQ is using LLMs as a complimentary distribution channel per se. Yeah.

21:09
Those are really great points. I think for business leaders,  the strategic takeaway is that artificial intelligence value is shifting. The first wave was about productivity, right? We just talked about that earlier, the tasks that it would help you with. Can we write faster? Can we summarize faster, code, support customers faster? But really the next wave is about better decisions. Can we design better products? Can we discover better materials? Can we reduce failed experiments?

21:38
improve operations. uh Can we optimize supply chains and strengthen our cybersecurity? Maybe it's even, you know, can we allocate capital more intelligently? And that's a much, yeah, it's a much bigger conversation for the business. Yeah.  And that, to that point, that kind of changed the data strategy. So if everyone has access to super powerful models, then the competitive advantage shifts towards, towards what? Towards proprietary data. Right.

22:08
workflow, specific workflow integration, specific domain expertise and validation.  And the companies that win will not just have access to AI, they'll have access to unique data, real world feedback loops, and the ability to connect AI in the way that uh their business actually operates.  I think that that's why  AI strategy can't only sit inside an innovation team, it needs the executive leaders, the CIO, CTO, CISOs,

22:38
uh, the data leaders, business unit leaders, everybody, and, and, uh, and specific domain experts. Right. And that's how you have to protect your moat from somebody to just come and compete against you because they don't have, you know, the access to all the proprietary information and  workflows that you've built over the history of your company. Yeah. Beyond the LLM to domain expertise. Right. Right.

23:07
So if we were advising  CIOs, ah what should they do with all this, Yeah, good question. I think probably just  try to keep it simple and stop thinking about  your typical AI interface as just a chatbot. They need to really inventory the expert systems inside the company or ones that the company needs.

23:33
know, the models, the databases, the simulations, workflows, and the tools that only specialists know how to use. And then they need to find places where an LLM could become a better interface into those systems. And then build a governed integration strategy where that involves MCP or another standard to integrate the LLM to those systems. And then of course, we talked about security, always imperative and becoming even more imperative.

24:02
you know, put security in from the beginning that includes the identity, access control, logging is critical, sandboxing, data minimization. And I think most importantly, prioritize the use cases where the LLM connects to a trusted source of truth. um So that you can avoid, you know, the LLM potentially  inventing an answer for you. Well, yeah,  I think that last point is critical. The best enterprise AI architecture is not

24:31
Using the LLM as the source of truth or at least I hope they're not it is using the LLM as the gateway to the source of truth Yep. Yeah. Yeah, I think that's a that's a perfect way to put it Thank you,  um,  and I think in in this case, you know Claude is the interface to sandbox a queues You know quantitative models and another company and LLM might be the interface to a pricing engine a claims model financial planning system a digital twin

25:01
talked about that on the show, ah or whatever it might be. But this is where artificial intelligence becomes much, much more powerful. Yep. All right. So let's bring it home. The Anthropic Sandbox AQ partnership is a signal that enterprise AI is moving beyond chat bots and co-pilots into something much more strategic. AI is becoming an interface into the systems that actually make these decisions.  Yeah. Yeah. I think that's the, that line is pretty much the takeaway.

25:31
So, Plot is the interface in this case with this partnership that we're talking about today. Plot is the interface  or  the  interface to the source of truth, as you said. uh MCP is the connection layer  and Sandbox AQLQMs, large quantitative models are the specialized engines underneath.  And the enterprise provides the data, the governance,  security, and then the human expertise.

25:59
And like you noted earlier in the episode, that architecture right there is going to start showing up across lots of industries.  And for CIOs, the question is no longer simply which LLM should we use? And you may still be at that initial stage, but the better question is how do we securely connect AI to the tools, data, models, and workflows that run our business? Yeah.  And then you could add to that, you know, where can

26:28
AI help the business make better decisions, not just produce more content. Yeah. And that could be, you know, research and development, manufacturing, finance, cybersecurity, whatever. And I think the companies that figure that out,  uh, will move  much more, much faster and compete in entirely new ways in, in the new AI economy as it replaces the old economy. Agreed. So first wave of AI helped us generate words.

26:58
The next wave will help us generate better decisions about the physical world. Thank you for listening to the MacRaei podcast. We appreciate you  as listeners and please like and subscribe  and send this out to your network if you find it useful.  Thank you.