Why Gemini 4 Argon Puts Google Back at the Frontier | [Sidecar Sync Episode 157]

Summary:

Most model releases don't earn a full episode, but Google's Gemini 4 Argon makes a strong case for one. Amith Nagarajan and Mallory Mejias spend this deep dive on what Google's new frontier model means for associations, from where Google stood before the launch to what Argon actually changes. They explain why its ability to pause and resume its reasoning across up to a million output tokens matters for long, complex jobs, and walk through three launch demos: 1960s Fortran rewritten into modern C++ and Rust, a decade of DEXA scans turned into an explorable digital twin, and a humanoid robot that learned cornhole from a single video. Amith shares how Argon made the MemberJunction framework 2 to 15 times faster in key paths without introducing errors, introduces the More Cheese Project, and weighs in on the mixed reviews of the Antigravity harness.

Timestamps:

00:00 - High School Chemistry and Argon Fever
03:00 - Google's Return to the Frontier
09:37 - Six Days Without a Mistake
14:05 - Escaping Legacy Code
23:45 - Digital Twins & The More Cheese Project
30:04 - A Cornhole-Playing Robot
34:04 - MemberJunction, Up to 15x Faster
41:18 - The Antigravity Harness Debate
44:20 - Takeaway: Test It on Your Own Work

👥Provide comprehensive AI education for your team

https://learn.sidecar.ai/teams

🤖 Join the AI Mastermind:

https://sidecar.ai/association-ai-mas...

🎀 Use code AIPOD50 for $50 off your Association AI Professional (AAiP) certification

https://learn.sidecar.ai/

📕 Download ‘Ascend 3rd Edition: Unlocking the Power of AI for Associations’ for FREE

https://sidecar.ai/ai

🛠 AI Tools and Resources Mentioned in This Episode:

Gemini 4 Argon ➔ https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon

Google Antigravity ➔ https://antigravity.google

NotebookLM ➔ https://notebooklm.google.com

Claude ➔ https://claude.ai

Claude Code ➔ https://code.claude.com

ChatGPT ➔ https://chatgpt.com

MemberJunction ➔ https://memberjunction.org

DB AutoDoc ➔ https://docs.memberjunction.org/packages/dbautodoc

Betty ➔ https://bettybot.ai

More Cheese Project ➔ https://morecheese.org

Artificial Analysis ➔ https://artificialanalysis.ai

Every ➔ https://every.to

Apptronik Apollo ➔ https://apptronik.com

GitHub ➔ https://github.com

👍Please Like & Subscribe!

https://www.linkedin.com/company/sidecar-global

https://twitter.com/sidecarglobal

https://www.youtube.com/@SidecarSync

Follow Sidecar on LinkedIn

⚙️ Other Resources from Sidecar: 

• Sidecar Blog
• Sidecar Community
• digitalNow Conference
• Upcoming Webinars and Events
• Association AI Mastermind Group

More about Your Hosts:

Amith Nagarajan is the Chairman of Blue Cypress 🔗 https://BlueCypress.io, a family of purpose-driven companies and proud practitioners of Conscious Capitalism. The Blue Cypress companies focus on helping associations, non-profits, and other purpose-driven organizations achieve long-term success. Amith is also an active early-stage investor in B2B SaaS companies. He’s had the good fortune of nearly three decades of success as an entrepreneur and enjoys helping others in their journey.

📣 Follow Amith on LinkedIn:
https://linkedin.com/amithnagarajan

Mallory Mejias is passionate about creating opportunities for association professionals to learn, grow, and better serve their members using artificial intelligence. She enjoys blending creativity and innovation to produce fresh, meaningful content for the association space.

📣 Follow Mallory on Linkedin:
https://linkedin.com/mallorymejias

Read the Transcript

🤖 Please note this transcript was generated using (you guessed it) AI, so please excuse any errors 🤖

[00:00:00:00 - 00:00:01:09]
 (Music)

[00:00:01:09 - 00:00:10:01]
Amith
 Welcome to the Sidecar Sync podcast, your home for all things innovation, artificial intelligence, and associations.

[00:00:11:12 - 00:00:12:09]
 (Music)

[00:00:14:15 - 00:00:25:07]
Amith
 Greetings everybody and welcome to the Sidecar Sync, your home for content at the intersection of AI and the world of associations. My name is Amik Nagarajan.

[00:00:25:07 - 00:00:27:16]
Mallory
 And my name is Mallory Mejias.

[00:00:27:16 - 00:00:36:11]
Amith
 And we are your hosts. And today we are going to bring back either fond memories or vivid nightmares of high school chemistry.

[00:00:37:13 - 00:00:48:23]
Mallory
 Meath, you know, I look, I finished high school chemistry with an A, okay? I made good grades, but I will say chemistry is not my area of expertise. Can you talk about some chemistry to our listeners, please?

[00:00:48:23 - 00:01:03:11]
Amith
 I cannot because I am terrible at chemistry. Okay, good. By far my worst subject in high school. I did nowhere near as well as you did. Let's just leave it at that. In terms of high school chemistry, yeah, I was terrible at it. I was generally a terrible high school student overall.

[00:01:03:11 - 00:01:08:01]
Mallory
 Man, I made better grades than Amith Nagarajan. This is big news.

[00:01:08:01 - 00:02:06:10]
Amith
 You definitely did. I can assure you you did and they probably had much better grades than me in college as well, I'm guessing. But I was most studious individual. I was starting a company at the end of high school and early college. So that affected me somewhat, but I can't blame that. It's more to do with me being kind of lazy or inert, which is actually the topic of our podcast today. Not laziness or inertness, but argon, right? And I can't wait to talk to you about that. I have been completely nuts about this AI model, which I don't normally use terms quite that strong. I've been going nuts about it for the last, I think six or seven days. I think people all over blue cypress are getting really sick of me coming around and saying, "Hey, are you using Gemini for argon yet? Did you switch to anti-gravity? Is it blah, blah, blah, blah, blah?" So it's been good fun. But as we go through the conversation today, I'll be happy to share some of my observations and realizations with this new tool. But it's like every day we wake up and there's a new model, but this one is pretty interesting.

[00:02:06:10 - 00:02:16:11]
Mallory
 Did you immediately know with the release announcement that this one was going to be exciting for you or was it that once you started testing it, you realized, "Oh, wait a minute. This is special."

[00:02:16:11 - 00:02:59:17]
Amith
 It was more the latter because I take with a grain of salt all of the press announcements, all the benchmarks. I pay attention to all of them. And when argon was in testing, it was blindly being tested by a bunch of people as they do on, I forget what the leaderboard is, we've talked about it here before, but it escapes me at the moment. But it's a leaderboard that shows new models before they're announced. And generally people are able to figure out what they are. But it was well received there. And then the benchmarks that Google was kind of alluding to pre-release were very, very interesting. But my experience has been really next level with this model. And that's in comparison to Fable and Opus 5.5 and even GPT 6 in high mode. So quite excited by it.

[00:02:59:17 - 00:04:39:24]
Mallory
 Well, we're going to get right into that. Amith, normally, as you all know, if you've listened to our podcast, we don't typically spend a whole episode on a model release, but I think there's a lot here to unpack. So we are talking about Google's Gemini 4 Argon. Amith, as you said, you've been deep in it for the past few days. We're going to start with the release and then we're going to talk about three of the demos Google shipped at launch, which are pretty impressive. And then we're going to close kind of with the general reception, which there is some mixed feedback on and your takeaways as association leaders. So starting off with the release for most of this year, Google looked a bit stalled on the frontier at least. They have released four flash models between May and September. Flash being the smaller, cheaper workhorse tier you'd run high volume tasks through, but no new flagship at all in those months. The newest pro model is still Gemini 3.1 Pro from February, still labeled right now as a preview. And then they released Argon, which skips the pro naming entirely. Google announced Gemini 4 Argon on September 30th, but didn't release it publicly initially. It went only to vetted cyber defenders through the fair win program. Google's partner track for security work. Public access came two weeks later, October 13th. Pricing is $2 per million input tokens and $10 per million output running through November 30th and then doubling on December 1st. In the Gemini app, you need an AI Pro or Ultra subscription. Amith, before we get into what it does, what's so impressive about it, how do you generally think about Google compared to other AI companies like OpenAI and Enthropic in terms of model quality and then their approach to competing in this market?

[00:04:39:24 - 00:05:03:02]
Amith
 It's such an interesting question, Mallory, and so many people in the industry for a number of years have just been bashing Google, saying how they're behind and how their progress has stalled out. And this is not the first time, this year is not the first time that happened in the past. For fair reasons, Google was seen as a laggard. You know, back when GPT was first a big thing and chat GPT,

[00:05:04:02 - 00:09:21:14]
Amith
 Google had this chat bot called Bard. I don't know if you remember Bard. Barely. It was not very memorable, but it was a really terrible product and it was essentially, you know, a tool that was not even as good as GPT 2, I think, but it was really terrible. And so that was not a good start to the public's perception of when AI began. Because if you think about it, most association folks and then by extension, the general public probably really thought of AI as starting in early 2023, late 2022 at the earliest. Obviously, everyone has an intellectual understanding that it's existed for a long time, and we're not really preceding that. But the first time it really mattered to most association leaders and to certainly that the broader public has been in the last three years. And so at that moment in time, OpenAI was far and away the leading company. And then, you know, Claude only released their chat, I think, several months into 2023. And the first version was pretty uninspiring as well. But Google was behind for sure. Now, remember, Google is an organization that has for a long time invested heavily in AI. They acquired DeepMind. They had the Google Brain Initiative. And Google's original algorithm actually is technically a branch of AI, a very different branch. But so they have deep roots in AI and they obviously have a lot of resources. But they've taken the approach before the last couple of years. That was a little bit more on the academic side with it. They didn't really drive hard to commercialize. They had concerns over safety, over privacy, things that you want Google to be worried about. And they were moving pretty slow. They also, of course, are Google and, you know, they generate more cash in a day than most organizations generate in a lifetime. So it's a great place to be in on the one hand, but they did have a bit of the innovator's dilemma issue in that chat applications can potentially kill classical search. So there was an element of that going on, certainly within Google. So all that background is to say that since that point in time, they've really pushed hard and developed a lot of great things. We've talked about Notebook LM here in the pod Mallory. I know you've worked with that a bunch. That's a pretty interesting innovation. They did a great job with the original Gemini 3 series late last year. They have an outstanding real time model. Listeners to the pod and people who know me know that I've been, you know, it's been hard to get me off of my little shoebox this year about how important real time and voice is to the association communities. It's about being able to go into much richer, deeper modalities than just text alone. And Google has an outstanding real time model. It's called Gemini 3.8 Flash Live. And there's another one called Gemini 3.8 Flash Live Extended Thinking, which is a different model. And so these models as clumsy as their names are, are in my opinion, the best or one of the best models out there for real time voice. So we can talk more about that at some point. But today when we're talking about Gemini for Argon, I would say to you to your question about how I think about Google. I think they have a diversity of AI tools. They also, by the way, have the nano banana family of really well-named fun names around image generation, Vayo, which is their video generator. They have a world model, which is kind of in that same family. Some really interesting stuff. But you're right that this year they fell behind because the frontier moved along with Opus and Fable and Methos and GPT 5.5 and then 6. And Google was nowhere to be found when it came to the frontier. They did introduce some new versions of Flash, which were outstanding. Our companies have been using Gemini 3.5, 3.6, 3.7 and now 3.8 Flash extensively in our product suite because you have to remember these are, in fact, near frontier intelligence and in excess of the frontier of not long ago. You know, we talked a lot about on this pod, you know, how like last year's frontier is this year's mini model, right? And so what I would say is that that's actually really where the bulk of AI inference is happening is in Flash level models. These are very smart models, but they're faster, cheaper for most use cases. They're great. So lots to be unpacked there. But ultimately to lead an AI, you have to be at the frontier. Google by no means was ever conceding that turf. And I think with Argon, they're squarely back in the game, if not in the lead in many respects.

[00:09:21:14 - 00:09:34:22]
Mallory
 Mm hmm. I think on the episode where we initially talked about notebook LM many months ago, we said something like Google sometimes shows up late to the party, but when they do show up, they are well dressed and ready to go. So this might be an example of that.

[00:09:34:22 - 00:09:36:11]
Amith
 Yeah, for sure.

[00:09:36:11 - 00:10:34:12]
Mallory
 Moving to Argon specifically, Argon reads a million tokens at once. And as a reminder, a token is about three fourths of a word, but so does GPT-6 Astra and Claude Opus 5.5. What's new is that Argon can resume. Astra and Opus 5.5 cap at 128,000 output tokens and stop. Argon caps lower per request around 65,000, but it saves its reasoning state and picks up where it left off up to a million tokens on one job. What that enables is a single long job that stays coherent instead of 20 shorter and the stitching is where things often drift. Google's examples are formal proofs and generating code across many files, which is to say developer work, whether it matters outside engineering is an open question. Amith, do you feel like Argon is a genuine step change over what Google already had or a better version of the same thing?

[00:10:34:12 - 00:12:56:05]
Amith
 It's a step change over certainly what Google had, possibly over even something at the frontier like Opus 5.5. For the reason that you mentioned, it's a subtle shift in thinking, but the idea of having a 10 or 20X larger output context is quite fascinating. And I can tell you more about that as we continue the conversation, Mallory, but I think I've been a direct beneficiary of that architectural shift along with a bunch of people on this team over the last several days. And it's quite exciting because that continuation of that reasoning and that to generate hundreds of thousands of tokens really does matter with very complex tasks. So in my case, to give a little preview of what I'll share more later in the conversation, I split my time roughly into half of my time I spend with association executives and boards, helping them think through the future of technology, the future of business in general, spending time speaking, things like that, writing a lot, that kind of stuff. And the other half of my time I spend with our Blue Cypress Labs team. And Labs for us is where we try to invent the future and try to obviously predict where things will help associations. So we're always inventing new things. The Labs team is the principal author behind the Member Junction open code data platform for associations and a whole bunch of apps. So there are probably 20 different repositories that are part of Member Junction's overall corpus now. And a repository for those unfamiliar is simply kind of like a library of code. We use a tool called GitHub, which is the most commonly used one. And so these 20 different repositories represent well over 40 million lines of code collectively. It's a lot of code. And so it's not only a lot to take in, but also to build meaningful output, you have to be able to continue and kind of continue that reasoning train of thought. So having the model have that step up in capability, the trailer to the movie overall is I have not seen Argon make a mistake yet. In the last five days. And that is not a thing I've seen with any other model. Shockingly different than every other experience I've had with AI. And we're dealing with a very complex, very sophisticated enterprise scale code base. So that's my most interesting outcome is no mistakes in like six days.

[00:12:56:05 - 00:13:05:01]
Mallory
 Wow, no mistakes is pretty darn impressive of me. Do you see there being a use case outside of code for something like this?

[00:13:05:01 - 00:14:04:07]
Amith
 Yeah, for sure. If you think about the knowledge bases of associations and the incredible depth and breadth of these knowledge bases, depending on the field, having Argon level reasoning woven into a knowledge agent could be very interesting. And now you have to think about where you use these tools versus, you know, using them everywhere. So in a product like Betty, which is a knowledge agent that we publish, Betty has a layered architecture where pieces of the architecture may use a frontier model like an Argon, but pieces might use something like a Gemini flash, which is smart enough for some of the basic functions. But when you use Argon level intelligence to kind of reason over this breadth of content, some new insights pop up. And so it's quite fascinating. I think associations will stand to benefit from this. I think more advanced analytics could be directly benefited from Argon. So it's not just coding. That's the world I spent a lot of time in. But I've also seen it come up with some really interesting output just in this conversation.

[00:14:04:07 - 00:15:48:15]
Mallory
 Okay, I definitely want to get more into that. But I want to shift first to the demos. Google shared a video on YouTube that was really helpful and exciting with 10 demos that we will link in the show notes. Most of these run through anti-gravity, which we haven't covered in a minute. But this is Google's agent first development environment launched late last year and expanded in May. Agent first, meaning the agent plans, writes, runs the thing and tests it rather than auto completing while you type. It also supports Claude and OpenAI's open models, not just Gemini, which goes straight back to the model monogamy conversation we had last week. If you haven't checked that out, I want to talk about a few of these demos and then we will discuss them. The first one being 1960s Fortran rewritten into modern C++ and Rust. So stay with me here because I didn't know this. A surprising amount of critical software still runs on Fortran written decades ago. This is weather models, physics research, scientific computing. Rewriting it by hand is an enormous task. The documentation is usually gone, modern tooling barely supports the language, and a change that looks trivial can break behavior someone depends on. Working in anti-gravity, Argon took an entire vintage code base, mapped how the pieces fit together and rewrote it into modern C++ and Rust. Not a line by line syntax conversion, it restructured the architecture along the way and the Rust version came out memory safe by default and 20% faster. I mean, that was the first demo they shared in the video and already I said, "Ooh, that is quite impressive." Can you extrapolate out of this and say legacy migration is tractable for associations with a model like this?

[00:15:48:15 - 00:17:07:03]
Amith
 It is and there's another element that's important to unpack here. One association that has a legacy system, perhaps written by now long departed internal developers or consultants. There's a lot of associations who still have these systems, in some cases for their core AMS or other systems, in some cases for websites that have been built for them or other things. Going back decades in time, in some cases, people still have websites that were built in a technology called ColdFusion, which goes back to, I think the late 1990s when that technology was first used, so 30-year-old technology. It's very hard to find people who have any idea about that type of code. In addition to that, some associations have software they're using in production that comes from technology vendors that have long since gone out of business. Old systems, it could be an AMS, but it might be like an event software, it could be something else. Now, software you don't have the source code for all of a sudden with AI is something you can actually make sense of. You've been able to do what's called decompiling code for a long time, where you can take the code from the machine readable format and turn it into source code. Source code is what humans can read. I'm not all humans, but programmers, right? A category, a subcategory of humans.

[00:17:09:23 - 00:17:09:24]
Amith
 The

[00:17:12:04 - 00:18:23:06]
Amith
 code that a decompiler would create is not really human readable in the same sense code that a professional developer would write with comments and descriptions and documentation. It would just be literally the raw code. There's a lot to be desired there. What AI can do is actually turn that decompiled code into really meaningful, useful, documented code almost instantly, and then can help you convert it. If you're stuck on a legacy system of some sort, be it an old website, or if you have some piece of software that's still in your value chain that's really critical, but you can't use it as a way to maintain it. AI may be able to help. AT gravity with Gemini 4 Argon might be able to help you decompile a legacy tool and turn it into something that you can port your data out of a cryptic, unknown format into a modern world. I do keep in mind, I'm not giving you legal advice on whether or not you're allowed to do these things. You have to think through that on your own and gain legal counsel if you're dealing with a licensed piece of software, but it is technically feasible to do these things, which is really important. That's something that even the prior models, including models that are still frontier from other labs, aren't quite as good at.

[00:18:24:15 - 00:18:39:12]
Mallory
 Okay, assuming it's legal, would you say that it is safe and that would you recommend for an association to hand over migration project data to a model like this? I guess in that sense, you can talk about the work you were doing with member junction and how you do it in a safe way.

[00:18:39:12 - 00:20:10:06]
Amith
 Yeah, I think, first of all, the safety question is always relative. It depends on what you rely on for this piece of software, what part of your business it's dealing with, closer it is to payments, the closer it is to PII or personal information, the more you have to be very thoughtful about it. But associations for a long time were afraid of having any software that they had to maintain themselves because it's expensive, it was hard to get talent. All these things require a rethink. I'm not suggesting that associations should be in the business of custom software development, but I do think it's possible that associations can be in the business of maintaining some custom code now in a very reasonable way because the AI will do it for you. You have to have certain guardrails and you have to have, I think, a setup around it that keeps you safe. And part of what you just mentioned with MJ or member junction, tools like that can help you establish an ecosystem or a framework that will help you stay within rails, which is very helpful too. The point here though is I really want people to open up their minds to the possibility of breaking free of shackles that may have once grounded them in the "we can't do this" mindset. And now we can do it. Now should we do it is a different question and that might require some professional advice to guide you in the right direction because there's always a "it depends" answer to questions like this. But it's now possible to do basically for free and very, very quickly with very high quality. So that to me is extremely exciting.

[00:20:11:16 - 00:20:31:10]
Mallory
 From a non-technical perspective, maybe someone listening has the same question too. When you're talking about large code repositories and the work that you were doing, Amith, with Argonne 4, how are you able to check the output? You said there were no mistakes. How can you be sure that what we get out of it is accurate and it didn't just create a whole bunch of mess?

[00:20:31:10 - 00:21:54:15]
Amith
 Totally. And that's an awesome question. This goes back to an article that I wrote on LinkedIn a little while ago about the need for frameworks and being able to observe and verify output. So Jensen Wong said this recently too, that the outputs are essentially free, verification isn't, knowing that it actually is what you want and it's correct is hard. For us, we have a bit of an advantage because we have a framework. We have an extremely strong set of test suites. Test suites are essentially pieces of software that test whether or not another piece of software is doing what it's supposed to be doing. And so we can do things like run entire multi-repo code optimizations, for example, to make software faster or more resource-sufficient or whatever the goal may be, and then run it through a suite of tests that tell us definitively and deterministically whether the software is still correct. Now, that's not perfect. It doesn't catch everything necessarily, but it's a very helpful tool. But this is also where human programmers that have really strong expertise are very helpful being able to look at the code and understand it. Now, you can also say, hey, you know, we're going to have Argonne do the work and we're going to have Opus 5.5 or GPT 6 review it. That's something we do here as well because model diversity gives you different thought processes just like dealing with multiple humans would do the same.

[00:21:54:15 - 00:23:28:17]
Mallory
 That makes sense. I want to move to Demo 2, which is turning a clinical scan into a digital twin of a body. Quick definition, a digital twin is a working model of a real thing built from its actual data that you can inspect and run scenarios against. Manufacturing has used them for years, a virtual copy of a jet engine or a factory line that you can test against instead of a real one. A DEXA scan measures body composition, things like bone density and where fat and muscle sit. The report you get back is numbers, percentages, and flat radiographic images. Argonne took these 2D images and the clinical measurements and built an interactive four-layer 3D model of the body, which is kind of like a digital twin of a person. You can strip the skin layer to compare muscle symmetry side to side. You can cut through the torso to look at internal regions and examine visceral fat and bone density directly. It pulled in 9 years of scans so you can move through time and watch what changed. There's also a simulator where you adjust future fitness goals and the anatomy reshapes to show that projection. Google is clear this is a visualization and not a diagnostic tool, but the pattern matters here. A decade of structured data, no one explored because it lives in reports, turned into a model you can move through and project forward. Amith, we've talked about digital twins on the podcast before. These have been a manufacturing and engineering idea for years. Do you think Argonne is a step in the right direction for digital twins becoming possible for associations?

[00:23:28:17 - 00:25:17:20]
Amith
 Yeah, it definitely is because the level of intelligence we're seeing here, even this is as sophisticated as what you're describing is, it's a very basic model, right, in terms of what it's actually trying to model. I think that what happens with associations and all businesses over time is there absolutely is a digital twin of the entire business. The way I see that happening is you have to start off with a digital twin of each constituent. That might be a member, it could be a board member, it could be a volunteer, someone in a committee, and you have to essentially have an AI that emulates what we think their predicted behavior might be on a go-forward basis. That's based on all sorts of signals, including everything you know about every individual, but also public information and other things that might be unique to your association, like your email history with that individual or conversation history. Then if you aggregate these tens or hundreds of thousands of individual digital twins of members or other constituents, then you can start to simulate the behavior of the system as a whole. What happens when this particular member chooses not to renew? Does that affect other members? How do these different... Think of it as like the world of The Sims or that kind of a video game where you have these simulated behaviors, but like a ultimately extremely realistic version of that. We're very close to being able to do those things. The biggest issue there actually isn't even so much frontier intelligence, but it's making the intelligence less and less expensive and more abundantly available, which we're obviously on the curve to do that. Because you're going to want really not quite an AI model, but you're going to want to be running an AI model continuously for every member. Then that model is going to continually be predicting what that individual is going to be doing, but it continually corrects and updates based on the history of data, which is a continuous feed. We're actually dabbling in this right now, Mallory. I don't know if I've told you about the...

[00:25:17:20 - 00:25:18:17]
Mallory
 No, I don't think so.

[00:25:18:17 - 00:25:20:24]
Amith
 Have I told you about the More Cheese Project?

[00:25:20:24 - 00:25:24:01]
Mallory
 No. Tell me, tell me.

[00:25:24:01 - 00:25:34:10]
Amith
 This is a fun one. Okay. Here in New Orleans, just down the street from the Blue Cypress Worldwide headquarters, we have a really great restaurant. I'm sure you've been there, Mallory. Hey, James. The St. James Cheese Company.

[00:25:34:10 - 00:25:36:01]
Mallory
 I knew you were going to say it. It's fantastic.

[00:25:36:01 - 00:27:28:01]
Amith
 Yes. We love this place. A lot of people here walk down there regularly. If you're ever in New Orleans and you want to have a Blue Cypress team member sighting, go to St. James Cheese Company uptown. They have two locations. They have one downtown. It's a little plug for them. We love their food. Great place. But one day, when several of us were having lunch out there, we were trying to come up with the name for a fictitious association that we could model in order to create a demo data set. We came up with the International Cheese Federation. We had to do some searches because there are actual cheese associations in the United States and abroad. There's nothing called ICF or the International Cheese Federation. Then we started, we said, "Well, why don't we get a website and actually make a complete association? Let's not just do demo data. Let's build a website, the blog. Let's create personas that represent the staff and the board and every other member of the association over time." If you go to morecheese.org, which is the website we have, so we actually call it the More Cheese Project. If you go to morecheese.org, you'll see our sample site and it's continually updating. The idea behind this is we use a whole bunch of different sophisticated AI to look at world events. Look at macroeconomic data. Look at what's happening. Every week, we generate blog posts and things that are happening in the world of the More Cheese simulation. We also generate structured data. That's like what's the activity that's happening. We have tools that are all part of the open code base in member junction. There's actually a data tool called Loom, which auto generates data in alignment with the AI's research. This is not down at the level of resolution I just described where every member is being simulated, but it's leaning in that direction. The thought process we have is, first of all, this is a great demo, but then it can lead directly to the idea of building technology that would enable an association, or anyone for that matter, to build a digital twin of their entire organization.

[00:27:29:01 - 00:28:03:02]
Mallory
 Wow. I'm on morecheese.org. Everybody if you're listening, just pop out Safari, Google Chrome, and go take a look. This is a pretty good website. Amit and wait, my favorite. Member organizations, Mongers Row, Nornbrook Dairy Systems. We've got Betty at the top. Oh, and also I love the little nod. I'm assuming that came from you. Where did I find it? Aisha Bell is the member spotlight, and Aisha was the name of the fictional character in the first Ascend book. This is like putting everything together. How long did this take, Amit?

[00:28:03:02 - 00:28:04:20]
Amith
 About three weeks.

[00:28:04:20 - 00:28:05:08]
Mallory
 Okay.

[00:28:05:08 - 00:28:12:12]
Amith
 I was going to say this is- Three weeks for one of our technology fellows to put it together. There's a lot of people who chipped in over time that the actual project was three-ish weeks.

[00:28:12:12 - 00:28:19:01]
Mallory
 Wow. That's pretty eye-opening, everybody. If you need a new website, just a couple weeks. I'm kidding. I know there's more to that.

[00:28:20:08 - 00:29:54:05]
Mallory
 Wow. Morecheese.org. All right. I want to move to the third demo of the day, which is a sketch turned into 3D printing. It's a very complicated hardware. Turning a rough drawing into a website is one thing. Turning it into a physical device is a different problem. Agent Gadget starts with a basic sketch. Argonne takes the mechanical engineering role. It searches for the exact 3D models of the internal electronic components the design calls for, works out across repeated passes how they fit together, and designs a custom 3D printable casing around them. Sketch to print-ready file to a gadget you can actually hold in your hands. The constraint is what makes it impressive, a nice-looking enclosure, easy enough. One that accommodates real components in a real arrangement is a genuine engineering problem and the model solved it from a drawing. There's also one more example of the model reaching into the physical world, and this one goes further than a printed casing. Argonne watched a video, a raw video, of a person playing cornhole, mapped the spatial coordinates, worked out the joint movements required, and wrote the simulation code. Put into Aptronics Apollo humanoid, the robot sank seven shots in a row. Google calls this video to simulation to real and says it compresses weeks of simulation modeling into hours. Amith, I feel like both of these examples are about understanding physical space and how parts fit together, how a body moves even. I feel like that's a lot different than reasoning about text. What do you think?

[00:29:54:05 - 00:30:08:17]
Amith
 Yeah, I agree. And Google has a pretty interesting advantage here when it comes to multimodality. That's been a strength of theirs. I didn't mention earlier when you asked me about how I think about Google, but Google's been multimodal since, I think, it might have been Gemini 1.5.

[00:30:09:17 - 00:30:25:14]
Amith
 But Google has such an incredible training corpus of video from YouTube, but also Waymo, which is real world video of all the Waymo vehicles racking up millions upon millions of miles, possibly also the videos inside the Waymo. I'm not sure if they use that for training.

[00:30:25:14 - 00:30:26:18]
Mallory
 I hope not.

[00:30:28:10 - 00:32:13:17]
Amith
 But the videos of the world outside the Waymo vehicle, I think, would be quite instructive for a lot of world modeling applications. So it doesn't surprise me at all that it's all of a sudden emergently good at these kinds of things. Google also has really incredible tools for video understanding and image understanding. So in our platform in Member Junction, we have embraced Google's embeddings model, which is an embeddings model for those that aren't familiar, essentially is an AI model that converts some type of human usable content, whether it's text or PDFs or video or images into numbers. It converts into what are called vectors. And then those vectors are comparable through very simple math. And the usefulness of that is hard to really overstate, but it's essentially the backbone of modern language models and language vision models and allows you to do contextual semantic search, many other things. And so by utilizing the new Gemini embedding model, which is multimodal, you can compare, let's say, the word cat to a picture of a cat, which sounds really trivial. But being able to cross modalities like that is part of what makes their models so incredibly smart. And I actually think that has relevance in all sorts of fields. One comment about association thinking around this whenever we talk about robotics or the physical world or science, association listeners sometimes may tune out a little bit thinking, hey, you know, we're office workers working with bits and bytes and data, but your members very likely are in the physical world in some way. And so this technology could be extremely valuable to make sure your members are aware of because, you know, your world is your members world by extension. So it's important to stay aware of this, even if you don't think this is directly applicable to you.

[00:32:13:17 - 00:32:33:06]
Mallory
 As we wrap up the demo section, I know you mentioned briefly at the top of the episode the work that you did with Argon 4, but you didn't get into detail. Is there anything you want to share with listeners that you specifically did? You said you worked with it for six days or so. What kind of is the most impressive standout activity that you were able to do?

[00:32:33:06 - 00:34:46:16]
Amith
 So one of the most challenging things to do in software architecture when you're at scale with a very large system is optimization. So optimization is a goal-oriented process where you're saying, I want to make the software a certain amount faster or I want to make it utilize fewer resources or take up, you know, less space, things like that. And so we were looking at the overall MJ architecture for the core framework, as well as a number of the apps that we make available. And we said, how can we improve the speed and lower the cost of execution simultaneously? One of the reasons this is so challenging is you don't want any regression and capabilities. So you can't allow, hey, we made it really fast, but now it only works three quarters of the time. It has to just work just as well as before. It has to do just as many things as before, but you're finding a way to kind of squeeze something out of it. It's really hard. And so I know plenty of software engineers that have been doing this for decades that really haven't cracked the code themselves and being terribly good anyway at performance optimization or resource optimization. It's a difficult subspecialty. It's also really difficult to do it across complex multi-repository or multi-project environments. So what I did is I said, hey, this is the first thing I threw at Gemini for Argonne. I said, hey, here are 20 different repositories that are all interconnected. They make up kind of the overall MJ corpus of tens of millions of lines of code. I want you to find opportunities to make the core system at least 50% faster in these five key areas, as well as evaluate the architecture and tell me where else we could be more efficient. Now, keep in mind, this is a product and technology that's been out there for many years. It's used by a lot of people. It's heavily exercised pieces of software that have a lot of really capable people behind it. So it's not like it was designed in some kind of haphazard, cobbled way. People building this framework, myself included, have been building advanced enterprise software architectures for decades. So that's the backdrop. Gemini basically told us we're all pretty stupid very quickly. So it highlighted a whole bunch of inefficiencies that just slipped under the radar. Some of it's the accumulation of things over time. Some of it's just missed opportunities that nobody thought of.

[00:34:48:05 - 00:36:41:18]
Amith
 And so it was extremely enlightening within a few hours. It ran for a few hours to come up with this analysis. So it's, I don't know how many millions of tokens or hundreds of millions of tokens it used, but it basically consumed the entirety of this massive code base. And it came back with a bunch of recommendations. I read the entire report. It was, I don't know, 20-ish pages of content that it generated that was super dense technical material and went through and gave it some notes. I gave it some opportunities for improvement on top of what it did. And then I had it executed. I had it, I said, "Hey, go execute this and I want you to do deep, you know, really deep benchmarking of everything across the board." And it ran for, I think about a day. And it came back and it had some areas that I wanted it to further improve, but nothing that was incorrect was really the amazing part. It ran through our complete set of benchmarks internally, our full suite of tests. And it was, the outcome of that day of work was flawless. And the headline is a number of capabilities in our core software are between 2 and 15x faster in certain key paths. And because MJ is a framework where everything kind of is leveraged on top of a core set of primitives, that has ramifications for everyone. So in member junction 6.2, which will be coming out for long-term support release in late November, all these capabilities produced by Gemini will be included in there. So it's pretty exciting. Wow, I was going to ask you to-- There's no way that Opus 5.5 or Fable 5.1 could do that. I've tried to do this project with every model release in the last 12 months or so. I didn't even bother before that. And none of them were able to get anywhere close to this. They had suggestions. They were like, "Oh yeah, you could do this or that." But a lot of it was stuff that we had done already. But Gemini 4 actually looked at what's called Git history. Git is the version control system. And it looked at all the things we had done over time and all the notes and everything. It was just pretty stunning.

[00:36:43:03 - 00:36:59:12]
Mallory
 Wow. OK, so many things to unpack there. I feel like this is a really good example of testing new models against your own benchmarks. Because as we're about to talk about, there's a little bit of kind of mixed review of Argon 4. But it's crazy to hear that for you. You've done this task with other frontier models. And this is the one that could get it done.

[00:36:59:12 - 00:37:20:21]
Amith
 Yeah, it's a tough task. I mean, it's kind of an unreasonable request. Right. When you think about this, this is something that you'd put incredibly capable top tier software engineers. And you'd give them probably at least weeks, most likely months, for a team of several really good software architects and software engineers to work on this. And they probably wouldn't come anywhere close to this.

[00:37:20:21 - 00:37:44:21]
Mallory
 OK, and then one more follow up question is, as you were talking, I was thinking, "Oh gosh, using Argon 4 for member junction." But then I thought, "Well, member junction is open source." So the code is available to anybody anyway. There's not really anything inherently secret that you were sharing with the model. Would you recommend that an association that has custom software do something similar to what you did? Or what do they need to think about?

[00:37:44:21 - 00:39:27:21]
Amith
 Yeah, totally. I mean, I think it's a great project to give it a shot, certainly with custom software. I think also, one of the things we're exploring right now, we haven't run this end to end yet, is using other tools powered by Argon that are specialists. For example, one of the other free things that we have inside the MJ project is this thing called AutoDoc or DB AutoDoc, which I think we may have talked about on the pod maybe six months ago or so. But this is a tool, if any of you listening have databases that are not properly documented or perhaps not documented at all, and you think your data environment is kind of messy, one of your big problems probably is you don't really know the structure, you don't really understand the database that well. So DB AutoDoc is a tool that we put out there. It doesn't require a member junction, by the way, it's just under the project kind of logo. But you can run this tool against any database of any kind and it will give you extremely good documentation within hours. This is again, something that would take months of time and tens of thousands of dollars, if not more, with a hired database analyst. With Gemini 4 Argon, our thinking is we'll get pretty close to 100% accuracy. Right now with prior models like Gemini 3.8, we got to about 96%, 97% accuracy. You might ask, "Well, what's the difference between an A and an A++?" The difference is making mistakes. So if we can get the database documentation truly to 100%, that's going to make everything downstream of that way smarter. So that's pretty exciting as well. And that's an example that would benefit every association person listening, is you probably have a data source or multiple data sources that are kind of unruly and you don't understand them. And this is a tool you can run against any database quite easily.

[00:39:27:21 - 00:40:30:11]
Mallory
 We will link that episode to in the show notes for DB Autodoc. The last part of this episode, I want to talk about some of the reception around Argon 4. Google's own scorecard has Argon winning 13 of 19 tests. Artificial analysis, which tests independently, puts it at about 53 on their intelligence index, tied with GPT-6 Astra and Claude Fable 5.1, with Claude Opus 5.5, five points clear at 58. The complaint from hands-on users isn't about the model. It's actually about antigravity. A panel of heavy users at the AI publication, Every, was unanimous that the harness undermines it. Dan Schipper, who runs Every, called Argon the best structural editor he's used and also called antigravity a terrible harness. Another general complaint, nobody can tell you which Google surface to use for what, the Gemini app, antigravity docs, workspace, etc. So, Amith, can you settle this for us? I know you're deep in this space. Obviously, you love Claude code. What do you think about their harness?

[00:40:30:11 - 00:41:15:04]
Amith
 Yeah, I think, you know, as tools go, I would say Claude code is better as a harness compared to antigravity. That's probably the fairest statement I can make. I think antigravity is good, but not yet great, but it's still able to do the job. And so, it lacks certain features that Claude code has, but I would suggest that people try it out for themselves. There is a little bit of a taste question here too, in terms of what do you prefer. Antigravity does some things very nicely that Claude code doesn't do, that some people may find preferable. It's more visual, for example. It has a very nice desktop application. Claude code is great in the terminal. It's Visual Studio plugin, things like that are okay. I mean, I think overall Claude code is still definitely the stronger overall product.

[00:41:16:13 - 00:42:30:00]
Amith
 But antigravity is going to gain ground quickly. So, the other thing to remember is Gemini 4 Argon, you could use it in other harnesses as well. So, you could take the model. The advantage to using antigravity is you can get a really inexpensive plan from Google. So, you can go sign up for a variety of different Google plans. I think it's called Pro and Ultra. I already was an Ultra subscriber with Google anyway, because I use a lot of their cloud storage. And so, I ended up getting the Ultra subscription. So, actually for the last several months, I've been using a lot of Google models because they're effectively free for me on an incremental basis. And for that reason, I've been trying antigravity a lot. It's improved too. They added remote control, for example. That was one of the critiques, is you couldn't control your desktop session from your phone or somewhere else. And Claude code has had that for a while. So, I suspect that as a practical matter for most of our listeners, first of all, it's plenty good. For what you're going to do. Second, you might like it even more, especially if you're not a developer. And third, whatever features Claude code has that antigravity doesn't, antigravity is going to have them in minutes or days or weeks and vice versa. So, they're all stealing from each other, which is kind of a healthy level of evolution for all these tools. So, I think all these harnesses really will converge into one capability set at some point.

[00:42:31:10 - 00:42:41:03]
Mallory
 So, do you think Argon 4 is kind of interchangeable with the other frontier models for the sake of our listeners? Or do you think there's really something special about it that our more technical listeners should try out?

[00:42:41:03 - 00:44:16:20]
Amith
 At a minimum, you should, if you're a little bit more on the technical side, try some of the stuff we've talked about in this episode. If you're not technical, I would go to the Gemini.Google.com site and try it for some of your workloads. Give it a shot. See what Argon can do for you directly. This is the thing I really suggest most people consider is there's lots of people, far more of them out there than probably we need, but there's lots of people out there sharing their opinion. Everyone's got opinions about stuff. Some are very well informed, others perhaps less so. But the point about opinions is they don't really matter if they don't affect you. And so, if you test some of your workloads, you should very quickly determine whether or not Gemini 4 Argon is a useful thing for your organization. And that goes back to your recommendation, Mallory, that you've made, I've made a number of times in the pod, which is have a library of several sample prompts and sample projects that you've run and you want to run in the future, in particular things that haven't worked. So, going back to the days of early chat GPT, we always would say, "Hey, listen. Hang on to those prompts that you couldn't make useful back in GPT 4. Try them again the next time the model comes out." This is a really important concept with AI that's got nothing to do with Argon, but everything to do with the pace of change. That which doesn't work today very likely will work in 6 to 12 months, possibly as soon as tomorrow. So, give it a shot and have a set of tasks that you think are your most amazing possible use cases of AI that you don't think are possible today. And you might be right. Some of them may not be possible yet, but keep trying them.

[00:44:16:20 - 00:44:24:07]
Mallory
 Well, everybody, we covered a lot in this episode. We know Argon rewrote 1960s Fortran into modern Rust and C++,

[00:44:25:08 - 00:44:40:05]
Mallory
 turned a decade of scans into a body you can explore, taught a robot to play cornhole from a video that is quite a frontier model release. And as Amit said, give it a try with your tasks. Everybody, thank you for listening and we will see you all next week.

[00:44:42:08 - 00:44:59:11]
Amith
 Thanks for tuning into the Sidecar Sync Podcast. If you want to dive deeper into anything mentioned in this episode, please check out the links in our show notes. And if you're looking for more in-depth AI education for you, your entire team, or your members, head to sidecar.ai.