Welcome to the podcast. I'm your host, Jayden Schaefer. Today, we have an absolutely packed lineup for the show. The biggest thing that I've been super excited about yesterday, Anthropic unveiled their hosted AI agents platform.
It's their cloud platform. I've been playing around with it a ton. There's so much there, and I think this is a really big shift. Basically, I've ended a lot of what OpenClaw did, but there's a nuance.
We're going to get into that. In addition, Meta just dropped their very first model that was built with Alexander Wang. Remember, that's formerly the CEO of Scale AI, but they kind of acquired him in. We also have a research team at Tufts that figured out how to cut AI energy consumption by a factor of hundreds, which is definitely a big deal if you think about how much power these data centers are burning through.
Eli Lilly just flipped the switch on what they're calling the most powerful AI human supercomputer in pharma, over a thousand blackwell GPUs, which are aimed at cutting drug development timelines in half. OpenAI published a set of policy proposals that include robot taxes and a four-day workweek, which is probably the most OpenAI thing I've read in a while, and Google released Gemma 4, which is their latest open source model that's getting a lot of attention for what it can do relative to its size. I mean, basically, this is an edge model that you can put on devices. So, a lot to cover in the show today.
Before we get into that, I want to mention AI Box, which is a tool I use every single day at this point. If you haven't checked it out, it gives you access to over 80 AI models in one place. This is my own startup, so instead of paying for separate subscriptions to Claude, ChaiGPT, Gemini, and everything else for, you know, 11 labs for audio or tons of image models, you get all of that on one platform for $8.99 a month, and we have annual plans to give you 20% off of that. It honestly pays for itself pretty fast when you're not stacking three or four different AI subscriptions.
The links in the description you want to try it out, you get access to every single top AI model in the world, basically all the best ones in one place, so you're not juggling tabs and juggling subscriptions. All right, let's get into the first story, which is Google Gemini. I want to talk about their new open source situation with Gemini 4. So, earlier this week, they released Apache 2.0 license, and basically, this is their latest family of open models built specifically for reasoning and agentic workflows.
What I think is really interesting about Gemini 4 is what Google's calling, you know, the best intelligence per parameter ratio in any open model right now. Basically, you're getting the frontier level capabilities of what you'd expect out of something like Claude, ChaiGPT without needing a massive hardware setup. You know, something like Llama 4 Maverick requires that huge hardware setup, and so you're basically getting around that. The model already has over 400 million downloads, and the community has spun up over 100,000 variants, which I think just kind of tells you how quickly developers are adopting this.
I think the significance is that it's less about kind of the benchmarks, and it's more about the trend, right? The gap between open source and closed source models is definitely shrinking, and I think that Gemini 4 is just another data point in that direction. If we want to get into kind of the licensing on this, the Apache 2.0 license is also really important because it means that companies can actually use this commercially without worrying about any sort of restrictive terms. I remember when Llama first came out from Meta, and they were like, look, it's like an open source model, and it's like, well, it's not really open source, it's just like, you know, open weight, and like, you can use it, but if you really want to use it for something commercial, you've got to let us know, and there was like all this kind of, I don't know, it was very unclear, and I think Meta is just getting right around this, right?
Anyone that's building agents are doing, you know, reasoning heavy work on their infrastructure, this is probably the most capable option that you can run locally right now, and I think given everything that we're going to talk about with Meta in a minute, the open source AI community really needs models like Gemini 4 to keep delivering. All right, let's talk about open AI. They just published a set of policy proposals this week, which they are outlining how they think wealth and work should be restructured in what they're calling the quote-unquote intelligence age. Now, maybe this is a big marketing thing.
It feels like, you know, open AI and Anthropik, they have like these big, huge, I don't know, visions for the future, and how AI is going to take over the world, and sometimes it feels like that's a bit of a marketing stunt. I think what's interesting, open AI is essentially saying, you know, AI is going to displace a lot of jobs, and here's how we think society should handle it, and they're kind of combining like traditional left-leaning ideas like wealth redistribution with a very kind of market-driven capitalist framework. It's pretty deliberate. I think they're kind of attempting to position themselves as a responsible player in all of this.
Now, whether any of this is actually going to have any sort of influence on policy, I think that's definitely a different question entirely. Tech companies publishing policy papers doesn't exactly have a strong track record for changing legislation, but do you think it matters that the company is building some of the most capable AI systems, and they're at least putting kind of a stake in the ground on the economic consequences? You know, there's a bunch of funny things in there, right? They're like, we should have like a four-day workweek, and it should be powered by AI productivity gains.
That's not the worst pitch in the world that I've ever heard. At the end of the day, though, I just don't know if it's going to get there, and personally, if I'm being 100% honest with my own experience, and I think a lot of people like myself, instead of doing four-day workweeks, I'm now doing, you know, six-day workweeks and 16 hours a day on cloud code and cloud code work, because I can get so much done, but maybe at some point the hype will die down, and I'll return to a more reasonable cadence. I hope. Who knows, right?
All right, Eli Lilly has inaugurated what they're calling LilyPod this week. It basically is a beast. It's about 1,000 NVIDIA Blackwell Ultra GPUs, and they're delivering over 9,000 petaflops of AI performance. So it's basically the world's first NVIDIA DGX SuperPod with DGX B300 systems, and it's the most powerful AI factory wholly owned by a pharmaceutical company, right?
So, of course, there's other AI companies that have more powerful systems right now, but for pharmaceutical companies, this is the most powerful one. I think the numbers that really were kind of shocking to me is that historically, even productive drug research teams can analyze about 2,000 molecular ideas per target year, and because every experiment requires physical synthesis and lab testing, Eli LillyPod removes all of that bottleneck, and they're creating what's essentially a computational dry lab, so they're doing this at a huge scale. Scientists can simulate and evaluate billions of molecular hypotheses in parallel before committing to physical experiments, right? So the AI is going to sort of simulate what happens when you merge these molecules together, and then it's like, look, these are the ones that look promising to cut the traditional 10-year drug development timeline about in half.
That's what Eli Lilly is trying to do right now. NVIDIA is kind of working on this. They've also invested up to a billion dollars over five years in an AI co-innovation lab in the Bay Area. To me, this is just a really big example of AI delivering really, like, legitimate, tangible value, you know, not just in the tech industry.
I think we talk a lot about, like, chatbots or coding assistants, but using AI to find new medicines faster, that is an application that I think could actually change millions of lives. If EliPod delivers even half of what Lilly is promising in terms of kind of accelerating drug development, then I think a lot of the downstream impacts on patients is huge, and of course, the profits for Eli Lilly are going to be massive, so they're going to be thrilled. And I'm just going to, I don't know, I don't like to be too pessimistic, because I'm really stoked about AI and healthcare, and I think there's a lot of benefit. Somehow, I have no hope in pharmaceutical companies, because I feel like there's a lot of solutions that they don't talk about if it doesn't make them more money.
So, anyways, they're going to probably discover some awesome drugs that solve a lot of things, and they're going to make a lot of money from it. So, pessimistically, I don't know. It's, I guess, good, and it, whatever. It is what it is.
Okay, let's talk about neuro-symbolic AI cutting energy use by 100x. I think this is from a bunch of researchers at Tufts University. It's one of the biggest stories in this list, I think, if you kind of look at this from a long-term perspective. The team, which is led by Matthias Schultz at the School of Engineering, they essentially built this neuro-symbolic AI system, and it achieves a 95% success rate on structured manipulation tasks while using just 1% of the training energy that standard vision-language action models require.
Okay, I know that sounds like a big mouthful. Let me explain it. In the most simple terms, 100x less energy and nearly triple the accuracy is what they've been able to achieve. So, when they're training these systems, when they're using these AI, they're able to get outputs for 100x less energy, and the accuracy is tripled.
Basically, the approach they took is merging some traditional neural networks with symbolic rule-based reasoning. Instead of throwing these massive compute pattern-matching, the system is kind of breaking the problem into a bunch of smaller logical steps and categories, and then they're going to—I mean, what's interesting, though, is this is basically a lot closer to how humans actually think through a problem, right? When you have a big problem, you think it'll get one of the steps to achieve that. You kind of break it down in your head, and that's literally where the term break it down came from.
And they're just teaching the AI model to do this in a more human way, and surprise, surprise, it saves energy. Obviously, our brains are designed to not burn too much energy when we're trying to figure stuff out. I think this matters because US data centers and also AI workloads now consume more than 10% of the entire country's total electrical output. Like, this is so much electricity, and I think that number is projected to double by 2030.
So if neuro-symbolic approaches can deliver this kind of efficiency gain across more domains, I think it's going to make a really big impact. Not just, you know, some people are like, oh, it's awesome for the environment. Yeah, it's also awesome for economics, right? If you can do this for 95 or 100 times, you know, more efficient, that saves you a ton of money for these companies.
Or, put another way, it makes this AI way cheaper for the user to use, which I think is really awesome. Right now, it's still a proof of concept, so we're not going to see this into production models today, but the direction, I think, of kind of where they're going with this is promising. I think kind of the broader industry is going to pay close attention to this, because if they're able to achieve this, like I mentioned, for, you know, the bottom line on all of these companies, it's going to be amazing. It still feels like they're falling behind, but they have this huge reorganization.
They brought him on as the CEO in June last year. They spent $14 billion to acquire scale AI and him, about 49%, a non-voting stake in the company, so essentially they kind of acquired it. But in any case, in terms of capability, Meta says that MuSpark is competitive with the leading models from OpenAnthropic and Google across a bunch of different things that rank fourth on the artificial analysis intelligence index with a score of 52. It's really good at figuring out an understanding and medical reasoning, but it does not surpass basically all the top models across the board.
So it's a good showing, right? Meta's like sort of in the races still, but I would say they're not really, because, you know, when they're like, hey, we came up with a new model and it's number four out of, like, basically what's happening is every three months when the top lab comes out with their newest model and it beats everyone in the benchmark so they can take a victory lap and say, I'm the best, right? And we saw this, it's not just, you know, Anthropic and OpenAI and Google, like Grok is also in the mix too, so somewhere between those four models, they're constantly kind of beating each other. And so if Meta comes out and is like, hey, look, guys, we're number four, like what, maybe slightly better than Grok and behind OpenAI, Anthropic and Meta, it's just like, yes, they can get integrated into all of the Meta kind of, you know, WhatsApp and Instagram and Facebook, they can get some users from that, but it's not like anyone's go-to model.
So, yeah, I mean, I think they're pushing forward, but it's not that impressive. What I do think is really interesting, though, is MuSpark is a closed model, so this is a huge pivot for Meta's kind of Lama strategy, which they were really aggressively open source for years, Lama was one of the most important things to happen to the open source AI community, and I think now Meta is going in the opposite direction, so the model's design code isn't going to be made public. I think it tells us a couple things. Number one, Meta clearly believes that they need a competitive closed model to keep up with OpenAI, kind of at the frontier, because in the past when they were doing all the open source, they were like, look, our model's not quite as good, but it's open source, and, well, sort of open, it's like open weights or whatever they were calling it, and they're like, yeah, you can like run it on your own computer and stuff, and that was cool, that did have a big draw a lot of people that didn't want to have to pay to run these models, were grabbing it or hosting it, and so that was cool.
Now that they're going closed source, I think they got to be a lot more, I mean, their model has to be a lot better. The open source strategy was really good for adoption and kind of developer goodwill, but it was not winning the race as far as the best company goes. Now there are other people working on open source, and I hope that open source models continue to get pushed. Not every task that you need an AI model to do needs like the absolute greatest, you know, frontier model.
You get there's plenty of tasks where you're sorting folders or files or searching for things or, you know, rewording something, and like, you don't need anything crazy, and so I think a lot of those open source models would be great, would save compute and energy and all that. And so I hope that this still gets pushed for it, and I also hope that we continue to make really powerful, really good open source models, but we'll see where this goes. It feels like Meta right now is kind of getting closed source, and it's also kind of a weird place in the industry, right? We have OpenAI who just came out today and was like, hey, we have a model also that's just as good as Anthropics, you know, Doom model that can, like, hack, you know, you can find vulnerabilities in every software platform, and so, you know, Anthropics is like, we can't give this to the general public, we're giving it to a handful of people, and so OpenAI is like, look, we have one of those too.
You also have to think, like, if Meta was like, yeah, we have one too, and it's open source, and anyone can use it, then everyone would be like, oh no, you know, China, Iran, Russia, they're going to take over the world with it. So I think at the end of the day, their argument is going to be that they've got to close sources because the models are getting so big, it's too dangerous to have, like, an open source version. People are already making that claim years ago, I remember Elon Musk in, like, what, 2014, 2024, something was like, you know, we've got to have, like, a six-month break on releasing any new models because they're too good with, like, ChatGPT4, and it's like, wow, how times have changed, those models seem so archaic now. Anyways, it's an interesting point, an interesting place, that's the argument we're going to make, but I'm sure in two years, we're going to look back at the models we had today and be like, oh man, those things are so bad, like, you can't just, like, talk to your phone and any software you could ever imagine gets built and you can go, like, use it instantly.
So I'm sure things will change a lot. Guys, thank you so much for tuning into the podcast today. If you enjoyed this episode, I would appreciate it to no end if you could leave a rating or review, any sort of comment you drop on Apple, really helps the show to get found by more incredible people like yourself. On Spotify, you have to hit the About tab, you can drop some stars, it helps out a lot.
And as always, make sure to try out AI Box, my own startup, there's a link in the description to get access to over 80 models and our automation builder, it's $8.99 a month. Anyways, I'll catch you guys all in the next episode.