Inside Google's AI Slump: Can Gemini 4 Argon Fix It?

Ejaz (0:00)

When it comes to AI, Google is a company that you would expect to see at the top. And yet, over the last six months, they've suffered a bit of a fall from grace. Their flagship model, Gemini, which peaked at the end of last year, has suffered delay after delay after delay, with companies like Anthropic, OpenAI, and even Meta running complete circles around them. I think 15 brand new AI models has dropped since Google's last one has. But the bad news doesn't stop there. Google has also recently lost key members of staff from their AI team, the most prominent one being Demis Hasibis, the AI CEO of DeepMind, who started the whole thing at Google. He's had to step away to focus on other things, but they also lost their chief AI scientist, Jeff Dean, who, for all intents and purposes, is the godfather of Google. He's responsible and the man behind 90 percent of Google's hit products, including the TPUs that power Gemini and power other models like Claude and ChatGPT, as well as having his hands in several other foundational research papers that has given birth to modern day AI as we know it. That guy also left, citing reasons that he can't do the work that he wants to work on at Google anymore. And so the public reaction was understandably catastrophic. The stock dropped 5 percent over a matter of hours, losing the company tens of billions of dollars. But recently, I think things have started to shift. Now, of course, Google in itself, on its own, without all the AI stuff, is a massively successful company. They're the third most valuable company in the world. I'm not negating that. But if AI is set to be the future, it's important that a company is able to pivot. When I look at the likes of Meta, when I look at the likes of Amazon, although they're operating in different spheres, they are able to focus and spend so much money and time and resources in trying to figure out the AI thing. And Google started very much in the same way. They've been arguably the longest company in this game, spanning back decades and decades. And so after this recent shift in staff, time and resources, it's important that they're able to get their foot back into the race, and we might be starting to see it. Just last week, they released Gemini 4 Argon, or at least they announced it. And if benchmarks are anything to go by, this model is pretty powerful and can compete with the best of the best from Anthropic and OpenAI. But they also recently, in fact today, as I'm recording this, released Nano Banana 2.1, their brand new image generation model. So we're starting to see more releases from Google. And on this episode, I want to dig into what's happened, where Google is at right now, and what the future might look like. All of this and more in today's episode of In the Loop brought to you by Qualcomm.

Now, very quickly, before we get into it, you might be wondering, Ejaz, this is a different setup and you're missing a co-host. There's a ton of updates that have happened for the show. Go and check it out on our video that we posted two days ago. It's on our channel and we'll explain everything there. But long story short is Limitless is evolving into a new brand refresh called In the Loop, but it's very much the same DNA of the show and we're going to be coming to you every single day of the week with the latest and greatest in AI and Frontier Tech. All right, so now setting the stage for Google as a company. Now, for the last seven months, Google has basically had their best model underperform every single flagship product that Anthropic and OpenAI have put up.

And I mean that for every single model, including Opus, including Sonnet, which are typically the cheaper, less impressive tiers from Anthropic, including Sol, including Luna and Terra, which again are the lowest equivalents coming from OpenAI. Their cheapest and technically worst models would still outperform Google's best. And that's weird because at the end of last year, Gemini was on top and they had a meteoric comeback where they kind of had the lead back in the day before GPT came out. And then GPT came out and took the world by storm, and Google was criticized for not being able to stage a comeback and actually catch up with all the research that they've done. And then they released Gemini, and the world's opinion changed on it. It became deeply integrated into every Google product that you could use, including Google Search. And it just became synonymous with using Cloud or Chat GPT. But something changed since then. And I don't know whether it's a research problem or whether it's a staffing problem or whether it's just generally a product company focus thing, but they weren't able to ship a competent model. Now, speculation behind that is at the same time that they started to announce delays, Cloud code hit an absolute exponential run. And this, of course, is the flagship product from Anthropic. And at that time, AI coding was kind of this mysterious, weird thing. Companies didn't really know whether they should focus on automating coding or not, or whether they should focus on automating research or other such things. And Cloud was the very first instance where it became clear coding and AI are a natural fit. And that is where the leading edge is going to be figured out. OpenAI figured this out about a month after Cloud code launched and reshifted everything to build their own competitor called Codex. And since then, they've been able to catch up. And you could argue that's why they've been able to build much better models that stay on par with Cloud. And now it's Anthropic versus OpenAI because they're using these models to build the next better versions of themselves, something called recursive self-improving. Google unfortunately watched all of this happen and thought, I think we can wait a bit. We can try and figure out all these other types of products that we've already committed to on our product roadmap. And we can come back to coding. And that ended up costing them pretty greatly. Which is a shame because Google has some of the biggest distribution on earth. You're seeing on the screen here, their market share value is around $4.1 trillion, making them, I think, the third most valuable company in the world. And they have around 3 billion users across all their different products. Now, it's very important to state that there is no other company like Google that has such a large product surface area across both general consumers, so the public like you and I, and enterprise customers that use their products pretty aggressively to document stuff on Google Docs, to create slides or to manage their databases via Google Sheets. So pretty much wherever you are on the internet, whether you're searching for something, whether you're trying to produce something, whether you're trying to analyze something, there's always a very competitive Google alternative, which is why then when AI exploded onto the scene, you would expect them to be able to pull off a similar stunt. And they actually were able to with Gemini, but they haven't been able to keep up. But don't take my word for it. I mean, look at the screen. When we look at the artificial analysis index, which is a really good general benchmark to try and test the intelligence of the latest model from these frontier companies, you see Claude Opus 5.5, which was released about a week and a half ago, score of 58 You've got GPT-6 Astra scoring a very close 53 And then you have Gemini 3.1 Pro, which was released back in February, which in AI years, funny I say years, which in AI timeline is years and years and years worth of time. So you wonder, why hasn't Google been able to ship an effective product? Well, the answer is I think there is a work culture problem because starting August 5th, but actually even before that, they started to bleed a lot of very important talent. And we covered this on previous episodes. We mentioned that John Jumper, who shared a Nobel Prize with Demisysibis for building AlphaFold, which is basically their protein sequencing DNA AI model, had left to join OpenAI. And then Jeff Dean, who I mentioned earlier on in the intro, who is kind of like the chief scientist and grandfather at Google for a lot of the other products that aren't just AI related, but just an absolute invaluable person at the company left to start his own AI lab. I think it's called Discovery Loop, which is focused on automating research first starting in AI and then other parts of research in physics and science. That is exactly what the frontier AI labs that are leading this entire field are doing and focusing on. So my question then becomes, why on earth did a loyalist like Jeff Dean feel the need to leave when he could have easily built this at Google? I think there's a work culture problem going on. So it led to this massive exodus. Demisys service had to step down and focus on something else. And so what we're left with is a completely bare team that needs to figure out a completely new strategy. But everything I'm saying, I admit, sounds very Duma-esque for a company that is the third most valuable in the world and is making so much money. Because when you look at the stats, right, they have over one billion in AI mode for search, past a billion in AI overviews, which have reached around 2.5 billion. And then when you look at the amount of tokens that they process for their AI products or across all of their products, it's 22 billion tokens a minute. They've also reported 82% growth in Google Cloud, which is their AI CapEx infrastructure business, where they're spinning up TPUs, which is their own version of Nvidia's GPUs and using it to facilitate selling training and inference costs to other AI labs or firms that want to train and raise their own models. And again, putting all of that aside, they have a humongous enterprise business, 8 million Gemini enterprise paid seats on their AI products alone. That's not even including some of their Google Cloud enterprise products. So for all intents and purposes, Google is an absolute cash generating machine, and I'm not negating that. They have one of the largest cash balances on earth for any company, but they've also been spending very aggressively to try and acquire and build as much compute data centers pay for the expensive CapEx costs that they're taking. But there's a bit of a mismatch. Google right now as they're positioned is focused to be a really good hyperscaler or infrastructure provider, akin to maybe a Neo Cloud that builds AI data centers or builds the chips, supplies them then to AI labs that can use them for inference costs or training their own AI models. And that is all fair and game, but Google at its core DNA is not just a research lab, they're this perfect hybrid of a research lab and a product company. That's why they've been able to become one of the biggest companies on earth, and that's why people love Google as a company and as a business. They haven't been able to translate this for AI yet. And you can argue that if they maintain this kind of product or working culture, then they're going to end up just remaining a hyperscaler or a neocloud, whilst companies that actually build the net new products like AI models, like AI agents that we've spoken about on previous episodes, will end up benefiting from the economic opportunity that just explodes over the next couple of decades. So with that said, now has never been a more important time for Google to shift gears and try and figure out an alternative strategy to get back in the AI race. I personally want to see Google win because they have such a huge advantage. They're not just a singular monolithic company. They have fingers in every different pie that you could think of when it comes to the internet. They're the most vertically integrated company that doesn't exist for any other AI lab. When you think of OpenAI and Anthropic, yes, fantastic models, but they lack the distribution. They don't have the billions of users. When you look at Meta, they have all the billions of users, but they don't have a bleeding edge model. Google, with all their wealth, resources and expertise, should be able to pull off a miracle and have the best vertically integrated AI product. They have all the data to mold that into the best consumer experience or enterprise experience, depending what kind of a product they want to build. But the result of what we've actually seen is very fractured teams being divvied up very specific resources that don't really make sense. This team gets compute, but the other team doesn't get compute, or another team gets way more compute, but it's spent on a product that maybe may not have as much impact on millions of users versus a couple hundred thousand. It's just this weird thing. And then they're selling compute to competitors when they probably should be using it to train their own model just to make cash in the near term so that it might look good on a quarterly earnings report versus trying to play the longer game. And so the question then becomes, okay, what moves will Google make under this new realm of losing half the AI talent and trying to figure out a new structure and approach? Well, we might see it with a brand new model. They announced Gemini 4 Argon, which as the Google post says, delivers frontier performance in complex workflows across real world software engineering, knowledge work and cybersecurity defense. Now, of course, the first thing that I looked at that Sundar Pichai, the CEO highlighted was their benchmark evaluations. Now, if you're watching the show, if you've watched it for a while, you know we're not big benchmark guys, it gives us an idea of what this model might be. And so the first thing that draws my eye is, okay, it's beating GPT-6 Astra and Claude Fabel 5.1, which are each respective company's flagship model, pretty well on a number of different things. But then you have to look at what that thing is. Now on the left here top, it's knowledge work. Talking about VALS index automation bench, knowledge work is basically, if you think about 70 to 80% of the tasks that are required for a software based job. And I'm not talking about software engineering, but I'm talking about anyone that has a 95, that uses a computer, that uses Slack and a bunch of these other different CRM tools or whatever that might be. You might need to access a spreadsheet or write up a strategy document or type messages to different workers in Slack or customer support. This model excels across any other frontier competitor model, which is fantastic, which honestly applies to a much larger percentage of the world than some of these other benchmarks and what they test for.