Is Google's Latest Drop Good Enough?

**Beth Lyons** (0:00)
Hey, everybody, welcome, welcome. Welcome to Wednesday. It is Wednesday, July 22nd, and this is episode 773, which means yesterday was not what I said it was. It was 772 Tomorrow will be 774, and this is the end of our math segment for The Daily AI Show. Thank you very much.
I am Beth Lyons. With me in the studio today is Andy Halliday.

**Andy Halliday** (0:29)
Good morning. I'm doing well. Thank you.
I know you're going to talk about OpenAI. I'm looking forward to that. But for me, it's Google Day. So I'll jump into Google.

**Beth Lyons** (0:42)
So let's start with Google. Absolutely. Because Google dropped some models.

**Andy Halliday** (0:47)
Absolutely. Well, yesterday, they released three new AI models. And we've been working with across the Gemini universe, we've been working with Gemini 3.5 Flash. And so they released the newest version of that, which is their sort of core, advanced intelligence model that powers Gemini throughout. And now they released yesterday, Gemini 3.6 Flash. It's the most powerful model that they've released, put into general release. And they also released something called Gemini 3.5 Flash Cyber. That goes back to 3.5 Flash, using that one. But they developed a fine-tuned cybersecurity model. And they released that to governments only.
And so this is because these more advanced models are really capable of finding what human cybersecurity evaluators have not been able to find, which are vulnerabilities, including zero-day vulnerabilities, being things that have never been seen before.
And so nobody's put up any defenses against these things. And I think that's one of the things you'll talk about with respect to OpenAI, because they mentioned something about this yesterday.

**Beth Lyons** (2:08)
Yes.

**Andy Halliday** (2:09)
So this shows that Google is actually going to compete with Anthropic and OpenAI on the cybersecurity front for governments around the world.
Anyway, the Gemini Flash 3.6 improves on Google's previous version of that Flash model. And by the way, Flash in the world of Gemini doesn't mean, oh, this is like a small and not so frontier model. It's pretty much a very, very good model for almost anything that you would work on. But behind the scenes, we've had 3.1 Pro.
And so rewind all the way back to Gemini 3.1, and Pro has been their most capable model behind the scenes. And they announced that, as part of these releases, they announced that Gemini 3.5 Pro is already in partner testing, meaning that they've given it to a number of major information technology companies to play with, and to see what they might need to sand off in respect of rough edges on the fine-tuning and final release of 3.5 Pro. And, not to be limited to 3.5 or 3.6, they revealed that Gemini 4 is in its pre-training run, it's in an ongoing pre-training run. So, never count Google out. You know, they have massive resources to do this sort of thing.
And, to reinforce that point, Ineffable Intelligence today announced strategic partnership with Google Cloud to provide the specialized infrastructure for Ineffable's pursuit of superintelligence. Now, Ineffable Labs, if you've forgotten, as I might have, is the startup that got $1.1 billion in its seed round, which is the world record for a seed round. It's in the hands of David Silver, who was the AlphaGo leader. He's a visionary of AI inside DeepMind. And so, their pursuit is for super learning. Let me tell you a little bit about that. But their connection to Google is that Google has the special capacities and systems to make this super learning possible. So, the pursuit of super intelligence is really Ineffable Intelligence, partnering with Google, Ineffable having sprung out of Google. And this is the kind of thing where Google often has done this in the past, where somebody inside Google, in this case, DeepMind, goes out and builds a start up, and then Google brings them back. Oh, sorry. I got excited there. Who brings them back in, you know, and reacquires that technology. So, they're not too far apart, and they're working very closely together. So, let me explain what super learning is about, at Ineffable Intelligence, this offspring of Google. Super learning capability is the ability to endlessly discover knowledge and skills without relying on human data. And it will be driven by the world's most powerful reinforcement learning algorithms, because reinforcement learning, frankly, was invented by Neve Bindt, and that's where the sort of the center of excellence has been always about reinforcement learning. But this is about super learner.
Super learner is expected to rediscover and then transcend the greatest inventions in human history, including language, science, mathematics, and technology. And they consider this pursuit to be towards something that will be a scientific breakthrough comparable to Darwin's understanding of how evolution has created intelligence and created all the many wonderful things that exist on our planet. And so, experience-based learning, I was talking about this yesterday, that the world models are experience-oriented because they simulate the actual causal and effect related transactions that happen in the real world. And so, that whole system places different and more extreme demands on the computing infrastructure than training on static data.