Imagine this for a second. Something that, you know, used to take years. Just painstaking, expensive research, tons of trial and error. Right. Like finding a needle in a haystack, basically. Exactly. And now an AI can do it in, what, two weeks? We're talking about designing brand new human antibodies from scratch. Yeah. That kind of speed, it's not just faster. It feels like it changes the whole landscape of what's possible in medicine.
Oh, absolutely. It's a huge leap. Traditional methods, they're so slow, they chew up resources. You're screening millions of things sometimes. Yeah. This just flips the script completely. Welcome to the Deep Dive. Today, we're going to jump into some really cutting edge AI breakthroughs. We've sifted through our sources to pull out
the really crucial bits of insight for you. We've got everything from these medical innovations like the antibodies to some really interesting stuff about how AI might, well, how it might think strategically. Yeah, that part's wild. Our goal here is just to unpack these key ideas, give you that shortcut to being really informed without, you know. getting buried in all the details. And it's going to be a pretty cool journey. We'll kick off with AI engineering and medicine,
specifically that. antibody design stuff. Okay. Then we'll shift gears a bit. Look at some surprising ways AI is popping up in different industries, some new tools people are using. Great. And finally, yeah, we'll tackle that big question. Can AI actually think strategically, like really think? All right, let's dive in then. First up, right into medical innovation. This new AI, it's called Chai2 from Chai Discovery. And interestingly, OpenAI is backing them. Yeah, that's a notable
connection. What seems really remarkable from the sources is its knack for engineering custom antibodies without needing like tons of previous examples to learn from. That feels different. It is. The clever part is how it does it. So you basically feed chai to the precise structure of what you want to target. Let's say a specific protein on a virus or maybe a marker on a cancer cell. And then the AI. It just goes and designs the antibody protein specifically to attack that
target. The analogy they used, which I kind of like, is thinking of it like Photoshop, but for proteins. Huh. Photoshop for proteins. Yeah. You've got a point where you want the antibody to connect. Yeah. And it designs something to stick right there. It sounds simple from the user end, right? Yeah. But the science underneath is obviously super complex. And the results, I mean, the numbers you mentioned earlier are just kind of wild, aren't they? They really are.
A 50 % hit rate on the targets they went after. And get this, they only need to test about 20 candidates. 20 potential antibodies for each target. Okay, wait. 50 % hit rate from just 20 tries? How does that stack up? It's massive. Like, the sources say it's 100 times better, 100x jump over traditional hit rates. Those are usually down around like 0 .1%. 0 .1%. Wow. Yeah. And think about the time. Biotech labs, they might screen millions of candidates. Takes months.
Sometimes years. Chai 2 did its thing in two weeks. Two weeks versus years. That's transformational. And a really key takeaway here is the impact, right? Economically and for patients, drug development is just notoriously expensive. Oh, yeah. Astronomical costs. Right. So expensive that companies often skip over smaller patient groups. Yeah. The R &D cost doesn't make sense for the market size. But if an AI like Chai 2 can churn out good candidates
this fast. Suddenly, those R &D costs plummet. And that means custom meds for rare diseases. They actually become financially viable. So treatments for conditions that were ignored before, suddenly they're potentially within reach. Exactly. For patients who maybe had zero options before, it kind of rejigs the whole economic equation of health care. OK, so with this incredible speed up in the design phase, it feels like we're right on the edge of something huge. But what's the
biggest hurdle now? What slows things down between the AI designing it and, you know, it actually helping a patient? Yeah, good question. It's likely still the usual suspects, regulatory approval and clinical trials. Those still take time. Right. The testing and safety checks. Makes sense. OK, so from the really intricate world of biological
engineering, let's kind of zoom out now. Let's look at how AI is appearing more broadly, sometimes in unexpected ways, and how it's starting to shift things across different industries, maybe even starting from the basement up. Absolutely. And the sources had some really interesting nuggets on this. First off, this idea of AI democratizing things, maybe shifting power balances. Mark Cuban was on a podcast. High performance, apparently. Yeah. And he predicted the first trillionaire.
It might not be a giant corp. It could be, quote, just one dude in the basement. Ah, one person. Yeah, someone who finds that sort of unseen use case for AI. Yeah. It kind of speaks to the power that can be concentrated now, maybe. Interesting. And we saw a bit of that user power with the Squid Game season three finale thing, right? Oh, yeah. That was funny. So apparently people really didn't like the finale. Which happens all along. Happens a lot. But fans, they didn't
just complain. They started using Google's VO3, that's one of the new video generation AIs, to make their own endings. No way. Really? Yeah. And these apparently became quite meme -y, as the source put it. It shows how people can use AI not just to consume stuff, but to actually reshape it, co -create almost. That's a different level of interaction. But it's not all smooth sailing with these new tools, is it? There was that issue with cursor software. Right, yeah.
That highlighted some of the bumps in the road. Users started reporting unexpected charges. One person apparently saw a $7 ,000 plan gone. $7 ,000? Whoa! Yeah. It seems like Cursor might have quietly changed its pricing tiers or something without being super upfront about it. Led to a bunch of people publicly saying they were canceling. It's a good reminder about, you know, transparency and ethics as these tools roll out. Definitely
need that clarity. Okay, but on the more positive side of accessibility, there was news about making apps easily. Yeah, a builder named Riley Brown put out a guide. How to create a mobile app from scratch under an hour. And get this, without writing any code. Under an hour, no code. Right. Just democratizing creation even further. Yeah. Giving more people the power to build stuff. That's pretty cool. And speaking of building things, what about the funding side? Harvey AI
came up. Oh, yeah. Big news there. Harvey AI, they do AI for legal stuff. Their valuation reportedly jumped from three billion. Hmm. Up to $5 billion. Wow. Quick jump. Yeah. After raising another $300 million from some big VCs, Kleiner Perkins, Sequoia, and the OpenAI Startup Fund was in there too. They're apparently serving like 337 legal clients now in 53 countries. Shows how fast these specialized AIs can scale, huh? And the value investors are seeing. Definitely. Big money flowing
into focused AI applications. Okay. Lots of activity. Let's maybe just mention one more of those new tools listed. Step fund diligence check. What caught my eye was the focus on verification. Yeah, that one sounds interesting. It offers AI search, but specifically with agent verified citations. Right. In this world of just information overload and sometimes misinformation, having AI tools that help verify sources seems, well,
pretty crucial. for trust couldn't agree more yeah and just one quick hit from the list someone apparently used chat GPT and saved themselves $3 ,000 just like that yeah presumably by using it effectively for some task or advice just shows the practical like real -world dollar impact it can have for regular users too so okay we've got new tools new funding new ways users are interacting but with so much popping up constantly what do you think is the biggest challenge for
people trying to navigate all this Honestly, probably just finding the right tool for the right job. There's so much choice now. Sponsor. All right. So we've seen AI designing antibodies, shaking up industries, even saving someone a few thousand bucks. Pretty practical stuff. But can it really, truly think strategically, you know, beyond just predicting the next word in a sentence? Yeah, the million dollar question,
isn't it? It really is. And this brings us to some fascinating research that was highlighted in the AI chart section of our sources. It tackled this head on. Right. And the way they tested it was pretty clever. They ran, get this, 140 ,000 games of Prisoner's Dilemma. 140 ,000. Wow.
OK. Classic game theory setup. Exactly. Using AI agents from... the big players open ai's models google's gemini anthropics claude and in each round the ai had to choose cooperate with the other ai or defect try to betray it standard prisoners dilemma choices but here's the kicker the really important part before every single move the ai models had to actually write out their reasoning why they were choosing to cooperate or defect. Ah, so they had to explain themselves.
Yeah, which allowed the researchers to track how they were making decisions. It gives us this little window into their quote -unquote thought process. Okay, I'm hooked. What did they find? Were they all just cold calculating machines? That's what you might expect, right? But the results were, well, the source described them as weirdly human. Weirdly human? How so? Well, they developed distinct styles. Gemini, for instance, turned out to be pretty ruthless. Calculated,
very reactive. If you betrayed it, it was quick to defect back. Okay, the pragmatic one. Sort of. Then GBT -4 was weirdly idealistic, was the phrase used. Often cooperative, like it kept trying to cooperate, even when the other AI was exploiting it. Huh, almost naive. Maybe. And then Claude emerged as the peacemaker. It was the most forgiving, apparently. Even after being metaphorically backstabbed, it was more likely to try cooperating again. That is weirdly human.
Different personalities almost. And that's the really wild part. Remember, they all trained on essentially the same foundational data, the same giant pile of text and code. Right. Same starting point. But despite that shared training, they developed these totally different approaches. Unique strategic fingerprints, as the source called it. It depended on how each model reacted internally to things like betrayal or success or building trust. So it wasn't just mimicking
patterns in the data. Seems not. This suggests that these LLMs, large language models, the AIs trained on text, they can actually do strategic reasoning, not just predicting the next word. Two sec silence. It's about making decisions
based on a consistent internal. a porch it shows that even with the same training they develop these kind of intrinsic cognitive styles whoa okay imagine putting them all on the same team or maybe more interesting having them negotiate against each other right but the implications there for like future ai systems especially autonomous agents yeah they feel really profound it's not just about what they can do but the style in which they do it these different styles of intelligence
emerging and how they'll interact Yeah. That's something else. Yeah. You know, I still wrestle with prompt drift myself. Sometimes that's where, you know, an AI's answers kind of subtly change over time, even if you ask the same thing. Yeah, I've seen that. So seeing these AIs develop such distinct, consistent strategies over 140 ,000 games, it's both amazing and honestly a little bit daunting. It suggests the level of internal coherence. It's, well, it's pretty sophisticated.
So thinking about building more complex AI systems down the road. What does this experiment tell us? What's the key takeaway? I think it tells us that future AI agents, they probably won't be interchangeable cogs. They might have their own styles. Right. Not just plug and play copies. Okay. So wrapping this up then, what we've really explored today, it feels like more than just AI getting faster or incrementally smarter. It
feels like a shift. How so? From AI just being a tool to maybe being more like a co -creator, like with the antibodies or the fan -made endings, or even an emergent intelligence with its own distinct style, like in the Prisoner's Dilemma games. It's not just about what AI can do anymore, but it raises questions about how we interact with it, how we regulate it, maybe even how we define these evolving digital minds. Yeah, I
think that's right. And to kind of reinforce that, AI is clearly delivering real economic value. We saw that, right? costs, potentially democratizing access to things like rare disease treatments. That's huge. And maybe the most intriguing part is what you just said. These systems are starting to show distinct personalities or at least distinct strategic approaches. Understanding those unique strategic fingerprints seems really important now because it'll shape not just what
AI can do, but how it'll do it. And that impacts, well, everything from business negotiations to how autonomous systems make decisions in the real world. So the final thought maybe. For you listening, what does all this actually mean for you? As AI starts developing these unique personalities and strategies, how might that change the way we interact with these systems day to day? Or maybe even how we think about intelligence itself in the coming years? It's definitely something
to ponder. We hope you'll consider the implications of these advancements. Keep asking questions. Stay curious. Keep exploring. Thank you for joining us on this deep dive. Out to you, music.