Latest / AI Fire Daily / 🎙️ EP 311: "AI 2040" Maps Out Future AI & Mistral’s Robostral Shakes Up Robotics
Transcript
- 0:00Welcome to the Deep Dive Learner. And we have
- 0:04some genuinely wild extremes to tease today.
- 0:08Yeah. We are talking about a 2029 global AI doomsday
- 0:12clock. Yeah. Human only ghost fonts. A robot
- 0:17that navigates the real world with just a single
- 0:19ordinary camera. It's a lot to process. I'm really
- 0:22glad you're sitting here with us today. I often
- 0:24just sit and, you know, reflect on this space
- 0:26quietly. The sheer velocity of AI development
- 0:29right now is frankly staggering. It really is.
- 0:32So here is the roadmap for today. We're going
- 0:34to start at the macro level. We'll map out a
- 0:36hypothetical 2040. Right. Then we'll zoom into
- 0:39the very real friction happening right now in
- 0:41the industry. We'll check out the grassroots
- 0:42developer rebellion. And we'll end with how AI
- 0:45is literally. growing legs. I love it. We're
- 0:47going top down future to bottom up reality. Exactly.
- 0:50So let's start with that top down view. Yeah.
- 0:52The AI Futures Project just launched something
- 0:53called AI 2040. Right. It's a massive stress
- 0:56test. Right. A detailed stress test of the next
- 0:58decade of the AI race. And they highlight the
- 1:01year 2029 as this crucial turning point. Yeah.
- 1:042029 is the big one. They basically tested five
- 1:07different scenarios for the globe. You've got.
- 1:10An all -out AI race. A one -sided delay where
- 1:14someone just stops. Global confrontation. Right.
- 1:18Cooperation. And finally, just a worldwide halt
- 1:21on everything. And the format is actually really
- 1:23accessible. It's a choose -your -own -adventure
- 1:25style structure. Yeah, it's super cool. But I
- 1:27want to impartially report on one specific scenario
- 1:29they call Plan A. Okay, Plan A. It proposes that
- 1:33the U .S. and China agree to pause frontier AI
- 1:36training by 2029. And frontier training simply
- 1:41means training the absolute most advanced models.
- 1:43Exactly. And this relies on something huge. It
- 1:46relies on strictly tracking global computer chips
- 1:49and forcing immense public transparency from
- 1:53AI labs. It's highly delicate geopolitically.
- 1:56Oh, totally. It's kind of like policing uranium,
- 1:58you know? Yeah, that's a good way to look at
- 2:00it. You can't see the software being coded, so
- 2:02you have to track the physical chips. Right.
- 2:04But a former OpenAI researcher actually co -authored
- 2:07this report. Oh, really? Yeah. And he explicitly
- 2:09noted that top labs won't like it, like at all.
- 2:13I can imagine. But why exactly is this 2029 timeline
- 2:18making companies so uncomfortable? Well, it's
- 2:21the tension between hyper growth and enforced
- 2:23transparency. By 2029, they expect to hit massive,
- 2:27massive profit milestones. And forcing them to
- 2:30open up their labs just completely ruins their
- 2:32competitive edge. The transparency mandate destroys
- 2:35their mode entirely. So controlling the hardware
- 2:38is the only real way to police the software.
- 2:40Got it. Exactly. Because software is just math,
- 2:43right? You can hide math. Yeah. But that is a
- 2:46very hypothetical 2040. Very much so. Let's transition
- 2:49to the messy reality of the present. Let's do
- 2:51it. Because right now the industry is just full
- 2:53of friction. Oh, it's so messy right now. Have
- 2:56you seen the Apple payout? I have. Apple is paying
- 2:59$250 million to iPhone buyers. Yeah, it's wild.
- 3:03It's over -delayed AI features. Right. People
- 3:05are getting $25 to $95 per device. It just shows
- 3:09how hard it is to actually ship this stuff. Right,
- 3:11and Microsoft is dealing with this too. The Microsoft
- 3:13CEO warned companies recently. He said they might
- 3:17end up paying twice for proprietary AI. Yeah,
- 3:21which makes helpful prompts incredibly valuable.
- 3:24Which is funny because I have to be honest here,
- 3:27I still wrestle with prompt drift myself. Oh.
- 3:30I mean, everybody does. You get it perfect, and
- 3:32then a week later, the model just forgets how
- 3:34to act. Yeah, it just wanders off. It's super
- 3:36frustrating. And then there's Anthropic. They're
- 3:38rolling out rupee pricing in India. Right, which
- 3:41is their second largest market. But ironically,
- 3:44they are completely missing the most common payment
- 3:47option. UPI. I know, it's crazy. Most of India
- 3:50uses UPI. It's a huge logistical stumble. But
- 3:53the money is still flowing. I mean, look at Pixverse.
- 3:55Right, they just raised $439 million. Yeah. They're
- 3:59valued over $2 billion now with 150 million users.
- 4:04It's staggering. And NOS's research has Hermes.
- 4:08Yeah, they're open source agent. And open source
- 4:10agents are free AI programs anyone can download.
- 4:13Right. And now they want a $1 .5 billion valuation.
- 4:17Everyone wants a billion dollars right now. Even
- 4:20OpenAI is making moves. Paid chat GPT users just
- 4:24got an extra banked reset. Yeah, celebrating
- 4:26their 7 million active codecs and work users.
- 4:29But users are also finding their own quirks in
- 4:31the system. Oh, definitely. Like the viral guides
- 4:33to bypass GPT 5 .6 cells limits. Right. The five
- 4:37hour limits. People hate limits. They will spend
- 4:3910 hours bypassing a five hour limit. Which brings
- 4:42us to Eric Liu. He's a developer who made something
- 4:45called ghost font. Oh, this is my favorite thing
- 4:48this week. It's a font that only humans can read.
- 4:50Yes. The smartest models, like Fable and GPT
- 4:54-5 .6 Sol Ultra, completely fail to decipher
- 4:57it. They just see static. They can't read a single
- 4:59letter. So how does this ghost font actually
- 5:02work to blind an AI while looking normal to us?
- 5:05So it's all about how they see versus how we
- 5:07see. Machine vision uses highly rigid mathematical
- 5:10pattern recognition. Okay. But human vision naturally
- 5:13filters out messy visual noise to find broader
- 5:16meaning. Basically, our brains handle visual
- 5:19noise way better than their rigid pattern recognition.
- 5:22Fascinating. Yeah, we just ignore the static.
- 5:24They get completely paralyzed by it. That is
- 5:26such a clever rebellion. We are going to take
- 5:31a short break right here. Mid -roll sponsor break
- 5:34placeholder. And we are back. Welcome back. So
- 5:38before the break, we talked about big tech stumbling
- 5:41over delayed features. Right. The corporate friction.
- 5:44But solo developers are moving incredibly fast.
- 5:47Oh, lightspeed. The grassroots developer rebellion
- 5:49is happening right now. Yeah. From the bottom
- 5:51up. Tell me about Osiris. Okay. Osiris is amazing.
- 5:55It's an open source AI agent running 100 % locally
- 5:58on Macs. So it's not in the cloud at all? Nope.
- 6:01It has persistent memory. And it is fully autonomous.
- 6:04But why does running a local agent like Osiris
- 6:07matter when cloud computing is so powerful? Two
- 6:10reasons. Ultimate privacy, because your data
- 6:13never leaves your laptop. And offline access.
- 6:16It has persistent memory, even if you lose your
- 6:18Wi -Fi connection. Local execution means total
- 6:20privacy and memory that doesn't wipe when the
- 6:22internet drops. Exactly. It's a huge shift in
- 6:25power. And there are tools to connect these agents
- 6:27to the real world, right? Yeah, like AgentKey.
- 6:30It connects agents like Claude Code or Codex
- 6:33to live external data. It's just one command.
- 6:36Right. And then you have things like the AI Media
- 6:38Buyer. Oh, the ad tool. Yeah. It automatically
- 6:41audits and scales meta, Google, and TikTok ad
- 6:45accounts. This is entirely on its own. Totally
- 6:48autonomous. No human needed. And for developers
- 6:51building these tools, there's Lumul. Yeah, Lumul
- 6:54is great. It lets creators easily monetize their
- 6:56API calls and tools. It's creating an entire
- 6:58micro economy. It really is. But what about security?
- 7:01With all these autonomous agents running around?
- 7:04Well, that is where February Clay Grant comes
- 7:07in. I saw this. It's offering over $100 ,000
- 7:09in weekly rewards. Yeah, to successfully hack
- 7:12live AI agents. It's completely gamified. It
- 7:15heavily reminds me of a digital bounty hunter
- 7:18board. That's exactly what it is. Because gamifying
- 7:20the hacking process actually builds better immune
- 7:23systems for future AI. Yeah. You find the holes
- 7:26now so the bad guys can't exploit them later.
- 7:28Makes perfect sense. two sex islands so we've
- 7:32talked about software agents on max yeah we've
- 7:34talked about the cloud but what happens when
- 7:36ai is literally growing legs the physical world
- 7:39this is the crazy stuff mistral ai just dropped
- 7:43a new 8b robotics model robostral navigate right
- 7:47and it allows robots to navigate complex unseen
- 7:51spaces. But the wild part is how it does it.
- 7:54It uses just one single ordinary camera. That's
- 7:57insane. No LiDAR. No LiDAR. No expensive multi
- 8:00-sensor arrays. Just a webcam, basically. And
- 8:02it achieved a 76 .6 % success rate. Yeah. Beating
- 8:06the expensive multi -sensor systems by 4 .5 points.
- 8:09Wow. And it doesn't calculate strict metric distances,
- 8:13right? Oh, it's more intuitive. It just kind
- 8:15of points to where it needs to go next. It's
- 8:17very similar to human walking. Exactly. We don't
- 8:19mathematically calculate geometry with our eyes.
- 8:21We just look and step. Yeah, we just intuitively
- 8:23know where the open space is. Whoa! Imagine scaling
- 8:27to a billion queries or putting this cheap webcam
- 8:31system into millions of delivery bots tomorrow.
- 8:33It completely drops the hardware cost to zero.
- 8:36Well, close to zero. And the pipeline was built
- 8:38entirely in simulation. Right. Which used 22
- 8:42times fewer tokens. Why is building the data
- 8:44pipeline... entirely in simulations such a game
- 8:48changer for the industry? Because the physical
- 8:50world is too slow. Collecting physical data means
- 8:53driving robots around for years. Right. In simulation,
- 8:56you just run it at a million times speed. It
- 8:59solves the biggest bottleneck. Simulations allow
- 9:02infinite trial and error without breaking any
- 9:04million -dollar hardware in reality. Exactly.
- 9:07And they finish it with reinforcement learning.
- 9:09And reinforcement learning is just teaching software
- 9:11through simple trial and error. Right. So if
- 9:13the robot bumps into something, it instantly
- 9:15recovers. It just shifts the immense burden of
- 9:18robotics from hardware right into software. Yeah,
- 9:21it's a total paradigm shift. We've covered a
- 9:23lot today. Let's try to recap the big idea here.
- 9:26I think we are watching a massive simultaneous
- 9:29squeeze. A squeeze. Yeah. From the top down,
- 9:33you have these geopolitical guardrails. Like
- 9:36AI 2040. Right. And you have corporate monetization
- 9:39schemes like Apple's payouts and Microsoft's
- 9:41warnings. They're trying to lock down the future.
- 9:44Exactly. But from the bottom up. The software
- 9:46is becoming too efficient. Yes. Yes. Local Mac
- 9:49agents, robots navigating with one cheap camera.
- 9:52The technology is actively resisting being boxed
- 9:56in. That is a brilliant synthesis. The top is
- 9:59trying to control it, but the bottom is just
- 10:01moving too fast. Yep, we squeeze this on. Thank
- 10:04you for this deep dive, Lerner. Please check
- 10:06the show notes. We've linked the AI 2040 Choose
- 10:09Your Own Adventure report there. It's super fun
- 10:11to play with. It really is. Geet. I want to leave
- 10:14you with one final thought. Think about the ghost
- 10:16font we discussed today. And then think about
- 10:19Mistral's robot learning to navigate the physical
- 10:21world with one eye. As machines learn to navigate
- 10:24our messy physical reality flawlessly, we humans
- 10:27are inventing new visual tricks just to hide
- 10:30from them. So over the next decade, who is actually
- 10:33going to be adapting to whom? Otero Music.