Latest / AI Fire Daily / 🎙️ EP 288: Policy on the AI Exponential & Perplexity’s Mind-Blowing AI Agent Data
Transcript
- 0:00An AI model just discovered high -severity zero
- 0:04-day vulnerabilities across every single major
- 0:07web browser. At the exact same time. Right. And
- 0:09meanwhile, a Japanese farmer with literally zero
- 0:12coding experience used AI to, well, launch a
- 0:17satellite crop tracker to monitor his fields.
- 0:19It's just two completely different realities,
- 0:21both of them happening today. Welcome to the
- 0:23Deep Dive. I'm really glad you're here with us.
- 0:24Today we are exploring this... extreme dual reality
- 0:29of AI, we are going to unpack Anthropic's terrifying
- 0:32new safety framework. Yeah, that one is an intensely
- 0:35heavy read. It really is. And I'll also look
- 0:37at the wild real world tools dropping right now,
- 0:39things like Claude Fable 5. And then we'll analyze
- 0:43a massive Harvard and perplexity study. Which
- 0:46basically proves that AI is fundamentally rewiring
- 0:49human ambition. We are looking at a total paradigm
- 0:51shift. Completely. So let's start at the macro
- 0:53level. AI is moving at this exponential speed,
- 0:56but policymaking is, you know, moving at dial
- 0:58-up speeds. And that gap is where the real danger
- 1:02lives. Exactly. So to close that gap, Anthropic
- 1:05dropped a two -part policy framework. And the
- 1:09trigger for this is actually really fascinating.
- 1:11Yeah, the preview model. Right. They ran a preview
- 1:13version of their next -gen model. It's called
- 1:15the Claude Mythos Preview. And it basically just
- 1:18went hunting. It did. It successfully found thousands
- 1:22of high severity software vulnerabilities. This
- 1:25included zero days across all major web browsers.
- 1:28Right. Which are unknown software flaws hackers
- 1:31exploit before developers fix them. Exactly.
- 1:33And the sheer scale of that discovery just terrified
- 1:36the researchers. I mean, it would terrify anyone.
- 1:38Yeah. So they established strict new internal
- 1:41rules. But these rules only hit what they call
- 1:43frontier developers. Meaning the massive corporate
- 1:46players. Right. It means companies. pulling in
- 1:48over $500 million in AI revenue, or companies
- 1:52spending over $1 billion on AI development. So
- 1:56they are specifically targeting the whales here.
- 1:58Yes, the big spenders. And the framework outlines
- 2:02four nightmare scenarios. They are actively trying
- 2:05to prevent these specific outcomes. First, AI
- 2:08lowering the barrier to creating biological weapons.
- 2:11Right, because suddenly anyone could synthesize.
- 2:14dangerous materials without any background in
- 2:17biology whatsoever exactly second massive automated
- 2:20cyber attacks on critical infrastructure third
- 2:23models completely losing control and acting outside
- 2:26of developer intent and fourth ai accelerating
- 2:29its own research loop into a runaway spiral Two
- 2:32sec silence. Those are incredibly heavy scenarios.
- 2:35Very heavy. And to stop them, Anthropic proposed
- 2:37some strict mandates. Frontier Labs would be
- 2:40legally forced to run rigorous internal testing.
- 2:43They also have to publish highly detailed risk
- 2:45reports. Yeah. And hand their models over to
- 2:47qualified independent evaluators. You just can't
- 2:50grade your own homework anymore. Which makes
- 2:52sense. Right. There is also a broader societal
- 2:54resilience plan in there. They want mandatory
- 2:57gene synthesis screening. To prevent that bioweapons
- 3:00scenario we just mentioned. Like if someone tries
- 3:02to print a virus, it flags the system. Precisely.
- 3:05They also want urgent patching for aging software,
- 3:08the legacy code that runs our critical power
- 3:12grids and our hospitals. Oh, man, that is a massive
- 3:15undertaking. Updating decades old infrastructure
- 3:18is incredibly slow work. It is. Anthropic also
- 3:22took a really firm stance on state laws. They
- 3:25warned that Congress shouldn't block powerful
- 3:28state level AI regulations. Like the ones currently
- 3:31proposed in California or New York. Right. Unless
- 3:33the federal equivalent is just as strong, they
- 3:36do not want a watered down national standard.
- 3:38Makes sense from their perspective. It does.
- 3:40But I really have to push back on this framework.
- 3:42It feels like, well, like installing blast doors
- 3:45on a bank vault, but leaving the back window
- 3:47completely open. Because the framework only targets
- 3:50the billionaires. Exactly. Regulating only the
- 3:53massive billion dollar spenders leaves a huge
- 3:56blind spot. I mean, what about open source rogue
- 3:58actors? Yeah, they don't need a billion dollars
- 4:00to cause chaos. Not at all. That is the ultimate
- 4:02tension in the industry right now. Because open
- 4:05source models are getting incredibly powerful.
- 4:07A rogue actor could theoretically download a
- 4:10model today. Then they just fine tune it for
- 4:14malicious purposes on a cheap server. Which bypasses
- 4:17these frontier regulations entirely. Exactly.
- 4:20Anthropic is trying to contain the absolute cutting
- 4:22edge. But the floor of what is possible just
- 4:25keeps rising for everyone else. That leads me
- 4:28to a crucial question about these risks. What
- 4:30is the actual timeline we are looking at for
- 4:33these nightmare scenarios? We're not talking
- 4:35about decades anymore. Given the speed of the
- 4:38mythos preview discovery, we're looking at a
- 4:40zero to three year window for these threats to
- 4:43become highly actionable. So the threat is already
- 4:45here and we're just playing catch up. Right.
- 4:47And that is exactly why they are panicking. Well,
- 4:50let's transition from anthropics macro fears.
- 4:52If that is what AI might do, let's look at the
- 4:54micro reality. What is AI already doing right
- 4:57now on the ground? Yeah, this is where the landscape
- 4:59gets really fun. It does. Let's talk about Cloud
- 5:02Fable 5. A leaked prompt reveals it is highly
- 5:06tool -heavy and incredibly safe. It is blowing
- 5:09minds less than a day into its release. It aced
- 5:12three of the internet's trickiest questions.
- 5:14These are logic puzzles specifically designed
- 5:16to fool AI. Like that famous how many R's in
- 5:19strawberry test. Right. It is highly info -aware.
- 5:23It understands context way better than previous
- 5:25versions. Meanwhile... OpenAI is trying to stay
- 5:29competitive in this exact space. They are considering
- 5:32major token price cuts. Which is the cost to
- 5:35process a basic unit of data. Exactly. Cheaper
- 5:37access means entirely new classes of people can
- 5:40build pools. And speaking of building, let's
- 5:43talk about the creators. Remember that Japanese
- 5:45farmer we mentioned at the start? I love this.
- 5:47That is honestly one of my favorite stories right
- 5:49now. He has absolutely zero engineering background.
- 5:51None. But he used codecs to build his own greenhouse
- 5:54automation system. He also built a physical farm
- 5:57bot to navigate his fields. And a satellite crop
- 6:00tracker from his laptop. It's wild. It is unbelievable.
- 6:03He essentially replaced an entire agricultural
- 6:06engineering department by himself. We're also
- 6:08seeing wild video generation tools democratizing
- 6:13production. Like Luma AI just released Ray 3
- 6:16.2. It lets you direct complex video. with simple
- 6:21text prompts. You get 16 keyframes and 8 face
- 6:24tracks to lock down consistency. And 20 -second
- 6:27clips rendered at full 1080p resolution. It even
- 6:30has a full API for developers. That's huge. Yeah.
- 6:33And then there's HeyGen. It's an official cloud
- 6:35connector. You literally create polished videos
- 6:38directly from a text chat interface. It features
- 6:4025 unique skills for editing scenes and adjusting
- 6:43motion. It's just wild to see this infrastructure
- 6:46scale so rapidly. I mean, look at a company like
- 6:48Standard Bots. They just raised $200 million.
- 6:50Pushing them to a massive... $1 billion valuation.
- 6:54They claim their AI robots can fundamentally
- 6:56boost U .S. manufacturing. And the historical
- 6:58context here is massive. U .S. manufacturing
- 7:01jobs peaked at 20 million back in 1979. Today,
- 7:04we only have about 13 million of those jobs left.
- 7:08And AI robotics claims it can aggressively reverse
- 7:11this downward trend. It's a bold claim. But the
- 7:14digital ecosystem supporting these agents is
- 7:17exploding everywhere. We have new infrastructure
- 7:20tools like Polra popping up. Yeah, that's a publishing
- 7:23API connecting to 10 different social platforms.
- 7:25Right. It gives text -based agents like Claude
- 7:28a full engagement loop. They can post, read replies,
- 7:32and adjust their strategy automatically. There
- 7:34is also Spotlight. It's a free tool that natively
- 7:37reads your code sessions. It shows you exactly
- 7:39what your agents actually did behind the scenes.
- 7:42TypingMind is another great one in this ecosystem.
- 7:44It brings the best models across 18 different
- 7:47providers into one single workspace. The speed
- 7:49of this change is honestly dizzying. I have to
- 7:54admit something. I still wrestle with prompt
- 7:55drift myself. Oh, we all do. It's the ghost in
- 7:58the machine. You get a model working perfectly
- 8:00for a specific workflow, and then... A week later,
- 8:03it just acts completely different. It's incredibly
- 8:06frustrating. Well, the underlying models are
- 8:08constantly updating and shifting their weights.
- 8:10Right. But there is a fascinating contradiction
- 8:13in all of this. Standard bots needs a $200 million
- 8:17war chest. They need that massive funding to
- 8:20revolutionize physical manufacturing. But on
- 8:23the other hand, a solo Japanese farmer replaces
- 8:26an entire engineering team. With just a standard
- 8:29laptop and a basic internet connection. Exactly.
- 8:31So here's my question. Why is there so much friction
- 8:35bringing AI to blue collar robotics when white
- 8:38collar software scales instantly for a single
- 8:41farmer? It all comes down to the unforgiving
- 8:43physics of the real world. I mean, software bits
- 8:46can be duplicated instantly for practically free.
- 8:48Yeah. But blue collar robotics. deals with physical
- 8:52atoms. You have to manufacture steel, deal with
- 8:54gravity, and manage complex hardware supply chains.
- 8:57Code copies for free, but you cannot copy paste
- 9:00solid steel. Precisely. And that physical capital
- 9:02requires massive upfront funding. Sponsor. Welcome
- 9:06back. So we have seen the raw power of the tools.
- 9:08We've seen the macro fears from companies like
- 9:11Anthropic. Now let's look at the psychology.
- 9:13This is arguably the most important part of the
- 9:15entire discussion. I completely agree. What are
- 9:19these tools actually doing to human behavior?
- 9:21We're moving from software capabilities to fundamental
- 9:25psychological shifts. Perplexity in Harvard Business
- 9:28School just dropped a massive... joint study.
- 9:31It is completely shifting how we view modern
- 9:34knowledge work. They systematically compared
- 9:36traditional internet search against perplexity's
- 9:38new computer agent platform. They rigorously
- 9:41analyzed 10 ,000 identical queries across both
- 9:45methods. That is a massive sample size for behavioral
- 9:47research. Let's break down the data they found.
- 9:50First, we have the time gap. Regular search usually
- 9:52feels pretty quick. You type a question, you
- 9:56get a list of links. But search leaves the heavy
- 9:58cognitive execution entirely up to you. When
- 10:02you factor in the human effort of reading and
- 10:03synthesizing, the traditional workflow took a
- 10:05really long time. How long exactly did the researchers
- 10:08clock it at? An estimated 269 minutes to complete
- 10:11a complex task. That is well over four hours
- 10:14of solid human effort. Right. But the agent workflow,
- 10:17it wrapped the exact same task up in just 36
- 10:20minutes. Whoa. I mean, imagine scaling that 230
- 10:25minute time savings across a billion queries.
- 10:28It is a staggering amount of unlocked human time.
- 10:31But time saved is only half the story here. Let's
- 10:34look at the creation gap. Right. Like what were
- 10:36people actually doing with all that newly saved
- 10:38time? 50 % of the tasks handed to the agent involve
- 10:41building something entirely new. Wow. Yeah, that
- 10:44is double the creation rate we typically see
- 10:47on regular search engines. So people are rapidly
- 10:49shifting from passively consuming to actively
- 10:52creating. Yes. And then we have the extra piece
- 10:55gap. This is easily the most fascinating data
- 10:58point in the entire study. The number of tasks
- 11:00falling completely outside the user's actual
- 11:03field of expertise jumped significantly. By how
- 11:06much of a margin did it actually jump? It jumped
- 11:08nine full points, up to 59 % of total tasks.
- 11:12Wait, really? More than half of the tasks were
- 11:14outside their wheelhouse. Exactly. Users completely
- 11:17trusted the agent with highly cognitively heavy
- 11:20work. They confidently asked it to generate complex
- 11:23code across multiple disciplines. They drafted
- 11:25complex legal or technical documents without
- 11:28hesitation. They built multilayered visuals and
- 11:32complex data structures, things they would never
- 11:35historically attempt alone. Think about it this
- 11:38way. Using regular search is basically like going
- 11:41to a massive public library. You still have to
- 11:44find the specific instructions yourself. Right.
- 11:46You have to read the books, take detailed notes,
- 11:49and try to build it yourself. But using an AI
- 11:51agent, that is like hiring a brilliant chief
- 11:54of staff. A chief of staff who already read all
- 11:57the books beforehand. Exactly. They read the
- 11:59books and built the working prototype while you
- 12:01were just, you know, having coffee. The real
- 12:03unlock here isn't just raw speed. It is about
- 12:07removing the paralyzing friction of grunt work
- 12:09entirely. Yes. When you remove that friction,
- 12:12it actually increases human ambition. People
- 12:14naturally start aiming a lot higher. Perplexity's
- 12:17data proves this perfectly. I mean, interacting
- 12:20with capable agents gives us the underlying confidence
- 12:23to handle vastly more complex projects. Projects
- 12:26we normally wouldn't even attempt to start. But
- 12:28this incredible data leads to a highly critical
- 12:31question. If agents unlock ambition and creation
- 12:34for literally everyone, will this lead to a massively
- 12:38oversaturated market of average AI -generated
- 12:42creations? It definitely will create an absolute
- 12:44flood of content. I mean, when creation is easy,
- 12:47the baseline of quality becomes completely ubiquitous.
- 12:50But that just means original taste and human
- 12:52vision become the rare, valuable commodities.
- 12:55When anyone can build anything, original human
- 12:58taste becomes the premium. Exactly. The uniquely
- 13:00human element ends up mattering even more. Let's
- 13:03weave all of this together. We started with Anthropic
- 13:05today. They are building massive societal blast
- 13:08doors. Driven largely out of fear of AI's exponential
- 13:12compounding power. Yeah. They see the zero days.
- 13:15They see the existential risk to critical infrastructure.
- 13:18But that exact same technological power is what
- 13:20empowers a single Japanese farmer. It lets him
- 13:23intelligently automate his physical fields entirely
- 13:25from his living room. It gives regular people
- 13:27the confidence to step way outside their established
- 13:30expertise. 59 % of tasks were far outside their
- 13:33normal comfort zones. It is a massive behavioral
- 13:36shift for the entire global workforce. The dichotomy
- 13:40of our era is perfectly balanced. We have existential
- 13:43risk matched exactly by an unprecedented explosion
- 13:46of human ambition. Beat. Which leaves you with
- 13:50a really fascinating question to ponder. AI agents
- 13:53are rapidly removing the historical friction
- 13:55of execution. They are doing the heavy cognitive
- 13:58lifting for us. The repetitive grunt work is
- 14:00basically disappearing. Right. Your ultimate
- 14:02value is no longer in your ability to do that
- 14:05grunt work. Execution is rapidly becoming nearly
- 14:08free. So in a world where execution costs absolutely
- 14:11nothing, how do you cultivate the wisdom to know
- 14:14what to build? That is the ultimate challenge
- 14:16for the next decade. Try a simple experiment
- 14:18today. Take just one task that is completely
- 14:21outside your normal comfort zone. Offload it
- 14:23entirely to an AI agent. And just see what it
- 14:25actually does to your personal ambitions. See
- 14:27if it changes how high you aim. Thank you for
- 14:29joining us on this deep dive. We will see you
- 14:31next time. Outiero Music.