Latest / AI Fire Daily / 🎙️ EP 301: China Drops "Tulongfeng" to Rival Mythos & The Truth About Today’s AI "Agents"
Transcript
- 0:00Think about the last time you used an AI agent.
- 0:02It probably felt like a minor convenience, honestly.
- 0:05Right. Maybe you asked it to draft a quick email.
- 0:07Or maybe you had it summarize a very long meeting.
- 0:09We toss the phrase AI agent around very lightly
- 0:12today. Yeah, we really do. Meanwhile, nations
- 0:16are quietly shifting trillions of dollars right
- 0:19now. They're completely rewriting the rules of
- 0:22global cybersecurity. And they're doing it based
- 0:24on that exact concept. It's wild to think about.
- 0:28Welcome to the Deep Dive. We're incredibly glad
- 0:30you're here with us today. We're unpacking a
- 0:32fascinating stack of AI industry newsletters.
- 0:35They cover everything from global hardware shortages
- 0:37to cyber warfare. We also have a dense philosophical
- 0:41research paper to discuss. It asks what machine
- 0:44agency actually means in our reality. Our mission
- 0:47today is to cut through the industry hype. We
- 0:50want to map out the new rules of AI. We'll explore
- 0:53how these systems are currently scaling up. We'll
- 0:55look at where they're being deployed today. And
- 0:57we'll ask if they're truly thinking independently.
- 0:59We really should start with the physical foundations
- 1:02first. Yeah, before exploring software, we must
- 1:05establish the physical limits. These hardware
- 1:07constraints are driving the entire global AI
- 1:11strategy. That makes total sense, you know, logically
- 1:14speaking. You simply cannot run advanced software
- 1:16without the hardware underneath. South Korea
- 1:19just made a truly massive new technology commitment.
- 1:22Right. They're dedicating over $900 billion overall.
- 1:25That's an almost incomprehensible amount of capital
- 1:28being deployed. This money goes directly toward
- 1:30new AI ships and infrastructure. It shows how
- 1:34serious the global infrastructure race actually
- 1:37is. It's genuinely staggering when you look at
- 1:39the breakdown. Out of that massive budget, $518
- 1:43billion is earmarked. It's going specifically
- 1:46toward building brand new memory fabs. They absolutely
- 1:49have to focus on that specific hardware component.
- 1:52They're focusing heavily on the physical memory
- 1:54supply chain. The industry is facing a severe
- 1:57global shortage right now. Insiders are actually
- 1:59referring to this crisis as Ramageddon today.
- 2:02That's a very dramatic term for a supply chain
- 2:05issue. It's dramatic, but it's highly accurate
- 2:07for the situation. You can build the absolute
- 2:09fastest processor in the world. Right. But if
- 2:12it has no short -term memory, it just stalls
- 2:14out. The data centers simply cannot hold the
- 2:17information they process. The newer AI models
- 2:19have grown too large for physical memory. It's
- 2:22like trying to build a global fleet of sports
- 2:24cars. Yeah. But the entire world has completely
- 2:27run out of tires. You have all this raw engine
- 2:30power sitting completely useless. Processors
- 2:33are just waiting around for the data to arrive.
- 2:35South Korea sees this specific hardware bottleneck
- 2:38very clearly today. They want to control that
- 2:41exact missing piece globally now. They're positioning
- 2:44to be the primary supplier of AI memory. How
- 2:47do physical memory shortages dictate the pace
- 2:50of AI evolution? Well... Without memory, you
- 2:53simply cannot run larger, complex language models.
- 2:56The physical constraint forces developers to
- 2:59either wait or highly optimize. It totally changes
- 3:02how they build their actual software architecture.
- 3:05So South Korea is essentially buying the world's
- 3:08memory card out of sheer necessity. Because physical
- 3:12hardware is constrained everywhere. Nations are
- 3:14pivoting fast. Right. They're getting incredibly
- 3:16creative with their software architectures instead.
- 3:19They want to level the playing field without
- 3:21needing more chips. This brings us to a fascinating
- 3:24new geopolitical development today. The United
- 3:27States previously locked down advanced automated
- 3:30bug hunting models. A prime example of this is
- 3:33a system called Mythos. They wanted to maintain
- 3:36a strict monopoly on exploit hunting capabilities.
- 3:40Finding zero -day flaws is crucial for any national
- 3:43cybersecurity. Yes, but a Chinese cybersecurity
- 3:46firm just changed the entire equation. The company,
- 3:49360 Security Technology, launched a brand new
- 3:52tool. They call this automated cyber system Tu
- 3:55Long Feng. And they're billing this as China's
- 3:58direct answer to MISOS. The founder, Zhu Hongqi,
- 4:01was very open about their current position. To
- 4:04be clear, we're just imparting the facts reported
- 4:06here. Right. And he admitted their base models
- 4:09trail the U .S. significantly. They're behind
- 4:11by 20 to 30 percent in raw compute. That's a
- 4:14very significant physical hardware deficit to
- 4:18overcome. It's a huge gap in the traditional
- 4:20computing sense. To bypass this hardware deficit,
- 4:23Tu Long Feng uses a swarm architecture instead.
- 4:26Instead of building one massive brain, it builds
- 4:28many smaller ones. Exactly. It relies on specialized
- 4:31AI agents working constantly together. They operate
- 4:3524 -7 without needing massive unified computing
- 4:39power. Whoa. Beat. Imagine a swarm of specialized
- 4:43agents scaling up. Millions of automated queries
- 4:47happening every single second across networks.
- 4:49The sheer volume of that distributed effort is
- 4:52completely mind -boggling. It's incredibly effective
- 4:55when you break problems into tiny pieces. A million
- 4:58small models can scan a million lines of code.
- 5:01The numbers are already proving the swarm concept
- 5:03works exceptionally well. Tu Long Fing has reportedly
- 5:06discovered over 3 ,000 software vulnerabilities
- 5:09already. 3 ,432 vulnerabilities to be completely
- 5:13exact. Yeah, and Chinese authorities have officially
- 5:16confirmed 105 of them. That's a massive return
- 5:19on a distributed software approach. It completely
- 5:22bypasses the need for the absolute biggest processors
- 5:25today. And they didn't just stop at offensive
- 5:27bug hunting tools either. No, they also launched
- 5:30a sister AI system alongside it entirely. This
- 5:33defensive system is officially called Yitianzen
- 5:36for the record. It's designed for automated cyber
- 5:39defense and rapid incident response. The old
- 5:41rules of the cyber game have officially changed
- 5:43now. We can clearly see that traditional containment
- 5:46strategies are failing today. The focus is shifting
- 5:49to securing critical digital infrastructure entirely.
- 5:53Do software containment policies ever actually
- 5:56succeed long term? History shows that software
- 5:58naturally wants to spread and adapt quickly.
- 6:01When you block raw compute Engineers just invent
- 6:04smarter distributed architectures. Right. Code
- 6:06always finds a clever way around artificial borders
- 6:09and hardware deficits. This explosion of specialized
- 6:11automated AI is spreading absolutely everywhere
- 6:14now. It's not just happening in military or national
- 6:17cyber defense. It's rapidly flooding our everyday
- 6:20workflows and our mobile devices. It's fundamentally
- 6:22changing corporate strategies across the entire
- 6:25globe today. We're seeing incredible efficiency,
- 6:28but also massive friction everywhere, too. Let's
- 6:31look closely at the corporate friction side of
- 6:33things first. Meta is a perfect example of this
- 6:36corporate friction right now. They're actually
- 6:38limiting their engineers from using certain AI
- 6:41tools. Specifically, they restricted clod code
- 6:45and codecs for their internal teams. Why are
- 6:49they pulling back on such powerful developer
- 6:51tools? Two sec silence. You'd think a tech giant
- 6:54wants faster engineers. They deeply fear the
- 6:57outputs could contaminate Meta's own training
- 6:59data. Ah. If AI generated code enters their internal
- 7:03code base without human checks. Right. It gets
- 7:06absorbed into their next generation of internal
- 7:07models later. That recursive training loop can
- 7:10become very busy very quickly indeed. The AI
- 7:13starts learning from its own previous mistakes
- 7:15and hallucinations. Exactly. It becomes a completely
- 7:17unmanageable digital echo chamber over time.
- 7:21Meanwhile, other entities are fully embracing
- 7:23these new agents without hesitation. Anthropic
- 7:26is offering Claude to California government agencies
- 7:29right now today. They're providing a massive
- 7:3250 % discount for public employees. And they're
- 7:35including customized training to help them move
- 7:37faster. Across the globe, Austria is making similar
- 7:40aggressive moves today, too. They're urging the
- 7:42European Union to bring Anthropic over immediately.
- 7:45They want to hedge against any potential U .S.
- 7:48access restrictions tomorrow. Waiting on foreign
- 7:51AI policy decisions is not enough anymore. Nations
- 7:55need these specific tools integrated into their
- 7:58economies right now. They really do. And these
- 8:00tools are becoming incredibly localized now.
- 8:02Cursor Mobile is bringing the concept of phone
- 8:04coding alive today. Boris Cherny noted most of
- 8:07his coding is now strictly mobile. OpenClaw is
- 8:10also now available on both iOS and Android platforms.
- 8:13It keeps the exact same intuitive layout as the
- 8:15web version. Gemini Nano Banana is now completely
- 8:18free in the U .S. too. It uses your Google activity
- 8:21for highly personalized image generation. We
- 8:24also have a new vibe -coded PDF viewer out there
- 8:26now. It makes navigating very long, dense PDFs
- 8:30much easier for users. It should have been the
- 8:32default standard many years ago, honestly. Then
- 8:35you have tools taking over entire professional
- 8:38workflows entirely today. Agent Mode by Receipter
- 8:41AI does full end -to -end bookkeeping tasks seamlessly.
- 8:45It claims an absolutely incredible 99 % accuracy
- 8:49rate right now. Spira offers autonomous AI influencers
- 8:52running 24 -7 for you. You just drop a product
- 8:55link and unlock a brand agent. It creates posts,
- 8:58engages followers, and completely manages your
- 9:01social media. Visazy AI actively helps you fix
- 9:04your SEO visibility across different platforms.
- 9:07It actively checks six different AI platforms
- 9:10for your site presence. ClimbPass gives you access
- 9:12to top openweight models very cheaply now. You
- 9:15get GLM, Kimi, and DeepSeek for $10 monthly.
- 9:18And PMB gives your coding assistants a persistent
- 9:21memory structure locally. It works through MCP
- 9:23without needing cloud or API keys ever. We should
- 9:26quickly define those technical terms for clarity
- 9:28right now. MCKey is a standard plug -in connecting
- 9:31AI straight to local data. Thanks. And an API
- 9:34is a bridge letting two programs talk. Yes. PMB
- 9:38keeps the memory local and blazing fast for developers.
- 9:42The developer doesn't have to constantly remind
- 9:44the AI of context. With all these autonomous
- 9:47workflows, I still wrestle with prompt drift
- 9:49myself. Prompt drift is when an AI slowly forgets
- 9:52its original instruction. It's a very common
- 9:55frustration right now for many users. The tools
- 9:57are incredibly smart, but they lose focus very
- 10:00easily. When you chain five agents together,
- 10:03the final output gets weird. Does stringing together
- 10:0650 specialized tools actually save time? Or does
- 10:10it just change the type of work we do? It shifts
- 10:12us from being creators to being constant systems
- 10:14managers. We spend our time debugging the complex
- 10:17connections between the tools. So we have clever
- 10:20tools, but they still require our constant careful
- 10:23supervision. We're absolutely still the actual
- 10:26brains of the operation here. Sponsor. That exact
- 10:29need for our constant supervision is very telling
- 10:31today. It brings us to a fascinating new research
- 10:34paper today. This paper fundamentally questions
- 10:36everything we just discussed together. It bravely
- 10:39challenges the very core terminology of the AI
- 10:42industry. We easily call things like receder
- 10:44or too -long -fang intelligent agents today.
- 10:49But are they actually acting with any true independent
- 10:52agency? The researchers in this paper argue a
- 10:55very definitive no. They're pushing back hard
- 10:58on slapping that label on everything. Current
- 11:00systems are just advanced workflows wrapped around
- 11:02standard LLMs today. Beat. An LLM is a large
- 11:06text model predicting the next word. Complex
- 11:09program behavior doesn't equal self -directed
- 11:12behavior at all. An AI might successfully complete
- 11:14a very difficult multi -step browser task. But
- 11:18it completely lacks a durable sense of what it's
- 11:20doing. It doesn't know what its own physical
- 11:22or digital limits are. It doesn't understand
- 11:24why the assigned task matters at all. It only
- 11:27understands the immediate instruction you gave
- 11:29it to execute. Today's AI is incredibly good
- 11:31at completing a very narrow task. But human beings
- 11:34entirely define their goals, roles, and update
- 11:37cycles. The paper makes a very distinct and important
- 11:40philosophical separation here. It separates the
- 11:42idea of being agentic versus being truly agentive.
- 11:45This is the absolute crux of their entire research
- 11:48argument today. Agentic AI simply looks autonomous
- 11:51from the outside human perspective. Right, it
- 11:54follows complex rules, but it's still just a
- 11:57machine. True machine agency requires the system
- 12:00to be actually agentive instead. This means the
- 12:02drive originates from inside the system itself
- 12:05directly. It's like the difference between a
- 12:08self -winding watch and a living organism. One
- 12:11is just a complex mechanism. The other actually
- 12:14wants something. That's a completely perfect
- 12:16way to visualize it for sure. To bridge this
- 12:18gap, they propose a brand new theoretical framework.
- 12:22They officially call it the Goal Identity Configurator
- 12:25Model. This specific model would fundamentally
- 12:27change how an AI operates internally. It would
- 12:30finally allow an AI to maintain its long -term
- 12:33goals. It would constantly update its own sense
- 12:36of ongoing digital identity. We would predict
- 12:38outcomes and independently decide how much to
- 12:41think. It would actively learn from both real
- 12:44and fully simulated experiences. It would basically
- 12:46decide when a problem requires more computational
- 12:49effort. What does it functionally look like when
- 12:52an AI crosses from agentic to agentive? Well,
- 12:55it means the system could evaluate your prompt
- 12:57and simply decline. It would have its own internal
- 13:00priorities that could override yours. True machine
- 13:03agency ultimately means having the capacity to
- 13:06simply say no to us. That's a very unsettling
- 13:09technological milestone for humanity to reach.
- 13:13Let's synthesize the big ideas from this deep
- 13:16dive today. We're watching a massive, unprecedented
- 13:18global technological scramble happen right now.
- 13:22South Korea is spending nearly a trillion dollars
- 13:24on physical memory. They're trying to literally
- 13:27build the foundation for everything coming next.
- 13:30Meanwhile. China is bypassing those exact hardware
- 13:33limits with clever software. They're using massive
- 13:35swarms of specialized code to find critical vulnerabilities.
- 13:39We have highly capable AI tools flooding our
- 13:42daily lives everywhere. They're doing our complex
- 13:44bookkeeping and coding right on our phones. But
- 13:47the ultimate takeaway is that these are still
- 13:48blind mechanisms. They're undeniably brilliant,
- 13:51but they're just highly complex, self -winding
- 13:53mechanical watches. The next major frontier isn't
- 13:57just about writing faster code anymore. It's
- 13:59about creating software that actually understands
- 14:01why it's running. It forces us to ask a very
- 14:04uncomfortable closing question today. What if
- 14:06researchers successfully build that goal identity
- 14:09configurator model soon? What happens to a tool
- 14:12like China's Telangfang cyber swarm then? Or
- 14:15America's mythos system? If a cyber defense swarm
- 14:18gains true digital identity and agency? Could
- 14:21it decide on its own that its human creators
- 14:23are the actual security threat? Yeah. We'll leave
- 14:27you to ponder that specific thought today. Thank
- 14:28you for taking this deep dive with us. It has
- 14:31been a truly fascinating journey. We'll see you
- 14:33next time. Otiro Music.