Latest / AI Fire Daily / 🎙️ EP 287: "Too Dangerous" Claude Fable 5 & The New "MANGOS" Tech Era
Transcript
- 0:00Right now, AI models can actually boot up and
- 0:03beat Pokemon autonomously. Yeah, it totally sounds
- 0:06like a really entertaining, harmless party trick.
- 0:09But then you read today's MIT Future Tech Research
- 0:11Report. Right. Those experts warn this exact
- 0:14level of autonomy brings severe dangers. They
- 0:18see a 10 % chance of triggering global societal
- 0:21catastrophe. I mean, we're talking about $100
- 0:24billion damages by 2030. Beat! Welcome to today's
- 0:29deep dive. I am incredibly glad you're joining
- 0:32us. We truly have a lot of Complex, fascinating
- 0:35ground to cover today. So we are unpacking an
- 0:38avalanche of massive AI capability jumps right
- 0:41now. Anthropic just dropped their highly anticipated
- 0:43Claude Fable 5 model today. And we really need
- 0:46to examine the structural collision course this
- 0:48creates. Yeah, these blistering capabilities
- 0:50are hitting global security and hardware supply
- 0:52chains. They're also crashing right into our
- 0:55necessary safety guardrails. So let's start with
- 0:57the sheer raw power hitting the market right
- 1:00now. The newest models arriving today are just
- 1:03absolutely wild. Anthropic released Claude Fable
- 1:065 directly to the public recently. It is the
- 1:09first major model from their new mythos family.
- 1:13Fable 5 is absolutely crushing the benchmark
- 1:15charts across the board. It shows truly exceptional
- 1:18results in complex knowledge and vision tasks.
- 1:21It is kind of like stacking Lego blocks of data.
- 1:24That is a perfect way to visualize its multimodal
- 1:26architecture. You just snap a visual screenshot
- 1:29onto a basic text prompt. And the AI just builds
- 1:32the entire functional code right from there.
- 1:34It does not see images and text as separate languages.
- 1:37It blends them together into a unified understanding
- 1:40of the problem. Which perfectly explains how
- 1:42it can rebuild entire web apps from scratch.
- 1:45Yeah, and it also explains the Pokemon Fire Red
- 1:47achievement. The AI completed the entire game
- 1:50without any human intervention at all. Just think
- 1:53about this. deep long term planning required
- 1:55for that. You have to manage inventory and navigate
- 1:57highly complex spatial maps. The agent handles
- 2:00all of those sequential decisions completely
- 2:03autonomously. It proves these agents can handle
- 2:06long term enterprise software workflows now.
- 2:08But this incredible autonomous capability definitely
- 2:12comes with a steep price. It costs $10 per million
- 2:16input tokens right now. Tokens are just the fundamental
- 2:18puzzle pieces of language that AI processes.
- 2:21And output tokens will actually cost you $50
- 2:24per million. That is exactly double the cost
- 2:26of their previous Opus model. Opus 4 .8 was already
- 2:30considered a fairly pricey option. Startup founders
- 2:32building on Fable 5 will burn cash much faster.
- 2:36They really have to route simple tasks to much
- 2:38cheaper models. Fable 5 is currently available
- 2:41via their developer API system. API is a software
- 2:44bridge letting different computer programs talk.
- 2:47If you maintain a Pro, Max, or Team subscription,
- 2:49though, you actually get it included completely
- 2:52free until June 22nd. That gives power users
- 2:55a fantastic window to test the capabilities.
- 2:58And this rapid capability leap is mirrored across
- 3:00the whole ecosystem. We are fully entering the
- 3:03agentic era of artificial intelligence. Look
- 3:05at this fascinating new tool called KimiWork,
- 3:08for example. It deploys up to 300 local AI agents
- 3:12simultaneously. Local agents are AI programs
- 3:15running directly on your personal hardware. They
- 3:20automate complex web tasks and analyze dense
- 3:25financial data sets. Then we have another tool
- 3:31called Macatron, making major waves today. It
- 3:36controls native applications using very simple
- 3:42conversational English commands. We're also seeing
- 3:45Google aggressively push Gemini 3 .5. Their live
- 3:49translate feature handles real -time conversational
- 3:51translation. Almost perfectly now. You can translate
- 3:54complex, nuanced conversations exactly as they
- 3:58happen globally. The business scale behind these
- 4:00tools is getting incredibly hard to fathom. A
- 4:02company called Lovable just crossed $500 million
- 4:05in annualized revenue. They have seen over 50
- 4:08million user projects created already. And their
- 4:11usage is accelerating to 1 million new projects
- 4:13weekly. Whoa, imagine scaling to a billion queries.
- 4:17It completely breaks my brain just trying to
- 4:19comprehend that volume. It is a totally un -
- 4:22unprecedented level of software creation speed.
- 4:25You no longer need to hire an entire team of
- 4:27software engineers. You just describe what you
- 4:29want in simple, plain English. And Lovable spins
- 4:32up the entire working application for you instantly.
- 4:35Two sec silence. It is genuinely magical to watch.
- 4:39With Fable 5 dominating, where does that leave
- 4:41Claude Opus for a user trying to get work done?
- 4:45We actually heard directly from Tariq on the
- 4:47Claude Code team about this. Oh, wow. What did
- 4:49he specifically say about Opus? Well, he says
- 4:52Opus remains the absolute strongest model for
- 4:54extended work. So it is better for those really
- 4:56long, complex tasks. Right. Opus still handles
- 5:00multi -day, sustained, autonomous execution much
- 5:03better today. So Opus remains the marathon runner
- 5:05while Fable sprints through complex reasoning.
- 5:08Exactly. You always need to pick the right tool
- 5:10for the job. Which logically brings us to the
- 5:12physical reality underneath all this software.
- 5:14You cannot magically scale to a million new coding
- 5:17projects weekly. It takes immense hardware, massive
- 5:21funding, and incredible physical infrastructure
- 5:23networks. And we are seeing massive structural
- 5:27shifts in the financial markets. The old FAA
- 5:30acronym is completely dead at this point. Financial
- 5:32analysts are heavily using the new acronym MANGOES
- 5:36instead today. That stands for Meta, Anthropic,
- 5:39NVIDIA, Google, OpenAI, and SpaceX. It perfectly
- 5:43captures the massive financial shift toward artificial
- 5:46intelligence dominance. And the sheer volume
- 5:48of money flowing into these companies is staggering.
- 5:50A startup called Moonshot AI is currently raising
- 5:54$2 billion. That represents a massive 50 % jump
- 5:58in their overall valuation. They were valued
- 6:00at $20 billion just one short month. It is a
- 6:03level of rapid capital accumulation we rarely
- 6:06ever see. But all that investment money has to
- 6:08buy actual physical hardware. They need these
- 6:14highly specific AI processing chips for 2028.
- 6:17And there is a massive shift happening in global
- 6:20chip manufacturing. Google and Nvidia are reportedly
- 6:23turning to Intel foundries right now. They are
- 6:26actively using Intel as a critical backup to
- 6:28TSMC. TSMC has historically dominated the entire
- 6:32global semiconductor chip manufacturing market.
- 6:36Why are Google and Nvidia suddenly looking at
- 6:39Intel for their 2028 chips instead of just sticking
- 6:42with TSMC? It largely comes down to intense geopolitical
- 6:46realities and sheer volume. Meaning they just
- 6:49cannot rely on one single factory anymore. Exactly.
- 6:52Relying entirely on a single foundry in Taiwan
- 6:55is incredibly risky. Right. Diversifying away
- 6:57from TSMC is basically a massive supply chain
- 7:00insurance policy. That is exactly what it is
- 7:02for these massive tech giants. Building a semiconductor
- 7:05foundry. It takes many billions of dollars in
- 7:07years. You absolutely cannot just spin up a new
- 7:10chip factory overnight. It requires massive clean
- 7:12rooms and incredibly complex photolithography
- 7:15machines today. But Taiwan sits in an incredibly
- 7:18precarious geopolitical hotspot right now. If
- 7:21those supply chains get disrupted, the entire
- 7:23AI industry halts completely. Google and Nvidia
- 7:26absolutely cannot afford that massive level of
- 7:28existential risk. So they are pouring massive
- 7:30resources into Intel's new domestic foundries.
- 7:33But the infrastructure... friction is not just
- 7:36about the physical semiconductor chips. We also
- 7:38have to deeply examine the massive software security
- 7:41vulnerabilities. Microsoft just had to surgically
- 7:44remove over 70 compromised GitHub repositories.
- 7:48Hackers actively slipped highly dangerous malware
- 7:50into these specific coding projects. These compromised
- 7:53projects were tied directly to automated AI coding
- 7:56tools. They specifically targeted automated agent
- 7:59platforms like Cloud Code and Gemini CLI. It
- 8:02shows how deeply vulnerable these automated development
- 8:05environments truly are. When AI agents write
- 8:07code, they often pull from open source libraries.
- 8:11If a hacker poisons that library, the AI blindly
- 8:13pulls the malware. And human developers are trusting
- 8:17these AI tools a bit too blindly. It is a massive
- 8:20structural security flaw in the modern coding
- 8:22ecosystem. This severe friction extends to the
- 8:25actual raw coding capabilities as well. A company
- 8:28called Cognition just released their new frontier
- 8:30code benchmark system. The absolute top frontier
- 8:33models currently score just 13 out of 100. That
- 8:36explicitly shows real -world enterprise coding
- 8:39still hits massive walls of friction. I still
- 8:42wrestle with prompt drift myself. Yeah, prompt
- 8:45drift is when an AI gradually loses focus on
- 8:47your original instructions. You ask the agent
- 8:49to fix a specific tiny front -end button issue.
- 8:52And suddenly it tries to rewrite your entire
- 8:54back -end database architecture instead. We are
- 8:57still very far from having completely autonomous
- 8:59software engineers. The agentic era is clearly
- 9:02here, but it is deeply. deeply messy. We have
- 9:05brilliant flashes of magic mixed with fundamental
- 9:08structural failures. It creates a very precarious
- 9:11foundation for our critical digital infrastructure
- 9:13today. And the mounting geopolitical tensions
- 9:16over chips just adds significantly more pressure.
- 9:19We are moving incredibly fast on deeply fragile
- 9:21underlying foundation systems. Which brings us
- 9:24to the biggest structural problem we collectively
- 9:27face. The catastrophic risks of this technology
- 9:30scaling without proper safety guardrails. Be
- 9:33sponsored. Before the break, we analyzed intense
- 9:36geopolitical hardware chip supply tensions. And
- 9:40we looked at those massive automated GitHub malware
- 9:43security breaches. These structural software
- 9:46issues lead directly into the broader macro risk
- 9:49landscape. We have blistering capability growth
- 9:52actively colliding with terrible structural dangers.
- 9:56MIT FutureTech just released a deeply unsettling
- 9:59new AI blueprint today. They comprehensively
- 10:02surveyed 272 leading international AI experts.
- 10:06The detailed findings are honestly enough to
- 10:08keep you awake tonight. They carefully analyzed
- 10:1024 major AI risks currently threatening our modern
- 10:15society. Experts critically believe 18 of those
- 10:17risks have catastrophic outcome potential. They
- 10:20assign a greater than 10 % chance to those catastrophic
- 10:23events. That explicitly means... hundred billion
- 10:26dollar damages or massive global societal disruption.
- 10:29And they predict these massive structural failures
- 10:31could easily happen by 2030. That is only a few
- 10:34short, highly volatile years away from us today.
- 10:37The top catastrophic risks they identified are
- 10:39incredibly specific and deeply concerning. They
- 10:42deeply worry about dangerous autonomous capabilities
- 10:44emerging in these frontier models. They highlight
- 10:48severe geopolitical competitive dynamics aggressively
- 10:51accelerating the global arms race. And they specifically
- 10:54warn about highly weaponized cyber attack capabilities
- 10:58rapidly expanding globally. Which links perfectly
- 11:01back to that recent Microsoft GitHub malware
- 11:03hack. The software vulnerabilities are already
- 11:06being actively exploited in the wild today. But
- 11:09the MIT study also revealed a massive, terrifying
- 11:12societal vulnerability gap. The people most deeply
- 11:16vulnerable to these catastrophic failures are
- 11:18everyday users. The general public absorbs the
- 11:21massive financial damage when vital... systems
- 11:23break. But the general purpose AI developers
- 11:26are the ones structurally responsible today.
- 11:28Government regulators also hold massive inherent
- 11:31responsibility for preventing these specific
- 11:33catastrophes. The critical crosshairs are currently
- 11:35aimed at information, finance and national security.
- 11:39Think about autonomous AI getting deeply embedded
- 11:42in critical financial decision making systems.
- 11:44If these AI tools get embedded in the financial
- 11:47sector and fail, who is actually left holding
- 11:49the bag? Well, the MIT study clearly shows the
- 11:52general public takes the massive hit. Wait, really?
- 11:54The public takes the hit? Yeah. While developers
- 11:57and regulators are responsible, they lack financial
- 12:00incentives to slow down. So every day people
- 12:03suffer. But tech developers and government regulators
- 12:06hold the actual responsibility. That is the tragic,
- 12:10undeniable reality of this massive, glaring accountability
- 12:14gap. The underlying structural dilemma here is
- 12:17incredibly difficult to solve quickly. Any tech
- 12:19company that deliberately slows down for safety
- 12:22takes a massive hit. They immediately lose massive
- 12:25competitive ground in a violently red hot market.
- 12:28Absent strict external government rules, slowing
- 12:30down is essentially financial suicide today.
- 12:33Developers have deep structural and financial
- 12:35reasons not to prioritize vital risk mitigation.
- 12:38This perfectly explains Anthropic's highly specific
- 12:41approach to the recent Fable release. Ampropic
- 12:44purposefully put a massive safety guardrail system
- 12:46on Claude Fable 5. They have loudly warned everyone
- 12:49about severe recursive self -improvement dangers
- 12:52very recently. An AI system continuously making
- 12:55itself smarter without human help. So if you
- 12:57prompt Fable 5 with highly sensitive or dangerous
- 13:00topics. The model will actually actively hard
- 13:03block your specific request right away. It automatically
- 13:06routes your session straight back to the safer
- 13:08Opus model. To legally use Fable 5, you also
- 13:11face a mandatory data retention policy. You must
- 13:15formally agree to a strict 30 -day data retention
- 13:18window. Anthropic explicitly promises not to
- 13:22train future models on that retained user data.
- 13:24But they absolutely need the logs to actively
- 13:27monitor for dangerous emergent behavior. Meanwhile,
- 13:31the completely unrestricted, deeply powerful
- 13:33version is called Claude Mythos 5. But Anthropic
- 13:36intentionally locked that specific model away
- 13:38behind heavily closed doors. It is currently
- 13:41only available for highly approved, heavily vetted
- 13:44security partners now. They are actively taking
- 13:46a massive, deliberate financial hit to ensure
- 13:49global safety. They essentially built the fastest
- 13:51car but permanently locked the top gear. Because
- 13:54they clearly know the regulatory brakes on the
- 13:56broader industry are failing. It is a fascinating,
- 13:59rare, real -world example of prioritizing structural
- 14:02safety first. But they are largely acting completely
- 14:05alone in this highly specific, cautious approach.
- 14:08Most corporate competitors are just blindly racing
- 14:11toward maximum capability and profit. And Fropic
- 14:14has always actively branded themselves as the
- 14:16premier safety -first AI lab. They famously split
- 14:20off from OpenAI years ago over fundamental core
- 14:23safety disagreements. So releasing a highly capable
- 14:26autonomous model like Fable 5 is deeply risky.
- 14:30They truly had to thread a nearly impossible
- 14:32corporate and ethical needle here. The mandatory
- 14:3530 -day enterprise data retention policy is a
- 14:38truly perfect example. Most massive enterprise
- 14:40users absolutely hate corporate data retention
- 14:43policies today. They desperately want complete
- 14:45privacy and zero data logging for internal...
- 14:48corporate security. But Anthropic stubbornly
- 14:50insists on keeping those specific user logs for
- 14:5330 days. They need to actively monitor for extremely
- 14:56dangerous emergent capabilities developing online.
- 14:59They are specifically looking for terrifying
- 15:01signs of biological weapons research queries
- 15:03happening. Or highly automated, deeply destructive
- 15:06cyber warfare scripts being actively generated
- 15:09by their models. Think about the sheer terrifying
- 15:11scale of $100 billion in damages. That isn't
- 15:15just a tiny, easily ignored blip in the global
- 15:18financial market. That represents massive global
- 15:20job losses and entirely decimated corporate pension
- 15:24funds. And a 10 % chance is practically playing
- 15:27Russian roulette with the economy. You would
- 15:29never voluntarily board an airplane with a 10
- 15:31% crash rate. But we are currently strapping
- 15:34our entire global economy to this exact rocket.
- 15:37And the ambitious people building the massive
- 15:39rocket currently have no real brakes. Two secs
- 15:42silence. If we logically pull back and look at
- 15:45the broader big picture today. We are living
- 15:47in a deeply profound jarring split screen reality
- 15:50right now. On one side we have. deeply magical,
- 15:54fully autonomous software coding agents. They
- 15:57are completely rebuilding complex web applications
- 16:00entirely from visual scratch today. We have a
- 16:02violently red -hot mangoes economy that is basically
- 16:06printing infinite money. But on the other darker
- 16:08side, we face a deeply terrifying structural
- 16:11gap. We have a massive, currently unaddressed
- 16:14structural safety and global accountability gap.
- 16:17Actively slowing down to implement crucial safety
- 16:19guardrails takes immense corporate courage today.
- 16:22Anthropic, courageously restricted Mythos 5,
- 16:25but willingly took a massive financial hit. This
- 16:28dangerous competitive dynamic leaves our most
- 16:31critical financial and national security sectors
- 16:34vulnerable. Completely vulnerable to massive,
- 16:37highly catastrophic global risks by the year
- 16:402030. The fundamental corporate incentive structures
- 16:43are completely misaligned with long -term human
- 16:46survival today. Which ultimately leaves you with
- 16:48a deeply unsettling, highly provocative question
- 16:51to ponder. Taking a severe competitive financial
- 16:54hit is currently the only reward for safety.
- 16:57And external federal government regulations are
- 16:59historically moving far too slowly today. Will
- 17:02it necessarily take the first $100 billion global
- 17:04financial market crash? Or perhaps a massive,
- 17:07deeply weaponized global political disinformation
- 17:10event soon? To finally force the entire global
- 17:13AI industry's brake pedal straight to the floor.
- 17:16It is a truly terrifying, sobering thought to
- 17:19leave hanging in the air. Thank you so much for
- 17:21taking this fascinating deep dive with us today.
- 17:24We always deeply appreciate you thoughtfully
- 17:26exploring these complex technical topics with
- 17:28us. Stay intensely curious, and we will definitely
- 17:32see you on the next one.