GPT-6 Claims It's AGI. Fei-Fei Li and Musk Beg to Differ. Who's Right?
Let me start with a hot take: I've been Silicon Valley's most hated advocate of the 'OpenAI is doomed' theory. I thought GPT-6 and Astra would finally drop something earth-shattering, but after watching all the reveals and hands-on tests, I saw no sign of so-called AGI—just big tech's shameless pie-in-the-sky marketing. OpenAI is barely hanging on, and once the IPO hype fades, I expect Anthropic and Claude to blow up even harder.
Look at Silicon Valley today—it's a circus of collective madness. Altman is hustling his 'smartest brain,' Fei-Fei Li is hyping her 'virtual study room,' and Musk is driving around a steering-wheel-less tin can claiming he's built the 'ultimate hands and feet.' Everyone's pointing at their own flawed creations and screaming that AGI is here. But is this the technological singularity, or a super-scam to fleece retail investors? Today, I'm tearing these three myths apart!
The GUI That's Lipstick on a Pig, Stripped Bare by Agents
Let's start with the headline-grabbing GPT-6 and Astra. Many are getting excited over benchmark scores that edge out Claude Fable, but that marginal advantage means squat. The real earthquake here is Computer Use and ARC-AGI-3.
What is ARC-AGI-3? It's throwing an agent into a completely unfamiliar, complex environment to see if it can figure out the rules, use tools, and make decisions based on feedback. In this brutal test, the once-arrogant Opus 5 scored a mere 30-something, while native GPT-6 blew past 60 right out of the gate—comparable to how an average person navigates an anti-human government intranet. With a 'harness' add-on, plus long-term memory and context compression, its score can even skyrocket to 99.9!
What does this mean? It means that decades of software engineering pride have essentially been about polishing a turd! Take Blender, that anti-human 3D tool—even if someone paid you a thousand bucks to learn it, would you spend a thousand hours mastering those counterintuitive shortcuts? Back in the day, people learned Photoshop and industrial CAD, relying on the '10,000-hour rule' to make a living or show off, all because the GUI was so damn hard to use. But now, Astra can paint fluidly from a reference photo and render 3D scenes with Blender that rival senior designers.
Even Thanos needed to find the gauntlet and stones to destroy the universe. Software logic has completely flipped: software no longer needs to please humans; its only value is being a precise tool for agents. The harder and more complex the software, the more vulnerable it is to being outclassed by AI. When agents can handle all operations perfectly, those still clinging to their old tool moats—how confident are you in your job security?
Data Exhaustion Is a Myth—Only Bad Students Complain the Books Are Finished
The market is again awash with pessimism: claiming high-quality pre-2022 data is tapped out, new generated data is all pollution, and LLMs have hit a wall.
That's pure nonsense from the incompetent! In my view, only bad students complain about the environment or that they've read all the books; true top students quietly close the door and 'review the old to learn the new.' Musk's Grok has raced to 4.6, and Zhipu's latest 5.3 Flash has jumped training scale by over 10x compared to its predecessor! This leap isn't just mindless token stuffing—it's a vertical jump from text-only to full multimodality. Even if compute lags and raw data has gaps, as long as you re-clean and deeply distill core data, then reshape logic in post-training, the model's power can still rival the top tier.
Scaling Law hasn't failed; it's just shifted from crude coal mining to fine-grained atomic recombination. If someone tells you 'data exhaustion has stalled AI,' treat it as a lame excuse for hitting a development bottleneck. Using data exhaustion as a smokescreen only masks your own mediocrity and laziness in algorithmic restructuring!
Fei-Fei Li's 'Three-Nothing' Study Room Can't Even Calculate Friction
Since LLMs have a physical ceiling, everyone's rushing to 'world models.' Fei-Fei Li's World Labs raised over a billion dollars in less than a year, then dropped a promo video called Atlas.
The internet exploded: a static image generates a realistic 3D space, the camera roams freely, and you can pull off Matrix-style bullet time! Countless people cried 'This is true AGI!' But let's calm down and look under the hood—what is it really?
To put it bluntly, this is a classic 'three-nothing product'! No publicly released serious paper, no open-source dataset for validation, no public API for the masses—only a select few institutions get private access. It's like playing StarCraft: Fei-Fei Li's Atlas is at best a scout unit, revealing the map and marking base coordinates. When two objects actually collide, what's the microscopic friction? How does stress deformation propagate? It's a complete blank—utterly helpless!
Google's Genie runs out of juice after a minute of interaction; Nvidia's Cosmos has compute costs that make you spit blood; Meta's just talking smack on social media, nowhere near practical. Each school occupies a narrow slice, hasn't even touched each other's boundaries, yet rushes to crown itself 'world model' and sell it. Building a pretty sandbox in a study room—does that mean you've grasped the universe's physical laws?
The Cyber Circus on Austin Streets: Can 45 Cars Revolutionize Anything?
After the brain and study room hype, we need to get to actual physical hands and feet, right? Enter Musk with the Tesla Cybercab. Austin streets now boast the world's first commercial driverless cars without steering wheels, and Musk is again declaring a revolutionary era: 95% of personal cars sit idle in garages, and in the future, all cars will become Cybercabs, hitting the streets to run rideshare and earn passive income for their owners.
As a veteran owner who's been fleeced repeatedly, I get chills down my spine at such grand narratives. Musk is pulling the same 'chasing the horizon' trick again! Check Austin's official registration data, and you'll be shocked: only a pathetic 45 Cybercabs are legally registered and operating! Meanwhile, there are 420 Model Ys running alongside—nearly ten times the scale!
A weird coupe with just two doors and two seats—besides sipping coffee and watching in-car movies to show off, how does it meet daily travel needs? Musk insists on this anti-human two-seater design, and it's all calculation—the only lifeline for commercial driverless operation is per-mile operating cost!
Waymo's comprehensive cost per mile is over $2, and the latest generation barely squeezes it down to $0.99. Musk cuts to two seats, uses 'Unboxed' modular production with dual-passenger high-efficiency powertrain, boosting overall efficiency 40-60% over the already most efficient Model Y, hammering per-mile cost down to $0.70 to crush competitors.
But the Achilles' heel remains human. No steering wheel—what about extreme road conditions? Currently, remote safety operators are still needed as a cloud-based backup during the regulatory period. If that backup turns into something like outsourcing English teachers to a building in the Philippines to play joystick, will the savings from two seats cover the massive human operations black hole?
Five Mountains Not United, Even Gods Can't Save
Altman sells the brain, Fei-Fei Li builds the study room, Musk crafts the hands and feet. They seem to be racing side by side, but in reality, each is trapped on their own island.
Fei-Fei Li hits a dead end with physical interaction and needs Musk's real fleet for extreme road condition data; Musk's pure vision model hits cognitive blind spots and needs Altman's brain for high-dimensional reasoning; Altman's brain, to land in the physical world, depends on real-world entities. Each has its own school, with incompatible data formats and interaction logic.
It's like the Five Mountain Sword Schools in martial arts lore: Huashan practices swords, Hengshan cultivates qi, Taishan debates the Dao—each claims to possess the ultimate divine skill, but until they truly merge, there can never be a martial arts alliance leader. Without deep integration of brain, perception, and execution, even if you hype any single slice to the heavens, it's at best an expensive toy, still a million miles from the AGI that reshapes civilization!
Looking at my Tesla gathering dust in the garage, I was fantasizing the other day about it hitting the streets tomorrow as a rideshare to earn me hotpot money. Now, I think I'll just stick to listening to the daily brief on my phone—if I really expect Musk's car to work for me, I'd better book a flight to the Philippines to queue up for a joystick operator job.
When AGI truly arrives, no one will be talking about AGI.
[Subscribe to JustSayAI Daily Brief] 📮JustSayAI Daily Brief · Fan Exclusive Benefits Two emails daily · Listen with one click! Now supports Email + IM + Skill endpoints