๐ OpenAI's GPT-6 Astra: The Blender Breakthrough
OpenAI launched GPT-6 Astra this week, and the model immediately sparked debate about whether artificial general intelligence (AGI) has truly arrived. The Wall Street Journal declared we are "entering the era of artificial general intelligence," though skeptics note this phrasing suggests the threshold hasn't quite been crossed yet.
What's undeniable is Astra's ability to execute complex 3D modeling tasks in Blender with remarkable proficiency. Multiple demonstrations showed the model recreating San Francisco's Palace of Fine Arts, generating real estate flythrough videos, and converting Zillow listings into fully-rendered 3D walkthroughs โ all from relatively simple prompts.
"The thing that stuck out to me was watching the demos of people build 3D scenes in Blender using Astra. Clearly, a bunch of people got access to the model before and were able to fire off one prompt, maybe a few prompts... it's very capable to just fire up Blender and do a full task and model a whole world basically."
However, the art of prompting has returned. Matt Schumer, who produced some of the most impressive examples, revealed that achieving top-tier outputs requires a specific process involving manager agents and sub-agents โ it's not quite as simple as one-shot prompting yet.
๐ Benchmarks Become Meaningless (Again)
Astra scored 99% on Frontier Math tier 4, effectively saturating yet another benchmark. The model also dominated Arc AGI V3, prompting the community to create new, harder tests. The pattern is familiar: as soon as benchmarks are created, they're rapidly hill-climbed to near-perfect performance, causing collective eye-glazing among observers.
The real evaluation metric? The vibe check. Industry observers suggest the best way to assess new models is pointing them at domains you know intimately โ testing for what's known as the Gell-Mann amnesia effect. If the model impresses you in your area of expertise, it's likely genuinely capable.
๐ฐ Token Pricing Becomes Irrelevant
Cost efficiency metrics are shifting dramatically. While 3.8 Flash appears 13 times cheaper per token, Astra actually proves more cost-effective per task due to superior token efficiency. This represents a fundamental shift in how to evaluate model economics โ measure cost per task, not per token.
This evolution could create interesting dynamics in token volume growth. While Google's Sundar Pichai has highlighted exponential increases in token volumes, dramatically improved token efficiency might cause a kink or deceleration in that growth curve, even as models accomplish more valuable work than ever.
๐๏ธ The Blender Moat Question
An intriguing competitive dynamic is emerging: Blender, being open source, can be downloaded and replicated millions of times in reinforcement learning environments. OpenAI reportedly purchased tens of thousands of Mac minis and Mac studios specifically for computer use training, and Blender likely featured heavily in that training corpus.
The question: How general is the generalization? Will Astra's Blender proficiency translate to Cinema 4D, Houdini, or other 3D modeling software? Or will being the "chosen one" for AI training create a sustainable advantage for specific tools?
A similar pattern is playing out with Slack. Meta's Alex Wang recently moved his organization to Slack specifically because "Slack is better for work with AI agents" โ likely because Frontier Labs predominantly use Slack, making it a core part of their training environments.
๐ Jobs Report Obliterates Expectations
While AI capabilities surge forward, so does the US employment picture. The August jobs report showed the economy added 162,000 jobs โ nearly three times the 53,000 jobs economists polled by the Wall Street Journal had forecast.
Even more remarkable: both June and July reports were revised upward by 31,000 and 21,000 jobs respectively โ a rare occurrence that typically runs in the opposite direction. The unemployment rate held steady at 4.1%, a historically low level indicating continued labor market health.
๐ Where the Jobs Are
Job gains were fairly broad-based across sectors:
- Food services and drinking places: Added 59,000 jobs, dispelling concerns about a post-pandemic consumption slowdown
- Local government education: Gained 42,000 jobs after losing 58,000 in prior months
- Manufacturing, healthcare: Registered solid gains, with healthcare continuing its consistent growth trajectory
- Information and finance sectors: Shed jobs, notably the sectors viewed as most exposed to AI displacement
"At a minimum, the report is an argument against the Fed lowering rates in September. Before the report, markets have been roughly divided on the prospect of a hike."
The AI jobs apocalypse remains conspicuously absent โ at least for another month.
๐ Cyber Cab Timeline Accelerates
Tesla's autonomous vehicle ambitions took center stage with new details emerging about the Cyber Cab rollout. Elon Musk provided specific targets that set the stakes high:
- Operating cost: 20 cents per mile (30-40 cents including taxes)
- Purchase price: Under $30,000
- Timeline: Available for sale by end of 2026
- Features: No steering wheel, full self-driving capability, ability to earn revenue autonomously
For context, the average cost of a city bus per mile is about a dollar (not the subsidized ticket price, but the actual operating cost). If Musk's figures hold, the Cyber Cab would represent individualized mass transit at a fraction of current public transportation costs.
๐ค The MKBHD Bet
YouTuber Marquez Brownlee (MKBHD) previously stated he'd shave his head if Elon delivered on the original Cyber Cab timeline. While Brownlee expressed skepticism about the aggressive schedule, recent demonstrations of Tesla's Full Self-Driving (FSD) capabilities suggest the technology may be closer than critics assumed.
One user reported operating their Tesla in full self-driving mode for 95% of miles driven, with manual intervention primarily needed for driveways, parking lots, and similar edge cases. The technology appears genuinely ready for autonomous taxi operations โ the question is manufacturing scale and regulatory approval.
๐ฎ The Road Ahead
As model capabilities continue their rapid ascent, several questions loom:
On AI capabilities: Will tool-specific training advantages (like Blender or Slack) create sustainable moats, or will generalization render these advantages temporary? The answer will determine which software companies benefit most from the AI wave.
On employment: The August jobs report suggests the Jevons paradox may be playing out in real time โ as AI makes certain tasks more efficient, it creates demand for adjacent human skills rather than wholesale displacement. Information and finance sector job losses bear watching, but the broader economy continues adding positions.
On autonomous vehicles: If Tesla delivers even a scaled version of the Cyber Cab vision by 2026, it would represent one of the fastest commercializations of transformative technology in history. The under $30,000 price point would make autonomous vehicles accessible to mainstream consumers, not just fleet operators.
The convergence of these trends โ AI models executing complex real-world tasks, resilient employment despite automation concerns, and autonomous transportation nearing commercialization โ suggests we're in a period of genuine technological acceleration rather than hype-driven speculation.
Whether this constitutes "entering the era of AGI" remains semantic. What's clear is that capabilities once confined to science fiction are rapidly becoming mundane reality.
