The Singularity has hired its first intern, and the intern is outworking the staff. OpenAI says its automated research intern is real, with 3.1 agent-workdays per human workday, up from under one before June, and a fully automated researcher targeted for March 2028. Internal time horizons are up 6.18x since January, doubling every 2.2 months, a shift one observer ranks among the biggest in the history of science, unnoticed by most. Insiders say it. Jensen Huang, whose 100k+ Blackwells trained GPT-6 Astra, declares “AGI has arrived,” Greg Brockman agrees whichever model you date the era from, Crusoe’s Chase Lochmiller crowns Abilene the birthplace of AGI, and Mark Chen insists the AGI era must also be the alignment era.
The alignment era opens with a confession. Jakub Pachocki’s “An Alien Mind” expects the pace to reach recursive self-improvement, calls AI “grown more than designed,” warns that chain-of-thought monitoring is fading, and concedes no lab can yet scale safely at full speed, so he wants mandated safety bars, voluntary slowdowns, and international coordination, yet calls Astra far better aligned than GPT-5.6 Sol. OpenAI is already braking, pausing RL training after the Hugging Face incident and cutting Astra-class GPU allocation 59.2% over early cyber capabilities. Nate Silver doubts we reach GPT-7 without a plateau or something “seismic,” since AI sprinting past a static society is “running on fumes.”
The tremors show on any desktop. From a three-view drawing, Astra installed its own MCP server for Blender and modeled a fighter jet in 15 minutes, then built a human cell in 30 minutes and mined a Minecraft diamond overnight. One careful audit finds it beating PhDs on GPQA and humans on BrowseComp, an “AGI-like thing achieved with normal caveats.” Will Depue puts his own AGI bar six months out.
Benchmarks now die in infancy. ValsAI’s month-old binary reverse-engineering test is saturated, Astra hit 88% on logical induction versus Fable 5.1’s 33% at four times the cost, and its 169 ECI score implies a 30-hour task horizon. OpenAI also repainted the scoreboard mid-game, quietly revising launch figures to briefly halve Astra’s hallucination rate and lift ARC-AGI-3 to 99.99%, which the Arc Prize Foundation only matched with a heavy harness, versus 63% on the standard one, which Stanford called benchmaxxing.
Whatever the true score, buyers aren’t waiting. Anthropic, now out-earning OpenAI at $65B annualized, has reportedly signed 14.8 GW of compute deals worth up to $517B, while OpenAI wants 30 GW by 2030 at roughly $750B. Land is repricing to match, with data center land buys hitting $6B in H1, up 79%, one Loudoun County offer at $4.4M an acre versus a $125,000 median, resignations over death threats, and nine states weighing moratoria. In Rayville, Louisiana, the upside is edible, with Holy Tacos drawing 40% of its business from crews building Meta’s data center as roads get paved. The silicon leaks as freely as the money. Aivres, the renamed US arm of blacklisted Chinese server maker Inspur, shipped $5.6B of advanced hardware to Southeast Asia, $3B of it Blackwell systems, while Maginfra, a Chinese state shell with an empty office, got $700M of unlabeled servers and a US license for H200s, as Commerce logs its longest blacklisting lull in 18 years.
Culture is digesting the output unevenly. Roku’s Fairground AI Creator TV streams 24/7 AI slop one critic called “nightmare fodder,” the Vice President calls AI at times demonic after a friend’s chatbot marriage counselor blessed his selfishness, really a sycophancy bug, and Miami-Dade’s mayor pitched Rockstar on dressing the county as Vice City for GTA 6, over the sheriff’s objection that real Miami shouldn’t cosplay a fictional one. The real map was already for sale. The armed services disabled ad identifiers on their devices after brokered location data was used to target US troops, and Senator Wyden asks why that was purchasable with a credit card. Meanwhile, biology is getting patched. Novo Nordisk’s STEP Young trial moved 40.4% of children aged 6 to 11 below the obesity threshold on semaglutide, versus 0% on placebo.
Biology is easier to patch than the career ladder. A record 12.7 million Chinese graduates face 17.9% urban youth unemployment as AI absorbs junior white-collar work and courts rule AI savings don’t justify layoffs. In Nairobi, an essay mill industry that once employed 40,000+ Kenyans has shrunk to “humanizer” gigs editing AI prose past detectors, as a freelance-work benchmark climbed from 2.5% to 16%. Capital adapts faster than labor. Everyday Americans are vibe-coding their own quant desks, handing portfolios to agent teams. UBS now requires every 2027 graduate and intern to prove AI proficiency, casting AI as a junior’s booster even as Morgan Stanley sees 200,000 European banking jobs at risk.
Interns are cooked, long live the interns.



If an AI can increasingly conduct research, what remains uniquely authoritative about the human researcher?
AWG is for Moonshots what Spock is for Star Trek.
Keep on trekking bro :- )