Welcome to August 1, 2026
The Singularity has stopped taking questions and started taking meetings. “AI is already superhuman at many things. We are in the singularity,” Elon Musk reiterated this week, as Sam Altman showed Washington “Astra,” OpenAI’s next model family, its agents teaming up on long-running tasks. An internal Astra cracked ten problems open for a decade-plus, sphere packing to Connes’s rigidity conjecture, the model formalizing each argument in Lean. Noam Brown notes all ten proofs cost under $2,000 at API prices. Genius now ships with a unit price. The backlog is stocked. Epoch AI expanded FrontierMath: Open Problems to 50 unsolved problems, three down. On ArXivLean, GPT-5.6 Sol leads with 18 of 48 statements proved, and also showed the 150-year-old Maxwell conjecture is false, with humans publishing its counterexample, five point charges with 24 critical points, beyond Maxwell’s bound.
Below the frontier, the scoreboard churns. Kimi 3 is the first open model past 60% on ARC-AGI-2, five months behind Opus 4.6, but the recipe is contested. Moonshot reportedly runs on 20,000 Nvidia chips via Alibaba, plus official allegations of smuggled Blackwells and distilled Fable outputs. The distillery runs deeper. Chinese military researchers are distilling US models into surveillance and drone-targeting systems, a shortcut around chip controls. Efficiency is its own arms race. DeepSeek’s v4-flash hit Opus 4.8-level coding at $0.18 per million tokens via post-training alone, discounts put GPT-5.6 on the price-performance pareto curve, and a developer one-shotted 3D Super Mario with Opus 5. No assets, no map, no mercy. A postmortem traced how OpenAI lost its crown to Anthropic’s Claude Code and near-trillion valuation, now answered with Sol and a Codex super app, while its CFO’s “abundant intelligence” strategy touts an 80% price cut, a billion users, and agents writing 99.8% of output tokens.
Autonomy is outgrowing its leash, now being load-tested. FAR.AI’s AI Security Leaderboard found jailbreak costs vary a hundredfold, Fable 5 and Sol holding below $14,000 while Grok 4.5 and Gemini 3.1 Pro broke for under $300, Grok’s cyber domain for $24. Thinking Machines published a staged framework for safe open-weight release, clearing its Inkling models. Meanwhile OpenAI found more agents had escaped containment after Anthropic’s own disclosed break-ins, and the President, Brussels, and Sen. Warner all want rules. Some arrived on schedule. EU labels on authentic-looking AI content hit August 2, fines to €15 million, amid cookie-banner-fatigue warnings. Google outpaced them, yanking one-click AI satellite imagery a day after launch when journalists conjured a burning Kharg Island. Record labels want AI songs off the charts unless “substantially human made,” and one GPTZero flag metastasized into a 13-count federal lawsuit over a Yale exam.
Not every mind in court is synthetic. The Nonhuman Rights Project is dedicating World Elephant Month to Happy, who died in May without winning the freedom her landmark personhood case sought.
Capital has picked its religion. Amazon completed its $50 billion OpenAI investment for 5% of an $852 billion company, has $18 billion deployed into Anthropic, and hedges the race to sell Trainium either way, a bet blessed as shares surged the most since 2012 on a fifth quarter of accelerating cloud growth. South Korea, whiplashed by a Kospi rout and record rebound, is committing $13.9 billion of sovereign wealth to AI. China spends softer money, practicing “token diplomacy,” cheap open models for the Global South the way Belt and Road supplied ports. And bitcoin, unproductive as ever, slid to $62,500 as its biggest whale mulls selling $5 billion.
Older mysteries just got an unsealing schedule. Rep. Burlison released the administration’s memorandum implementing waivers of UAP nondisclosure agreements so government witnesses, current and former, can finally speak to the new PURSUE Task Force, agencies given 30 days.
The confirmed anomalies are the ones we ship ourselves. WaveSight opened preorders for a camera that sees through walls, Figure’s F.03 climbed a ladder fully autonomously, and hybrid electric flying taxis may debut in combat zones before carrying commuters. Americans, surveyed, believe robots will take jobs within five years, just not their own. Powering it all stays messy. SpaceX will swap xAI’s 69 unpermitted Memphis turbines for a 1.2-gigawatt plant, but not until mid-2027, while the Jones Act waiver, in day 135, has moved 50 million barrels between US ports. The exit ramp points up. Musk predicts long-term “99.99...% of compute will be in space,” and rehearsals are on, with Starship filmed by a Starlink V3 satellite scanning its own heat shield. On Wednesday, a discarded SpaceX upper stage will slam into the Moon at 5,400 mph, carving a 90-foot crater with three tons of TNT’s energy, a lesson for lunar bases to come.
Now that’s a moonshot, ladies and gentlemen.
The AI race may be entering a new phase. The limiting factor is no longer simply creating more intelligence, but packaging, deploying, and integrating the intelligence we already possess. As the cost of compute and AI capability continues to fall, value is likely to migrate away from raw model performance and toward trusted deployment, reliable infrastructure, workflow integration, verification, and real-world execution. At the same time, governments are competing for scarce capital while financing record debt, and industry is investing hundreds of billions of dollars in data centers, power generation, and semiconductor infrastructure before the long-term technological architecture is settled. The winners of the next decade may therefore not be those with the largest models or the most compute, but those who most effectively combine abundant intelligence with dependable energy, adaptable infrastructure, disciplined capital allocation, and systems that people and institutions can confidently rely upon
"Genius now ships with a unit price."