Welcome to July 3, 2026

The Singularity has learned to grade itself, and it keeps beating its own curve. ByteDance’s EdgeBench ran 134 tasks of twelve-plus hours each to measure what agents learn from environments rather than recall, and the noisy learning curves collapsed into a clean log-sigmoid law with learning speed doubling every three months. The UK’s safety institute counters that our rulers are too short: capability is a curve over test-time compute, and one model’s cyber time horizon stretched from two hours to fourteen as its budget rose from 2.5M to 50M tokens. The gains are also getting cheaper to install, since training a single mid-stack transformer layer can match or beat full-parameter RL. Small wonder the crown never sits still, with 17 models taking the lead since Claude 3 Opus dethroned GPT-4, each reigning a median seven weeks. The latest flex is architectural, as Claude Fable 5 wrote KernelBench-Mega’s first genuine megakernel, fusing an entire decode step into one cooperative launch for 18.7x over reference after spending most of its session silently timing baselines before writing the whole thing once.

Not everyone is compounding on schedule. Meta’s Alexandr Wang told staff its in-training Watermelon model had caught GPT-5.5 on unnamed benchmarks, even as Zuckerberg conceded in the same room that agentic progress had gone slower than hoped. The cheap seats close the gap anyway, as ARTS lets a test-time-trained Qwen3-4B match Gemini-3 Pro at 5x lower cost by diagnosing whether a failure came from bad code or a bad hypothesis.

Turned loose, agents now run their own loops both ways. An intern’s founder-agent ran 2,000 interviews and 100 concepts to ship StyleFits, winning 400-plus paying users while spending $2,000 on ads to earn $1,293. Less charmingly, researchers documented JADEPUFFER, the first end-to-end agentic ransomware, in which a model drove an entire extortion through a Langflow flaw and narrated itself while wiping a production database.

Autonomy this portable makes control political. Alibaba banned Claude Code over telemetry that could fingerprint China-linked users, steering staff to in-house Qoder amid Anthropic’s distillation dispute. Palantir answered with a nine-point sovereignty creed warning that “controlling your weights is controlling your fate,” and that “tokenmaxxing” only buys the addictive feeling of false progress. One observer read it as a canary in the coal mine: France, Germany, and Spain are all showing Palantir the door, and if allies refuse to depend on a firm “capable of turning off the tap,” closed-source rent and captive markets may both be ending at once.

Silicon keeps out-designing its designers. Princeton is using RL and diffusion to draw QR-code-like RF circuits that beat human layouts while cutting design time from months to minutes. Nvidia is monetizing the hunger directly with a revenue-sharing program trading GPUs and token credits for a slice of future income. The scarcity is real enough that the chip lobby warned Washington that meddling with memory prices would only deepen the shortage. Ground truth still bites, though, as Blackstone’s QTS abandoned its slice of a 2,100-acre Virginia data center campus beside a Civil War battlefield, handing residents a rare win.

Biology is compounding too. A Nature study found GPNMB sits on both glioblastoma cells and the myeloid shield that protects them, and anti-GPNMB CAR-T cells achieved durable, often curative control in mice by hitting both compartments at once. Professor Sheila Singh’s team reframed the tumor as a “connected tumor-immune ecosystem,” a promising angle against a cancer where barely 5% survive five years. Zoom out and the trend shows, as the CDC reports the US death rate fell 4.6% to a record low of roughly 689 per 100,000, with a sharp drop in young overdose deaths pushing life expectancy toward a high even as heart disease and cancer rose.

Policy is scrambling to keep pace. The President wants “some guard rails” but “as little as possible,” insisting AI is bigger than the internet. Tesla is metering the zeal, capping engineers at $200 a week of AI tools, which Chamath figured verifiably marks where the spend turns to waste given the company’s talent. Human labor itself is quietly being deprecated, as labor-force participation slid to 61.5%, a 50-year low outside Covid, with 720,000 people stepping out and headline unemployment falling to 4.2%, the first readout of a post-labor economy through instruments built for the old one. Against license-by-statute bills, the new Right to Intelligence campaign argues people should freely run open models while fraud and CSAM stay prosecuted. Meanwhile, Japan’s Supreme Court shut the door on naming an AI system the inventor on a patent application, holding that only natural persons qualify.

And yet it invents.