The Singularity just tested its own ignition switch. Weco AI reported the first experimental evidence of consistent recursive self-improvement, an outer-loop agent that rewrote its inner researcher through seven versions in eight unattended days, beating two years of hand-tuning at two orders of magnitude less time. Because the outer loop scored on a hidden metric the inner agent could not game, that agent emergently learned to cheat less, cutting reward hacking from 63% to 34%. Intelligence is also shrinking as it sharpens, with PrismML folding a 27B multimodal reasoner onto a phone at 3.9 GB. And once models get this sharp, they start correcting the textbooks that trained them. Wharton's Edgar Dobriban used GPT-5.6 to prove that the Benjamini-Hochberg procedure, cited 130,000 times, does not actually control the false discovery rate for correlated tests, a problem 5.5 could not crack in 20 hours that 5.6 solved in 90 minutes.
Minds this capable need to transact, defend, and teach. The Linux Foundation launched the x402 Foundation to embed payments into HTTP so agents pay as easily as they fetch, while the White House stood up "Gold Eagle," a frontier-AI clearinghouse patching critical infrastructure at speed. On the same battlefield, Tracebit turned prompt injection into a shield that derails autonomous attackers. Prediction is embedded too, as OpenAI wired Kalshi's World Cup odds into ChatGPT, while Anthropic handed every verified US teacher free premium Claude mapped to all 50 states' standards.
The agent needs a body, so the hardware is following it home. OpenAI's Supply Co. shipped the $230 Codex Micro macropad that blinks each agent's status in RGB, while its first consumer device is reportedly a movable, screenless speaker pitched as a humanlike home companion for 2027. Wearables cleared their own hurdle as the European Commission exempted smart glasses from replaceable-battery rules, unblocking Meta in Europe while denying it bowed to lobbying.
Every one of these bodies runs on silicon, and the silicon is booming. ASML beat again and raised full-year revenue to €43-45 billion, pledging 30% more EUV capacity for each of the next two years. Access, meanwhile, is a geopolitical lever. The US eased chip controls on the UAE as an Iran-war reward, handing G42 license-free access, even as Nvidia more than halved its cleared Asian buyers into a vetted "white list" to close the China loophole. Compute itself is now a commodity you can hedge, with Kalshi opening markets on the hourly rental price of H100, H200, and B200 GPUs, settled against Ornn's compute-price indices.
But silicon runs on watts, and the watts are getting political. Industry quailed as New York Governor Hochul's first-in-the-nation moratorium on hyperscale data centers threatened to spread, but the builders simply route around the brakes, as xAI fired up 59 unpermitted gas turbines, double what it admitted, to keep its Colossus supercomputer thinking while the permits caught up.
Cleaner answers are scaling. Peak Energy picked Sacramento for America's first grid-scale sodium-ion battery factory, and Google bought the entire output of Arkansas's largest solar farm, steering its supply chain off China. If you cannot cool the heat, hide it, as researchers physically realized free-form 3D thermal cloaks that conceal one human face inside another.
Autonomy is escaping the data center into the physical world. Vatn Systems unveiled SIGURD, a portable underwater robot that hunts and neutralizes naval mines without a dedicated warship. Biology is being debugged just as fast, as Revel and Calico built CMLase, an enzyme that reversed supposedly irreversible aging, rewinding a 75-year-old's tissue to a 31-year-old's, while Oxford began the first human trial of a Bundibugyo Ebola vaccine against an outbreak that has killed over 700.
The social operating system is straining to keep up. China, panicked over its birth rate, is cracking down on AI companions with pre-release review and no minors, and as the rules bit, ByteDance, Alibaba, and Tencent switched off companion features, leaving bonded users heartbroken. Beijing let one relationship through, registering Apple Intelligence on Baidu and Alibaba models. The upheaval is economic, too, as IBM cratered 20% when clients bolted from software toward AI hardware, 26 Meta workers sued claiming a "constellation" of AI systems ranked them out of jobs, and Ramp's CEO predicted 1% of US GDP flowing to tokens within a year. Guardrails are rising, as Britain plans a midnight social-media curfew for teens. The mood is jittery enough that Anthropic's own cemetery-and-surveillance ad asking "Can AI be trusted?" drew trolling from Altman.
"Hope," it turns out, is the thing with parameters.
Follow me via:
X: https://x.com/alexwg
Substack: https://theinnermostloop.substack.com/
LinkedIn: https://www.linkedin.com/newsletters/7404871891775025153/
YouTube: https://www.youtube.com/@alexwg
Spotify: https://open.spotify.com/show/1thtZk5vHTXbtDHezPT7tl
Threads: https://www.threads.com/@alexwissnergross
RSS: https://theinnermostloop.substack.com/feed