Welcome to March 29, 2026

The Singularity is about to shift into a higher gear. GPT-5.5, Claude 5 Mythos, and DeepSeek-V4 are all expected to drop in April, a triple release that commentators warn could make frontier intelligence too expensive for most of humanity to afford as massive training runs become table stakes, putting pressure on rate limits and pricing. The compression game continues regardless. In OpenAI’s Parameter Golf competition, a challenge to train the best language model that fits in a 16MB artifact on 8xH100s, the best claimed result is now 42.7x better than the baseline, proving optimization finds a way at any scale. That optimization is already eating mathematics alive. Last year, models scored below 5% on USAMO 2025, but GPT-5.4 just scored 95% on the 2026 exam, saturating the benchmark in twelve months flat.

Biology is becoming a compile target. Anthropic is testing Claude Operon, a desktop mode for biology, from phylogenetic trees to CRISPR knockout screens. We are reading the source code of senescence. Researchers have demonstrated the first integrated framework for how epigenetic regulation controls aging, while Northwestern scientists created HOBIT, a “living pharmacy” implant that kept engineered cells alive for a month inside rats while dosing three drugs at once, including a GLP-1 medication. Neko Health, founded by Spotify’s Daniel Ek, plans to open its first US body-scan clinic in New York this spring. The garage biohackers are keeping pace. The man who developed a custom mRNA immunotherapy for his dog’s cancer using Gemini, Grok, and ChatGPT is now starting a company to end cancer for dogs.

The atoms are catching up to the bits. Northwestern researchers introduced autonomous modular legs, single-degree-of-freedom links that learn complex behaviors and snap together to form acrobatic multilegged machines at the meter scale. China’s Agibot is about to ship 10,000 humanoid robots, double its milestone from just one quarter ago, while Cyan Robotics unveiled “Amoo,” an under-80-cm “embodied character” companion with sub-30-ms synchronized eye and body language. The geopolitical immune response is activating. Two US senators plan to introduce the American Security Robotics Act to ban government use of Chinese robots. Meanwhile, on four wheels rather than two legs, the Tesla Model Y has emerged as the world’s best-selling car for the third consecutive year.

The infrastructure buildout is going vertical. Sam Altman shared that the first steel beams went up at the Michigan Stargate site. Intelligence needs watts before it needs weights. South Korea now mandates solar panels for public parking lots with 80 or more spaces, and the DOE’s reactor pilot program could see three to four microreactors reach criticality by July 4 for America’s 250th birthday. Above the atmosphere, the Space Biostasis Coalition has launched to mobilize $1 billion into engineering human cryosleep for space travel, while NASA announced it will send Space Reactor-1 Freedom, the first nuclear-powered interplanetary spacecraft, to Mars before the end of 2028, carrying a fleet of tiny “Skyfall” helicopters. Below the atom, IBM and collaborators demonstrated that a superconducting quantum processor can produce meaningful comparisons with neutron-scattering measurements of a canonical Luttinger liquid system on up to 50 qubits, a milestone for pre-fault-tolerant material simulation.

Back on Earth, the capability overhang is collapsing into deployment. Anthropic researcher Nicolas Carlini demonstrated autonomous 0-day discovery on stage, prompting predictions that the US will gain a major cyberoffense boost that, unlike nukes, will be deployed daily. At Google, a principal engineer reports that PMs are running circles around SWEs with vibe coding tools. Linux kernel maintainer Greg Kroah-Hartman says AI-driven security reporting has “really jumped” across all open source projects in the past month. The market agrees. Claude’s paid subscriptions have more than doubled this year.

Wherever the optimizer converges, human behavior diverges to compensate. Chess grandmasters, all training on the same perfect AI moves, are now winning by playing suboptimal moves their opponents haven’t practiced. Research finds that large language models elevate expert consensus and moderate views, in sharp contrast to the populist polarization of social platforms. And Samsung’s Family Hub refrigerators have started serving full-screen ads on their 32-inch doors, one of which lit up with a plea to a fictional TV character: “We’re sorry we upset you, Carol.”

E Pluribus, Singularity.