Welcome to New Year's Eve 2025

The timeline to Superintelligence has new dates. The updated “AI Futures Model” from the authors of “AI 2027” now forecasts a 2x gap between ASI and peak human capability by July 2034. At the same time, the training runs are decentralizing. Epoch AI extrapolates that decentralizing training compute is growing 20x annually, far outstripping the 5x annual growth of frontier training compute, and is on track to catch centralized labs by mid-2031. This convergence coincides with the moment orbital compute has been predicted to become cheaper than ground compute, pointing the vector of progress directly at a Dyson Swarm.

Mathematics is cooked. GPT-5.2 Pro has scored 29.2% on FrontierMath Tier 4, a result that has forced professional mathematicians like Bartosz Naskrecki to admit that “2026 will be a hell of a year.” The panic is productive. Naskrecki is now using AI to take “intermediate steps” toward solving the Langlands Program, while Terry Tao concedes that “the definition of a mathematician will broaden” as proofs become a collaborative synth-human output.

We are beginning to understand how the machine thinks. Researchers have verified for the first time that Transformers implement Bayesian inference geometrically, with residual streams acting as belief substrates, feed-forward networks performing the posterior updates, and attention providing content-addressable routing. Efficiency is also jumping. Berkeley’s new ZIP-RC architecture reuses logits to guide reasoning without any inference overhead.

The priesthood of code is being secularized. Stanford computer science graduates are finding their degrees no longer guarantee employment, as firms replace 10 juniors with 2 seniors and an AI. Meanwhile, Shaquille O’Neal has reportedly completed seven “vibe coding” projects, and Claude Code autonomously built a $30 bird feeder camera with custom firmware. The creative stack is merging. Adobe has partnered with Runway to bring the Gen-4.5 video model directly into Firefly. Knowledge is being re-indexed. xAI’s Grokipedia now hosts 1.8 million articles, roughly a quarter the size of the English Wikipedia.

The industrial base is pivoting to power the intelligence explosion. Elon Musk’s xAI has acquired a third site, “MACROHARDRR,” in Tennessee to reach its 1-million-chip goal. The energy sector, especially, is retooling. Caterpillar’s power segment is now its fastest-growing business, while jet engine shop FTAI Aviation has launched a division to convert aircraft turbines into data center power sources. Capital is fully committed. SoftBank has fully funded its $40 billion OpenAI investment, and Nvidia is in talks to acquire Israeli LLM builder AI21 Labs for $3 billion.

Hardware is dissolving boundaries. Chinese memory maker YMTC is developing “high-bandwidth flash” to fuse storage directly to GPUs in 3D. The air itself is becoming a sensor. The new IEEE 802.11bf standard turns Wi-Fi into a native motion detector. Meanwhile, supply chains are hardening. China now mandates 50% domestic equipment for chipmakers, while Intel confirms its 14A node will debut the world’s first High-NA EUV.

Robots are taking over the physical service economy. Cava and Chipotle are deploying Hyphen’s automated makelines that produce a bowl every 10 seconds. In China, humanoid robots are directing traffic, and at least 36 humanoid vendors are preparing for CES 2026. The tunnels are opening. The Boring Company has started driverless airport loops in Las Vegas.

We are mining the latent space of physics. ML-guided simulations have identified new 2D materials for proton-conducting membranes. Environmental remediation is advancing. ETH Zurich developed an electrolysis process to neutralize DDT. Biology is being debugged. Researchers linked ADHD to circadian rhythm dysfunction, suggesting scheduling as the latest therapeutic vector.

Societal violence is plummeting. Baltimore recorded its lowest homicide rate in 48 years.

On the eve of 2026, the Singularity is no longer a prediction, it’s a procurement order.