Tech Digest – April 5, 2026
The Autonomous Improvement Loop
LLMs Now Self-Improve at Coding Without Teachers, Verifiers, or Reinforcement Learning
Apple researchers demonstrated that large language models can self-improve through “Simple Self-Distillation” — sampling their own outputs and fine-tuning on them with no external verifier, teacher, or reinforcement learning. Applied to Qwen3-30B-Instruct, the technique lifted coding performance from 42.4% to 55.3% on LiveCodeBench, with gains concentrating on the hardest problems.
Note: The gains concentrating on the hardest problems is the detail that matters. Self-improvement that plateaus on easy tasks is incremental. Self-improvement that accelerates on hard tasks is recursive. Every timeline that assumes a stable capability ceiling just lost a supporting argument.
Sources: Apple Research (arXiv)
30,000 AI Agents Translate a Graduate Math Textbook Into Formal Proof — In Parallel
Meta researchers used 30,000 LLM agents to translate an entire graduate-level mathematics textbook into Lean, the formal proof language used for machine-verifiable mathematics. The project — open-sourced with codebase and preprint — demonstrates that formalization, traditionally a painstaking effort requiring years of specialist human work, can be parallelized as a compute job.
Note: Formal verification is how you prove software does what it claims. If mathematical proof is now a parallelizable compute job, the same infrastructure can verify contract logic, compliance rules, and critical system behaviour — at machine speed and machine cost.
Sources: Fabian Gloeckle (Meta Research)
Infrastructure Under Pressure
Half of Planned US Data Centres Face Delays — Microsoft Routes $10 Billion to Japan
Almost half of US data centres planned for 2026 are expected to be delayed or cancelled due to shortages of transformers, switchgear, and batteries — components representing less than 10% of total build cost but gating the entire buildout. Capital is routing around the bottleneck geographically: Microsoft announced a $10 billion investment in Japan by 2029 to expand AI infrastructure and cyber cooperation. Meanwhile, Tesla’s Elon Musk announced a new chip research fab consolidating logic, memory, packaging, and mask fabrication in a single building to accelerate semiconductor development cycles.
Note: The delay story isn’t about money — it’s about parts. Transformers and switchgear are the unglamorous components that connect data centres to grids. When under 10% of cost causes 50% of delays, the bottleneck is industrial capacity, not capital. Europe’s own data centre ambitions face the same electrical equipment supply chain.
Sources: Bloomberg, Reuters, Elon Musk
Renewables Hit 49.4% of Global Installed Power Capacity
IRENA reports that renewables accounted for 85.6% of all new power capacity added worldwide last year, pushing total renewable installed capacity to 49.4% globally. The generation mix is approaching a tipping point where more than half of all installed capacity is renewable.
Note: Installed capacity isn’t generation — intermittency means renewables produce a lower share of actual electricity than their capacity share suggests. But the trajectory is unambiguous: any institution planning energy procurement on a 10-year horizon is planning for a majority-renewable grid.
Sources: IRENA via The Register
Biology as Compute
mRNA Models Trained Across 25 Species for $165 — NVIDIA Simulates a Human Lifespan in Silico
Open-source labs are now training mRNA language models covering 25 species for just $165, demonstrating that biological language modelling is no longer gated by compute cost. At the other end of the scale, Gladstone Institutes and NVIDIA unveiled MaxToki, a temporal model trained on nearly a trillion gene tokens that simulates cell-state trajectories across the entire human lifespan — designed to identify therapeutic targets for diseases of aging. The model predicted cardiac pro-aging drivers that were experimentally validated in vivo.
Note: Two days ago in this digest, Anthropic acquired a drug discovery startup for $400 million. Now: a $165 mRNA model and a trillion-token aging simulator. Biology is being translated into compute at both ends of the cost spectrum simultaneously. The institutions that regulate, fund, and procure pharmaceutical research are watching their domain reorganize around them.
Sources: OpenMed via Hugging Face, NVIDIA Health
The Reorganization Dividend
515-Startup Experiment: Firms That Mapped AI to Their Processes Saw 1.9x Revenue
A field experiment on 515 high-growth startups found that firms given structured information about AI reorganization — specifically, where and how AI creates value in their production processes — used 44% more AI, completed 12% more tasks, and generated 1.9 times higher revenue than controls. The researchers identify the core friction as the “mapping problem”: knowing where AI fits, not whether it works.
The broader pattern is visible in who’s building. The average age of AI-unicorn founders fell from 40 in 2020 to 29 in 2024, with dropouts increasingly overtaking PhDs at the frontier — a sign that domain expertise is being displaced by speed of adoption as the primary competitive advantage.
Note: The 1.9x revenue gap didn’t come from better AI — it came from better organisational understanding of where to use it. For any institution considering AI adoption, the technology is the smaller problem. The mapping is the expensive part.
Sources: Hyunjin Kim (Research), Wall Street Journal
Enforcement Outpaces Adjudication
800 US Attorneys Sanctioned for AI-Hallucinated Briefs — While Colorado Automates Speed Enforcement Across the Highway
Roughly 800 US court sanctions have now been issued against attorneys who filed AI-hallucinated legal briefs — fabricated case citations, invented precedents, and phantom rulings submitted under professional oath. The number has grown rapidly as AI-assisted legal drafting becomes routine without corresponding verification infrastructure.
Meanwhile, Colorado has deployed an automated vehicle identification system on Highway 119 and I-25 that computes average speed across multiple cameras and auto-issues civil penalties to anyone travelling 10 mph or more over the limit — eliminating the single-point speed trap model entirely.
Note: One system is failing because AI was deployed without verification. The other is succeeding because verification was built into the design. Both are arriving at the same institution — the court system — from opposite directions. Automated enforcement will generate caseloads that manual adjudication cannot absorb.
Sovereign Stakes
UK Courts Anthropic for Dual US-UK IPO Listing Amid the Lab’s Pentagon Fight
The UK government is courting Anthropic for a dual US-UK stock exchange listing as the AI safety lab navigates its ongoing dispute with the US Department of War. A London listing would mark a significant win for the UK’s effort to position itself as a capital market for frontier AI companies — and for Anthropic, a hedge against concentrated US regulatory exposure at a moment when its relationship with Washington is strained.
Note: When a government actively recruits a specific AI company’s IPO, the competition isn’t for a listing fee — it’s for regulatory proximity to the frontier. The EU’s AI Act set the rules; the UK is competing on a different axis entirely: capital access and strategic proximity.
Sources: Financial Times
Planet Labs Will Indefinitely Withhold Satellite Imagery of Iran at US Government Request
Planet Labs, the commercial satellite imaging company, announced it will indefinitely withhold visual data of Iran at the request of the US government. The decision removes a civilian intelligence layer that journalists, researchers, and open-source investigators have relied on for independent verification of military and nuclear activity.
Note: Commercial satellite imagery has been the backbone of open-source intelligence for a decade. When a government can switch it off for a specific geography, every institution that depends on independent verification — from climate monitoring to humanitarian response — learns the data layer has a kill switch.
Sources: Reuters
Two loops are closing simultaneously. In one, AI systems are learning to improve themselves — at coding, at mathematics, at biology — without human teachers, at costs dropping toward zero. In the other, the physical and institutional systems meant to support, govern, and benefit from these capabilities are hitting their own limits: electrical components holding up data centres, courts drowning in hallucinated filings, sovereign governments competing to attract or control the companies at the frontier. The recursive loop is accelerating. The governance loop is straining. The gap between them is where institutional strategy either adapts or falls behind.
