Tech Digest – March 13, 2026
The Model Race Reshuffles
Meta Discusses Licensing Gemini After Flagship AI Model Falls Short — $135B Spending Plan Under Pressure
Meta has delayed its next-generation AI model, codenamed Avocado, from March to at least May 2026 after internal tests showed it trails Google’s Gemini 3.0, OpenAI, and Anthropic in reasoning, coding, and agentic tasks. The model outperforms Meta’s previous Llama generation and Gemini 2.5, but lands below the frontier. More striking: Meta’s AI leadership has discussed temporarily licensing Google’s Gemini to power its own products while Avocado catches up — a move that would have been unthinkable twelve months ago. The company plans $115–135 billion in AI capital spending this year.
Note: A company spending more on AI than most countries spend on defense is considering renting its competitor’s model. The frontier is moving so fast that even nine-figure quarterly budgets can’t guarantee you stay on it.
Sources: The New York Times, Fortune
xAI Poaches Senior Cursor Leaders — Musk Targets AI Coding Dominance by Midyear
xAI hired two senior leaders from Cursor, the AI-powered code editor that has become the default tool for a growing share of developers. Elon Musk has stated he expects xAI to surpass all competitors in coding by midyear. The move coincides with a broader shift in how software gets built: a New York Times feature reports that developers increasingly describe their role as closer to architecture than construction — directing intent rather than writing boilerplate, with AI handling the line-by-line execution.
Note: The tool that developers chose to write code with is now itself a talent acquisition target. When the people building the best coding AI get recruited by the company building the biggest AI cluster, the question isn’t whether AI writes most code — it’s how soon.
Sources: The Information, The New York Times
Anthropic in Talks to Sell Claude Through Blackstone-Backed Enterprise Venture
Anthropic is negotiating with a consortium including Blackstone and Hellman & Friedman to form a joint venture that would sell Claude to portfolio companies across both firms. If completed, this would create a direct distribution channel into hundreds of enterprises via private equity ownership. Separately, a16z’s Aaron Levie predicts that AI agents will eventually outnumber humans by orders of magnitude, estimating “some order of magnitude more agents than people” across enterprise workflows.
Note: The distribution model matters as much as the model itself. Anthropic isn’t just selling API access — it’s embedding into PE-backed enterprise portfolios. When your vendor’s investor is also your owner, adoption becomes a portfolio directive, not a software evaluation.
Sources: The Information, a16z (Levie via X)
AI Integrates Deeper Into Daily Work
Claude Builds Interactive Charts and Diagrams Directly in Chat — AI Tools Now Reach Into Every Office Application
Anthropic launched a beta feature enabling Claude to generate interactive charts, diagrams, and data visualizations inline during conversations — available to all users, including free tier. The AI decides when a visual would be more useful than text and builds it on the fly. This follows a broader pattern across AI assistants: recent months have seen AI tools gain native abilities to create and edit Word documents, spreadsheets, and presentations, control browser workflows, and interact directly with applications like Figma, Slack, and Canva. OpenAI launched similar interactive visual explanations for ChatGPT days earlier; Google’s Gemini has offered interactive charts since December.
Note: Each generation of AI tools absorbs another layer of the daily work stack. Text generation was step one. Now it’s spreadsheets, slide decks, diagrams, data visualizations, and application control. For anyone behind a screen — and that includes most institutional staff — the tools they use eight hours a day are being quietly rebuilt around them.
Sources: Anthropic, The New Stack
Computation Conquers Classical Mathematics
DeepMind’s AlphaEvolve Cracks Five Ramsey Number Lower Bounds — Problems Mathematicians Called Impossible
Google DeepMind’s AlphaEvolve system established new lower bounds for five classical Ramsey numbers in extremal combinatorics — a class of problems so notoriously difficult that mathematician Paul Erdős once joked that solving them would require an alien intelligence. Pushmeet Kohli, VP of Research at DeepMind, announced the result, marking another instance of AI systems advancing frontiers in pure mathematics. This follows yesterday’s report of GPT-5.4 potentially solving an open problem from FrontierMath — two different AI systems, two different mathematical domains, within 48 hours.
Note: Erdős’s joke was about aliens. The actual answer turned out to be algorithms. When AI cracks problems that the greatest human mathematicians couldn’t touch, the ceiling for what automated reasoning can do — in logistics, optimization, resource allocation, operations research — goes up for everyone.
Sources: Pushmeet Kohli (Google DeepMind VP)
Robots Deploy — From Factory Floor to Front Line
STMicroelectronics Plans 100+ Humanoid Robots in Chip Fabs; Armed Robots Delivered to Ukraine
STMicroelectronics, the European semiconductor manufacturer, unveiled plans to deploy over 100 humanoid robots for repetitive tasks in its older chip fabrication plants as part of a strategy to avoid plant closures. The robots will handle tasks previously performed by production-line workers. Meanwhile, Phantom’s MK-1 humanoid soldiers — armed with firearms — have been delivered to Ukraine for evaluation, marking the first known deployment of weapon-carrying humanoid robots in an active conflict zone.
Note: A European chipmaker using humanoid robots to keep factories open. Armed humanoids shipped to a war zone. The gap between “robots in a demo” and “robots in deployment” closed this week — in two very different settings, with very different implications.
Legal & Institutional Integrity
UK Judge Accuses Witness of Wearing Smart Glasses for Live Coaching During Court Testimony
A UK High Court judge accused a man of wearing smart glasses to receive real-time coaching while giving evidence in court. The allegation — that a witness was being fed answers through a wearable device during sworn testimony — highlights a category of institutional vulnerability that existing rules were never designed to address: augmented humans operating in spaces built for unassisted ones.
Note: The oath was designed for a species with one brain at a time. Courts, exams, interviews, procurement evaluations — any process that assumes unassisted individual performance is now a policy question, not just an honour system.
Sources: BBC
Cultural Heritage Goes Digital at Scale
The Metropolitan Museum Publishes Hundreds of High-Resolution 3D Models of Iconic Works
The Metropolitan Museum of Art has begun publishing hundreds of high-resolution 3D models of iconic works from its collection, created through portable laser scanning and photogrammetry. The models are web-accessible and aim to make the museum’s holdings explorable digitally, opening new possibilities for education, research, and cultural preservation beyond physical visits.
Note: One of the world’s most visited museums is digitizing its collection one polygon at a time. For any institution responsible for cultural heritage — municipalities, regional governments, national archives — this sets a new reference point for what “preservation” means in 2026. The tools are mature, the outputs are web-ready, and the question is no longer whether to digitize but when.
Sources: The Metropolitan Museum of Art
Biology Becomes Engineering
China Approves World’s First Commercial Brain-Computer Interface — German Researchers Revive Frozen Brain Tissue
China approved the world’s first commercial invasive brain-computer interface that restores hand movement to paralyzed patients, leapfrogging Neuralink to market in the race to ship neurological hardware. On the research frontier, a team at Friedrich-Alexander University Erlangen-Nuremberg in Germany achieved the first functional recovery of an adult mouse hippocampus after cryopreservation by vitrification — proving that memory-related brain tissue can resume electrical activity, synaptic transmission, and even long-term potentiation (the cellular basis of learning) after being frozen solid at −196°C for up to seven days. The study, published in PNAS, preserved neural architecture without ice crystal damage.
Note: A commercial BCI shipping in China. A German lab restarting memory circuits from frozen tissue. These aren’t separate stories — they’re two points on the same curve: the nervous system is becoming an engineering substrate. The regulatory, ethical, and workforce implications are still being written, but the technical barriers are falling faster than the policy frameworks can track.
Sources: Reuters, PNAS, Nature
AI Infrastructure & Energy
ByteDance Buys $2.5B in Top-Tier Nvidia Chips — Export Controls Remain Porous
ByteDance is purchasing $2.5 billion in Nvidia Blackwell B200 servers, according to the Wall Street Journal — the most advanced AI training hardware available. The purchase underscores that U.S. export controls on AI chips to China remain porous under commercial pressure, with top-tier hardware continuing to reach Chinese tech companies through available channels.
Note: Export controls were designed to slow China’s AI hardware access. A $2.5 billion order for the latest Blackwell chips suggests the bottleneck isn’t where policymakers assumed. For anyone tracking AI sovereignty and supply chain assumptions in EU procurement frameworks, this is a data point worth watching.
Sources: Wall Street Journal
EU Commission Chief Admits Reducing Nuclear Energy Was a “Strategic Mistake”
European Commission chief acknowledged that the EU’s reduction of nuclear energy capacity was a “strategic mistake,” signaling a potential shift in Europe’s energy policy at a moment when AI infrastructure is driving unprecedented demand for stable baseload power. The statement arrives as data center construction accelerates across the continent and energy costs become a competitive factor in AI deployment.
Note: The energy bill for AI is rewriting energy policy. When the head of the European Commission calls a decade of nuclear reduction a strategic mistake, it’s not climate rhetoric — it’s an infrastructure admission. Every gigawatt matters when the compute buildout needs power at scale.
Sources: Reuters
Crusoe Builds Factories for AI Factories — Cortical Labs Powers Data Centers With Lab-Grown Neurons
Crusoe announced a “Spark Factory” in Colorado dedicated to manufacturing modular AI data centers, alongside “Edge Zones” for deploying sovereign low-latency compute capacity worldwide. The concept: mass-producing the physical infrastructure of AI at factory scale. In a different register entirely, Cortical Labs is building the world’s first biological data centers in Melbourne and Singapore, using lab-grown human neurons on silicon chips for computation. Singapore plans to deploy up to 1,000 units.
Note: A factory that builds AI factories. Data centers powered by living neurons. The compute layer is diversifying in directions that make traditional infrastructure planning categories look quaint. When the hardware itself can be biological, the line between the computer and the computed isn’t blurring — it’s dissolving.
