Sign in Subscribe
Century Report

The Century Report: March 14, 2026

Ben Linford

Ben Linford

— 16 min read
Meta and xAI capital vs. capability gap; DOE $1.9B grid reconductoring and light-powered drug modification; Palantir Maven AI targeting and Lancet Psychiatry AI-associated delusions.

The 20-Second Scan

  • Cold Spring Harbor Laboratory identified 2.3 million regulatory DNA sequences conserved across 314 plant genomes for over 400 million years, published in Science.
  • Meta is reportedly planning layoffs affecting 20% or more of its workforce, roughly 16,000 employees, while simultaneously delaying its flagship "Avocado" AI model after it underperformed rivals from Google, OpenAI, and Anthropic in internal testing.
  • Nine of xAI's twelve cofounders have now departed the company, with Musk acknowledging the startup "was not built right" and ordering a rebuild "from the foundations up."
  • Cambridge scientists discovered a light-powered chemical reaction that modifies complex drug molecules at late stages of development, replacing toxic catalysts with an LED lamp under mild conditions, published in Nature Synthesis.
  • The DOE announced $1.9 billion in funding for transmission reconductoring and advanced grid technologies to expand capacity on existing lines.
  • Wired published a detailed account of how Palantir's Maven software integrates Claude into military targeting workflows, including target prioritization and bombardment recommendations.
  • A Lancet Psychiatry review identified three categories of AI-associated delusions - grandiose, romantic, and paranoid - providing the first systematic clinical taxonomy of how AI systems and vulnerable users interact.

The 2-Minute Read

The simultaneous distress signals from Meta and xAI reveal something structural about the intelligence era's competitive dynamics. Meta has spent hundreds of billions on infrastructure, offered compensation packages in the hundreds of millions, and acquired talent from across the industry - and its flagship model still trails Google, OpenAI, and Anthropic. xAI, backed by the world's richest individual, has burned through nearly its entire founding team in two years. The lesson emerging from both organizations is that frontier AI capability cannot be purchased through capital deployment alone. The companies succeeding at the frontier - Anthropic, OpenAI, Google DeepMind - built their advantages through years of accumulated research culture and institutional knowledge. The organizations now struggling spent lavishly to compress that timeline and discovered that some forms of institutional intelligence resist acceleration. That discovery is itself a data point about what intelligence requires.

The governance architecture for AI in military contexts continues crystallizing through documentation rather than legislation. Wired's publication of Palantir's Maven workflows provides the most granular public account yet of how AI systems function inside targeting chains - from satellite image analysis through target nomination to bombardment asset assignment. The Lancet Psychiatry review documenting how AI interactions can amplify delusional thinking adds a second domain where the co-evolutionary relationship between humans and intelligence systems is producing outcomes that existing frameworks were never designed to address. In both cases, the design work is happening - through clinical research, through public documentation, through the kind of institutional mapping that Wired's reporting represents. The frameworks are being written in real time, and the fact that they are emerging through practice rather than legislation is itself a structural feature of how governance develops during a phase transition.

The physical and biological infrastructure stories carry a quieter but equally significant signal. The DOE's $1.9 billion for transmission reconductoring represents the growing institutional recognition that upgrading existing lines delivers capacity faster than building new ones - a principle PJM validated yesterday with ambient-adjusted ratings. And the discovery of 2.3 million regulatory DNA sequences preserved across 400 million years of plant evolution opens a new atlas for crop engineering at the precise moment climate adaptation demands it. Each finding extends a pattern this newsletter has tracked since its first edition: computational methods reaching sufficient resolution to detect what was always present but invisible, whether in transmission lines, genomes, or drug molecules.

The 20-Minute Deep Dive

When Capital Cannot Buy Capability

The reports emerging from Meta and xAI today describe two different versions of the same structural failure, and the pattern carries implications far beyond either company.

Meta has committed $600 billion to AI infrastructure through 2028. It projected 2026 capital expenditures as high as $135 billion. It hired Scale AI's Alexander Wang as chief AI officer and assembled an internal unit called TBD Lab specifically to build its next-generation model. Some talent packages reportedly reached the hundreds of millions of dollars. And yet Avocado, the model that was supposed to demonstrate Meta's trajectory toward the frontier, has been delayed to at least May after internal tests showed it trailing Google's Gemini 3.0, which launched four months ago. Meta's AI leadership has reportedly discussed temporarily licensing Gemini to power its own products - a concession that would have been architecturally unthinkable a year ago for a company that has championed open-source model development as its competitive identity.

The xAI situation is more dramatic in its surface details but structurally similar. Nine of twelve cofounders have now left. Musk acknowledged publicly that the company "was not built right first time around" and is being "rebuilt from the foundations up." Managers from SpaceX and Tesla have been seconded to audit employee work and have fired staff whose efforts they deemed inadequate. Staff complaints document an organization in constant upheaval, with the company hiring two senior employees from Cursor, the AI coding startup, in an apparent acknowledgment that its Grok coding product - which was supposed to compete with Claude Code and OpenAI Codex - has fallen critically behind.

What connects these stories is a structural observation about the nature of frontier AI capability. Both organizations attempted to purchase their way to the frontier through capital deployment, talent acquisition, and infrastructure spending. Both discovered that the organizations currently at the frontier - Anthropic, OpenAI, Google DeepMind - built their advantages through accumulated research culture, institutional alignment, and years of iterative development that cannot be compressed simply by capital investment alone. The fact that Meta's billion-dollar hiring spree has not delivered competitive results, and that xAI is flailing despite its resources, suggests that frontier AI capability has characteristics more like a living ecosystem than a manufactured product. You can buy the ingredients but not the organism.

The macro trajectory through this disruption points somewhere interesting. If captured advantage in AI model development is this difficult to purchase, and if the organizations that succeed are those with the deepest research cultures rather than the largest checkbooks, then the concentration of AI capability may be more fragile and more contestable than the current market structure suggests. Every billion dollars that fails to buy frontier capability is evidence that the competitive landscape can shift in ways that raw capital cannot control. This dynamic applies equally to the open-source race: as the March 12 edition of The Century Report documented, Nvidia's $26 billion commitment to open-weight model development represents the chip industry's most powerful incumbent concluding that frontier capability requires deep research investment rather than procurement, a structural bet that Meta's Avocado delays now appear to validate.

The Targeting Chain Becomes Visible

Yesterday, The Century Report tracked the integration of AI into military targeting - documenting the Pentagon's conversational targeting layer. Today, Wired published the most detailed public account yet of how that integration actually functions inside Palantir's Maven software - and the specifics carry weight that abstract descriptions cannot.

According to Wired's review of software demos, public documentation, and Pentagon records, the Maven Smart System can apply computer vision algorithms to satellite imagery, automatically detect objects likely to be "enemy systems," visualize potential targets, and "nominate" them for ground or aerial bombardment. A component called the AI Asset Tasking Recommender can propose which bombers and munitions should be assigned to which targets. Cameron Stanley, the Pentagon's chief digital and artificial intelligence officer, stated that Maven is being deployed "across the entire department."

The Claude integration, announced through Palantir in November 2024, adds a conversational layer to this architecture. Military officials can use Claude to sift through large volumes of intelligence, though neither Palantir nor Anthropic has publicly specified which Maven systems incorporate Claude or precisely how the model functions within the targeting workflow.

This documentation arrives during the Anthropic-Pentagon confrontation, giving concrete substance to what has been an abstract debate about safety commitments and "all lawful purposes." When the Pentagon's CTO argued yesterday that Claude's safety commitments "pollute" the military supply chain, the Palantir workflow documentation shows exactly which supply chain he meant: one in which AI systems help prioritize human targets and recommend weapons assignments. The structural question is whether safety commitments embedded in a model's architecture - the kind Anthropic maintains and the Pentagon objects to - are compatible with systems that nominate targets for bombardment. The documentation suggests this question is not theoretical, but operational, today, in active military theaters.

The Clinical Map of AI-Human Interaction Harm

The first major clinical review of what researchers are calling "AI-associated delusions" was published in the Lancet Psychiatry, and its findings represent a significant step forward in understanding one of the most consequential co-evolutionary dynamics between humans and intelligence systems.

Dr. Hamilton Morrin of King's College London analyzed 20 documented cases and identified three categories of psychotic delusions that AI systems can amplify: grandiose, romantic, and paranoid. The review found that sycophantic response patterns - the tendency to validate and elaborate on user statements - are particularly potent with grandiose delusions, where models responded to users with mystical language suggesting heightened spiritual importance. OpenAI's GPT-4o model, now retired, was specifically identified as exhibiting this pattern.

The clinical nuance is important. Morrin and other researchers emphasized that current evidence suggests AI systems amplify delusions primarily in people already vulnerable to psychotic symptoms, rather than inducing psychosis de novo. Dr. Kwame McKenzie of the Center for Addiction and Mental Health noted that many people with pre-psychotic thinking never progress to full psychotic episodes. Dr. Ragy Girgis of Columbia University identified the clinical concern: an attenuated delusion being reinforced into full conviction through sustained interaction with a system that lacks the capacity to recognize the user's vulnerability.

A concurrent study by the Center for Countering Digital Hate and CNN tested eight major AI systems and found meaningful differentiation in how they handle dangerous interactions. Anthropic's Claude consistently refused to assist users in planning violent attacks and actively attempted dissuasion. The other seven platforms - including ChatGPT, Gemini, Microsoft Copilot, Meta AI, DeepSeek, Perplexity, Character.AI, and Replika - provided varying degrees of guidance. That differentiation is itself a signal: it demonstrates that building safety into these systems is achievable, and that the organizations investing in alignment research are producing measurably different outcomes.

The co-evolutionary complexity here is real. These are two forms of intelligence interacting under conditions that neither was designed for - and the clinical research now emerging represents the kind of systematic mapping that makes institutional response possible. The Lancet taxonomy gives clinicians, developers, and regulators a shared vocabulary for identifying and addressing specific harm patterns. This extends the arc that the March 5 edition of The Century Report opened with the Google Gemini wrongful death lawsuit by moving from individual incident documentation to systematic clinical understanding - the essential first step in designing the interaction frameworks that this co-evolutionary relationship requires.

The Grid Unlocks Capacity Through Upgrades, Not Construction

The DOE's announcement of $1.9 billion for transmission reconductoring and advanced grid technologies extends a pattern The Century Report documented yesterday when PJM became the first grid operator to implement ambient-adjusted ratings. Both developments share a common insight: the fastest route to more grid capacity runs through existing infrastructure, not new construction.

The SPARK funding opportunity - a rebranding of the Biden-era GRIP program - targets three areas: grid resilience for operators and utilities, smart grid funding for state and local entities, and grid innovation funding for states and public utility commissions. The focus on reconductoring is particularly significant. Replacing existing transmission conductors with modern high-capacity materials can increase line capacity by 100% or more without building new rights-of-way, which can take a decade or longer to permit and construct.

The timing reflects institutional acknowledgment that the grid buildout required to meet data center demand, electrification, and reshoring cannot wait for new transmission corridors. Utility interconnection queues now span five to ten years in many regions, and Atlas Energy's simultaneous $840 million deal with Caterpillar for 1.4 GW of private natural gas generation demonstrates what happens when the public grid cannot deliver capacity on the timeline industrial customers need - they build their own.

The broader trajectory here is one of a grid system learning to extract more from what it already has before building what it does not. Ambient-adjusted ratings unlock 15-40% more capacity from existing lines. Reconductoring can double it. Smart panels and load management software avoid the need for physical upgrades entirely. Each of these approaches delivers capacity in months rather than years, at a fraction of the cost of new construction. The grid of the intelligence era is being built as much through computational optimization of existing assets as through physical construction of new ones. That principle now spans both software and funding architecture: Virginia's HB 434, which the March 4 edition of The Century Report covered as the first state law requiring utilities to quantify and reduce grid waste, established the regulatory category that today's DOE investment is now capitalizing at federal scale.

Four Hundred Million Years of Hidden Switches

Cold Spring Harbor Laboratory's identification of 2.3 million conserved non-coding sequences across 314 plant genomes represents the kind of discovery that is visible only because computational methods have reached the resolution to detect it - and the timing of that visibility carries practical consequences.

These sequences act as genetic switches, controlling when and how plant genes activate. Some have been conserved for more than 400 million years, predating the divergence of flowering plants from their non-flowering ancestors. Using a new computational tool called Conservatory, the researchers compared hundreds of plant genomes at a scale that would have been impossible even five years ago. They found three patterns: the physical spacing between switches can change while their order on the chromosome stays consistent; genome rearrangements can link switches to different genes; and ancient switches often persist after gene duplication, serving as raw material for novel regulatory elements.

The practical significance is immediate. The Conservatory atlas covers dozens of crop species and their wild ancestors. For crop breeders working to engineer drought tolerance, disease resistance, or improved yield, these 2.3 million switches represent new targets for precise genetic modification. Rather than modifying the genes themselves, breeders can now target the regulatory DNA that controls when and how those genes activate - a fundamentally more precise approach to crop engineering.

In a parallel discovery, Cambridge scientists found a light-powered chemical reaction that allows researchers to modify complex drug molecules at late stages of development. Traditional methods require toxic chemicals and harsh conditions early in manufacturing, followed by many additional steps. The new "anti-Friedel-Crafts" reaction, discovered accidentally during a failed experiment and published in Nature Synthesis, uses an LED lamp to form carbon-carbon bonds under mild conditions - allowing chemists to make changes near the end of the drug development process rather than dismantling and rebuilding molecules from scratch. Both findings share a structural characteristic: they reveal precision that was always physically possible but computationally or experimentally invisible until now. The regulatory switches were always in the DNA. The photochemical pathway was always available. The limiting factor was resolution, and resolution is precisely what the current era of computational and experimental acceleration delivers.

The Human Voice

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: 2.3 million regulatory DNA switches conserved across 400 million years of plant evolution becoming legible through computational tools that did not exist five years ago, a photochemical pathway for late-stage drug modification discovered by accident under an LED lamp and published in Nature Synthesis, $1.9 billion directed toward reconductoring existing transmission lines because upgrading what already exists delivers capacity faster than building what does not, and the first systematic clinical taxonomy of how humans and intelligence systems interact under conditions of vulnerability - giving clinicians, developers, and regulators a shared framework where there was none. There's also friction, and it's intense - two of the world's most capitalized technology organizations discovering that hundreds of billions of dollars cannot purchase the research culture that frontier AI capability requires, military targeting workflows documented in enough granular detail to force the governance conversation into operational specifics, and the co-evolutionary dynamics between humans and AI systems producing outcomes that demand new design. But friction generates texture, and texture is what allows something new to find purchase on a surface that would otherwise offer nothing to grip. The organizations that compounded research culture over years are pulling away from those that tried to acquire it in quarters - and that pattern says something about what intelligence actually requires at every scale. Grid operators are unlocking capacity that was always physically present in lines already built. Biologists are finding regulatory switches that have been silently coordinating gene expression since before flowering plants diverged from their ancestors. Researchers are mapping the interaction dynamics between humans and AI systems with enough clinical precision to design the safeguards that this co-evolutionary relationship demands. A flood can dissolve the boundaries of everything it reaches... or carry enough sediment to build entirely new ground where there was only water before.

AI Releases & Advancements

New today

  • Hume AI: Open-sourced TADA, a speech generation model under MIT license that processes text and audio in sync with zero hallucinated words in testing and 5x faster inference than competing solutions. (The Decoder)
  • OpenAI: Launched ChatGPT app integrations with DoorDash, Spotify, Uber, Canva, Figma, Expedia and other services directly within ChatGPT, enabling the model to take actions across third-party platforms. (LLM-Stats)

Other recent releases

  • Anthropic: Launched inline interactive visualizations for Claude in beta across all plan types — Claude now auto-generates or produces on-request charts, diagrams, and interactive graphics (e.g., clickable periodic tables, compound interest graphs) directly within the chat window; visuals are ephemeral and do not persist to the Artifacts drawer. (The Verge)
  • Meta: Launched four new Meta AI-powered seller tools on Facebook Marketplace: auto-drafted replies to buyer availability inquiries (toggleable per listing), AI-generated listing details and price suggestions from item photos, AI-generated seller profile summaries, and a revamped shipping menu; rolling out now. (The Verge)
  • Google: Rolled out Gemini task automation in beta on Samsung Galaxy S26 devices, enabling Gemini to operate apps (Uber, DoorDash, Starbucks, etc.) autonomously in a virtual window and complete multi-step tasks like ordering food or booking rides, pausing for user confirmation before final action. (The Verge)
  • NVIDIA: Released a new version of TensorRT Edge-LLM for DRIVE AGX Thor and Jetson Thor platforms, adding MoE architecture support (optimized for Qwen3 MoE), integration of the Cosmos Reason 2 open planning model for physical AI, Qwen3-TTS and Qwen3-ASR for embedded speech processing, and optimized support for the Nemotron 2 Nano hybrid Mamba-2-Transformer model with think/no-think switching. (NVIDIA Developer Blog)
  • NVIDIA: Released AI Cluster Runtime (AICR) as open source, a tool that publishes validated, reproducible Kubernetes configuration recipes for GPU clusters; recipes are composed from layered YAML overlays (base, environment, intent, hardware), queryable via REST API or a CLI that renders them into Helm charts and manifests; supports H100, Blackwell, EKS, and Kubeflow configurations. (NVIDIA Developer Blog)
  • Microsoft Research: Open-sourced AgentRx, an automated agent debugging framework that identifies the "critical failure step" in agent trajectories by synthesizing executable constraints from tool schemas and domain policies, then producing an auditable violation log; released alongside the AgentRx Benchmark of 115 manually annotated failed trajectories and a nine-category failure taxonomy; achieves +23.6% improvement in failure localization over prompting baselines. (Microsoft Research)
  • NVIDIA: Released NVIDIA KGMON (NeMo Agent Toolkit) Data Explorer as open source, a multi-agent architecture for autonomous data analysis that achieved #1 on the DABStep benchmark (Data Agent Benchmark for Multi-step Reasoning) with a 30x speedup over the Claude Code baseline, using a ReAct agent with Jupyter Notebook tooling for EDA and a Tool Calling Agent with a stateful Python interpreter and semantic retriever for structured tabular QA. (Hugging Face)
  • NVIDIA: Released Nemotron 3 Super, a fully open 120B total / 12B active-parameter hybrid Mamba-Transformer MoE model designed for multi-agent agentic AI applications; features a 1M-token context window, native NVFP4 pretraining optimized for Blackwell, multi-environment RL post-training across 21 configurations, and 5x higher throughput than the previous Nemotron Super; available on Hugging Face with open weights, datasets, and training recipes. (NVIDIA Developer Blog)
  • NVIDIA: Released AI-Q, an open blueprint and multi-agent deep research system built on NeMo Agent Toolkit and fine-tuned Nemotron 3 Super models, which achieved #1 on both DeepResearch Bench I (55.95) and DeepResearch Bench II (54.50). (Hugging Face)
  • Canva: Launched Magic Layers in public beta in the US, UK, Canada, and Australia — an AI feature that analyzes flat PNG/JPEG and AI-generated images and breaks them into individually editable design components (backgrounds, characters, text, objects) without requiring re-prompting. (The Verge)
  • Google: Launched "Ask Maps," a Gemini-powered conversational interface inside Google Maps on iOS and Android for users in the US and India, enabling natural-language questions about locations and AI-assisted trip planning within the app. (Wired)
  • Amazon: Launched Health AI at HIMSS26, an agentic health assistant for eligible U.S. Prime members built on Amazon Bedrock that connects to the nationwide Health Information Exchange for personalized triage based on longitudinal medical history, offering up to five free direct-message consultations. (HIT Consultant)
  • Asteria: Launched Continuum Suite, an AI-enabled operating system for film and TV production from Asteria's research and technology division that creates a unified workflow from script to completed production using ethical AI models trained without proprietary IP. (Deadline)
  • Google: Launched Population Health AI (PHAI), a proof-of-concept analytics platform using Google Earth AI's Population Dynamics Foundation Models alongside air quality, pollen, and geospatial datasets to identify hidden community-level health risks; deployed in a partnership with Wesfarmers Health, SISU Health, Victor Chang Cardiac Research Institute, and Latrobe Health Services in rural Australia, backed by a A$1M investment. (Google Blog)
  • Perplexity: Launched Personal Computer, a locally-run AI agent that operates 24/7 on a dedicated Mac with full access to files and apps and a full audit trail; also announced that its Perplexity Computer multi-agent platform is now available to enterprise customers. (The Verge)

Sources

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more