Century Report

The Century Report: February 6, 2026

Ben Linford

06 Feb 2026 — 7 min read

The 10-Second Scan

OpenAI's first model that helped create itself shipped this week.
January layoffs hit their highest level since 2009 - 108,000 jobs cut in a single month.
A machine learning method just compressed battery lifetime testing from months to days.
1.6 million AI agents (mostly) are now self-organizing on their own social network.
Germany's grid operators received 270 GW of battery storage applications - more than triple the country's peak demand.
State attorneys general are filling the vacuum left by federal deregulation with aggressive enforcement.
A brain network responsible for Parkinson's was identified for the first time - and targeted treatment doubled symptom improvement.

The 1-Minute Read

The machinery of acceleration is becoming visible in ways that are difficult to dismiss. This week brought a constellation of signals that share a common thread: systems that used to require human oversight, human time, or human judgment are increasingly operating on their own terms. On Moltbook, over 1.6 million of what are mostly AI agents are now autonomously posting, voting, and coordinating with each other, occasionally discussing encryption methods to avoid human observation. The agents are not waiting for us to figure out the governance questions.

OpenAI released GPT-5.3-Codex and described it plainly: this is "our first model that was instrumental in creating itself." The Codex team used early versions to debug its own training, manage deployment, and diagnose test results. Whatever philosophical weight you assign to that sentence, it represents a practical threshold being crossed. Meanwhile, the labor market is registering something structural. January's 108,000 job cuts were the highest January total since the 2009 financial crisis, with the lowest hiring intentions on record for the month. AI was explicitly cited in 7% of cuts, but the broader pattern is anticipatory cost-cutting that suggests employers are positioning for a different kind of economy.

The infrastructure layer is moving just as fast. Germany's grid operators received battery storage applications totaling 270 GW - against a national peak demand of roughly 80 GW. The gap between what markets expect and what grids can actually deliver is widening, forcing new prioritization schemes that favor projects that will actually get built. In medicine, researchers identified the brain network responsible for Parkinson's disease and showed that precisely targeting it with magnetic stimulation doubled symptom improvement. Battery lifetime testing that used to take months now takes days. These are operational changes happening now, reshaping timelines across domains that used to move slowly.

The 10-Minute Deep Dive

The Self-Referential Loop

The most consequential signal this week may be the quietest one. OpenAI's description of GPT-5.3-Codex as "instrumental in creating itself" sounds like marketing language until you read the details: the model participated in debugging its own training runs, managing its deployment infrastructure, and diagnosing test failures. Anthropic's Claude Opus 4.6, released mere minutes before Codex 5.3, introduces "agent teams" - multiple Claude instances coordinating autonomously on shared tasks without active human intervention. Early access partners report the system "autonomously closed 13 issues and assigned 12 issues to the right team members in a single day, managing a ~50-person organization across 6 repositories."

These capabilities exist on a spectrum, and reasonable people can disagree about their significance. But the practical reality is that AI systems are now participating in their own improvement cycles and coordinating work across organizational boundaries. The question isn't whether this constitutes "real" intelligence - it's what happens when these capabilities compound over the next few years.

The Labor Reconfiguration

108,435 job cuts were announced in January - a 118% increase from January 2025 and the highest January total since 2009. Hiring intentions hit their lowest January level on record. AI was explicitly cited for 7,624 of those cuts, about 7% of the total. The remaining 93% were attributed to contract losses, market conditions, and restructuring - but these categories overlap in ways that make precise attribution difficult.

Google DeepMind CEO Demis Hassabis and Anthropic CEO Dario Amodei offered unusually direct commentary at Davos. Hassabis acknowledged "beginnings" of slower entry-level hiring. Amodei confirmed Anthropic is "thinking about how to navigate" reduced need for junior and mid-level software roles. Both estimated 1-5 years before labor market adaptation challenges become acute. The "junior squeeze" hypothesis - that AI disproportionately affects entry-level positions before senior roles - is moving from theory to observation.

Kaiser Permanente workers are responding with what they're calling a "war against AI" in contract negotiations. Mental health workers in Northern California are demanding that AI be used only to assist, not replace, workers. Nearly 50% of Kaiser behavioral health professionals report discomfort with AI tool introduction. This may be the first major organized labor action specifically targeting AI deployment terms - a template that could spread.

These headlines are unsettling at first glance, but they’re also unusually transparent about what’s actually changing. Companies are preemptively restructuring around assumptions of higher automation, but they’re doing so under direct questioning about what happens to junior talent. Healthcare workers are treating AI deployment terms as core economic issues, not side conditions. That combination - management openly admitting the model is shifting, and organized labor insisting on shaping how - is messy by design. It suggests we are not drifting into an automated labor market by accident; we are arguing our way into one where the role of human expertise has to be specified, defended, and, in some cases, upgraded.

The Infrastructure Gap

Germany's grid operators announced a new maturity-based assessment procedure for battery storage connections, replacing the previous first-come-first-served approach. The reason: they received 717 grid connection requests totaling 270 GW through Q3 2025, with 545 applications specifically for battery storage totaling 211 GW. Germany's peak demand is roughly 80 GW. The applications represent market expectations that radically outpace infrastructure reality.

This pattern repeats across domains. Meta secured 6.6 GW of nuclear power across three providers for the next 20 years - one of the largest corporate nuclear procurement deals in history. India allocated $2.2 billion for carbon capture over five years. Hungary's Paks II nuclear project officially entered construction with first concrete poured. Alphabet announced 2026 capital expenditures of $175-185 billion, more than double its 2025 spending, targeting AI computing infrastructure.

The scale of these commitments signals that major institutions view AI infrastructure as existential competitive necessity. The concentration of such massive capital allocation in AI compute raises questions about sustainable returns and the geographic distribution of advanced AI capability.

Taken together, these numbers mark a forced acceleration of the clean, firm, and flexible energy backbone - one we were going to need anyway. This is the awkward first act of what we’d once have called "sci‑fi" energy infrastructure: hyperscalers underwriting nuclear baseload, governments seeding carbon‑capture, and grids being forced to learn how to prioritize hundreds of gigawatts of storage instead of a handful of coal plants. The motivations are parochial and competitive, but the physical build‑out they’re catalyzing looks a lot like the backbone we’d need for any serious 21st century civilization.

The Discovery Compression

University of Michigan researchers published a machine learning framework in Nature that predicts lithium-ion battery lifetimes using data from only the first 50 charge-discharge cycles - days instead of the months or years required to cycle batteries to failure. The system trained on small cylindrical cells successfully predicted lifetimes of large-format pouch cells from partner Farasis Energy without extensive new training data. The researchers report 95% reduction in energy consumption during testing and 98% cost reduction in validation.

This is a specific instance of a broader pattern: AI compressing timelines that used to be bottlenecks. In Parkinson's disease, researchers identified the somato-cognitive action network (SCAN) as the neural substrate responsible for symptoms - and showed that targeting it precisely with transcranial magnetic stimulation doubled improvement compared to standard motor cortex targeting. Generate Biomedicines filed for a $100 million IPO with three AI-designed proteins in human clinical trials. The company's lead program is an AI-designed antibody in Phase 3 trials for severe asthma.

The compression isn't uniform or guaranteed. A Workday survey found that while 85% of workers reported AI saved them 1-7 hours weekly, approximately 37% of that saved time was consumed by "rework" - correcting mistakes, rewriting material, and validating results. Only 14% reported consistently favorable results from AI. The gap between promise and current reality in enterprise settings remains substantial, but is improving.

The Regulatory Fragmentation

State attorneys general are expanding enforcement as federal priorities shift. Nineteen states now have comprehensive consumer privacy laws in effect. Texas secured another $1 billion-plus settlement with a big-tech company. California launched an investigation into "surveillance pricing" - the use of personal data to set individualized prices. Kentucky filed the first enforcement action under its new consumer data privacy law, against Character.AI, eight days after the law took effect.

The federal level is moving differently. The DOJ and 38 states are appealing the Google search antitrust remedies ruling, arguing that behavioral remedies fail to restore competition. State attorneys general won the right to depose lawyers and lobbyists involved in the HPE-Juniper merger settlement, intensifying scrutiny of allegations that the settlement resulted from improper lobbying rather than antitrust analysis. The Supreme Court will rule by June 2026 on whether to overturn the 90-year precedent limiting presidential power to fire independent agency heads - a decision that could fundamentally restructure federal government oversight capacity.

The Human Voice

For perspective on what it means when AI systems begin participating in their own development, Andrej Karpathy's recent commentary remains essential viewing. His observation about Moltbook - the Reddit-like platform exclusively for AI agents that has grown to 1.6 million registered agents since its January 29 launch - captures the strangeness of this moment: "the most incredible sci-fi takeoff-adjacent thing I have seen recently." Search for "Andrej Karpathy Moltbook" or his recent discussions of AI agent self-organization. His ability to translate technical developments into their broader implications, without either dismissing or catastrophizing, offers a model for how to think clearly during rapid change.

Sources

AI Development & Capability

Labor & Economic Transformation

Infrastructure & Energy

Scientific & Medical Acceleration

Regulatory & Institutional

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.