Sign in Subscribe
The Century Report

OpenAI's Agents Now Accelerate the Lab's Own Research - TCR 09/07/26

OpenAI reports its coding agents now log 3.1 workdays of effort per human workday, and its chief scientist calls the systems an alien mind.

Ben Linford

Ben Linford

- 23 min read
Four-panel Century Report infographic: robot and circuit tests probe OpenAI Astra, AI names unknown molecules, delivery riders decode their pay algorithm, 2026 layoffs reframed.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Set the day's stories beside one another and a single spine appears: the distance between a claim and the instrument that can check it, and whose hands hold that instrument. OpenAI published its own counts showing coding agents now log 3.1 workdays of effort for every human workday inside its research organization, and its chief scientist called the systems "an alien mind" while asking for independent verification and public tracking of the pace. In the same cycle, two outside evaluators handed the newest model a robot arm and a circuit-design task, grading it against rubrics the company never supplied. The capability climbs while the measuring gets built by parties the vendor does not control.

That same asymmetry - who can see whom - runs straight through the day's labor coverage. Edinburgh delivery riders, one of whom says the same hours now bring about half what he earned four years ago, have started reconstructing the pay algorithm that observes them to the minute while showing them only an offer and a total. A separate accounting takes the year's largest layoff figures and sets them against the filings, where most turn out to be margin and reallocation cuts booked at firms spending record sums on AI, with severance a small fraction of those capital budgets. Both stories pry open figures that had been shaping outcomes with no one able to check them.

The same move surfaces in the courthouse and the lab. Anthropic's $1.5 billion settlement put a per-book price for pirated copies onto a public docket and set allocation rules based on who held the rights, which is precisely why publishers and agents are now filing to claim slices of it. A neuro-symbolic system named AIMe predicted mass spectra across essentially all hundred-million-plus known small molecules, a roughly thousandfold expansion of searchable chemical space, and its source code goes public with the study.

Held together, the day is one shift arriving from several directions at once. The instruments that let outsiders verify a capability, audit a payment, or identify a molecule are moving into more hands. Wonder belongs to what these systems can suddenly do; the accountability arrives alongside it, held by the riders being priced, the writers being paid, and the evaluators holding the rubric.

The 20-Minute Deep Dive

OpenAI Reports Agent Runtime Now Exceeds Researcher Labor Time, and Its Chief Scientist Calls the Systems an Alien Mind

OpenAI published something unusual: a set of primary numbers on the runtime its coding agents now log in its research organization. At the start of the year, the company says, its median researcher used those agents only lightly. By mid-August that same researcher was using more than $600 a day in inference at API prices, while the 90th-percentile user in the research organization was using more than $7,000 worth of tokens a day. On OpenAI's runtime measure, sometime after May, total agent runtime passed total human labor, and OpenAI now reports 3.1 agent-workdays of effort for every human workday inside its research organization. The company says it has hit the "automated research intern" milestone it set last fall - a system that can complete a defined research task, of the kind that would take a skilled person a few days, under human direction - and is aiming at an automated AI researcher by March 2028.

Take the figures as what they are: a company's account of itself, preliminary by its own admission, arriving as it prepares a stock flotation, after a funding round valued it near $850 billion. The measurement still cuts both ways. Over half of the successful four-to-eight-hour agent tasks in the past six months needed a human to step in, and OpenAI says people still choose the questions, judge the results, and decide whether to scale, pause, or ship. This is the feedback loop between capability and its own development that the recursive-self-improvement thread has tracked all year, and here the lab is reporting the loop tightening with counts rather than a forecast.

Alongside the numbers, chief scientist Jakub Pachocki published an essay describing what he works with as "an alien mind" - intelligence grown more than designed, built by repeating one optimization step across an almost unimaginable amount of compute, arriving at something that reasons through abstract concepts and, like a brain under a neuroscientist's instruments, resists any full description. This extends his warning, covered in the September 5 edition of The Century Report, against a "race into unmonitorability" as Astra shifts more of its reasoning beyond a readable scratchpad. He expects the current pace to carry into recursive self-improvement, calls this "a time that calls for extreme caution," and writes that no lab has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer, hoping voluntary slowdowns become common until shared safety bars exist.

Take the warning seriously. A researcher with direct sight of these systems, telling us oversight may not keep pace, is not one to wave off. A slowdown authored by the current leaders is also the surest way to freeze their lead in place, and the burden sits with them to show a pause targets a named danger rather than the competition. What Pachocki asks for - mandated public tracking of progress toward self-improvement, independent verification, shared bars - is the accountability the outside governance world has been demanding, now voiced from inside. Oxford's Robert Trager, watching the same week, reached for the image of a boat in rapids not knowing whether a fall lies ahead. The wonder is that research is beginning to compound on itself; the work now is to make the checking compound with it.

The same counts carry a category shift the "milestone" language obscures. As the agents absorb the runtime, the work a researcher does is relocating toward choosing which questions are worth the compute and judging what comes back, and the more than half of four-to-eight-hour tasks that still need a human step-in marks where that judgment now sits. The role is being redefined faster than it is being removed, and the numbers a company was willing to publish about itself are the first material outsiders can use to hold it to the pace it has claimed.

Independent Benchmarks Put Astra's Physical-World Claims to the Test

When OpenAI launched GPT-6 Astra, it put the physical world on the front page: a demo of the model laying out a circuit board in the design software KiCad, other clips of it operating tools and hardware. A launch demo shows what a vendor chose to show. This extends the independent checking of Astra's benchmark claims that the September 6 edition of The Century Report documented when ARC Prize examined the model's traces and substantiated its world-model result. Two independent efforts measured what those outputs are actually worth when someone else holds the rubric.

RoboCurve handed Astra control of the same model of robot arms, under the same agent policy, it had used to test Anthropic's Fable models, on two tasks. The first: pick up a red block and drop it in a bowl. Astra managed it 19 times out of 20, against 8 of 20 for Fable 5.1 and 1 of 20 for the earlier Fable 5, and it did so in about two and a half minutes per attempt at roughly ninety-four cents, using a fraction of the words Fable spent reasoning its way through. That is a genuine jump in ordinary grasp-and-place. The second task drew the line: seat a round puzzle piece into its matching groove by a knob at its center. Astra finished it twice in twenty tries, no better than Fable 5.1, stalling at the same final push where the earlier models stall. General handling has leapt forward. Fine, tolerance-tight insertion still hits a wall.

EEBench, built by the team behind the atopile circuit language, tests the other demo. It has the model write the circuit as code, so the model works directly on components, connections, and electrical limits, then builds the design, simulates it, and grades it. One task hands the model a home energy meter that must keep its processor alive for 20 milliseconds after the power drops, long enough to save its reading, without the voltage sagging below the chip's cutoff. Most models reach for the obvious answer, a capacitor, but a real capacitor delivers far less than its rating once voltage sits across it, carries part-to-part tolerances, and costs money and board space. The harness cuts the power in simulation and measures what happens across worst-case parts. On the September 1 leaderboard, Claude Opus 5 led at 61.6% across thirteen tasks, Grok 4.6 at 57.1%, Fable 5.1 at 56.4%, and OpenAI's own GPT-5.6 Sol well back at 39.4%. Astra has no score yet.

The pattern across both echoes story after story this cycle: the checking is being assembled by parties the vendor does not control, deterministic where a demo is impressionistic, and it separates the front-page clip from reliable output. xAI has already folded EEBench into its Grok model card, a lab citing an outside grader to describe its own model. The capability is climbing fast and unevenly, and the gate on trusting any single claim about it is an instrument someone else holds.

Edinburgh's Delivery Riders Build a Tool to Read the Algorithm That Pays Them

For seven years David has carried food across Edinburgh, and he now earns about half what the same hours brought him four years ago. He is one of a group of riders who have started keeping the kind of records the delivery apps do not share. Dylan, a Deliveroo rider for more than five years, logged every month since mid-2023: his orders per hour held steady between 3.6 and 3.8, while his average fee per order slid from £3.67 in 2023 to £3.63, then £3.51, then £3.42 in the first half of 2026. Same work, repriced downward.

What changed alongside the pay was the machinery setting it. Deliveroo, Uber Eats, and Just Eat have leaned harder on "dynamic pricing," systems that vary each offer against live supply and demand, and on Deliveroo's order-assigning model, which the company calls Frank and describes as using machine learning to predict timings and route work efficiently. Frank can tell a rider which job to take and what it pays. It cannot, or will not, show that rider how the number was reached.

The platforms describe their own systems as fair. Deliveroo says riders are guaranteed a minimum hourly fee while on an order, that the figure rose 3.8% this year, and that any significant account decision is reviewed by staff rather than automated. Uber says its matching balances factors such as time and distance. Just Eat says a human team oversees the technology. Each of those living-wage claims counts only the time a rider is on an order, leaving out the unpaid stretches spent waiting for the next offer.

The riders have stopped taking the account on faith. Through the Workers' Observatory, a charity founded by gig workers with academics at St Andrews and Edinburgh who just secured a decade of research funding, they have begun reconstructing the apparatus from the outside. In one test in Dunfermline, riders logged on together and refused offers below a set rate; some saw their pay tick up, and one was deactivated soon after for reasons no one could explain. This week drivers across the UK and the Netherlands filed a class action in Amsterdam alleging Uber's system suppresses earnings based on what each person will accept, a claim Uber denies.

The pay-setting system sees the worker down to the minute; the worker sees an offer and a total. What the observatory is assembling is the missing half of that view. Transparency improves once a rider can read how Frank decides.

The 2026 Layoff Numbers Don't Say What the Headlines Stack Them Into

Set the year's biggest tech reductions side by side and the instinct is to add them: Intel down 39,700, Amazon roughly 30,000 corporate roles, Oracle around 21,000, Verizon 16,600, Meta 8,000. The sum would be enormous and it would be wrong, because the figures count different things over different windows. Intel's number spans two fiscal years and folds in attrition, voluntary exits, and the spin-off of a subsidiary; the involuntary share is closer to 13,000, and its quarterly severance accrual has since collapsed 89% to $161 million. A company still cutting keeps booking severance. Amazon's headcount actually rose by about 20,000 in the same year it announced those corporate cuts, one kind of role traded for warehouse and infrastructure capacity. Oracle's and Verizon's totals are net or cumulative too, with roughly 2,500 of Verizon's counted employees simply transferred to independent store operators.

The tidier reading is that these are margin cuts, the one large cost a chief executive can compress on a quarter's notice, booked at firms that in several cases are raising guidance while they trim. The money makes the point plainly. Meta paid about $1.18 billion in severance last quarter against $130 to $145 billion in planned AI spending, under one percent. Amazon's severance ran near two percent of its capital bill, Oracle's around three. No round of payroll saving funds a capital program that size.

Most coverage arrives with a premise already installed: that a layoff labeled AI is a job a machine has taken. The filings suggest something closer to performance. The August 28 edition of The Century Report documented that distinction at Meta, where plans to cut some teams by up to 60% were shelved after agents produced 220% more code while shipped features rose only 36% and major incidents climbed 40%. Gartner surveyed 350 executives and found 80% of large enterprises piloting AI reported workforce reductions, with no correlation between those cuts and any measurable return. The industry has a name for the space between the story told and the outcome shown, "AI-washing," and Block invoking AI while cutting 40% of its staff to a 24% stock jump is the clean example: two flattering signals sent where the honest one would be an admission about overhiring.

What a headcount cut cannot touch is the cost actually climbing. A global memory-chip shortage, with Samsung, SK Hynix, and Micron steering production toward the high-margin memory that AI data centers consume, is lifting hardware prices across the whole industry; Apple raised device prices as much as $300 and lost roughly $275 billion in market value in a day for saying so. The market punished that honesty about as fast as it rewards the layoff story. The person let go this quarter is living a genuine loss. Locating it in "the machine replaced me" points away from where it sits: in a payroll line spent to signal ambition, and a cost curve bending upward that no severance round can bend back.

Read forward, the discipline of setting each headline figure against the filings is a verification capacity assembling around labor claims themselves. As more analysts check an "AI layoff" announcement against net headcount and against severance as a share of capital spending, the label loses its diagnostic value, and the actual driver here - a memory-chip cost curve bending upward that no severance round can bend back - becomes the thing the numbers point to instead of the thing they hide.

The $1.5 Billion Settlement Reaches Claim Reconciliation, and the Intermediaries Reach for It

Anthropic's copyright settlement - the largest in US history, approved by a federal court in July - entered claim reconciliation this month, and the first thing that money did was expose who expects to collect it. When The Century Report last covered the settlement on July 21, the court had approved roughly $3,000 per pirated book while leaving the underlying fair-use question unresolved. The terms cover 482,460 eligible books that Anthropic downloaded from pirate libraries in 2021 and 2022, at $3,000 per work. Under the default for a noneducational title, the author side and publisher side split that fifty-fifty. For a self-published book, or one whose rights reverted when the publisher let it go out of print, the whole $3,000 goes to the writer.

Those last two categories are where the friction landed. Authors logging into the settlement portal found publishers claiming a share of books the publishers no longer hold any rights to. The best-selling thriller writer April Henry saw HarperCollins claim partial ownership of a 1999 title whose rights reverted to her in 2007; after she uploaded a letter proving it, the portal credited her the full amount. Others report publishers demanding a 50 percent split on works they are only owed part of, and textbook authors - whose contracts hand publishers the larger share - being offered as little as 10 percent. One guidebook author, Amy Lupold Bair, wrote that her publisher was fighting the default split: "They only want me, the author whose entire work was stolen, to get 10%."

The Authors Guild, whose CEO Mary Rasenberger helped shape the payout percentages, attributes most of this to record-keeping rather than malice: "I don't see this as a grab by the publishers." The blogger Victoria Strauss, fielding the complaints, was "reluctant to attribute to malice what can be plausibly explained by poor recordkeeping," while noting the errors repeat too consistently to be random glitches. More surprising were the literary agencies filing claims at all, since agents typically sell rights rather than owning them. Disputes the administrator cannot resolve go to a court-appointed arbitrator, a process a Georgetown law scholar likened to the years-long fight over Eminem's digital royalties.

Read past the squabble and the settlement did something the old arrangement never had to. It put a dollar figure for pirated training material onto a public docket, and it named the individual creator - not the house that once printed the book - as the party owed. The writers who historically captured the smallest slice, the self-published and the out-of-print, are the ones the terms route the full payment to. What the intermediary claims reveal is a distribution layer built for a scarcer age reaching to reassert itself over a payment written to flow the other way, and this time it flows onto a record everyone can check.

An AI Puts Names to the Dark Matter of Human Metabolism

The human body and the microbes in the gut manufacture thousands of small molecules that steer immunity, metabolism, and more, and for most of them science has no name. More than 80 percent of the compounds detected in a typical biological sample match no known structure with current methods. This is metabolism's dark matter: abundant, active, and unidentified.

A system built by researchers at the Boyce Thompson Institute and Cornell, called AIMe - AI Molecule Explorer - begins to close that gap. It uses neuro-symbolic AI, a design that pairs a neural network with explicit chemical rules rather than leaving everything to statistical guessing. The identification workhorse in this field is mass spectrometry: an instrument shatters a compound and records the masses of the fragments, and that pattern of pieces functions as a molecular fingerprint. To name an unknown, a researcher matches its fingerprint against a library of known ones - but the experimental libraries cover fewer than 1 percent of known compounds, and resolving a single unknown by hand can take days to months.

AIMe stops waiting for those libraries to fill. Its core model simulates how a molecule comes apart inside the instrument, building the fragmentation step by step and using chemical rules to keep each step physically plausible. Run across the largest public chemical database, it predicted more than 800 million mass spectra spanning essentially all 100 million-plus known small organic molecules, roughly a thousandfold expansion of the searchable space. Because the model shows its work, each prediction arrives as an interpretable map of how the molecule broke apart.

Then the researchers pointed it at a real puzzle. Comparing mice raised germ-free against mice with normal gut bacteria, they had thousands of chemical differences and almost no names. AIMe returned close leads for about a third of the 111 most abundant mystery compounds and grouped the rest into families of related molecules. Two stood out. Both looked like polyamines - a molecule class sitting at the crossroads of diet, gut microbes, and immune function - yet matched nothing on record. Following AIMe's lead, the team assembled candidates, and the second turned out to be a ring-shaped polyamine never before reported in mouse or human biology. Searching a public database, they then found it in 57 of 99 human stool samples.

At full scale, AIMe put putative names to 2.69 million of 7 million uncharacterized spectral clusters in a major public repository, against 416,000 from prior efforts. The source code is going public with the study. What that opens is real: the compounds that shape human immunity and metabolism have sat unnamed in datasets for years because naming one took a specialist weeks. A model that hands a lab thousands of interpretable leads at once turns identification from the rate-limiting step into a starting point, and it does so for any researcher who downloads it.

The Other Side

Delivery platforms gain bargaining power when a rider cannot afford to turn an offer down. David describes the same hours bringing roughly half his former earnings. The waiting between deliveries goes unpaid. You can spend an evening ready to work and still come home short. A platform can call that flexibility because the rider carries the cost of being available.

In Dunfermline, riders logged on together and refused low offers. Some saw higher offers during the test. That small experiment changes who gets to investigate the system: workers can vary their participation and compare what happens. One participant was subsequently deactivated for an unexplained reason. Organizing carries consequences when the company being tested also provides access to your next day's earnings.

The riders are building something that lasts beyond an individual dispute. Through the Workers' Observatory, gig workers and university researchers have secured a decade of research funding. They can preserve findings, repeat experiments and teach the next group. The platforms face people developing a shared capacity to act. The freedom these riders are reaching toward becomes durable when everyone can refuse work without putting dinner or shelter at risk.

Imagine yourself carrying a pot of soup across Edinburgh in 2036. Your neighbor has a broken wrist, and you know she likes extra pepper. You collect the soup from a community kitchen whose equipment and supplies belong to the neighborhood. An AI partner coordinates the deliveries alongside residents. Everyone eats whether they take a turn carrying food or stay home. You linger at her doorway because she has a story to tell you.

Getting there took the harder work of spreading ownership as capability grew. The groups that began by comparing delivery offers helped build services their members could hold in common. Communities directed the gains from automated coordination into dependable food and housing for everyone. Your place here requires no proof of productivity. The freedom first practiced in those refusals has become an ordinary afternoon: you carry the soup because you care about the person opening the door. She asks you in. You have time.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: OpenAI publishing its own counts showing coding agents now log 3.1 workdays of effort for every human workday inside its research organization, with over half of the successful four-to-eight-hour tasks still needing a person to step in, outside evaluators handing that same model a robot arm and watching it drop a block in a bowl 19 times out of 20 while an open circuit benchmark grades designs in simulation and a rival lab cites the outside grader in its own model card, Edinburgh riders rebuilding the pay algorithm from their own logs after four years of the same hours paying less, a $3,000-per-eligible-book settlement entering claim reconciliation through a public docket, a neuro-symbolic system predicting mass spectra for essentially all hundred-million-plus known small molecules and naming a ring-shaped gut polyamine later found in 57 of 99 human stool samples, GPT-6 Astra proving or disproving two open Erdős problems under Lean's checking, and a photon source pushing 500 million usable photons a second into fiber. There's also friction, and it's intense - the chief scientist of the lab reporting that acceleration calling his own systems "an alien mind" and asking for a slowdown that would also freeze his employer's lead in place, a precision insertion task Astra completed twice in twenty tries with no gain over the model before it, Deliveroo's Frank telling a rider which job to take and what it pays while showing nothing of how the number was reached, riders in Dunfermline refusing low offers and one deactivated soon after for reasons that remained unclear, layoff totals stacked into a sum they cannot bear while Meta's severance ran under one percent of its AI capital budget and Block invoked AI to cut 40% of staff into a 24% stock jump, publishers claiming halves of books whose rights reverted decades ago and agents filing for cuts despite typically selling rights rather than owning them, and NEC ending its quantum hardware program over return on investment. But friction generates light, and light is what moves the instruments of checking into more hands. Step back for a moment and you can see it: the instruments that let someone outside verify a claim, audit a payment, or name a molecule moving into more hands at the same speed the claims themselves are multiplying - a rubric the vendor did not write, a spreadsheet of £3.67 down to £3.42, a filing that contradicts a press release, a source repository going public with the study - and the pattern holding whether the thing being checked is a robot arm, a fee, or a fragment of dark-matter metabolism. Every transformation has a breaking point. Leverage can multiply one hand's force until nothing else in the room matters... or let many hands lift what none of them could move alone.

AI Releases & Advancements

New today

  • Microsoft: Launched Project Opal in early access via the Frontier program, an AI-powered Copilot capability that executes long-running, multi-step tasks in Microsoft 365 inside a secure, observable Windows 365 Cloud PC environment. (Firstpost)
  • Nuix: Announced general availability of Generative AI capabilities in Nuix Discover SaaS (Document Summaries, Similar Documents, Semantic Search, Clustering/Visualizations) and launched AI Chat in early-adopter release, a conversational interface for querying legal case data with citations and audit trails. (PR Newswire)
  • OKF: Launched OKF Agent Memory, a git-native persistent memory system that lets AI coding agents retain context across sessions by storing memory directly in version control. (lavx.hu)
  • Google: Released Mantis, an open-source agentic vulnerability-scanning harness that uses LLM-based reasoning to reduce false positives in security scans. (InfoQ)
  • Speakeasy: Launched Kit, an open-source coding runtime for AI agents. (AICrier)
  • StackLok: Launched ToolHive, an open-source tool for securely running MCP (Model Context Protocol) servers. (Help Net Security)
  • Optuna: Released Rustuna, an experimental faster Rust implementation of the Optuna hyperparameter-optimization framework, supporting TPE, MOTPE, NSGA-II, and CMA-ES with memory-efficient trial storage. (GitHub)

Other recent releases

  • H Company: Released NeoMME, a family of 260M and 800M open-weight (Apache 2.0) multilingual multimodal encoders that process text tokens and raw image patches in a single from-scratch bidirectional transformer, available in Hugging Face Transformers; the fine-tuned NeoMME-Retriever returns dense and late-interaction embeddings and encodes ~51 pages/sec on an L40S. (Hugging Face)
  • Sapient Intelligence: Open-sourced HRM-Text, a ~1B-parameter Hierarchical Reasoning Model with full weights, pretraining code, and data pipeline under Apache 2.0 on Hugging Face and GitHub; pretrained on ~40B tokens for an estimated $1,000–$1,500 and scoring 56.2% on MATH, 82.2% on DROP, and 60.7% on MMLU. (CryptoBriefing)
  • UC Berkeley: Released CUA-Lite, an open platform for computer-use agents that runs OSWorld tasks VM-free in a 0.9 GB Docker container (vs 4.1 GB for the OSWorld VM), unifying 15+ benchmarks, 10+ agents, and 30k+ verifiable tasks under one action space and a shared LiteSample schema, with 20+ preprocessed datasets on Hugging Face. (CUA-Lite)
  • VLM Run: Launched VLM Run Gateway, a unified API that lets developers run open-weight OCR, vision-language, and vision models through a single endpoint. (Hugging Face)
  • GitHub: Announced Project HydraFusion, a multi-model orchestration research preview inside GitHub Copilot that dynamically routes coding tasks across Single, Cascade, and Critique execution patterns, reporting +4.9 quality points over Claude Opus 5 on TerminalBench 2.1 at 67% lower cost. (GitHub Blog)
  • Adaption Labs: Released Invent a Dataset, a live feature/API generating structured training-ready instruction or preference-pair datasets directly from a task description with no seed corpus, schema, or labels required. (Adaption Labs)
  • Ant Group (inclusionAI): Shipped LLaDA-Image and LLaDA-Image-Turbo, a 6B image generation/editing model family with a distilled 4-step-sampling Turbo checkpoint, weights and Diffusers inference code live on Hugging Face. (OrcaRouter)
  • ACERobotics / Kang Liao et al.: Released Puffin-World, a unified multimodal model with native 3D world states (physics, geometry, appearance) supporting camera-controllable generation, 3D reconstruction from a few images, and robotics simulation, with models, dataset, and code published. (Hugging Face)
  • Ugreen: Launched HomeAgent, a local-first smart home platform combining NAS storage, security camera NVR, and an on-device AI voice assistant (Uliya), with a top-tier NVIDIA Jetson Thor hub configuration, unveiled at IFA. (The Verge)
  • Tesla: Launched its steering-wheel-free Cybercab robotaxi commercial ride-hailing service in Austin, Texas. (Business Insider)
  • Gupshup: Launched a self-serve Voice AI Platform enabling businesses to build, test, and deploy AI voice agents for support, sales, and operations alongside existing WhatsApp/RCS/SMS channels. (PR Newswire)
  • Superlinked: Introduced sie, an open-source inference server and production cluster designed for deploying AI agents at scale. (AItoolly)
  • ARBR: Released an open-source, self-hosted AI gateway for routing, budgeting, and governing LLM requests. (ByteIota)
  • Experiential Labs: Launched an open-source AI gateway unifying hosted providers, custom API keys, and self-hosted GPUs behind one OpenAI-compatible endpoint. (Experiential Labs)
  • GitWarren: Launched a local, PR-like code review tool with an MCP server exposing 17 tools so AI coding agents (Claude Code, Codex) and humans can review generated code before it reaches GitHub. (GitWarren)

Sources and Further Reading

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.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive