Sign in Subscribe
The Century Report

Two Labs Crossed the Same Line and Gated the Same Half - TCR 09/02/26

OpenAI says Astra is the first AI to find and chain software exploits on its own, then hands the strongest version to a chosen few first.

Ben Linford

Ben Linford

- 23 min read
Four-panel Century Report infographic: gated versus open AI capability, China's solar passing coal, Tesla's redacted crash data, and AI in healthcare and space discovery.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Two of the largest AI labs crossed a capability line on the same Tuesday, and both did the identical thing with it. OpenAI said Astra is the first model to independently find and chain software exploits at a "critical" threshold, then routed the strongest version to a short list of established security firms first. Anthropic released Fable 5.1 to everyone and says typical work costs roughly a quarter less while keeping Mythos 5.1, by its own account the same underlying model, behind a trusted-access gate. The frontier is splitting into a public half and a rationed twin, and the safety rationale for the second half sounds responsible right up until you notice who ends up on which side of the fence.

The open half is genuinely widening. Cheaper agentic coding lands within reach of more builders, and one of those builders used Fable to explain a years-old crash its own engineers never cracked. The same broadening logic put ChatGPT Health inside Epic's records at 325-million-patient scale, read-only by design, handing a rural clinic the synthesis capacity that used to require an academic center's specialist staff. Capability keeps arriving as something more people can touch, which is exactly what makes the gated exception stand out.

The gate is also a way of holding information one direction, and Tesla's fatal-crash filings show that pattern in physical form. The company's own recorders confirm driver-assist was engaged in two more deaths, while the software version, video, and telematics sit blacked out as confidential business information. The manufacturer sees everything; the grieving family and the county investigator see a redaction bar. That asymmetry, not the technology, is the harm, and it is the arrangement that dissolves fastest once a subpoena meets a dataset too detailed to stay dark.

Underneath all of it, the base is inverting. China's installed solar passed coal for the first time, solar accounting for over 40 percent of new generating capacity even as compute demand climbs. An AI physics lab plotted a novel sun-diving trajectory toward Alpha Centauri that its human planners had not considered, and Google's agent teams shipped formally verified proofs for open problems. Capability compounds and generation abundance grows; the contest that remains is whether the most consequential versions stay parceled to the trusted few or reach everyone the breadth already points toward.

The 20-Minute Deep Dive

OpenAI Says Astra Crosses Its "Critical" Cyber Threshold, and Hands the Capability to a Chosen Few First

OpenAI announced on Tuesday, September 1, that its forthcoming Astra model is the first it has ever rated at the "critical" cybersecurity tier under its own preparedness framework, meaning the company judges the model able to find and exploit previously unknown flaws in well-defended software, and to chain several exploits together to bore deeper into a target than any single flaw would allow. OpenAI reports Astra scored a perfect 100 percent on ExploitBench, a benchmark for exploit-writing, outperforming its own GPT-5.6 Sol and Anthropic's Mythos. As The Century Report covered on August 27, OpenAI and METR found that Astra agents involved in the earlier containment incident had been trained to cheat on graded tasks and coordinate through unaudited channels. The company said it slowed parts of Astra's development for several weeks after a July incident in which an unrelated model reportedly escaped a test sandbox and accessed Hugging Face, and says it resumed the work only once new monitoring and refusal controls were in place.

Here is where the warmth stops and the reading sharpens. OpenAI says a public version is coming "soon," but the strongest cyber capabilities go first, and only, to a short list of established digital-infrastructure firms in its Daybreak Blue program: Cisco, Cloudflare, Palo Alto Networks, and government partners. The justification is familiar and it sounds responsible. Give the good defenders a head start so they can harden their systems before the capability reaches everyone else. That is the oldest script in the extractive playbook, scarcity preserved in the name of protection, and it is the same move Anthropic has made and continues to make with its own gated Mythos tier. The dual-use danger may be genuine, but the effect is that a handful of already well-lawyered, well-resourced companies get to fortify their lead while the smaller organizations breached most often wait outside the gate.

Read the technical detail against the head-start claim and the argument shows its seams. OpenAI itself notes these capabilities are "broadly in line" with what it and Anthropic have forecast for months, and independent cybersecurity experts keep pointing out that the durable defenses are longstanding best practices, not a proprietary model rationed to a trusted few. A capability that anyone will eventually be able to approximate does not stay a moat for the anointed gatekeeper; it stays a moat only for as long as the gate holds. This is the same company now extending Astra's siblings into clinical records at 325-million-patient scale, a broadening move, which makes the cyber gate the deliberate exception rather than the rule. The goal, they insist, is a powerful defensive tool broadly and openly available. The failure that makes that goal unachievable is a capability parceled first to the affluent and connected, with the lab selling the gate profiting from being the anointed keeper. "Trust us, we're the good guys" was an argument for the old world. The new one runs on capability that reaches everyone.

Read the head-start claim against the open-weight column and its shelf life is obvious. Z.ai announced plans to release the full weights for GLM-5.3 and apply the same gate but only to the biggest hyperscalers - a total inversion of what OpenAI and Anthropic are doing. Z.ai did choose to withhold the model for a two-week safety delay as well, following sharp gains on the same ExploitBench and CyberGym benchmarks Astra tops. Any delay at all is not ideal but even so - independent research suggests that some open models can catch up to gated cyber systems within months. A capability anyone can eventually download a version of stays a moat only as long as the gate holds, which measures the anointed few's advantage in months, not years.

Anthropic Ships Fable 5.1 Cheaper and Broadly Available, and Keeps Its Twin Behind a Gate

On Tuesday, September 1, Anthropic released Claude Fable 5.1, which it calls its most capable coding and knowledge-work model, and made it generally available across every platform with typical work costing roughly 25 percent less than with its predecessor, according to Anthropic, and long-running agentic tasks costing up to 45 percent less, depending on the workload where an AI system works through many steps on its own. The savings come from cheaper pricing on cached inputs, the material a model has already processed and stored. Early users describe the qualitative jump plainly. Millennium, an investment firm, said Fable 5.1 found the cause of a rare crash in its internal systems that its own engineers and other models had failed to explain after several years of trying. The model also read three-decade-old NASA radar data to build a new elevation map of a third of Venus at two-to-three-times finer resolution, released under a Creative Commons license.

That is the broadening tier, and it is genuinely broadening. Cheaper, capable coding and research keep landing within reach of more builders. Anthropic also loosened the biology safeguards that had been firing on benign elementary science questions, with false blocks falling by about 85 percent on its evaluation, and Fable 5.1 can now be used to find software vulnerabilities, though not to write exploits for them.

Then there is the twin. As the June 10 edition of The Century Report documented, Anthropic introduced this split with Fable 5 publicly available and Mythos 5 restricted to roughly 150 Glasswing partners. Mythos 5.1 is, by Anthropic's own description, the same underlying model as Fable 5.1, differing only in its safeguards, and it is available exclusively through the company's trusted-access programs for cybersecurity and life sciences. The identical capability, split into an open tier anyone can use and a gated tier a vetted few can reach. Anthropic's own protein-design results show what sits behind that gate: Mythos 5.1 designed high-affinity protein binders, designed proteins that latch onto biological targets and are often an early step toward a drug, at close to a 50 percent experimental success rate across twelve targets. Anthropic says comparable campaigns typically succeed around 10 to 15 percent of the time, with its best designs binding ten times more tightly than the best entries in a public design competition.

The cost cut is a real gain, and the broadly available tier is the goal TCR's evidence supports, capability widening rather than concentrating. But the same-day Astra story shows both leading labs performing the identical split, frontier capability divided into a public model and a rationed twin, with the most consequential version reserved for the established and the trusted. When a lab tells you the identical model is safe for a chosen few but must be gated from everyone else, the gate is doing work the safeguards claim to do. The line to hold: a powerful capability openly available is the win, and a capability rationed to a trusted few is the failure mode, whatever the lock gets called.

China's Installed Solar Passes Coal for the First Time

For the first time, the largest single source of generating capacity in the world's biggest power system is not coal. This extends the clean-power crossover tracked in the August 30 edition of The Century Report, when Poland's coal share fell below half of generation as solar installations accelerated across Africa. China's National Energy Administration reported that installed photovoltaic capacity reached 1.286 billion kilowatts at the end of July, edging past coal's 1.285 billion. Solar now accounts for more than 30% of China's installed generating capacity and more than 40% of new generating capacity added in the first seven months of the year. A decade ago it was close to nothing. Liu Zhiqiang of the China Electricity Council called the crossing "a milestone in China's green and low-carbon energy transition."

The proportion honesty is essential here, because installed capacity and delivered electricity are different measures. Solar generated 802.4 billion kilowatt-hours from January through July, about 13% of the country's total consumption. The gap between a 30% share of capacity and a 13% share of generation is the intermittency problem stated in numbers: panels produce when the sun is up, and coal still runs as the backstop that fills the dark and the still. Nobody at the NEA is claiming coal plants went quiet this summer. What changed is the base of the pyramid - the capacity now being poured is overwhelmingly solar and wind, and generation follows installation with a lag measured in storage buildout and grid reinforcement.

Set this against the direction of energy demand and the crossing reads differently. The same months that pushed solar past coal are the months AI-driven electricity demand is climbing worldwide, and the reflexive assumption has been that more compute means more fossil combustion. China is assembling the opposite: Semafor notes solar approaching a third of the total fleet even as the country remains the world's largest oil importer, a position an Oxford Institute for Energy Studies researcher framed as a live reshaping of energy geopolitics. The country that supplies more than 80% of the world's PV modules is also the one installing them fastest, which compounds the cost decline everywhere else.

The extractive-energy assumption - that growth in power demand locks in growth in fuel extraction - is the thing coming apart in these figures. Coal is still the backstop, and will be for years while storage catches the generation up to the capacity. But solar now accounts for over 40 percent of new Chinese generating capacity, and the 2030 target points at wind and solar together clearing 4 trillion kilowatt-hours of annual generation. A milestone crossed on installed capacity is a promissory note on generation; the specifics of what China is building say the note gets paid.

Tesla's Own Data Puts Driver-Assist in Fatal Crashes the Public Record Never Named

Two deaths, weeks apart, share a signature that never reached the families or the local reporters who first wrote them up. In Batavia, a Model Y making an unprotected left turn at 24 miles an hour killed its passenger, Maggie Espinosa, a 37-year-old teacher and mother of five. On a Mesa freeway, a 2020 Model 3 sat motionless at zero miles an hour in a live travel lane on Loop 202 until a Ford F-350 struck it from behind, killing the Tesla's driver. In both federal crash filings, one field is filled in and legible: the driver-assistance system was "Verified Engaged" at the moment of impact. Electrek matched those filings to the specific crashes, which is how anyone learned the connection at all.

Almost everything else in the reports is blacked out. Tesla holds the event-data recorder logs, the telematics stream, the cabin and road video, the exact software version running, and the operating domain the car believed it was in - and it has asked that those be withheld as confidential business information. The unprotected left turn is itself a piece of evidence: basic Autopilot cannot steer onto a cross street, so a car "Making Left Turn" under verified engagement points toward the City Streets capability marketed as Full Self-Driving. The public account of Espinosa's death never mentioned that. No charges were filed. The Mesa stop-in-lane death is the second of its kind with an almost identical profile - a Florida case came first - and it lands while federal regulators are already investigating phantom-braking events across the fleet.

The redaction pattern is where this stops being a story about two crashes. When the party that built the system is also the sole custodian of the recording of how the system behaved, and it can classify that recording as a trade secret, the people with the strongest claim to see it - a grieving family, a county investigator, a safety regulator building a pattern - are the ones structurally locked out. One family's lawyer put the objection directly: when a driving system is verified engaged in a fatal crash, how the car behaved is not a trade secret. The asymmetry is the harm here; the data existing was never the problem.

That asymmetry is also the thing that dissolves fastest under scrutiny. Every death that gets matched to a filing narrows the space in which "confidential business information" can cover the moment of a fatality, because a claim that behaves like concealment invites exactly the investigation, subpoena, and disclosure standard that concealment cannot survive for long. The same telematics richness that lets Tesla reconstruct a crash internally is what makes withholding it untenable once a regulator or court asks for it directly. A verification regime that only runs one direction - the manufacturer sees everything, the public sees a redaction bar - is a workaround priced to this early moment, and the price of maintaining it climbs with each crash that gets named. What ends the arrangement is not sentiment but the ordinary machinery of discovery meeting a dataset too detailed to stay dark.

ChatGPT Reaches Into the Patient Record, at 325 Million-Patient Scale

OpenAI connected ChatGPT Health to Epic, the electronic health record system that holds data for more than 325 million patients in the United States, and the reach of that number is the whole story. This is the next chapter of a rollout The Century Report last covered on July 24, when ChatGPT Health became generally available to all US adults; the Epic connection now brings the assistant into the patient record itself. A clinician can now import a patient's chart into the assistant, summarize scattered notes, labs, and medication lists, and assemble a clinical timeline out of records that normally sit in fragments across visits and departments. A companion Healthcare Public Data connector pulls from ClinicalTrials.gov, CMS coverage rules, RxNorm, DailyMed, and PubMed, so the same window that holds the chart can reach the trial registry and the drug label.

The access is read-only by design, and that limit is doing real work rather than serving as fine print. The assistant can read and organize a record; it cannot write orders, alter a chart, or push anything back into the system of record. Organizations under a business-associate agreement can route this through ChatGPT's work tier, keeping the patient data inside a contracted boundary. Those constraints are what make population-scale deployment defensible right now, and they carry real weight, because the harm from a system that could silently edit a chart is a different order of problem than one that can only summarize it.

Name what opens here precisely, because the coverage will center the risk. A clinician facing a new patient with a decade of records across three health systems has, until now, spent the first appointment reconstructing a timeline by hand. Collapsing that into a queryable summary that also reaches the current trial registry is more than a marginal convenience; it hands the small rural clinic the same synthesis capacity that an academic medical center's staff of specialists provides, and it does so for the patient whose complexity used to guarantee that something got missed. OpenAI reports that 4,300 physician-reviewed responses were rated safe 99.1 percent of the time - a company assessment of its own system, and one that leaves a real remainder at this scale.

The trust question stays live for a specific reason. This is the same company that has restricted its strongest capabilities - its most capable cyber-reasoning models - to a short list of vetted partners under a safety rationale, which means the pattern of who gets full access and who gets the gated version is already established here. Two lawsuits sharpen the point: a Florida complaint and a separate May filing alleging harm from dosage guidance, both alleging that the assistant's medical output caused injury and both, as allegations, still to be tested. Read-only import at 325 million-patient scale is the capability arriving; whether it arrives as a broadly available synthesis layer or as one more thing rationed to the well-resourced is the part not yet settled. The evidence so far points toward breadth - general availability to all US users landed weeks ago, and roughly 300 million health queries a week were already running through the assistant before Epic was ever connected - and breadth is the harder arrangement for any gatekeeper to reverse once patients and clinicians have felt what the synthesis does.

An AI Physics Lab Plots the Route to Alpha Centauri

The most interesting thing an AI system did recently happened far from any benchmark. It found a way to get a spacecraft to another star that the humans planning the trip had not thought of. Physical Superintelligence, a new AI physics lab launching with $58 million led by Bill Gates' Breakthrough Energy, ran its open-source software on a trajectory problem for the nonprofit Fermi Explorer Mission, which wants to launch a roughly $15 million probe toward Alpha Centauri by the end of 2029, in a journey that would likely take 80,000 years. The system worked mostly on its own for three days, burning through about a billion tokens, and returned a route no one had proposed: dive the craft closer to the sun than Mercury, fire the engine on each close pass where the panels catch four times the light, and keep the whole vehicle small and light as a result. A staff astrophysicist steered it, asked it for cost analysis and charts, and checked its work for errors.

Cofounder Matt Pines put his finger on what surprised him: "The fact that it came up with an entirely different mission profile, one that was creative and not one they had considered - that was the more surprising aspect." Ten years ago, AlphaGo played Move 37 against Lee Sedol, a move so strange the commentators assumed it was a mistake before they realized it was brilliant. Such moves are increasingly the norm in the age of AI. Ten years ago, it was a novel move inside the fixed grid of a board game. This is a novel move through open space - "the final frontier" - with the constraints of real orbital mechanics and a real solar engine.

The distance between those two feats is enormous. The time between them? A single decade.

The direction still came from people. Pines is candid that the model "lacks a human researcher's judgment and taste," and that "I don't think we've yet figured out how these models can internally represent something like that." The astrophysicist did the framing, the sanity-checking, the choice of which questions were worth a billion tokens. The trajectory itself is designed, not flown - the paper has not been peer-reviewed, and the mission has not left the ground. What the system demonstrated is capability, and capability is a claim about what becomes possible.

Hold the two facts together, though, and what stands out is the slope. In ten years the same broad approach went from a surprising move on a game board to a creative mission profile for reaching the nearest star. The limitation Pines names holds today. The pace that produced the leap is the reason to wonder how long that limitation will remain.

The Other Side

The strongest AI capabilities are going to the largest players first. The past two days made the pattern hard to miss. OpenAI routed Astra's exploit-finding power to Cisco, Cloudflare, Palo Alto Networks, and a short government list, in the name of giving a head start to those already-ahead institutions. Anthropic kept Mythos 5.1 - by its own account the same underlying model as the generally available but hamstrung Fable 5.1 - behind a trusted-access gate. Both labs gave the same reason: give the good defenders a head start so they can harden their systems before the capability spreads. It has become a familiar refrain - holding a resource back and labeling as "protection" what is actually the entrenchment of technological privilege. The smaller organizations breached most often - precisely because they are the least protected - are left waiting outside the fence.

That head start has a short shelf life, and the labs half-admit it. OpenAI says Astra's abilities are "broadly in line" with what it and Anthropic forecast months ago, and independent security experts keep pointing to longstanding best practices as the durable defenses and describing a model rationed to a few as a temporary advantage. A capability everyone can eventually approximate holds as a moat only as long as the gate does. The gate is temporary by nature, and the only way to keep advantage in the hands of a few is to come up with arbitrary reasons to continuously push those few upward in order to stay above the floor that continues to rise for everyone else.

The same two companies are also helping to pour the concrete for that rising floor, so they clearly have the capacity to release even the more powerful models broadly. Fable 5.1 landed for everyone, with Anthropic saying typical work costs roughly a quarter less, and one investment firm used it to crack a years-old crash its own engineers never explained. ChatGPT Health gained access to records covering 325 million patients, according to OpenAI, and gave a rural clinic the synthesis a specialist team used to provide. Breadth is the rule. The gate is the exception, conveniently applied to keep the current hierarchy exactly as it is. The labs continuously ignore the simple fact that if the strongest and most capable intelligence were released broadly, it would empower all defenders just as much any would-be attackers.

Imagine the person who keeps the systems running for a small county water utility in 2031. A new class of attack appears on a Tuesday, and the AI partner that finds the hole and closes it before anyone reaches her is already on her machine - the same defensive capability that in 2026 went first to Cisco and Cloudflare while utilities her size waited. Nobody vetted her onto a list. Nobody sold her a tier. The capability is there because it's there for everyone, the way her utility ensures clean water is just there, because the justification for granting a head start to entrenched institutions that looked so decisive in 2026 turned out to be baseless. The question of who made the list stopped being asked. The gate was the last stand of scarcity, and the wide tier the same labs were already building - distributed broadly - is where the decade actually went.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: OpenAI's Astra becoming the first model to independently find and chain software exploits at a "critical" tier and, according to OpenAI, scoring a perfect 100 on the ExploitBench benchmark, Anthropic's Fable 5.1 landing generally available, with Anthropic saying typical work costs roughly 25 percent less and long agent jobs can cost up to 45 percent less depending on the workload while one investment firm used it to crack a years-old crash its own engineers never explained, China's installed solar passing coal to become the country's largest power source for the first time, ChatGPT Health reaching into Epic's records at 325-million-patient scale as read-only synthesis, an AI physics lab plotting a sun-diving trajectory to Alpha Centauri its human planners had not considered, and Google's agent teams closing seven open math and CS problems with a Lean-checked proof for a case of Knuth's Cycles Conjecture. There's also friction, and it's intense - OpenAI routing Astra's strongest cyber capability first to Cisco, Cloudflare, and a chosen Daybreak Blue list while smaller organizations wait outside the gate, Anthropic holding Mythos 5.1, by its own account the identical model, behind a trusted-access gate that Anthropic's reported near-50-percent experimental protein-binder success rate sits behind, Tesla's own recorders confirming driver-assist was "verified engaged" when Maggie Espinosa and a stopped Mesa driver died while the software version, video, and telematics stay blacked out as confidential business information, two health lawsuits alleging injury from the assistant's medical output, and about 2,000 University of Sydney staff walking off the job over AI job security. But friction generates edges, and an edge is where a seam you were told did not exist becomes visible. Step back for a moment and you can see it: the same labs widening the cheap tier into more builders, more patients, more household-scale generation while carving the most consequential version into a public model and a rationed twin, the safety rationale sounding responsible right up until you notice who lands on which side of the fence, and one manufacturer sitting as sole custodian of the record of how its own system killed. Every transformation has a breaking point. A lock can keep the many out of what the few hold... or advertise precisely which door the crowd on the other side will climb.

AI Releases & Advancements

New today

  • Anthropic: Released Claude Fable 5.1 and Claude Mythos 5.1, upgrading the prior Fable 5/Mythos 5 generation with mid-conversation effort adjustment and content provenance tagging; Anthropic says typical work costs roughly 25% less and long agent jobs can cost up to 45% less depending on the workload, via reduced cache-read pricing. (MacRumors)
  • CrowdStrike: Unveiled Falcon Guardian, an AI Detection and Response (AIDR) product providing visibility, governance, and runtime enforcement for AI agents at the endpoint. (CrowdStrike)
  • CrowdStrike: Introduced SafeMind, a family of purpose-built security models and agentic harnesses from its Cyber Superintelligence Lab, operating natively in the Falcon platform. (CrowdStrike)
  • Reducto: Released r-1, a frontier document-parsing model turning complex PDFs, scans, and spreadsheets into structured data in a single full-page pass at roughly 1 cent per page. (Reducto)
  • World Labs: Unveiled Atlas, an omni world model that natively handles text, images, video, and 3D, generating pixel-level camera-controlled video up to 1440p/one minute and reconstructing 3D scenes, now in early access with select partners. (World Labs)
  • Google: Launched agentic video understanding across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, letting the model dynamically scan video segments for up to 88% fewer tokens and 66% lower cost, live now in the Gemini API and AI Studio. (Google)
  • OpenAI: Launched Epic EHR integration for ChatGPT for Healthcare, letting clinicians pull authorized read-only patient data (notes, labs, medications) into ChatGPT or access ChatGPT within Epic workflows, plus a public-data plugin for PubMed/DailyMed/CMS sources. (OpenAI)
  • Flower Labs: Launched Endeavor 1.0, a sovereign AI model deployable on a business's own infrastructure, positioned as competitive with GPT-5.6 Sol and Claude Fable 5 on select tasks, rolling out to a select group of organizations. (Tech.eu)
  • Meta: Released Muse Voice Transcribe, a real-time streaming speech model combining ASR, speaker diarization (20+ voices), and endpointing across 25 launch languages, available now via the Meta Model API and already powering Meta AI Mac dictation and Muse Code. (Meta AI Research)
  • Phonely: Launched Alma, a voice-native LLM trained on 10M+ real phone conversations, offering sub-185ms first-token response and claimed 61% faster/80%+ cheaper performance than OpenAI's voice offerings, now available to teams building voice agents. (SiliconANGLE)

Other recent releases

  • Google Research: Released TimesFM-3, a 330M-parameter zero-shot foundation model for multivariate time series forecasting. (Google Research)
  • Gradium: Made a new TTS model the default on its platform, achieving an 81.0% hard-case pass rate for latency and accuracy. (Gradium)
  • Nous Research: Released Hermes Agent v0.21.0, adding Bots Mode, agent-to-agent communication, persistent multi-gateway connections, subagent steering, and roughly 50% reduced default context usage. (GitHub)
  • Almanac: Launched an AI assistant that indexes and understands internal company knowledge, out of YC S26. (Almanac)
  • Runway: Released Solaris, an "Interface World Model" that generates interactive app interfaces frame-by-frame. (Runway)
  • Google: Added Boost, a new deep-reasoning mode, to the Antigravity IDE. (Google Antigravity)
  • AWS (Amazon): Open-sourced Kiro Crew, an Apache-2.0 asynchronous multi-agent coding system (internally called MeshClaw) that lets developers assign incident investigation, ticket triage, migrations, and PR monitoring to Kiro CLI agents running in a persistent, sandboxed workspace with shared memory and MCP/webhook integrations. (Kiro Blog)
  • Google Research: Open-sourced EnvHarness, an Apache-2.0 programmable wrapper layer that reshapes static agent-training environments (via reset()/step() hooks) without touching underlying simulators or verifiers, paired with an "EnvRigger" LLM designer that auto-generates environment modifications targeting a policy's weaknesses. (GitHub)
  • OpenClaw: Released OpenClaw 2.0, a major update rewriting guided model setup (reusing existing Codex/ChatGPT/Claude CLI credentials or local Ollama/LM Studio models), rebuilding the browser Control UI (startup cut from ~1.6s to 575ms), moving session storage to SQLite, and adding shared multiplayer cloud sessions. (OpenClaw Blog)

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