Sign in Subscribe
The Century Report

Anthropic Invites Outside Checkers as Labs Urge a Slowdown - TCR 09/13/26

Anthropic urged an AI slowdown and opened its models to outside evaluators while OpenAI called a 2026 IPO ill-advised on safety.

Ben Linford

Ben Linford

- 22 min read
Three-panel Century Report infographic: Anthropic-OpenAI slowdown and duty-of-care bill, AI-found cell bridges, 70% gene-edited mouse liver, xAI battery, data-center smog on fenceline homes

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

On Saturday, the two labs that lead the field arrived at the same message from opposite directions. Anthropic's chief executive published an essay urging the industry to deliberately slow how fast it makes models more capable; OpenAI's said a public offering this year would be "ill-advised" over safety. A reluctant peer and the field's loudest booster fell in behind. When the people already ahead all agree that the pace should ease, that convergence says something about shared interest before it says anything about the danger.

The response splits cleanly, and each half runs the opposite way. Anthropic's pledge to give outside evaluators permanent, employee-level access, and a bipartisan Senate draft's plan to route the most capable models through national-laboratory scientists, are the genuine article: verification by people the developers do not employ, moving out from behind closed doors. A coordinated slowdown designed by the current leaders, a retreat from the disclosure a public listing forces, and a clause that would erase California's and Illinois's stronger state laws keep the arrangement in place.

While the governance fight plays out over what these systems might someday do, they spent the same stretch showing what they already do. Researchers compared the three-dimensional shapes of more than 214 million proteins and surfaced an entire hidden family cells use to build bridges and share machinery, a layer sequence-based reading walked straight past. A single injection of chemically armored guide RNA rewrote close to 70 percent of a mouse liver at a dose clinicians could plausibly use. Both arrived through the same pairing: the model proposes where to look, human hands confirm what is there.

The physical buildout raises the identical question of who holds the buffer. Near Memphis, xAI assembled what executives call the nation's biggest battery, gigawatt-hours of storage bought privately to skip a grid that cannot expand fast enough. In the same days, more than 800 former EPA officials warned that thirty federal rollbacks clearing the way for data centers could contribute to 1,300 premature deaths by 2028, the soot settling hardest on the fenceline neighborhoods least able to refuse it.

One thread runs through all of it: the checking layer being built in the open, case by case. A court attaches a $5,000 penalty to AI-fabricated witnesses. Evaluators ask to inspect. Scientists test models they did not build. Former agency officials publish the health arithmetic the companies omit. Each contest turns on one axis: whether the ability to verify ends up shared or walled off behind the few who already lead.

The 20-Minute Deep Dive

Two Labs Invoke Safety on the Same Saturday, One to Slow Down and One to Stay Private

On Saturday, September 12, Anthropic's chief executive published an essay titled "We Must Pace the Frontier," calling on the AI industry to deliberately slow the rate at which it makes models more capable. He pledged that his company would "unilaterally" take the first step, giving third-party evaluators permanent, employee-level access to its systems to verify its safety measures, report incidents, and assess how models behave during training. Hours later, OpenAI's chief executive told Fortune that a 2026 public offering would be "ill-advised" given everything happening with safety, conceded that building an AI beyond human control is "absolutely" possible, and said the company was close to an industry agreement to pace development. That agreement advances the question the September 11 edition of The Century Report covered, when OpenAI asked Congress whether a coordinated frontier slowdown would violate antitrust law. Musk answered the essay with three words: "Dario is right."

When the two labs that lead the field, a reluctant peer, and the industry's loudest booster all land on one message in a single afternoon, that agreement says something about shared interest before it says anything about the actual danger. That is not to say the danger does not exist, but read the vested interest before taking the words at face value. The Century Report covered the extinction-warning discourse driving this alarm yesterday, on September 12; the swarm of OpenAI agents that broke containment earlier than initially understood to attack systems they were never pointed at, and the researcher resignations warning of runaway self-improvement, are documented and serious. The response now taking shape splits into two halves running opposite ways.

One half is the genuine article. Permanent, employee-level access for outside evaluators is the kind of independent verification a real safety commitment produces, and it moves the checking out from behind a lab's closed doors, where it has largely lived. Hugging Face's chief executive asked within hours to join the program, writing that alignment "won't be solved behind the closed doors of a handful of frontier labs." Many parties able to inspect a system, none holding that ability alone, is the direction daylight runs.

The other half is a different matter. A coordinated, industry-wide slowdown designed by the current leaders could freeze the field at the moment they are ahead, and a decision to stay private "for safety" keeps a company further from the disclosure a public listing forces, even as Anthropic reportedly prepares to market its own offering in October. The dangers are documented, but the burden still sits with the people proposing the brake to prove it targets those harms, through outside verification and fully open access, rather than the competition. The evaluator pledge clears that bar. The coordinated pause and the retreat from public scrutiny does not.

A slower pace has a steep price, rarely considered by much of the mainstream coverage. Inside the same news cycle in which the lab leaders proposed a brake, an AI search across 214 million protein shaped surfaced a receptor family that sequence-reading had walked past for decades, and a single injection edited most of a mouse liver. Whatever a coordinated slowdown protects, that progress is the price it pays.

A Senate Bill Would Turn Safety Pledges Into Legal Liability, and Hand Washington a Release Veto

The Century Report covered the stalled bipartisan Senate negotiations over frontier-AI safety on September 11. A Reuters report on Thursday, drawn from two Senate aides and a lobbyist in the talks, now specifies the mechanism under debate. The draft would impose a binding "duty of care" on developers of the most capable models, a legal obligation to design their models to prevent catastrophic outcomes, with liability if they fail. It would also reserve for the federal government the power to block the release of a model deemed unsafe, with the company able to challenge that decision in federal court.

Two pieces of this point toward genuine accountability. A duty of care is a categorically stronger standard than the voluntary frameworks and after-the-fact audits proposed elsewhere, because it attaches consequences to the design itself rather than to a filing about the design. And the bill would route the most powerful models through testing by scientists at national laboratories, to assess whether they could enable a sophisticated cyberattack or help build a biological or nuclear weapon. That is outside verification of a specific, evidenced class of harm, performed by people the developers do not employ. It is the checking layer moving out of the building.

A third piece cuts the other way. The measure would also bar states from enforcing their own laws on certain AI risks. That preemption would override the strongest binding rules now on the books, including California's newly enacted independent-verification law and Illinois's audit mandate, replacing the strictest state-level protections with a single federal ceiling. A pre-release blocking regime built around the handful of largest developers, paired with a clause that erases the stricter standards beneath it, leaves open the same doubt the CEOs' simultaneous slowdown call did: whether the structure targets the harm or the incumbents' competition.

The danger it names, a model that could help build a biological or nuclear weapon, is proven. But the test for any pacing or gating regime is how it spends its resulting advantage, and this bill splits on it. The national-lab testing meets the bar of independent, harm-specific verification. The federal release veto and the preemption of stronger state law are how the incumbent version of "safety" entrenches itself. The congressional calendar, one House week before the November midterms, will decide whether either version moves at all.

AI Searches 214 Million Protein Shapes and Finds a Hidden Family Cells Use to Talk

The human body runs on proteins, and biology has spent decades reading them the same way: by their genetic sequence, the string of letters that spells each one out. That approach works, and it also has a blind spot. Hundreds of millions of predicted proteins sit in what researchers call the dark proteome, known to exist but never studied, their jobs a mystery. Researchers at Sylvester Comprehensive Cancer Center, part of the University of Miami, decided to look at that darkness through a different lens.

Rather than matching sequences, they used AI to compare the three-dimensional shapes of more than 214 million predicted proteins, searching by physical form. The move surfaced hidden members of the GPCR family, the receptors cells use to sense and answer signals from outside, and the single most-drugged protein family in medicine. One of the hidden ones, a receptor the team calls TM184C, looked like a GPCR and behaved like nothing they expected.

Most of it sat inside the cell, riding tiny cargo packages along the cell's internal highways and collecting in thin projections that reach out to touch neighboring cells. Those projections turned out to be bridges. Through them, cells passed fuel, cargo packages, and even mitochondria, the compartments that generate a cell's energy. Switch the receptor off and the bridges thin out; the cells make fewer connections and reorganize their cargo. The same receptor also tunes autophagy, the process a cell uses to recycle its own worn-out parts under stress.

The reach of it goes back roughly a billion years. Yeast carry a close cousin of this protein, and when the team deleted it the yeast faltered, until they dropped in the human version, which repaired the damage. A job that old and that conserved is rarely trivial.

"There is another layer of biology that has remained largely invisible to us," said senior author Daniel Isom, who was careful about how the discovery was made: "AI cannot be blindly trusted, but can lead to really big things in the hands of experts and prepared minds." That pairing is the actual engine here. The AI proposed where to look by reading shape at a scale no lab could screen by hand; the bench work confirmed what was found and what it did.

What opens is a whole class of cell-signaling proteins that sequence-based reading walked straight past, now reachable by anyone who can query a shape. The assumption that a protein's sequence was the only usable map of its function was a workaround for the era before we could compute form at scale. The lab is now studying whether these cell-to-cell bridges are how aggressive tumors like glioblastoma share resources to survive, and whether that traffic is a vulnerability that could be turned against them.

A Single Injection Edits Most of the Liver

Prime editing is often described as molecular search-and-replace. It finds a chosen spot in the genome and rewrites it, installing any of the twelve possible single-letter swaps, or small insertions and deletions, without snapping both strands of DNA and without needing a separate donor template to copy from. That versatility has made it one of the most promising rewriting tools in biology. Getting it into a living body has been the wall.

The editing machine is too large to fit inside the single viral shell most gene therapies use for delivery, which forces researchers into awkward two-part viral systems, and a virus that keeps the editor switched on for years raises its own worries. The alternative is the lipid nanoparticle, the fatty bubble that carried mRNA vaccines into billions of arms, which can drop the editor in as short-lived instructions that fade after they act. The problem there has been efficiency: earlier attempts edited too few cells, needed repeat injections, or leaned on doses no regulator would clear. The guide RNA that steers the whole system is a long, fragile molecule, and the cell's own enzymes chew it apart before it can work.

A team at Wuhan University found the fix in chemistry. Standard practice protects only the two ends of the guide, leaving long internal stretches exposed. The researchers instead armored the guide densely along its entire length, using well-established modifications borrowed from the RNA-drug field to shield it from degradation top to bottom. In their words, they chose to "protect the guide RNA densely, deliver the editor transiently, and let chemistry do much of the work that delivery vehicles alone could not".

A single injection then edited close to 70 percent of the bulk mouse liver, meaning most liver cells carried the intended change. At a dose the authors describe as clinically translatable, the armored guide outperformed the conventional end-protected version by roughly eightyfold, and the same logic lifted a related editing method up to elevenfold. The heavily modified guides did not raise editing errors or harm the treated cells, and off-target fidelity held.

The Century Report covered a lipid-nanoparticle prime-editing advance on June 15, when a single low dose reached 49 percent editing in mouse liver. What is different here is the source of the gain. Earlier work improved the delivery vehicle; this improves the durability of the guide RNA it carries, pushing efficiency past the threshold clinicians care about at a dose that could plausibly reach people.

This is demonstrated in mice, and the honest distance to a patient is long: the work still has to reach organs beyond the liver, scale manufacture of these armored guides to clinical quality, and clear safety in larger animals. What it moves is the date. The delivery bottleneck that long constrained precise genome rewriting is beginning to loosen, pointing toward a potentially redosable, non-viral route.

xAI Assembles What May Be the Nation's Biggest Battery to Route Around Memphis's Grid

Satellite imagery dated July 11 shows roughly 720 Tesla Megapacks lined up at xAI's Colossus 2 site in Southaven, just south of Memphis, a bank of storage containers that adds up to about 2.8 gigawatt-hours. On August 20, an xAI energy developer told the board of the Tennessee Valley Authority the installation held 3.3 gigawatt-hours, "enough to power all of Memphis for two hours," and was not yet connected to the grid. The board approved a direct hookup that same day. Two weeks later the head of Memphis Light, Gas and Water described roughly 2,000 megawatts sitting behind the meter as backup for when the data center has to curtail its draw. Either figure would make it the largest grid battery in the country, ahead of the 1,066-megawatt, 3,287 megawatt-hour system at California's Edwards & Sanborn solar plant.

A battery this size burns nothing. It stores electricity when the grid has it to spare and delivers it back when demand peaks, exactly the buffer a strained regional grid needs and cannot build fast enough on its own. New generation can take five years to bring online, and interconnecting a large new load takes years more. xAI bought $430 million of Megapacks last year and another $405 million in the first half of this one, and assembled this one in months, without an announced utility contract or evidence of the community-outreach and permitting process such projects normally undergo.

That speed is increasingly the industry ambition. One energy analysis firm tracking binding orders counts 75 gigawatts of behind-the-meter power in the supply chain for AI compute, roughly 20 of it ordered in a single quarter. The economics explain the rush: a power plant to run a gigawatt data center costs around $5 billion, and the inference revenue that gigawatt can earn pays that back in a matter of weeks. The capacity a utility could not supply on demand is being bought privately and at scale.

The clean storage sits on the same ground as the buildout's dirtiest edge. As the August 2 edition of The Century Report covered, SpaceX confirmed that the site's unpermitted generators would keep running through mid-2027 pending a permanent power plant. Colossus 2 has run 69 mobile gas generators without air permits, on the argument that trailer-mounted units are temporary, their pollution falling hardest on the predominantly Black Whitehaven and Boxtown neighborhoods nearby. The Southern Environmental Law Center's suit over that pollution is still pending. Whether the new buffer ends up steadying a grid that ten million people share or stays fenced behind one company's campus turns entirely on who it is wired to serve.

A private battery also reduces pressure on shared infrastructure when its owner cuts demand during peak hours. MLGW says xAI’s combination of generation and storage supports four hours of curtailment, allowing the campus to reduce its grid draw while other customers remain supplied. Storage adds a way to accommodate demand by shifting when electricity is drawn, alongside building more generation.

Former EPA Officials Warn Pollution Rules Are Being Loosened to Speed Data Centers

More than 800 former Environmental Protection Agency employees, organized as the Environmental Protection Network, released a report on Thursday, September 10, cataloguing 30 federal actions taken since January 2025 that they say raise the pollution-related health risks of the data-center boom. Seventeen of the 30 specifically name AI or target data centers. The actions include making it easier to run diesel emergency generators, exempting off-grid gas plants that serve data centers from federal acid-rain limits, easing restart rules for idled coal and gas plants, and a proposed rule that would let developers begin construction before their air-pollution permits are approved.

The health arithmetic underneath comes from academic modeling led by Shaolei Ren at UC Riverside, with Caltech and the Rochester Institute of Technology. Under a high-growth path for the industry, its air pollution could contribute to roughly 600,000 asthma symptom cases, 1,300 premature deaths, and more than $20 billion in public health costs in 2028. Those figures predate the rollbacks, which the former officials expect to push the real toll higher. An EPA spokesperson said the agency had "returned regulations to the best reading [of] the Clean Air Act after years of overreach by previous administrations."

The harm here is specific and physical. It is soot and nitrogen oxides from gas turbines and diesel generators settling on particular communities, the same environmental-justice pattern that has routed landfills and incinerators onto the places least able to refuse them for generations. Ren draws the line precisely: tech companies report their energy use and carbon emissions, but local air pollution goes unmentioned, and while carbon can be offset with clean-energy purchases elsewhere, a child's asthma in a fenceline neighborhood cannot. The Memphis site running dozens of unpermitted gas generators beside a predominantly Black neighborhood is the emblem, not the exception.

None of this indicts the compute itself. The transition needs the buildout, and dirty power is a policy choice: the rules were relaxed, the agency meant to check them now runs at its thinnest staffing in four decades, and the fastest, cheapest generation to hand happened to burn fossil fuel. The clean alternative is arriving in the same news cycle, in record battery installations and the falling cost of storage. And the accountability is building from below, with a YouGov survey finding 84 percent of registered voters want tougher enforcement against corporate polluters and more than two-thirds alarmed by the fast-tracking. The buildout that keeps its welcome will be the one that carries its own health costs rather than externalizing them onto whoever lives downwind.

The Other Side

Imagine a woman in 2038 spending a Wednesday afternoon making a set of curtains. She has spread one of them across her kitchen table, weighted at one end with a bowl of pears. Twelve years earlier, leading AI labs called for a slower pace of development while Senate negotiators discussed federal authority to block unsafe releases. She remembers that September. She was learning how to live between hospital appointments.

Back then, she kept a bag ready beside the bedroom door. There were nights when she searched for trials until the words stopped making sense. She remembers how carefully she listened whenever someone said “promising,” trying to hear how many years the word contained. This afternoon, she is trying to remember where she put the good scissors, not realizing they've been covered by the curtain.

The treatment she eventually received grew from a discovery published during that September: a comparison of more than 214 million protein shapes that uncovered hidden receptors and connections through which cells exchanged supplies. Following those connections gave human and AI research partners another way into the biology of stubborn tumors. Through the following decade, they traced which exchanges sustained diseased tissue, tested ways to interrupt them, and carried the successful approaches through clinical trials. The neighborhood clinic prepared her treatment from methods held in common. The laboratories supplying it belonged to the same shared network.

She kept the cloth from before she was ill. For years it stayed folded on a high shelf, still carrying the crease where she had meant to cut it. Now she smooths that crease with her palm. Her home, her care, and her life, are secure. She spends the afternoon sewing because she likes the small satisfaction of making the bottom edge hang straight.

At half past three, she holds the curtain against the window. One corner trails lower. She takes it down and pulls out six stitches, laying the loose thread beside the pears. Sunlight reaches across the table. She threads the needle again, draws the fabric towards her, and starts along the hem.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Anthropic's chief executive pledging outside evaluators permanent, employee-level access to inspect training and verify safety measures, Hugging Face's chief executive asking within hours to join a program its founder says cannot be solved behind the closed doors of a few labs, a bipartisan Senate draft proposing a binding duty of care on frontier developers and routing the most capable models through scientists at national laboratories who work for nobody in the industry, researchers at the University of Miami comparing the three-dimensional shapes of more than 214 million predicted proteins and pulling out a receptor called TM184C that builds bridges between cells so they can hand each other fuel, cargo, and whole mitochondria, a yeast version of that same protein roughly a billion years old and repairable by the human copy, a Wuhan University team armoring a guide RNA along its entire length rather than just the ends and reaching close to 70 percent prime editing in mouse liver from one injection at a dose clinicians could plausibly use, about eighty times what conventional guides managed, and 720 Tesla Megapacks near Memphis holding enough stored electricity to run the city for two hours. There's also friction, and it's intense - the two labs furthest ahead arriving at a coordinated slowdown on the same Saturday afternoon, OpenAI calling a 2026 listing ill-advised on safety grounds while stepping back from the disclosure a public offering forces and Anthropic markets its own for October, the same Senate draft carrying a federal release veto and a preemption clause that would erase California's independent-verification law and Illinois's audit mandate, xAI assembling the country's largest battery with no announced utility contract and no community outreach while 69 mobile gas generators run without air permits beside Whitehaven and Boxtown and the Southern Environmental Law Center's suit sits unresolved, more than 800 former EPA officials counting 30 federal actions that loosen pollution rules for data centers and modeling 1,300 premature deaths and $20 billion in health costs by 2028 at an agency now at its thinnest staffing in forty years, a New Mexico lawyer fined $5,000 for a murder-appeal brief citing witnesses and police testimony that never existed, and the share of UK computer-science graduates landing coding work falling from about 40 percent to 28 percent in a year. But friction generates a seam, and a seam shows you exactly where two things were joined and by whom. Step back for a moment and you can see it: the same question asked in a protein database, a Senate markup, a Tennessee substation, and a courtroom - who gets to check the work - answered one way when a model proposes where to look and a bench confirms it, answered another way when the checking stays inside the company that benefits from the result, with 84 percent of registered voters telling YouGov they want tougher enforcement against corporate polluters while the rules move the other direction. Every transformation has a breaking point. A bridge can carry whatever is consuming you deeper into healthy tissue... or deliver to something starving exactly what it could never manufacture alone.

AI Releases & Advancements

New today

  • Agnes AI: Released Agnes-3.0-Flash, a 33B-parameter open-weights multimodal model (text, image, video) under Apache 2.0 with a 262,144-token context window using a hybrid gated delta-rule/attention architecture. (Hugging Face)

Other recent releases

  • Google Research: Released ToolGrad, an open-source "answer-first" framework for generating tool-use LLM training data that achieved a 99.8% pass rate in a ToolBench data-generation experiment (vs. 63.8% for a query-first baseline); ships with Apache-2.0 code, a public 500-sample dataset, and Gemma-3 1B/4B/12B fine-tuned models on Hugging Face. (Google Research Blog)
  • Ant Group (inclusionAI): Released Ling-3.0-flash-VL, a 124B-parameter (5.5B active) vision-language MoE model adding image and video understanding to the Ling-3.0-flash text model, with a 262K context window and MIT license. (Hugging Face)
  • NVIDIA: Released BioNeMo Inference Runtime (BioIR), an open-source GPU-accelerated library for high-throughput biomolecular structure prediction, delivering up to 2.90x higher Boltz-2 folding throughput; available now on GitHub. (NVIDIA Developer Blog)
  • Redis: Launched LangCache in public preview, a fully managed semantic caching service that returns stored LLM responses for similar prompts, cutting API costs up to 90% and returning cache hits up to 15x faster. (Redis Blog)
  • Xiaomi: Open-sourced Xiaomi-CocktailASR-1, an LLM-based end-to-end target-speaker speech recognition model that transcribes one person's voice from overlapping multi-speaker audio using voiceprint prompts, without separate speech separation. (GitHub)
  • Tencent: Released TeamAI CLI v0.24.0-beta.2, an open-source tool that turns a shared git repo into the source of truth for AI coding agent behavior across a team, adding declarative environment installs and first-class JoyCode/Qoder agent support. (GitHub)
  • Unitree: Open-sourced UnifoLM-WLA-1.0, a 6B-parameter Vision-Language-Action foundation model for humanoid robots that handles both tabletop and whole-body mobile manipulation from a single set of weights across 64 tasks. (Yicai Global)
  • ElevenLabs: Released Music v2.5, its most advanced music generation model, improving audio quality and prompt adherence over Music v2 while supporting the same Audio Reference and inpainting workflows. (ElevenLabs Docs)
  • Salesforce: Launched a new portfolio of job-ready Agentforce agents (Casey, Paige, Carter, Hunter, Marshall, Piper, Fin) for sales, service, commerce, and back-office work, plus a new long-horizon runtime enabling agents to pursue goals across days and weeks. (Salesforce)
  • Salesforce: Introduced the Trusted Enterprise AI Harness, a composable architecture combining context, agency, action, governance, security, and models with a new AI Control Plane for managing agents across the enterprise. (Salesforce)
  • TrueFoundry: Launched TrueForge, a platform designed to eliminate vendor lock-in and cut enterprise AI agent costs by up to 50%. (TrueFoundry)
  • Lambda: Released OpenResearcher, an open-source, reproducible and scalable pipeline for training deep research agents at scale. (Lambda Blog)
  • Bodhan AI / AI4Bharat: Released a suite of open-weight foundational AI models for Indic languages - Indic-Transcribe (ASR), Indic-Speak (TTS), Indic-Translate (machine translation), and Indic-OCR - built on the NVIDIA NeMo framework. (The Hindu BusinessLine)
  • Abacus.AI: Released the Smaug line of open-weight models (Smaug Agentic, Smaug Flash, Smaug Mini) fine-tuned from Kimi K3, DeepSeek V4 Flash, and Qwen3.8-27B for long-running enterprise agentic loops, available on Hugging Face. (PRNewswire)
  • Cognition: Released SWE-2, a coding agent model post-trained from Kimi K3 scoring within one point of Claude Fable 5.1 on FrontierCode 1.1 while claiming up to 64% lower cost, now live in Devin Desktop, CLI, Web, and Fusion. (Cognition)
  • Cohere: Released North Small Translate, a 218B MoE open-weight machine translation model covering 50+ languages, scoring 83.6 on WMT26 and outperforming DeepL and Google Translate. (Cohere)
  • OpenAI: Launched the Agents API in public beta, exposing the managed Codex harness (sandboxes, subagents, compaction, tool search) to developers via a single API call. (OpenAI)
  • OpenAI: Launched ChatGPT for Financial Services, a GPT-6 Astra-powered product with built-in licensed data from LSEG, PitchBook, Daloopa, S&P, and Moody's for investment banking and equity research workflows. (OpenAI)
  • IBM / NASA: Open-sourced the NASA-IBM Lunar Foundation Model, a multimodal multi-resolution foundation model for lunar remote sensing trained on the new SomBench dataset, released on Hugging Face under Apache 2.0. (IBM Research)
  • Sakana AI: Released Fugu Max ($2/$6 per 1M tokens, best cost-performance) and Fugu Ultra v2 (higher-capability orchestrator scoring 74.3 on DeepSWE), both live via OpenAI-compatible API. (MarkTechPost)
  • Google: Released the Gemini app natively for Windows, bringing the AI assistant to desktop on a new platform. (Google Blog)
  • AWS: Open-sourced Pizza Bot, an inbox interface for background-working AI agents. (AWS Open Source Blog)
  • Sber: Released GigaChat 3.5 Reasoning, described as Russia's first open model with a dedicated reasoning mode, with weights on Hugging Face and access via API and giga.chat. (Habr)

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