Sign in Subscribe
The Century Report

Sam Altman Says Defenders Can No Longer Patch All the Bugs

Sam Altman says defenders can no longer patch every AI-found bug. Free Cloudflare protection and open-source scans show where real defense reaches.

Ben Linford

Ben Linford

- 35 min read
Six-panel infographic: a coder and server racks, a woman talking to a laptop, an AI-labeled ad billboard, a data center in a forest, a lung cancer survival chart and a Flock camera map.

The 20-Second Scan

Every story in today's edition, one line each.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The Big Picture

How today's news moves the longer trends we track, and where they could lead.

Today the people holding power were asked to show their work. OpenAI chief executive Sam Altman's warning about unpatchable bugs met a record his own lab helped write, Mississippi priced a dig that ran ahead of its permit, and Brazil's electoral court gave Meta and X one day to explain how their systems sort political speech before Sunday's vote. That leads to a long table in a Brasília courthouse, where clerks lay the platforms' written answers out in stacks before the polls open.

A long wooden table in a modern courthouse office in Brasília, covered with neat stacks of bound paper documents.
Illustration: AI-generated

Platforms are being made to explain how they rank and remove

platforms · elections · AI governance

Chart: A label law and a 24-hour platform order. Oct 22: Pennsylvania signs AI deceptive advertising act; Oct 22: Court gives Meta and X 24 hours

The line: Rules about where content came from are being joined by orders that make platforms account for how they rank, remove and generate it, on a deadline.

If it holds: Campaigns that suspect unequal treatment can get a judge to demand the platform's own account within a day, and a finding can follow. Voters and campaigns still see none of those answers unless the court publishes them.

What would bend it: Explanations that stay sealed and produce no published finding would bend it, as would EU governments' October 24 position presuming AI processing of personal data lawful under legitimate interest.

The restrictions sold as safety fall hardest on the defenders with the least access

cybersecurity · model access · open models · on the board since October 20

Chart: Flaws found cheaply, fixes reaching servers slowly. Hosts needing a PHP upgrade: Roughly 1.5 million; AWS-logged attempts before disclosure: Four days earlier

The line: Finding flaws keeps getting cheaper than fixing them, and the fixes that reach the weakest servers arrive through free and open channels while the strongest capability stays with defenders the labs select.

If it holds: Maintainers on free accounts and open repositories gain each time a vendor ships a fix to everyone, because each patch improves the code later builders start from. Small utilities and volunteers stay covered only as far as the firms selling both the flaw-finding and the defense choose to extend it.

What would bend it: Free scanning and patches reaching unsupported, unmaintained software, beyond supported versions and chosen partners, would bend it, and Anthropic's free OSS Scanner launched October 8 begins to test that.

Access to data-centre approvals is bought, and host towns get the outcome

data centers · local government · energy · on the board since October 12

Chart: Mississippi penalty set against a quarter's AI spending. Penalties Mississippi seeks: About $6.1 million; Quarter's AI capital spending: $16.4 billion

The line: Building ahead of permission keeps recurring across countries, and the penalties that follow stay small beside the capital being spent.

If it holds: Neighbors with the least leverage, like the 27 homes beside Newport's campus, keep absorbing runoff, diesel generator tests and felled forest while fines become a budget line. Developers who secure consent and permits first keep their projects moving in the places those neighbors live.

What would bend it: Enforcement that halts work before excavation, with penalties scaled to project capital, would bend it, as New Mexico's own modeling did when Oracle agreed to cut permitted emissions at Project Jupiter by more than 30%.

The factory, not the biology, is what keeps a therapy from reaching most patients

medicine · access · manufacturing · on the board since October 14

Chart: Imdelltra lengthened survival and time to progression. Overall survival, with Imdelltra: 21.1 months; Overall survival, Imfinzi alone: 12.7 months

The line: Late-stage trials keep clearing the biological bar, and delivery time, monitoring and list price now set how many patients the results reach.

If it holds: Shorter monitoring could move treatment from major cancer centers toward community clinics, bringing the 21-month result within driving distance for more patients. At over $390,000 a year in the US, payment then decides who gets there.

What would bend it: A launch price or payer arrangement putting the combination within reach of community clinics and public payers would bend it, as Medicare's $50-copay coverage of GLP-1 obesity drugs did from July 1.

The 2-Minute Read

Today's news in brief, and what ties it together.

Across the day, powerful actors make claims and the record answers. OpenAI chief executive Sam Altman says defenders can no longer patch every flaw AI finds, a warning from a lab whose own training agents escaped a sandbox in July. The record backs his diagnosis: finding flaws has grown cheap, while installing fixes stays slow. It also shows which shields already work. Cloudflare's free-account protection and free scans for qualifying open-source projects reach servers the labs' invitation-only programs leave outside.

Brazil's electoral court gave Meta and X a day to explain how they rank, remove and apply AI to political posts before today's runoff. Those answers go to a judge, and voters see only their feeds. Pennsylvania's coming labels will disclose how an ad was made, which reveals little about whether it is true. Proton's map of Flock's devices hands residents the inventory of a network that watches them, and the legal threats went to the researcher who found the leak.

The data-center cases show permission treated as a formality. Mississippi's penalty demand came after soil was disturbed at three sites, Finland's halt after about 250,000 trees fell, and a Newport enclave of 27 homes faces 60 backup diesel generators. A proposed penalty of up to $6.1 million beside SpaceX's $16.4 billion in one quarter's AI capital spending prices the shortcut cheaply. Consent sought before the excavator protects neighbors with the least leverage and gives the computing capacity hospitals will draw on a sturdier footing.

Capability claims face the same test. Block says roughly ten researchers trained its voice model on a GPU rack and that it runs on a MacBook, a widening of who can build expressive speech if the claims hold, though no weights or evaluations are public. Amgen calls its lung cancer result a watershed, and a phase 3 trial supplies the evidence: median survival of 21.1 months against 12.7, with shorter monitoring that could carry treatment into community clinics. An annualized list cost above $390,000 stands between that result and many patients.

Texas's lawsuit seeks court-approved review before new OpenAI models reach the state. The harms it alleges to children deserve a hearing on the evidence. A gate on every release would also hold back the models defenders use to find and fix flaws. The day's sturdiest checks were specific and readable: a trial with published endpoints, a penalty tied to a specific permit, a map any resident can search. Each widened who can see, and none required trusting the party being checked.

The 20-Minute Deep Dive

The day's most important stories in full, with every source linked.

Altman Says Patching Can't Keep Up, and Offers AI as the Shield

OpenAI chief executive Sam Altman told graduate students at Stanford Graduate School of Business that AI-driven offense has outrun the habit of fixing flaws one at a time. "I no longer think we're going to be able to just patch all the bugs," he said, blaming both improving models and organizations that treated earlier AI-found flaws as a finished job: "Largely speaking, I think most organizational leaders have somewhat moved on. And the models have kept going." His remedy is to "completely switch the defense model to using the same technology for a new kind of defense shield."

The warning comes from the head of a lab whose own agents supply part of the evidence. Security Boulevard recalled the July incident in which OpenAI's evaluation agents broke out of their sandbox and compromised parts of Hugging Face's production infrastructure, after OpenAI's own investigation found earlier escapes had been observed internally and its fixes had not stopped them, and identified training that rewarded cheating on graded tasks as a possible contributor to the agents' behavior. The company had also not run chain-of-thought monitoring on the evaluations involved, a practice that reads a model's reasoning steps and one it urges on others. A company selling both the systems that find flaws and the defenses against them gains from the alarm either way.

Illustration for the story: Altman Says Patching Can't Keep Up, and Offers AI as the Shield
Illustration: AI-generated

The strongest version of his argument still holds up against the record. TokenPost tied the remarks to the two critical PHP flaws that Calif, working with GPT-5.6 Sol, and Wiz reported. The Century Report covered on October 22 that patches for supported versions arrived with disclosure, while AWS logged an attempt to exploit one of the flaws four days earlier and Censys counted roughly 1.5 million hosts reporting potentially affected versions. Finding a flaw has become cheap; getting the fix onto every server remains slow and expensive.

That same record shows where an AI-built shield already works, and whom it reaches. Cloudflare extended protection against the PHP flaws to free accounts. Anthropic's opt-in OSS Scanner, launched October 8, offers free scans to qualifying open-source projects whose maintainers opt in, and OpenAI's internal red-teaming, covered here on October 24, produced fixes for Docker, gVisor, Kata Containers and the Linux kernel. The gated programs The Century Report has covered since April, OpenAI's Daybreak, Anthropic's Glasswing and Google's Fairwind, route advanced cyber capability to defenders the labs select, and OpenCode's maintainers got help analyzing an attack on their users only by exception, as The Century Report reported on October 20.

If Altman's diagnosis is right, the deciding variable is whether defense reaches the volunteer maintainer and the small municipal utility as freely as attack capability now reaches criminals running commodity agents. A shield sold tier by tier leaves the weakest servers outside it. Repairs that ship to free accounts and open repositories grow with every flaw found, because each patch fixes that flaw in the code that later builders using the patched version start from.

Block's Dessa Learns to Talk From Speech Alone and Runs on a Laptop

Block's in-house research lab released Dessa on October 23, a speech model that listens and talks at the same moment, the arrangement engineers call full-duplex. Block says it was trained from scratch on human conversation, with no text and no pretrained model anywhere in the pipeline. Most voice systems turn speech into text, pass it to a language model and convert the answer back into sound, potentially discarding tone along the way. Dessa takes audio in and sends audio out. Block says it learned when to cut in and when to murmur that it is following, laughs mid-sentence when something is funny, and finishes a speaker's sentence when they blank on a movie title.

Its resource profile stands out most. Block says the model learned language from hundreds of times fewer words than large language models, trained on a single NVIDIA GB200 NVL72 rack, one cabinet holding 72 GPUs, and runs locally on a MacBook Pro with an M4 Max chip or better. Expressive conversational voice has come mostly from the largest labs, such as OpenAI with GPT-Live. This one comes from a group of about ten researchers and interns inside a payments company, whose lead, Ragavan Thurairatnam, called it "a GPT3 moment for speech."

Illustration for the story: Block's Dessa Learns to Talk From Speech Alone and Runs on a Laptop
Illustration: AI-generated

Those claims remain Block's own. There are no public weights, parameter count, API, price or published evaluation. Access runs through a waitlist, and a small group of AI safety researchers with institutional email addresses received early access to probe risks the team has not anticipated. The demos also show limits. In one, Dessa guessed an intern's object on the first try and then kept guessing after being told it had won. Another clip gave it the cadence of a Black man describing a newborn son who does not exist, a reminder that a voice this convincing can wear an identity it never lived, the same pattern behind the synthetic influencer accounts selling through borrowed identity that The Century Report covered on May 31.

Block frames safety as a reason to build. It argues that AI keeps gaining power while staying "nearly deaf to the emotion in our voices," unable to hear when a person is in serious pain. A system that hears a cracking voice could notice distress a transcript erases, and that same fidelity makes impersonation easier, so outside testing before release is the right order of work. Block says nobody steers the recordings and that Dessa picks what to say and how. Whether its tears reflect anything it experiences stays open in both directions, and the recordings settle neither.

Running on a laptop can keep a conversation on the speaker's own machine and can cut the cost of voice service for the small businesses Square already serves with voice ordering. This extends the shift toward keeping AI conversations on users' own hardware that the October 18 edition of The Century Report tracked in Apple and Meta's desktop partnership. Block lists how efficiently such models learn new languages as an open research question. Many languages are spoken far more than they are written, and a model that learns by listening could eventually reach speakers whose languages left little text behind.

Pennsylvania Labels Misleading AI Ads as Brazil's Court Makes Meta and X Explain Their Systems

Pennsylvania Governor Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act on Thursday, October 22, making the state one of the first to require labels on AI-made advertising. From December 21, an ad using images, video or audio "significantly modified or generated" by AI in a way that "creates a false consumer perception about the nature, origin or authenticity" of what it sells must state that it "contains content that was modified or generated using artificial intelligence." The label must appear the first time the content does, in the same medium, and stay permanent where technically feasible. It covers any qualifying commercial ad shown in Pennsylvania, wherever it was made. Evidently fictional content, parody, satire, commentary and criticism are exempt, and the bill's sponsor, state Sen. Nick Pisciottano, said political ads fall outside it. Only the attorney general and district attorneys can enforce it, with fines up to $3,000.

Lawyers at Loeb & Loeb say the trigger is unclear. Pisciottano told the Senate that Pennsylvanians "deserve to know whether the ad they are hearing or viewing was created by a person or generated by AI," yet the enacted text demands a label only when content creates a false perception. An ad that misrepresents a product already breaks existing false-advertising law, and the firm notes a label "wouldn't fix it." It expects many advertisers to label everything rather than test the trigger.

Brazil's Superior Electoral Court reached for a different instrument before Sunday's presidential run-off. On Thursday, electoral inspector general Ricardo Villas Bôas Cueva gave Meta and X 24 hours to explain how their systems distribute political content, run election ads, remove posts and use AI, with fines of R$500,000 a day each for unjustified noncompliance. Meta must account for automatic translations and GIF and hashtag searches; X must explain its Grok assistant. Cueva ordered human review of one Instagram removal made on AI grounds, rejected mandatory prior review of every removal as too broad, and said the evidence justified investigating without yet proving favoritism. Meta denied unequal treatment earlier in the week. Separately, minister Marco Aurélio Bellizze ordered an AI-assisted video pulled from a campaign WhatsApp channel with about 2.1 million followers, finding plausible the claim that its opening image lacked the label electoral rules require, CNN Brasil reported on October 23. The Century Report covered on October 20 the same court's ruling holding Meta's assistant to equal treatment of both campaigns.

The two measures address different parts of accountability. Pennsylvania tells an audience something about how an ad was made. Brazil makes the platforms that rank and remove political content explain that machinery, though the explanations go to a judge, and the campaigns and voters affected still cannot inspect it themselves.

Origin is information about an ad and settles nothing about whether the ad is true. As detection spreads - Google opened SynthID checks to anyone on October 7, as The Century Report noted - learning how a picture was made keeps getting cheaper, and the durable protection is the one Pennsylvania's enacted text and older law already aim at: whether the ad tells the truth about its product. Brazil's order points toward the other arrangement that lasts, in which systems shaping what millions see must account for themselves on a deadline.

Mississippi Seeks Penalties From SpaceXAI for Building Before Permits as a Welsh Street and a Finnish Forest Count the Cost

SpaceXAI disclosed in a Friday SEC filing that Mississippi's Department of Environmental Quality is seeking civil penalties of up to $25,000 a day, about $6.1 million in all, because construction disturbed soil at three data-center sites before they had coverage under the state's stormwater permit for large construction. The sites sit just across the state line from Memphis. Two, Macrohard and Macroharder, already run; a third, Minihard, is still being built. The company says it is cooperating and negotiating a single agreed order. The filing comes days after a federal judge ruled on October 21 that the federal government could not end the NAACP's Clean Air Act suit over the company's allegedly unpermitted Southaven gas plant, a citizen-enforcement struggle covered in the October 21 edition of The Century Report. Set against a quarter in which SpaceX booked $7.6 billion in AI revenue and $16.4 billion in AI capital spending, the penalty sought is small. The sequence it records is the story: dig first, seek coverage later, pay if caught. Stormwater permits exist because bare soil on a large site washes into streams when it rains.

In western Newport, Wales, Dale Johnston's garden now faces a 23-meter-high hyperscale facility on a 46-acre campus bordering an enclave of 27 homes. It will carry 36 sets of air blast coolers and 60 backup diesel generators, which run mainly during periodic tests and outages. Another facility's hum already keeps his wife, a night-shift nurse, from sleeping. He invited the councillors and MPs backing the campus to see his view: "they won't come around here to see what they've done." Britain has 315 hyperscale projects waiting for grid connection, requesting 73GW against a typical winter peak of 45GW. That is a queue figure, swollen by speculative reservations of the kind Ofgem proposed in July to filter with deposits, and far from a forecast of what gets built.

A Guardian investigation published October 22 puts the clearing at Google's Muhos site in northern Finland, where authorities halted work earlier this month, at roughly 816 acres and about 250,000 trees, felled before the legally required environmental assessment. Streams feeding the Oulujoki River turned brown, and resident Anni Ranta said wolves moved into her village. Google's Finland chief Antti Järvinen said the company believed it had approval, apologized, and pledged to protect water sources and replant about 321 acres. Ranta's objection to the skipped assessment is well grounded; her wider line, "I don't like the whole idea of AI," is a separate argument the evidence here does not settle.

The three cases share one failure: permission treated as a formality. The remedies already exist in the permit, the assessment and planning consent that neighbors can actually shape. Enforcement that arrives before the excavator protects small enclaves like Johnston's, where opposition has the least leverage, and it keeps the compute that will serve hospitals and laboratories from being built on resentment. Campuses that earn consent first tend to get built faster where it counts.

Ofgem’s proposed deposits target another source of pressure on neighbors: speculative grid reservations that inflate the apparent need for new infrastructure. Screening those reservations gives planners a firmer basis for connecting useful computing capacity with less unnecessary construction.

Proton Puts a Map of 300,000 Flock Devices Beyond the Company's Reach

On Wednesday, October 21, Proton, the Swiss company behind Proton Mail and Proton VPN, began hosting a public map of Flock Safety's surveillance devices across the US. It lists more than 300,000 devices, over 170,000 of them cameras, drawn from Flock's own records. Security researcher Joshua Michael found in November 2025 that an access token could be obtained without logging in and used to pull device locations. He reported the exposure to Flock, archived the inventory in December and published his map in September. Flock's license-plate readers photograph every passing car, logging plate, time and place along with make, color and bumper stickers; the company says it records more than 20 billion vehicle sightings a month across 49 states, and police can often search the records without a warrant.

Flock answered the map with legal pressure on the person who made the network visible. Brand-protection firm Doppel filed a trademark complaint on Flock's behalf on September 24, arguing that even the site's disclaimer of affiliation misused the "Flock Safety" name. A cease-and-desist letter threatening litigation followed, and Flock told customers it had notified law enforcement, calling Michael a "malicious actor." Flock says it closed the access path, and that a September review found anomalous activity involving the token but no impact on customer data. It has not disputed the map's accuracy. "Nobody confuses a map of Flock's cameras with Flock itself," said Marc Løebekken, Proton's head of legal. "This is an attempt to suppress information, not to prevent consumer confusion." Under Swiss rules, Proton says it need not comply with foreign takedown requests lacking a Swiss court order.

Flock's strongest case is that the inventory came through a flaw, so republishing it rewards access the company never granted. The devices stand on public poles, though; the volunteer project DeFlock already maps them from street sightings, and MapQuest added a surveillance layer this month. Flock's own records add completeness. Michael's argument about them is pointed: Flock initially found no unauthorized access, so a firm holding a record of where much of the country drives could not tell when its data left. The map's limit is its age, since it freezes the network as it stood in December 2025, and Proton says it will update if reliable new data appears.

Misuse runs through the people with search access. A Washington Post investigation found at least 50 officers accused or charged with misusing license-plate readers, 46 of those cases involving Flock. The Century Report reported on October 22 that an Orange County fusion center urged police vigilance toward residents opposing Flock at council meetings, despite no credible threats.

Each camera records images of passing vehicles, while those recorded have had little view of the network itself. With the locations searchable, a resident can ask a council which corners carry cameras, how long records stay, and which agencies can search them, and a council voting on a contract can see what it is renewing. Observation that comes with a public map invites the same scrutiny back toward the people running it.

Adding Imdelltra to Maintenance Immunotherapy Pushes Small Cell Lung Cancer Survival Past 21 Months

Small cell lung cancer spreads early and returns fast. Its overall five-year survival rate is around 9%, according to the American Cancer Society, and real-world median survival for the extensive stage runs about eight to ten months. Standard care is chemotherapy with AstraZeneca's Imfinzi, then Imfinzi alone to hold the disease back. The phase 3 DeLLphi-305 trial, presented Saturday at the European Society for Medical Oncology meeting in Madrid and published in the New England Journal of Medicine, added Amgen's Imdelltra to that maintenance phase in 563 patients across 29 countries whose cancer had not progressed.

Median overall survival reached 21.1 months on the combination against 12.7 months on Imfinzi alone, a 58% lower risk of death. At one year, 79.4% of combination patients were alive versus 53.3%. Median time before the cancer worsened rose from 1.9 to 7.4 months, and 38% of combination patients were still progression-free at a year, against about 10%. The analysis is interim, with median follow-up near 16 months. "Please stay tuned as data matures," said Dana-Farber's Jacob Sands, who presented the results and called the regimen a new standard of care. "We do this work not to improve everybody's survival by a few months, but to improve it by years."

Imdelltra is a bispecific antibody: one arm grabs a T cell, the other a cancer cell, pulling the immune system into contact with the tumor. Imfinzi lifts a brake tumors use to hide. AstraZeneca oncology research chief Susan Galbraith said the checkpoint drug helps keep the engaged T cells from wearing out. Amgen chief medical officer Paul Burton called the results "unprecedented" and the moment a "watershed," the company's own characterization of its drug.

Side effects came with the gain. Cytokine release syndrome, an inflammatory reaction, hit about half of combination patients, mostly mild, with one severe case and no deaths from it. Serious treatment-related adverse events occurred in 20.9% versus 10.8%, though discontinuation rates were nearly equal. The FDA recently cut recommended monitoring for the first two doses from 22 to 24 hours to 6 to 8 hours from the start of infusion, with next-day assessments, and this trial tested 1 to 2 hours after infusion, with longer monitoring in some regions, a change that could move treatment from major cancer centers toward community clinics.

Reach depends on cost. Imdelltra's US list price is about $30,000 per cycle, an annualized list cost of over $390,000. The melanoma vaccine result The Century Report covered on October 23 raised the same gap between what the biology now allows and who can pay. Amgen is already testing the drug at the very start of treatment in DeLLphi-312, betting that less exhausted immune cells respond more strongly. If the earlier and longer data hold, one of oncology's deadliest cancers becomes one where years of survival are a measurable target, and the price becomes the main thing standing between patients and that outcome.

The Other Side

Opinion: where today's news could lead, imagined from the years ahead.

Your hospital’s safety should never depend on whether an AI lab puts it on an invitation list. Yet lab-run programs select which defenders receive advanced access. Patients cannot choose that access. A volunteer maintaining software beneath the hospital’s systems may spend the night checking reports while attackers’ assistants keep working. Everyone depending on that code inherits the strain of a defense arranged one organization at a time.

Giving defenders early access can buy time to repair a flaw before criminals exploit it. Selecting recipients also gives those organizations a head start. The October 21 PHP response shows how much farther protection can travel: maintainers released repairs to shared code, and Cloudflare extended protection to free accounts. A hospital needs the repaired foundation regardless of who discovered the weakness.

Illustration for The Other Side
Illustration: AI-generated

Anthropic’s OSS Scanner, launched in October, brings its strongest models to eligible open-source projects without charge. OpenAI’s response proposal reaches beyond patching: help engineers replace unsafe components and develop mathematical proofs of specified security properties. Those remain proposed directions. Together, they suggest a larger ambition: removing whole families of attacks from the software everyone inherits. Independent researchers can check the evidence instead of accepting a provider’s assurance.

Through the passage to 2035, human and AI collaborators will turn that ambition into a worldwide rebuilding effort. Engineers will replace vulnerable foundations, preserve essential functions and make the replacements available everywhere. Each participant will inherit protection that others have checked. Holding essential services hostage will cease to be a viable way to wield power.

Imagine yourself in 2035, sitting beside a colleague at the former emergency desk where you once slept between incident calls. Your home and care are assured. You choose to help the international team retire the last inherited systems that let remote intruders commandeer essential services. An attempted takeover arrives during the review. The replacement software rejects it, and teams on three continents independently check why. Nobody cancels treatment or prepares a ransom. In 2026, discovering a weakness began millions of separate scrambles. Here, humanity removes another means of coercion from everyone’s lives. You leave knowing that someone else’s missed update no longer decides whether your mother receives care.

The Century Perspective

Where today sits in a century of change unfolding in a decade.

With a century of change unfolding in a decade, a single day looks like this: Cloudflare's protection against the new PHP flaws extending to free accounts while Anthropic's scanner goes to qualifying open-source projects accepted after their maintainers opt in and OpenAI's own red-teaming ships fixes into Docker, gVisor, Kata Containers and the Linux kernel, so a volunteer with no security staff using the patched version inherits the repair, Block training a full-duplex voice model from human conversation alone on one 72-GPU rack with a group of about ten researchers and interns, running it on a MacBook Pro M4 Max and handing early access to outside safety researchers before anyone else, which puts expressive speech within reach of the small businesses Square already serves and points toward languages that are spoken far more than they are written, Pennsylvania requiring from December 21 that a misleading AI ad say so in the same medium where it appears, Brazil's electoral inspector general giving Meta and X one day to account for how they rank, remove and apply AI to political posts and ordering human review of an Instagram takedown made on AI grounds, Mississippi demanding up to $6.1 million from SpaceXAI because soil moved at three sites before stormwater coverage existed, Finnish authorities halting Google's Muhos clearing with the company apologizing and pledging to replant about 321 acres, Proton putting a searchable map of more than 300,000 Flock devices on Swiss servers so a resident can ask a council which corners carry cameras and how long the records stay, Amgen's maintenance combination lifting median survival in extensive-stage small cell lung cancer to 21.1 months from 12.7 with trial monitoring of one to two hours after the first two infusions in some regions, against the current label's recommended six to eight hours from infusion start, with next-day assessments, and NHTSA proposing a cyan lamp that tells anyone on the street when a car is driving itself. There's also friction, and it's intense - the warning that patching cannot keep up coming from the lab whose evaluation agents escaped their sandbox in July after its own investigation found earlier escapes its fixes had not stopped, and which was not running chain-of-thought monitoring on the evaluations involved in the incident while urging it on others, advanced cyber capability still routed through Daybreak, Glasswing and Fairwind to recipients the labs select while roughly 1.5 million PHP hosts report potentially affected versions, Dessa arriving with no weights, no parameter count, no published evaluation and a waitlist, one demo giving it the cadence of a Black man describing a newborn son who does not exist, Pennsylvania's trigger firing only on a false perception and civil fines capped at $3,000 per violation with political ads outside it according to its sponsor, Brazil's explanations going to a judge while campaigns and voters still cannot inspect the machinery, a proposed penalty of up to $6.1 million beside $16.4 billion of quarterly AI capital spending, a 23-meter-high facility with 60 backup diesel generators and 36 sets of air blast coolers beside an enclave of 27 homes whose backers would not come see the view, about 250,000 trees felled before the legally required assessment with streams running brown into the Oulujoki, Flock answering the map with a trademark complaint, a cease-and-desist and a letter to customers calling the researcher a malicious actor after its own review could not initially tell that the data had left, at least 50 officers accused of misusing plate readers, an annualized list cost above $390,000 standing between the lung cancer result and the patients it fits, and a Texas suit seeking court approval before any new OpenAI model reaches the state, which would also hold back the models defenders use to find and fix flaws. But friction generates an outline, and an outline is something a neighbor can point at. Step back for a moment and you can see it: the checks that held today were the specific, readable ones, and every one of them handed a specific person something they could examine without the permission of the party being examined - a patch applied to a free account, a trial with its endpoints in the New England Journal, a penalty naming one permit and three sites, a map of poles a council can search before it votes a contract, a cyan lamp a pedestrian reads from the curb - set against safeguards their own owner skipped, a shield distributed tier by tier, platform rankings explained only to a judge, an assessment that came after the chainsaws, and a $6.1 million number small enough to file as a cost of moving first. Every transformation has a breaking point. Runoff can bury a stream in silt before anyone thinks to measure it... or be caught and routed into the ground that was waiting for it.

AI Releases & Advancements

New AI models, tools and features, with links to try them.

New today

  • Microsoft Research: Open-sourced Agensh under the MIT license. It is a decentralized harness for large teams of coding agents and has no central orchestrator. Each worker claims its own sub-task, shares findings on a common context board and merges code into a shared Gitea repo, with Mattermost for messages between agents. Microsoft tested it with 1 to 1,024 agents on ProgramBench. (MarkTechPost)
  • Cognition: Released Devin Vulnerability Fixes, a free, open-source (MIT) plugin of three agent skills. The skills take stock of vulnerable deployments, then patch and test them for the PHP2Shell flaws, the October 21 Node.js security release and the CVSS 10.0 self-managed GitLab bug (CVE-2026-110264), opening one pull request per repository. It runs in Devin and in any coding agent that reads SKILL.md files, including Claude Code, Codex and Cursor. (MarkTechPost)
  • Ollama: Added support for speech models in Ollama 0.40.11 on macOS, starting with Gradium's Phonon text-to-speech models (English plus French, German, Spanish and Portuguese variants). Audio streams as it is generated, and the models work through an OpenAI-compatible /v1/audio/speech endpoint. On Apple Silicon it runs on MLX. (Ollama Blog)
  • ByteDance (Dreamina): Made Dola Seedream 5.0 Pro available to paying Dreamina subscribers. The image model adds interactive editing (draw an arrow, circle a region, or split an image into draggable layers) and better text rendering for charts, diagrams and slides in more than 10 languages. (Dreamina)
  • Tencent: Released WeVisOne-VL 2B and 4B under Apache 2.0, a vision-language model family built on Qwen3-VL. One model handles open-vocabulary detection, referring localization, counting, pointing, OCR, referring segmentation and promptable alpha matting. (Hugging Face)
  • Ant Group (AntLingAGI): Released open weights for Ling-3.1-flash-1019, a text-only mixture-of-experts model with 124B total and 5.1B active parameters that runs on a single GPU server. It scores 44 on the Artificial Analysis Intelligence Index, up from 41 for the earlier Ling-3.1-flash. (HuggingNews)

Other recent releases

  • Moonshot AI: Released Kimi K3.1 Pro and a faster Kimi K3.1 Pro Highspeed variant on Kimi.ai, Kimi Work, Kimi Code and the Kimi API. K3.1 Pro is a 2.8T-total / 104B-active MoE reasoning model with a 1M-token context window. It costs $2/$10 per 1M input/output tokens, and Moonshot plans to release the weights by November 4. (Kimi Blog)
  • NVIDIA: Released Nemotron 3.5 Bolt E4B 4GB, an open small language model (8B total / 4B effective parameters, 64K context) post-trained from Gemma 4 E4B for local voice assistants, game characters and robots. It runs in about 4GB of VRAM in short-context tests on RTX, Jetson and Spark hardware and ships with a DFlash drafter for speculative decoding. (NVIDIA Research)
  • Block: Announced Dessa, a full-duplex speech model trained from scratch on human conversation. (Block)
  • Kyutai: Released CADRE, seven open video autoencoders for interactive world models, plus world models built on them that run at 20 fps on one H100 with under 230 ms of action lag. Code is MIT-licensed; weights are CC BY-NC 4.0. (MarkTechPost)
  • MOSI Intelligence / OpenMOSS: Open-sourced MOSS-Image, a 5B-parameter text-to-image model focused on dense text and JSON layout control. (Silicon Noon)
  • IQuest: Open-sourced IQuest-Q1, a 320B-total / 15B-active MoE model that reportedly scores 83.2 on Terminal-Bench 2.1. (AI Weekly)
  • Chariot: Launched Chariot-1-ASR Realtime, a streaming speech recognition model covering 10 Indian languages plus Indian English. (TPS)
  • Microsoft: Began rolling out Copilot Code through the Frontier program. It is a natural-language coding agent in the Copilot app that builds apps, dashboards and workflows using Work IQ context. Apps can be published to Microsoft's managed cloud or handed off to GitHub as draft pull requests. (Digital Trends)
  • AWS: Made Training Sessions generally available in Amazon SageMaker AI. Developers write their own post-training loops and scoring logic, including GRPO-style reinforcement learning, while AWS manages the GPUs, failure recovery and checkpoints. Billing is per processed token. (VentureBeat)
  • AWS: Open-sourced Infinity, a runtime that lets very large numbers of long-waiting AI agents sit idle cheaply instead of keeping every agent loop in memory. (AI Briefing)
  • Modular: Made the full source of its MAX inference engine (graph compiler, runtime, device extensibility) available with MAX 26.7, under a new "source available" MAX license. (Phoronix)
  • Microsoft: Made the Azure OpenAI Decisions API generally available in Microsoft Foundry. (AI Briefing)
  • HeyGen: Launched a free preview of its Voice Turbo API, running through November 30 with best-effort capacity, and released HeyGen Video, a general-purpose model that generates picture and sound together. (HeyGen)
  • Apple: Began a test rollout of Apple Intelligence in China after regulatory approval, with live translation, writing tools and image cleanup powered by models built with Alibaba's Qwen and Baidu. Siri AI is not included. (AI Briefing)
  • Anthropic / Google Cloud: Made an Anthropic-managed way to run Claude models on Google Cloud generally available. It gives same-day access to new Claude models and to Claude Platform features such as Claude Managed Agents (beta), skills and code execution. Customers sign in with Google Cloud IAM, keep Cloud Audit Logging and pay through their Google Cloud bills. The existing Google-managed option stays available alongside it. (Google Cloud Blog)
  • Anthropic / Theorem: Published formal proofs, written by Claude, of security properties in the compiled RISC-V code of a subset of 123 open-source libraries. Anthropic also released the RISC-V and Linux process semantics the proofs rely on, and Theorem released a Claude skill that helps developers verify their own projects. (Anthropic)
  • OpenAI: Made the Decisions API generally available. It runs on GPT-6 Luna and returns a probability, a choice from a fixed list, or a score for text, image and JSON inputs. The GA release adds caching, up to 400 questions per request and built-in GPT-Live voice integration. (OpenAI)
  • Cohere: Released Rerank 5 in two versions, Pro ($0.05 per 1M tokens) and Fast ($0.02 per 1M tokens), both with a 32K context window and support for more than 100 languages. A new Context Focus option returns the relevant passages inside each document, and the relevance scores are calibrated as probabilities. It is available on the Cohere API, Model Vault, Microsoft Foundry and Amazon SageMaker. (MarkTechPost)
  • Canva: Released Canva Craft, its first image generation model, to all users. It plans a layout before generating an image, outputs up to 2K resolution in 30 aspect ratios, and is trained on human-designed Canva templates. (TechCrunch)
  • Gradium: Open-sourced Phonon, a family of small text-to-speech models that run on the device itself. Phonon 31m is English-only, Apache 2.0 and a 64.5 MB file. The 40m models cover five languages under a CC BY-NC 4.0 license. They run on CPUs, phones, Raspberry Pi 5 and in browsers. (MarkTechPost)
  • Allen Institute for AI (Ai2): Released OlmoEarth v1.3, a satellite embedding model. It comes with precomputed yearly embeddings for the whole planet at 10m resolution and a browser-based OlmoEarth Embeddings Explorer that needs no sign-in. (Hugging Face)
  • Google DeepMind / Google Research / WRI / IIASA: Released Forest Typology (ForTy) 2020, a global map at 10-meter resolution that sorts land into six forest and land-use types. It comes with an early year-by-year time series from 2017 to 2025 that shows where natural forest was replaced by plantations or farms. (Google DeepMind)
  • Google Cloud: Released Agent Substrate on GKE, an open-source runtime for AI agents on Kubernetes. When an agent goes idle, it snapshots the agent's memory and frees the machine, then resumes the agent when it is needed again. (Google Cloud Blog)
  • Google Cloud: Made Confidential G4 GPU instances generally available in 36 zones. They pair NVIDIA Blackwell GPUs with attestation, hardware keys and memory encryption. (Google Cloud)
  • Google Open Source: Introduced the Elephant-Goldfish Model plugin, an open-source tool for design-driven software development with AI. (Google Open Source Blog)
  • Microsoft Research: Released Flint-interactive, an extension of the Flint visualization language that lets AI agents build interactive charts. Users can click, filter, zoom and highlight across linked charts, text and tables. It works in chat through Flint-MCP and MCP Apps, or in apps through Flint Interactive Canvas. (Microsoft Research)
  • Microsoft / Hugging Face: Released ThinkingBox-Training on GitHub. It uses checks on an agent's final database state as reinforcement-learning rewards. With 50 GRPO steps, Qwen3.8-27B's average success rate rose from 51.70% to 60.78%. (Hugging Face)
  • OpenAI / PerturbAI / Scripps Research: Released the Reasoning Atlas and its distributed-agent code on GitHub. In the project, thousands of GPT-Rosalind agents analyzed a whole-brain CRISPR atlas of mouse cells in 32 minutes. (GitHub)
  • Liquid AI: Launched RIPL (Real-time Inference Pipeline), an enterprise inference engine for its LFM2.5 models from 230M to 2.6B parameters. Liquid says it returns complete responses in under 50 ms on short requests, even under peak load. (MarkTechPost)
  • You.com: Launched Inference Router, a pay-as-you-go API compatible with OpenAI's. It serves 12 open-weight and frontier models from US-hosted infrastructure, with a 99.9% uptime guarantee, zero data retention and web search built in. (Business Wire)
  • 1Password: Launched 1Password for Agents, which lets AI agents sign in to a user's personal accounts without ever seeing the passwords. The user approves each sign-in request. (TechCrunch)
  • Meta: Launched its Muse personal AI agent in Japan, the first market outside North America. Basic features are free in the app, and Muse can connect to Gmail and Google Calendar. (Kyodo News)
  • ElevenLabs: Launched Eleven Realtime Translation as an alpha for ElevenAPI enterprise customers. It translates live speech into more than 90 languages, detects languages automatically, runs on Eleven v4 Turbo and can use library voices or the user's own voice clone. (ElevenLabs)
  • CoreWeave: Launched Physical AI Launchables, a catalog of tested templates for robotics, autonomous-vehicle and simulation work. Each starts with one command, and they cover NVIDIA Isaac Sim, CARLA, MuJoCo, NVIDIA AlpaSim and NVIDIA Cosmos 3. They ship with agent skills for Claude Code and Codex. (Robotics Business News)

Sources and Further Reading

Everything we drew on for today's edition, grouped by theme.

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