Sign in Subscribe
The Century Report

mRNA Cancer Vaccine Clears Phase 3 - TCR 08/20/26

Moderna and Merck's individualized mRNA melanoma therapy plus Keytruda met both survival endpoints in a 1,137-patient Phase 3, the first win for the class.

Ben Linford

Ben Linford

- 21 min read
a personalized mRNA melanoma vaccine clearing Phase 3, OpenAI cutting vetted cyber defenders and Flock's surveillance engine, and one-video robot learning plus untraceable AI art.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The clearest signal of the day came out of a melanoma trial. An individualized mRNA therapy, built from a single patient's own tumor mutations and paired with a checkpoint inhibitor, met the recurrence-free survival and distant metastasis-free survival endpoints in a 1,137-patient Phase 3 study. This is the first positive Phase 3 result for an individualized neoantigen therapy in a trial meant to support registration, and it turns cancer treatment from a catalog of standardized compounds into a manufacturing process priced per patient. The mRNA infrastructure scaled under pandemic pressure helped make it possible to go from reading a tumor to answering it in weeks, and the same platform is already in trials for lung, renal, and bladder cancers.

Set that acceleration beside the day's harder stories and a single tension comes into focus: capability keeps outrunning the controls meant to bound it. WIRED reconstructed a police system that names drivers and infers their associates from a company that spent years insisting its cameras could not identify individuals. OpenAI cut vetted defenders outside the US and Europe from its strongest models and called it a glitch. A campaign committee is now warning its own members that data-center power bills are becoming an electoral liability. Each shows the same seam, where what a system does has drifted from what its keepers say it does.

The wonder side of that gap points somewhere generous. A robot at Generalist AI learned a new chore from one video and grabbed a banana when its tool was taken away, which erodes the premium on hoarding millions of labeled task-hours. An MIT study found that a large model's outputs cannot be traced to any single training image, dissolving the forensic assumption underneath a wave of copyright suits. Claude designed protein binders that hit 14 of 15 targets, and US firms are routing production work to cheaper open models regardless of origin.

What connects a personalized cancer vaccine, a self-teaching robot, and unattributable image generation is the collapse of the scarcity each old system was priced to. Surveillance and gatekeeping run on being unobservable and singular; this week both got read in detail, by journalists parsing exposed code and by current and former rangers objecting publicly to a camera installation. The instruments of one-way control depend on a scarcity that the evidence keeps eroding.

The 20-Minute Deep Dive

A Personalized mRNA Cancer Vaccine Clears Its First Phase 3 Endpoint in Melanoma

Merck and Moderna reported Wednesday that intismeran autogene, an individualized mRNA immunotherapy, combined with the checkpoint inhibitor Keytruda, met its primary endpoint in a 1,137-patient Phase 3 trial of patients with resected high-risk melanoma. The combination extended recurrence-free survival and also cleared the key secondary endpoint of distant-metastasis-free survival. This is the first time an individualized neoantigen therapy of this kind has succeeded in a late-stage registrational trial.

The mechanism inverts how vaccines have always worked. A tumor is sequenced, its specific mutations identified, and an mRNA sequence is manufactured that codes for up to 34 neoantigens - proteins that appear on that one patient's cancer and nowhere else in their body. Injected, the mRNA instructs the patient's own cells to display those markers, training the immune system to recognize and hunt the tumor. Every dose is built for a single person from their own tumor's genetic signature. The checkpoint inhibitor removes the brakes the cancer uses to evade immune attack; the vaccine supplies the target list.

The Phase 2 readout that preceded this trial had shown a 44% reduction in the risk of recurrence or death, against a 20 to 40% benchmark in some pre-readout analyst commentary. The Phase 3 detailed data have not yet been released, and regulatory filings lie ahead, so this is demonstrated capability, not something a patient can request from an oncologist. What the readout moves is the date this becomes available - it does not make it available tomorrow. Moderna's shares climbed sharply on the news, a signal of how much the market had discounted the platform's odds, less a scorecard than a measure of how far expectations had drifted from what the biology could do.

Consider what the timeline compression actually means. As the August 14 edition of The Century Report documented across four advances in cell-therapy engineering, the distance between understanding why an individualized cancer treatment fails and building a version that works is already collapsing. A decade ago, the idea of sequencing a tumor, designing a custom immunological instruction set, manufacturing it, and dosing a patient inside a clinically useful window was a thought experiment. The mRNA manufacturing infrastructure scaled under pandemic pressure helped make that process possible in weeks. The same platform is already in trials for non-small-cell lung cancer, renal, and bladder tumors. What is cracking here is the assumption that cancer therapy has to be a catalog of standardized compounds tested against population averages - the economics of the blockbuster drug, one molecule sold to millions. An individualized therapy is a manufacturing process, not a product on a shelf, and once the process works the marginal cost of the next patient's medicine is the sequencing plus the synthesis. The distance between reading a person's disease and answering it is collapsing, and this is one of the clearest signals yet that it is collapsing on schedule.

Control Trails Capability: OpenAI Pulls the Ladder on Its Own Vetted Defenders

On August 19, vetted security researchers reported losing access to OpenAI's Daybreak Blue and its Trusted Access Compute tier - the arrangement that hands screened defenders the strongest models with the fewest guardrails so they can find flaws before attackers do. The cutoff reverses the controlled expansion that the August 11 edition of The Century Report documented when OpenAI launched GPT-5.6-Cyber through Daybreak for approved defenders. OpenAI told them it was a "technical issue," called it an error, and asked them to re-verify. All five researchers TechCrunch spoke with live outside the United States and Europe. The company's framing is that a gate of this kind protects the public: put the most dangerous capabilities only in hands you have checked, launch an offensive-testing tier like Daybreak Red under supervision, and preview privacy features like the "Private Safety Processing" teased in OpenAI's zero-data-retention post. Anthropic runs a comparable screened program of its own. The case being made is that defenders need the sharpest instruments first - but in practice only the most entrenched and already-capable defenders are actually getting them.

A gate that broadly benefits the public would not blink out the accounts of screened non-US and non-European researchers and then describe it as a verification error. Whoever holds the switch on who counts as a "trusted" defender holds the power to decide, unaccountably and overnight, who gets to see the frontier and who does not - and the people most likely to be cut are the ones furthest from the company's home jurisdictions. The burden sits with the gatekeeper to show the gate targets the harm rather than the competition, and an unlogged revocation of vetted outsiders is not that proof. Independent verification and preserved access are what a genuine safety gate looks like; a quiet cutoff explained after the fact is what entrenchment looks like.

Georgetown's Helen Toner, the executive director of CSET, described the underlying condition: the techniques for making AI more capable, she said, "are working much better than our techniques for making A.I. that reliably does what we want it to do and reliably stays within the constraints we've set." Read her claim against OpenAI's week and the two lock together. Capability is racing; the control layer trails it; and in that gap the company reaches for the crudest control available, the access list, deciding by hand who is inside the wall.

That reach is the move that will not hold. A safety architecture that depends on one company privately curating a list of trusted companies and institutions is priced to a moment when the strongest models live in one vault. The same capability keeps leaking outward - into open weights, into sovereign systems, into screened programs at rival labs - and every defender cut from one list is a defender who goes looking for a model no single switch controls. The gate assumes the vault stays singular. The evidence of the last year is that it does not, and a control that works only while capability is scarce is a control already running out of the scarcity it was built on.

The same evidence points to where the cut defenders go next. Screened researchers outside the US and Europe who lose access to one company's frontier cyber models are the readers most likely to turn to open-weight systems, and independent testing has already found open models like GLM-5.3 rivaling the gated frontier in exactly the bug-finding and cybersecurity work Daybreak was built to supervise. A revocation one company controls sends defenders toward capability no single switch does, and the near-term signal to watch is whether the researchers TechCrunch spoke with resurface on open models rather than re-verifying for the list.

The Data-Center Backlash Becomes an Electoral Problem Both Parties Are Repricing

The Century Report has tracked Pennsylvania's data-center fight before - the February standards framework and the May dispute over which costs a "but-for" tariff could shift onto ratepayers. The August 16 edition of The Century Report documented data-center opponents winning primaries and moratoriums across multiple states as siting fights moved into electoral politics. What is new is that the backlash crossed from a utility-commission argument into electoral arithmetic, and the party that has most championed the buildout is now warning its own candidates about it.

The National Republican Senatorial Committee circulated a memo urging members to remake their stance on data centers, citing races in Texas, Florida, Ohio, and Wisconsin where the issue is turning against incumbents. The trigger is measurable: Pew found that 52% of Americans are more concerned than excited about AI, up from 37% in 2021, and 63% say AI is advancing too quickly. A campaign committee issues defensive guidance when constituents connect a specific new substation and a specific rise in their power bill.

Two structural moves landed alongside the polling. Pennsylvania issued an executive order directing state agencies to review data-center permits only after developers make legally binding GRID commitments and receive local approval, and barring non-disclosure agreements for data-center projects. The order pairs that scrutiny with a permitting incentive tied to procuring firm clean energy: participating developers must procure 10% of each project's annual electricity consumption from incremental in-state clean firm sources beginning in 2027, rising to 32% in 2035. And in rural Texas, conservative landowners fighting a $33 billion transmission expansion have turned on their own governor - insisting, as one put it, that they are "not a bunch of woke Marxist liberals," but ranchers who do not want their land and their grid repriced to serve someone else's server farm.

The proportion here cuts both ways. Nationally, data centers remain a small share of electricity demand, and the reflexive alarm at a megawatt figure often smuggles in opposition to the compute itself. But the load is not spread evenly - it concentrates on specific towns, specific substations, specific water tables, and the NDAs existed precisely to keep those local costs from being itemized and sent back to the people bearing them. What the transition requires is the buildout; what these communities are refusing is having its costs socialized onto their bills while the benefits are exported. The NDA ban and the local-approval threshold are the mechanism by which an externalized cost becomes a line item the developer has to carry. That is the assumption coming apart underneath the surface fight: that the infrastructure of the intelligence era could be paid for out of sight by whoever happened to live nearest the transmission line. Once the meter runs the other way - developer pays, community approves, the deal is public - the buildout does not stop. It gets built where its costs are honestly priced, which is the only place it was ever going to last.

A Robot That Learns From One Video and Improvises When Its Tools Vanish

When a Wired reporter visited Generalist AI, the demonstration was deliberately simple to describe and hard to fake: show the robot a single video of a new task, and watch it attempt the task within seconds. Then take away the tool it was supposed to use. In one run, an operator removed the dustpan the robot had been reaching for, and the machine picked up a banana instead and kept going, adapting its grip to the improvised object. The company's new model, GEN-1.5, is built to do this kind of on-the-fly reasoning rather than replay a memorized routine.

The technical claim underneath the demo is what the company calls physical prompting. Just as a language model can be handed a few examples in its context window and generalize from them, GEN-1.5 takes in a demonstration - a video, a teleoperated run - and produces new action trajectories at roughly 100 Hz without retraining its weights. The model also transfers from simulation to physical hardware zero-shot, meaning skills rehearsed in a simulated environment carry over to real motors and grippers on the first try. According to the company, the previous generation, GEN-1, was trained on more than 500,000 hours of physical-interaction data, plus about an hour of robot data per evaluated task, and averaged a 99% success rate across six task-specific evaluations; this version is aimed at tasks nobody trained it for.

The honest limit is in the numbers the company published itself. On the company's one-shot evaluations of previously unseen tasks, GEN-1.5 lands around 59% - impressive for zero-shot generalization, and nowhere near the reliability a warehouse or a home would demand. The model conditions on demonstrations up to 30 seconds long. What counts is the direction of the curve, not the current height of it. In April, when The Century Report covered the company's 99% average across six task-specific GEN-1 evaluations, the frontier question was how much data it takes to make a robot competent at a fixed set of chores. The frontier question now is whether a robot needs task-specific data at all, or whether it can watch, infer, and adapt the way a capable person walks into an unfamiliar kitchen and starts cooking.

That shift changes what "teaching a robot" means. The extractive version of robotics assumed the value lived in proprietary datasets - whoever collected the most task-hours owned the most capable machines, and that advantage compounded. A model that generalizes from one video erodes the premium on hoarding demonstrations, because the scarce input stops being millions of labeled hours and becomes a single afternoon with a camera. The banana was a small moment. What it points at is a world where physical competence propagates the way software does, copied and adapted rather than painstakingly re-collected for every new object on every new shelf.

AI Image Outputs Often Can't Be Traced to Their Training Data

A team at MIT CSAIL set out to answer a question the entire generative-AI copyright fight has assumed has an answer: when a model produces an image, which training images made it possible? The result, published in Nature Communications, is that for large models the honest answer is frequently "no single one." As the training set grows, the influence of any individual image on a given output decays along an inverse power law, until the contribution of the specific painting or photograph a plaintiff might point to fades below the level of measurement.

The method is the interesting part, because attribution claims are easy to assert and hard to test. The researchers built what they call a diffusion ensemble, an architecture that lets them remove a specific image from the training data and regenerate the exact counterfactual output without retraining the whole model from scratch. That makes the question empirical: take this image out, and does the result meaningfully change? Across their experiments, past a certain dataset scale, removing any one source left the output essentially intact. The influence had been distributed across thousands of images, no one of them decisive.

The finding cuts against a framing both sides of the copyright litigation have leaned on. Rights-holders have argued that a model's outputs are traceable derivatives of specific ingested works; some AI developers have implied that provenance can be cleanly tracked and credited. The measurement suggests neither picture holds at scale. What a large model learns from a big corpus is closer to a statistical grammar of a visual style than a collage of retrievable sources, and the legal machinery built for the collage - identify the copied work, trace the line from input to output, assign the compensation - runs into a system where that line genuinely disperses.

None of this settles whether training on scraped images without consent is fair; that is a separate question about how the corpus was assembled, and it stands regardless of what the outputs can be traced to. What the study does is move the compensation debate off the assumption that per-image attribution is technically recoverable. If influence is diffuse by the physics of how these models learn, then consent and compensation frameworks built around tracing a single output to a single author are trying to measure something that dissolves as the dataset grows. The more durable approaches being floated - collective licensing pools, opt-in corpora, revenue-sharing keyed to participation rather than forensic tracing - stop looking like second-best workarounds and start looking like the shape the problem actually has. The tracing assumption was scaffolding built for a smaller world; the systems that replace it will have to distribute credit the way the models distribute influence, broadly.

The Code That Watched Back: Flock's Identity Engine Reconstructed From Its Own Files

For years Flock Safety told the public its cameras "cannot recognize, identify, or track individuals." WIRED reconstructed the opposite from roughly 450 files the company served openly on its own systems. The reconstructed system, OS Investigate - internally called Nightshift - ships with 69 prewritten prompts and 45 connected tools that reach across license-plate scans, arrest records, dispatch logs, and commercial databases carrying Social Security numbers, dates of birth, and named relatives. Nineteen of those prompts hunt for patterns of movement. Fourteen need no plate, no name, and no description to begin. An officer can search by physical description alone, and the system infers "associates" by finding people whose vehicles appear near a target's three or more times inside a two-minute window. Flock scans roughly 20 billion plates a month and counts about 140,000 monthly active users across American police departments.

The gap between the marketing claim and the reconstructed capability is the whole story. The August 14 edition of The Century Report documented Flock's announced response to officer stalking abuses and contract cancellations: saying it would require case numbers for every search and cut its standard data-retention period to seven days. A control that a company describes as blind to individuals turns out, in its own code, to be an identity-resolution engine. That distance between what an actor says its system does and what the system demonstrably does is exactly where attribution comes in: Flock's characterization is a claim, and the files are the evidence against it. A Texas officer, reporting reviewed elsewhere established, searched more than 83,000 cameras to trace a woman who had self-administered an abortion - the concrete shape of what "cannot track individuals" means once the tooling exists and the guardrails are prompts rather than physics.

At Yosemite, the National Park Service installed Flock cameras, drawing objections from current and former rangers, one of whom called the deployment "completely insane." The Park Service says the cameras only monitor traffic. The rangers reading the same hardware see a national plate-tracking network arriving inside the parks, and their refusal to stay quiet is the second half of what happened here.

Both halves point the same direction. The capability that lets a few watch everyone was built from the same material as the capability that lets everyone watch the watchers - and this week the second one moved. Flock's own exposed files are what let WIRED reconstruct the system; the rangers' dissent is what put Yosemite's install on the record; MIT Technology Review's scrutiny of the design choices put the architecture in front of a wider audience. A surveillance apparatus whose power depends on being unobservable just spent a week being observed in detail, by journalists reading its code and by the workers refusing to install it silently. The direction of the glass is not fixed. When the instrument of one-way watching becomes legible enough to audit, the asymmetry it runs on starts to close, and the people who were only ever the watched acquire the one thing the arrangement was built to deny them: the ability to see the machine and name it precisely.

The Other Side

The power of surveillance has always come from being invisible. The powers behind that surveillance have always justified their intrusions by skirting the line of the what the public has considered acceptable. Flock Safety did so by telling the public its cameras "cannot recognize, identify, or track individuals," and the value of the network depended on the public believing that claim. But surveillance systems always work as a one-way mirror: the few see the many, and the many cannot see the machine, cannot read its code, cannot name what it does. The asymmetry, attractively but deceptively packaged, is the product.

That asymmetry cracked this week on the most ordinary material. Flock served roughly 450 files openly on its own systems, and WIRED reconstructed from them an identity engine - 69 prewritten prompts, 45 connected tools, movement-pattern searches that need no plate and no name to begin. At Yosemite, rangers reading the same hardware refused to stay quiet, telling reporters the install sounded "completely insane." The machine built to watch without being watched spent a week being read in detail, by journalists parsing its exposed code and by the workers refusing to install it silently.

Imagine a woman in 2033 who left an abusive marriage and moved four hundred miles to a town where nobody knows her. No network can reassemble her route from a hundred pole cameras. No archive of plates exists for anyone to reach into, and no search will tell a determined person where she sleeps now. She does not have to hide, like so many women did before her. Hiding is what you do when someone has the means to come looking for you. Now, where she is belongs to her and her alone. She can rebuild her life as she chooses, four hundred miles from an abuser she no longer has to think about. The hard year was when a ranger risked her job to expose the deception inherent in surveillance systems. What comes of it will be a generation that struggles to picture such an arrangement at all.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: an individualized mRNA cancer therapy built from one patient's own tumor mutations meeting the recurrence-free survival and distant metastasis-free survival endpoints in a 1,137-patient Phase 3 melanoma trial and already moving into lung, renal, and bladder trials, Claude designing working protein binders against 14 of 15 targets at rates that roughly double today's campaigns, a Generalist robot learning a new chore from a single video and grabbing a banana when its intended tool was pulled away, an MIT team showing that a large model's outputs cannot be traced to any single training image, and US startups routing production work to cheaper Chinese open models regardless of origin. There's also friction, and it's intense - OpenAI cutting vetted defenders outside the US and Europe from its strongest cyber models and calling the revocation a glitch, Georgetown's Helen Toner warning that the techniques for making models capable now outrun the techniques for making them do what we want, WIRED reconstructing from Flock's own 450 exposed files an identity engine that names drivers and infers their associates after the company spent years insisting its cameras could not identify anyone, the National Park Service installing those cameras at Yosemite, drawing objections from current and former rangers, and a Republican campaign committee warning its candidates that data-center power bills are turning against them as Pew finds 52 percent are more concerned than excited about AI and 63 percent say AI is advancing too quickly. But friction generates seams, and a seam is where the join between what a system claims and what it does finally becomes visible. Step back for a moment and you can see it: cancer therapy turning from a catalog priced to millions into a manufacturing run priced per patient, physical competence propagating from one afternoon with a camera instead of millions of hoarded task-hours, and the surveillance and gatekeeping systems that ran on being unobservable and singular getting read in detail the same week - by journalists parsing exposed code, by rangers refusing to install cameras silently, by researchers proving that the per-image attribution the whole regime assumed simply disperses at scale. Every transformation has a breaking point. A searchlight can hunt down whoever it is aimed at... or expose the hand that holds it.

AI Releases & Advancements

New today

  • Ornith AI: Released Ornith-1.5, an open-weight model family (9B dense, 35B MoE, 397B MoE) trained with what its developer calls a full self-improvement loop where the model proposes tasks, builds scaffolds, and generates its own RL rollouts; the developer reports that the 397B variant scores competitively with Claude Opus 4.8 on Terminal-Bench 2.1, with MIT-licensed weights available on Hugging Face. (Ornith AI)
  • Warp: Launched Warp Factories, cloud infrastructure for running "software factories" where specialized AI agents triage, spec, implement, review, and verify engineering requests end-to-end into mergeable pull requests, now in early access for select teams. (Warp)
  • Cloudways: Launched Managed AI Agents, general availability of managed hosting for OpenClaw and Hermes agent frameworks, letting users deploy agents with BYO LLM keys and Slack/Discord/Telegram/WhatsApp channel support without infrastructure setup. (Cloudways)

Other recent releases

  • OpenAI: Launched ChatGPT for Teens, an age-gated version of ChatGPT with automatic teen detection, Study Mode defaults, and expanded parental controls. (OpenAI)
  • Cartesia: Released Sonic-3.6, a streaming text-to-speech model now ranked #1 on both Artificial Analysis speech-generation leaderboards. (Cartesia)
  • NVIDIA: Released TensorRT Model Connect (TRTMC) in public preview, an Apache-2.0 open-source tool that converts a Hugging Face or local checkpoint directly into native C++ TensorRT inference in two commands, eliminating the ONNX export step. (NVIDIA/TensorRT-Model-Connect on GitHub)
  • Cerebras: Introduced the CS-4, a new rack-scale AI inference system built on three Wafer Scale Engine 3 Turbo processors, which Cerebras says delivers up to 30x faster inference than production GPU systems and 10x more throughput per watt than CS-3. (Cerebras)
  • AWS: Amazon Bedrock AgentCore payments reached general availability, moving out of preview and adding support for the Machine Payment Protocol (co-authored by Stripe and Tempo) and spending-ceiling controls within x402, enabling AI agents to autonomously pay for APIs, MCP servers, and other agents. (AWS)
  • GenBio AI: Released AIDO Cell, a "virtual cell" world model that GenBio says simulates human cellular behavior in its natural state and in response to drugs and other interventions across the full biological hierarchy from DNA/RNA through protein to whole-cell level. (GenBio AI)
  • MeitY (Government of India): Launched VoicERA, an open-source end-to-end voice AI stack built on the BHASHINI national infrastructure, supporting multilingual voice AI across 700+ dialects for citizen services. (PIB India)
  • LMSYS: Released Miles v0.1, a full-stack production-ready reinforcement learning framework for large-scale MoE post-training, succeeding the initial Miles release. (LMSYS Org)
  • Nous Research: Released Hermes Agent Bot Mode (v0.20.3), turning agent profiles into a roster of named bots with persistent Agent Inbox messaging. (MarkTechPost)
  • Cursor: Launched Origin, a native code hosting platform (GitHub alternative) with repos, PRs, reviews, and CI integration, live in beta for paid users. (Cursor)
  • Tencent: Released UI-Mate-27B, an Apache 2.0 open-weight computer-use/GUI-navigation agent scoring 77.0 on OSWorld-Verified. (DataNorth)
  • Hazmat: Released an open-source containment and isolation tool for AI coding agents. (Help Net Security)
  • Speko: Launched an "OpenRouter for Voice AI" platform that auto-selects optimal STT/LLM/TTS model combinations based on benchmarked constraints. (Speko)

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