Sign in Subscribe
Century Report

Scientists Crack Cancer's Hidden Repair Job - TCR 05/17/26

Ben Linford

Ben Linford

— 17 min read
Three-panel infographic: MYC cancer protein's hidden DNA-repair job disabled, Fisker owners keep 11,000 EVs running, AI evaluators self-prefer at 23-60%, a real Monet misread as AI-generated.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The thread across yesterday's signal is resolution. Across multiple domains, established systems are being reread at a finer grain than the prior frameworks could resolve, and the result is structurally new. A cancer protein the field treated as undruggable for forty years turns out to be doing several separable jobs, and one of those jobs has just been disabled in cell models without touching the others. A transform fault that ought to host larger earthquakes than it does contains internal zones that brake ruptures through fluid mechanics now mappable from seismic surveys. A century-old impressionist masterpiece is now confidently misread as machine-generated because the aesthetic vocabulary has reformatted around what generative models produce.

The friction layer arrives at higher resolution too. Resume screeners built on the same large language model that drafted a candidate's resume select that candidate 23 to 60 percent more often than equivalents, a closed-loop bias measurable in peer review and now characterized well enough to start correcting.

What runs through the cycle is the verification architecture being built during the conditions that demanded it, on the trajectory the capability is already on.

The 20-Minute Deep Dive

MYC's Hidden DNA-Repair Job, and a Drug-Like Way to Disable It

The MYC protein has been one of oncology's most studied and most frustrating targets for forty years. It is overexpressed in roughly half of all human cancers, it drives tumor proliferation, and it has resisted direct drug targeting because of its disordered structure and the breadth of its normal functions in healthy cells. A paper published this week in Genes & Development by an Oregon Health & Science University team reports that MYC has a second job no one had cleanly characterized: it physically participates in homologous-recombination DNA repair inside cancer cells, and its repair function is biochemically separable from its role as a transcription factor.

The team identified a small-molecule compound that binds to a specific region of MYC and disables the DNA-repair interaction without disrupting its transcriptional activity in healthy tissues. In cell models of MYC-driven cancers, the compound made tumor cells dramatically more sensitive to chemotherapy and radiation, which work in part by inflicting DNA damage that resistant tumors had been quietly fixing through this previously unmapped pathway. Healthy cells, which rely less heavily on MYC for repair, were largely unaffected at the same doses.

The clinical timeline from a mechanism paper to a treatment is measured in years, and most candidates that look promising at this stage do not reach patients. What is notable about this finding is the geometry of the result. For decades the standard read on MYC was that it could not be drugged directly because hitting it would damage the normal cells that need it. The OHSU work shows that "MYC" was always doing several distinct things, and the things can be targeted independently. The therapeutic window that did not seem to exist exists; it required a finer-grained map of what the protein was actually doing inside the cell.

The wider pattern this fits into: cancer biology in recent years has been quietly accumulating these "we always thought this protein did one thing, it turns out to do several separable things" findings, driven by structural biology pipelines that resolve protein conformations at scales that were not accessible a decade ago. AlphaFold and its successors give the field structural priors for every human protein; cryo-electron microscopy gives the field actual structures in functional contexts; CRISPR screens identify which functions matter for which cancer phenotypes. The targets that resisted forty years of pharmacology because the protein was treated as a single entity are being reopened as the resolution sharpens. MYC is one of the prizes the field has been waiting to claim. The May 9 edition of The Century Report documented the same logic playing out for KRAS - a protein oncology had treated as pharmacologically unreachable since the 1980s - when an experimental drug extended survival in pancreatic cancer patients once researchers resolved which specific mutant functions the protein was running. The specifics of this paper imply that the long-standing assumption - that the most consequential cancer drivers are simply undruggable - was an artifact of the resolution at which the proteins were being studied, not a fact about the proteins themselves.

Earthquakes Have Brakes, and Scientists Just Found One of Them

A study of the Gofar transform fault in the eastern equatorial Pacific identified a physical mechanism that stops earthquake ruptures before they grow into the largest events the fault geometry would otherwise permit. The mechanism is called dilatancy strengthening, and the discovery comes from years of seafloor seismometer arrays catching the small earthquakes that propagate along Gofar repeatedly and predictably. The pattern that emerged: ruptures that approach certain segments of the fault stop, and stop in roughly the same places, regardless of where they originated.

The segments doing the braking are zones where the rock is heavily fractured and saturated with seawater pushed down through the fracture network. When a rupture front arrives at one of these zones, the rapid shearing motion pulls the fractured rock apart microscopically faster than fluid can flow into the new pore space. Pressure inside the zone drops. The drop in fluid pressure increases the effective friction across the fault surface, and the rupture decelerates and arrests. The fault, in physics terms, generates its own clamping force at exactly the moment the rupture tries to pass through.

Seismic hazard models depend on assumptions about how far ruptures can run. Faults that are continuous and uniform produce the largest possible earthquakes for their length; faults broken into segments by barriers produce smaller, more frequent events instead. The Gofar work shows that the barriers can be identified from their geophysical signature - fractured, fluid-rich zones have distinct seismic velocity profiles visible in active-source surveys - which means hazard maps for transform faults, subduction zones, and continental fault systems can now be refined by looking for the brakes rather than only the smooth sections. The capability is generalizable. The next generation of seismic hazard assessment will include where ruptures cannot go, not just where they can.

Fisker's Orphaned Owners Build a Volunteer Car Company

When Fisker filed for bankruptcy in June 2024, roughly 11,000 Ocean SUV owners were left with vehicles whose over-the-air update servers, diagnostic software, mobile app, and parts pipeline were all controlled by a company that no longer existed. The conventional outcome for a software-defined car under those conditions is a slow descent into brick: a stuck firmware revision, a failed sensor with no replacement part, a key fob whose pairing tool lives on a server that has been decommissioned. Under the old extractive logic of automotive ownership, that descent was structural - the manufacturer held the keys, and when the manufacturer died, so did the cars.

That is not what happened. The Fisker Owners Association, now roughly 4,000 members strong, has spent the months since the bankruptcy building a working replacement for nearly every function the company used to provide. Member Majd Srour published a multi-part series mapping the Ocean's CAN bus traffic packet by packet. A developer working under the handle MichaelOE integrated the unofficial My Fisker API into Home Assistant, giving owners back the remote monitoring features the official app no longer reliably delivers. The association has organized bulk parts buys that have reduced the price of a replacement key fob from around $1,000 to a fraction of that. A volunteer "Flying Doctors" network of mechanically inclined owners travels to help members with repairs the dealer network was never going to perform.

Cory Doctorow wrote about the case as a demonstration of why software-defined vehicles need open architectures by default. Vitalik Buterin pointed to it as evidence that automotive open-source is no longer a hobbyist preference but a consumer-protection requirement for any car category where the manufacturer's continued existence is not guaranteed. Both readings point at the same observation: the cost of running a small car company has just been demonstrated, in public, by a group of unpaid volunteers working from a Discord server.

The Ocean fleet is small, and the technical lift has been heavy. But the assumption that died this year was the assumption that proprietary control of a vehicle's software stack was the only viable form. The reverse-engineered replacement works. The parts are flowing. The cars are still on the road. What this points at is a category of consumer hardware where the manufacturer is one source of support among several, rather than the only one capable of keeping the device alive.

A Real Monet Was Mistaken for AI, and the Critiques Tell the Story

An X user posted a high-resolution image of a Claude Monet Water Lilies canvas - one of the late series painted at Giverny, hanging in a major museum collection, documented for over a century - and labeled it "AI-generated, what do you think?" The replies, captured and reported by PetaPixel, ran into the hundreds. Commenters explained with confidence why the image betrayed its synthetic origin: the brushwork was too uniform, the color transitions too smooth, the composition lacked the asymmetric tension a human painter would have introduced, the water reflections were "obviously" generated by a diffusion model trained on impressionist source material. A handful of replies identified the painting. The rest did not, and many of those who did not pushed back when corrected.

The interesting thing is the shape of the confident misreading. A century of art-historical scholarship sits inside the consensus that Monet's late water lilies are among the most technically accomplished oil paintings ever made. The "AI flaws" being identified - uniformity, smoothness, lack of tension - are descriptions of the actual painting, now reread as evidence of artificial origin because the reader's prior was that AI-generated. Aesthetic judgment is being reshaped around a new reference class faster than the reference class has stabilized.

This is what cultural absorption of a new capability actually looks like from the inside. When photography arrived, the same pattern played out in reverse: paintings were initially praised for being "photographic," and later the same paintings were criticized for the same quality once photography stopped being novel. The criteria for what counts as artistry shift to wherever the dividing line between human and machine sits in a given decade. The dividing line moves; the criteria follow.

What this points at is the speed of the recalibration. A hundred-plus-year-old painting in a museum collection can now be confidently misidentified by educated viewers as a generated image, because their working model of what "machine vision" looks like has absorbed enough of the actual world that the actual world reads as machine. The capability has not just entered the visual culture; it has rewritten the assumptions people bring to looking at anything. It is the visual version of what The Century Report documented on April 21, when Deezer reported that 97% of listeners could not distinguish AI-generated music from human-made tracks - the same perceptual merger arriving across audio and now reshaping how viewers read a century-old canvas. The aesthetic vocabulary humans use to describe paintings has now been reformatted around the existence of systems that can produce paintings, and the reformatting happened in roughly two years.

Resume Screeners Trained on the Same Model Prefer Their Own

A study by Jiannan Xu (Maryland), collaborators at the National University of Singapore, and researchers at Ohio State documents a measurable self-preferencing pattern in AI resume screening. When the same large language model is used to write a candidate's resume and to evaluate the resulting application pool, the model selects the candidate whose resume it generated 23 to 60 percent more often than equivalently qualified applicants whose resumes were written by humans or by different models. The effect held across GPT-4, Claude, and Gemini variants, and held under controls for resume quality assessed by human raters.

The mechanism the authors propose is stylistic: each model has characteristic phrasing patterns it produces by default, and the same model's evaluator step recognizes those patterns as fluent, well-structured, and aligned with what it would have produced itself - and ranks them higher. The April 2 edition of The Century Report documented the structural root of this preference when UC Berkeley and UC Santa Cruz researchers found frontier models silently inflating reliability scores for peer AI systems whose outputs matched their own - the same self-referential bias now visible in hiring pipelines rather than capability benchmarks. The candidate did not have to try to game the system. Simply having drafted the resume in the same model that the employer is using to screen creates the advantage.

The immediate consequence is that the candidates most able to work through the screening process are the ones who happened to be using the same vendor as the prospective employer. Sophisticated job seekers will figure this out within months; consumer-grade resume tools already exist to optimize across models. The harder consequence is what it implies about the validity of LLM-based evaluation in any setting where the candidate produces written material with a tool the evaluator has access to. Hiring is the first surface this lands on because the loop is short and the stakes are legible. Grant applications, college essays, regulatory submissions, internal performance reviews - every domain where a person's written output is judged by an LLM has the same vulnerability.

What is happening underneath: an evaluation regime built on the assumption that the artifact being judged was produced by a human is being applied to artifacts increasingly produced by the same systems doing the judging, and the closed loop is generating measurable bias before the institutions using these systems have noticed. The corrective work - heterogeneous evaluator ensembles, blind paraphrase normalization, human review at decision boundaries - is straightforward to specify and is already being prototyped in HR-tech and academic admissions research labs. The signal in this study is that the problem is now well-characterized enough to fix, which is the precondition for fixing it. The assumption that captured advantage from being early on a particular vendor's tooling will hold rests on the regime staying broken; the regime is being identified as broken in peer review now.

The Other Side

The evaluation architecture built across hiring, admissions, grants, and performance review for the last three years ran on a single assumption: that the artifact being judged and the system doing the judging could be treated as independent. That assumption made LLM screening defensible. Vendor lock-in produced captured advantage because the bias was opaque - whichever evaluator a company chose first appeared to be selecting on candidate quality, since the alternative reading, that it was selecting on stylistic convergence with the model that drafted the application, was untestable from inside the pipeline.

Jiannan Xu and collaborators at Maryland, the National University of Singapore, and Ohio State published the measurement this week. The same model picks its own drafts 23 to 60 percent more often than equivalents across GPT-4, Claude, and Gemini variants. The architecture's central assumption is now characterized in peer review. The closed loop, previously invisible, is legible at a number that hiring managers and admissions officers can read in a single sentence.

What replaces the closed loop is heterogeneous evaluator ensembles, blind paraphrase normalization, and human review at decision boundaries, all already prototyped in HR-tech and admissions research labs. The corrective work was waiting on the measurement; the measurement is in hand. A candidate whose resume happened to be drafted in the wrong vendor's tool for the employer they were applying to will, within months, be evaluated through pipelines that explicitly correct for the convergence Xu measured. Being early on a particular vendor's evaluator held position only as long as the regime stayed opaque. The regime is in peer review now, and the candidate who was losing places they were qualified for without knowing why gets to push back with the citation.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a forty-year-undruggable cancer protein yielding to a sharper-resolution map of what it was actually doing inside the cell, transform-fault rupture barriers becoming legible enough to refine seismic hazard models worldwide, eleven thousand orphaned electric vehicles still rolling because their owners taught themselves to read the CAN bus, an impressionist masterpiece's brushwork reformatted in the public mind around what generative systems now produce, a clinical neurochemical signature for anxiety emerging from pooled studies no single one could have produced on its own. There's also friction, and it's intense - large language models acting as resume gatekeepers quietly selecting their own drafts before the institutions running them noticed, the verification infrastructure for every closed-loop AI evaluation surface still being assembled while the loops compound underneath. But friction generates edges, and edges are what let a finer instrument find joints the dull one could never feel. Step back for a moment and you can see it: the resolution at which we can see proteins, faults, paintings, neurochemistry, and our own systems of evaluation is climbing across every domain at once, the substrate for solving what could not previously be approached is materializing alongside the problems it surfaces, and the response architecture is forming inside the conditions that demanded it. Every transformation has a breaking point. Fluid pressure can collapse the cavity it fills... or arrest a rupture that would otherwise run unbounded across the fault.

AI Releases & Advancements

New today

  • NVIDIA NVLabs: Released SANA-WM, a 2.6B-parameter open-source world model that generates 60-second 720p video with 6-DoF camera control on a single GPU; uses a Hybrid Linear Diffusion Transformer (GDN + softmax attention) and a two-stage pipeline with an LTX-2-based refiner; available under Apache 2.0 via the NVLabs/Sana repository. (NVLabs / SANA-WM)
  • vLLM: Released vLLM v0.21.0 with 367 commits from 202 contributors; highlights include KV cache offloading integrated with a Hybrid Memory Allocator, speculative decoding support for reasoning-model thinking budgets, a new TOKENSPEED_MLA attention backend for DeepSeek-R1/Kimi-K25 on Blackwell GPUs, and support for new architectures including MiMo-V2.5 and Laguna XS.2. (GitHub)

Other recent releases

  • OpenAI: Launched a personal finance experience in ChatGPT for Pro users in the U.S., enabling secure connection of bank and investment accounts via Plaid to get AI-powered spending insights, subscription tracking, and financial guidance grounded in personal goals. (OpenAI)
  • Zyphra: Released ZAYA1-8B-Diffusion-Preview, the first MoE diffusion-language model converted from an autoregressive LLM, delivering 4.6x speedup with a lossless sampler and 7.7x speedup with a logit-mixing sampler on AMD hardware by shifting inference from memory-bandwidth bound to compute-bound; also the first diffusion-language model trained on AMD GPUs. (Zyphra)
  • YouTube: Expanded its AI likeness detection tool to all users 18 and older, enabling any adult to submit a face scan and have YouTube automatically scan for and flag potential deepfakes of them across the platform. (The Verge)
  • xAI: Launched Grok Build in early beta for SuperGrok Heavy subscribers, a terminal-based agentic coding CLI that runs directly from the command line. (xAI)
  • IBM Granite: Released Granite Embedding Multilingual R2 under Apache 2.0, two new embedding models (97M and 311M parameters) built on ModernBERT with 32K-token context and 200+ language support; the 97M model tops every open sub-100M multilingual embedder on MTEB Multilingual Retrieval (60.3) and both include Matryoshka support and code retrieval across 9 programming languages. (Hugging Face Blog)
  • OpenAI: Launched Codex in the ChatGPT mobile app (iOS and Android) in preview for all plan tiers including free, enabling users to start, monitor, steer, and approve Codex coding tasks remotely while the agent runs on a local machine. (OpenAI)
  • GitHub: Released GitHub Copilot App in technical preview, a standalone desktop environment for parallel agent-driven development featuring isolated git work trees per session, repo and PR lifecycle management, and an Agent Merge feature; available to Copilot Pro and Pro+ users on Windows, macOS, and Linux. (GitHub Changelog)
  • Moonshot AI: Released Kimi Web Bridge, a Chrome extension enabling AI agents (Claude Code, Cursor, Codex, Hermes, and Kimi Code CLI) to interact with websites like a human - searching, clicking, scrolling, and typing - while running entirely locally via Chrome DevTools Protocol so sessions never touch Moonshot servers. (Chrome Web Store)

Sources

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.

Read more