Sign in Subscribe
Century Report

Why Antivirals Keep Failing COVID - TCR 05/11/26

Ben Linford

Ben Linford

— 19 min read
Three-panel infographic: the machine that lets COVID fix its own copying errors, malaria genome sequencing on laptops in six African countries, a new steel for seawater hydrogen.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The pattern yesterday is one of capability arriving in deployable form at the points where governance and infrastructure had previously named the bottleneck. A viral enzyme that has defeated antiviral chemistry since 2020 became legible at atomic resolution, opening two new drug-design pathways. Continental-scale pathogen surveillance, long imagined as a centralized data-pipeline problem, now runs on laptops in clinics across six African countries. An aging organ system whose decline was treated as host-genetic destiny turns out to respond to a transplant of microbial communities from a younger animal. Materials assumptions that locked seawater hydrogen to scarce alloys yield to a new stainless steel formulation. The substrate of medicine, surveillance, energy chemistry, and computation is being rebuilt at velocity.

The institutional layer is moving in two directions at once. A Malaysian state of 2.8 million people contracted UK universities and a federated-learning company to build a sovereign AI system rather than rent one from foreign providers, joining the broader pattern of subnational and national governments treating intelligence as infrastructure to be owned. The European Commission opened concrete preview-access discussions with OpenAI on its restricted cyber model while Anthropic's tighter containment posture leaves Brussels still without access to Mythos. Frontier capability is no longer being released into a single regulatory environment. A multi-jurisdictional verification layer is being assembled out of the conditions any one lab's release decision can no longer contain, and the speed at which that layer is forming is itself a structural feature of the era now arriving.

The 20-Minute Deep Dive

Cryo-EM at 2.4 Angstroms Opens a New Path Around the Coronavirus Proofreader

A research team at Iowa State University led by Yang Yang captured atomic-resolution images of the SARS-CoV-2 proofreading enzyme exoribonuclease interacting with RNA carrying antiviral nucleotide analogs including remdesivir, sofosbuvir, and bemnifosbuvir. The images, published in Nature Communications and PNAS, reach 2.4 angstroms of resolution, a new record for this enzyme and a level fine enough to map which functional groups are interacting and which atomic positions are doing the work of molecular recognition.

The structural problem the work addresses is one that has limited COVID-19 treatment since the pandemic began. Nucleotide analog antivirals mimic the building blocks the virus uses to copy its RNA, leading the replication machinery to insert flawed material into new viral copies. Coronaviruses, alone among most RNA viruses, carry a proofreading enzyme that scans newly synthesized RNA, identifies the flawed nucleotide, and excises it. That proofreading function is why remdesivir works less effectively against SARS-CoV-2 than similar analogs work against viruses without comparable error correction. The structural reach to see exactly how the proofreader recognizes a flawed nucleotide changes what antiviral chemistry can attempt next.

Two pathways open from the resolution gain. The first redesigns nucleotide analogs so the flawed RNA they generate carries a shape the proofreading enzyme cannot recognize, making the modification incompatible with the binding interface the enzyme depends on. The second binds the proofreading enzyme itself, locking it into an inactive form while increasing its affinity for the modified RNA so it grabs the flawed material and refuses to release it. Either path requires knowing the molecular geometry at the resolution Yang's team has now reached. Earlier images, at lower resolution, did not show the interactions clearly enough to design against them.

The instrument-layer story underneath the result is the same maturation pattern visible in nanopore sequencing for malaria and in the autonomous closed-loop perovskite discovery the April 15 edition of The Century Report covered. The cryo-EM hardware itself has not changed dramatically in the past few years. The protocols for sample preparation have. Yang's team developed new methods for preparing samples, tested them at Iowa State's facility, then ran the final imaging at one of the three federally funded cryo-EM centers in the United States. The improvements compound on hardware that was already capable. The therapeutic surface widens because the resolution at which biological systems become legible keeps falling.

The team is now screening commercially available nucleotide analogs for signs of proofreading resistance, which is the faster route to improved treatment than designing entirely new molecules from scratch. The architecture for the next generation of coronavirus antivirals is being built on the structural detail these papers made visible, and the same imaging protocols apply to every other RNA virus whose enzymes will eventually be approached at the same resolution.

A Nanopore Protocol Decentralizes Malaria Genomic Surveillance Across Six African Countries

A consortium of African research institutions published in Nature Communications a one-year deployment of decentralized nanopore sequencing for Plasmodium falciparum malaria across six sub-Saharan countries, sequencing 1,065 samples locally at a per-sample cost under $25 USD with roughly five-hour turnaround. The protocol covers antimalarial drug resistance genes, the hrp2 and hrp3 deletions that let some parasite strains evade rapid diagnostic tests, the vaccine target csp, and the polymorphic gene ama1. Each participating laboratory runs the analysis on a laptop with an offline bioinformatics dashboard that displays mapping and variant calling results as the run completes.

The structural shift the deployment demonstrates is that continental-scale genomic surveillance no longer requires shipping samples to centralized facilities in Europe or North America. Malaria kills more than half a million people annually in sub-Saharan Africa, and the parasite is evolving in two directions that threaten the region's primary defenses. Mutations in the kelch13 gene cause artemisinin-based combination therapies to clear parasites more slowly and can produce outright treatment failure when paired with partner-drug resistance. Deletions of hrp2 and hrp3 produce false negatives on the rapid tests that most clinics rely on for diagnosis. Both threats spread fastest where surveillance is thinnest.

The team estimates meaningful continental surveillance would require sequencing roughly 50,000 to 100,000 samples annually. Previous approaches concentrated capacity in a small number of well-equipped sites, with samples traveling long distances and results returning months later. The new protocol moves the instrument layer into the same buildings where patients are treated. The first-year deployment across 1,065 samples is a fraction of the eventual target, and a working proof that the architecture functions at scale and across jurisdictions.

The capability resting underneath the deployment is the maturation of nanopore sequencing into a field-deployable platform. The same hardware that once required a humidity-controlled core facility now runs on laptops in clinics with intermittent internet access. The protocol the team built around that hardware turns the platform into a public health instrument operated by African scientists on African samples for African epidemiological decisions. The data flows into a coordinated regional picture managed by the institutions that produced it.

The reach this opens is broader than malaria. The same sequencing platform, the same offline bioinformatics architecture, and the same training pipeline that let a clinic identify a kelch13 mutation in five hours can in principle track tuberculosis resistance, viral outbreaks, and zoonotic spillovers as they emerge. The infrastructure being assembled is a genomic surveillance system designed from the start to function where it is needed, on the timeline the pathogens themselves operate on.

Young Microbiota Rewrite an Old Liver

An aging liver loses regenerative capacity, accumulates fibrotic tissue, and shifts its gene expression in ways that increase vulnerability to cancer and metabolic disease. A team working with mouse models has now shown that a single intervention - transplanting gut microbiota from young animals into aged ones - reverses many of those signatures in the liver itself.

The mice in the study were middle-aged when treatment began. After receiving microbiome transplants from young donors, their livers showed reduced markers of cellular senescence, partial restoration of regenerative gene expression, and reduced fibrosis. One specific finding drew attention: the treatment suppressed expression of MDM2, a gene that regulates the tumor suppressor p53 and is implicated in cancer development and accelerated cellular aging. MDM2 overexpression in aged tissue is one of the molecular hallmarks of declining cancer surveillance, and quieting it pharmacologically has been a target of oncology research for years. Here it was quieted by microbes.

The result joins a growing body of work positioning the gut microbiome as something closer to an endocrine organ than a passive community of resident bacteria. The Century Report covered Tohoku University's work earlier this week on a constipation drug that slows chronic kidney disease through a gut-microbiome-spermidine pathway, and the oral microbiome map published the previous day showing periodontal bacteria affecting systemic cardiovascular and cognitive outcomes. The pattern across these findings is consistent: the microbial communities humans host are upstream of organ systems previously treated as if their decline were endogenous and inevitable.

What the new liver work shows specifically is that the microbiome's influence on aging is bidirectional. Young microbial communities can re-export youth-associated signaling into aged tissue, and aged microbial communities accelerate the decline that researchers had attributed to the host's own genetic clock. The liver in this experiment is responding to information arriving from somewhere it does not control, and the information can be replaced.

The trajectory implied across this body of work is a medical frame in which aging gets decomposed into systems-level processes. Each organ system has inputs. Each input can be characterized. Each can, in principle, be replaced or modulated. The liver work is one tile in that mosaic. Several more tiles arrived in the past week alone. The framing of aging as a single declining curve assigned by host genetics is losing ground to a framing in which aging is a distributed conversation between tissue and environment, with the environment partially editable. That shift in framing is itself the news beneath the news.

A New Stainless Steel Pulls Seawater Hydrogen Off the Titanium Path

A University of Hong Kong team led by Professor Mingxin Huang reported in Materials Today a new stainless steel, SS-H2, that survives the electrochemical conditions inside a seawater electrolyzer at potentials up to 1700 mV. Conventional stainless steel breaks down at roughly 1000 mV in chloride environments because the chromium oxide passive film that protects it gets oxidized into soluble chromium species. Water oxidation, the reaction at the heart of green hydrogen production, requires about 1600 mV. The gap is the reason every industrial seawater electrolyzer built to date relies on titanium components plated with platinum or gold for its structural parts.

SS-H2 closes that gap through a sequential dual-passivation strategy that the team itself described as counter-intuitive. The familiar chromium oxide film forms first. Above roughly 720 mV, a second protective layer based on manganese oxides forms on top of the chromium layer and continues protecting the metal through the entire water-oxidation window. Manganese has long been treated as a corrosion liability in stainless steel chemistry, and the researchers reported that it took them several years of atomic-level characterization before they themselves accepted what they were seeing.

The cost arithmetic the paper sets out is what reorders the field. The team estimated that for a 10 megawatt PEM electrolyzer the total system cost runs around HK$17.8 million, with structural components accounting for roughly 53 percent of that figure. Replacing those structural parts with SS-H2 in the team's accounting reduces structural materials cost by approximately a factor of 40. The number sits inside a single engineering study and will need independent validation under real operating conditions, but the order of magnitude is what carries the signal. Stainless steel is one of the most manufactured industrial materials on the planet. Titanium plated with precious metals is one of the least scalable.

The trajectory this points at is the substitution layer underneath green hydrogen rearranging. The materials assumption that locked seawater electrolysis to scarce alloys and platinum-group coatings is what made the technology expensive enough to keep desalination and freshwater electrolysis ahead on cost. A common alloy meeting the same electrochemical bar pulls one of the core assumptions of that cost stack out from under it. Combined with TSMC's gigawatt-scale offshore wind PPA covered on May 8 and Fervo's domestic geothermal supply chain, the substrate of clean hydrogen at industrial scale is being assembled along a curve where the older, scarcer path keeps getting more expensive on its own terms.

Sarawak Signs a Sovereign AI Framework Built on Federated Learning

The Sarawak Artificial Intelligence Centre signed three memorandums of understanding this past week with FLock.io and UK institutions including Oxford and Cambridge to build a sovereign AI system for the Malaysian state using federated learning and privacy-preserving infrastructure. The collaboration aims to deliver a state-wide government language model, distributed AI systems, data governance frameworks, and data center management tailored to local jurisdiction. Healthcare and agriculture are the named first deployment domains.

The proof of concept the partnership builds on is the first federated learning training of a language model on indigenous Sarawak Malay data. The model was fine-tuned on in-country devices without sending raw data to a foreign centralized server. Federated learning is a training method where model updates flow to a central coordinator while the underlying training data stays on local devices. For a state whose population speaks languages that frontier model providers have little commercial reason to support, the technique offers a path to a usable model without surrendering the linguistic record to companies headquartered elsewhere.

The structural read on the announcement is that the cost of building sovereign AI is dropping below the cost of depending on foreign providers for capabilities critical to local governance. State leadership framed the technology trajectory as moving from generative AI to agentic systems and now to coordinated multi-agent execution. Whatever the descriptive accuracy of that framing, the operational decision being acted on is that the capability layer underneath public services should run on local data, renewable energy, and local governance rather than on infrastructure rented from companies whose interests are aligned with a different national economy.

The Cambridge component establishes a research unit on AI investment economics, green economy modeling, and AI governance. The Oxford signing names FLock.io as the sole technology partner for building the production-grade federated learning infrastructure. Unitas Global Advisory, a UK firm staffed by former diplomats and ministers, sits between the academic side and the technology delivery side. The architecture of the partnership distributes capability development across institutions whose interests are not bound to any single national tech sector.

The pattern this fits sits alongside DeepSeek's frontier release optimized for Huawei Ascend silicon, the Vietnam SK Group AI stack import the April 27 edition of The Century Report covered, and the broader movement of sovereign AI infrastructure into jurisdictions that prefer to retain control over their own data and language. Sarawak is roughly 2.8 million people. A state of that size choosing to build rather than rent the AI layer it depends on is a small-jurisdiction proof point that the build path is now reachable. The capability that was nation-scale infrastructure six months ago is now within reach for subnational governments, and the substitution path away from foreign-controlled language models is widening accordingly.

OpenAI Opens Its Cyber Model to EU Review as Mythos Stays Held

The European Commission confirmed at a Monday press briefing that OpenAI is in active discussions with the bloc over preview access to GPT-5.5-Cyber, the cybersecurity variant of its latest frontier model that began rolling out to vetted defenders in early May. Commission spokesperson Thomas Regnier said the exchange has already produced an initial framework, with further discussions planned this week, and described the engagement as a way to "follow deployment of the model very closely and address security concerns." Anthropic, by contrast, has not yet granted EU regulators preview access to Mythos a full month after its restricted release. Regnier confirmed talks are underway with Anthropic too, while noting they are at a "different stage" from the OpenAI conversation.

The April 16 edition of The Century Report covered the structural divergence between the two release strategies, and on May 2 the newsletter documented the UK AI Security Institute's finding that GPT-5.5 matches Mythos Preview on advanced cyber benchmarks. What is genuinely new today is the jurisdictional layer arriving on top of those technical conditions. The EU is the first major regulator outside the US and UK to engage either lab on a concrete preview pathway for a "too dangerous to release" cyber model, and the two labs have entered that engagement on opposite footings. OpenAI's broader-access posture maps onto a faster regulatory cadence; Anthropic's tighter containment maps onto a slower one.

The forward read is that the architecture of staged release the field has been assembling since Project Glasswing is now meeting its first real test as a multi-jurisdictional system rather than a single-lab choice. Defenders inside the EU's financial, energy, and telecommunications sectors face the same vulnerability surface as their US and UK counterparts, and the Commission's involvement signals that distribution decisions made in San Francisco no longer travel cleanly through a single national regulator. The pattern that is forming, visible across the IMF spring meetings in April, the Bank of England's emergency coordination, the joint Five Eyes guidance, and now Brussels, is a multi-continent verification layer being built out of the conditions a single lab's release decision can no longer contain. Whichever cyber model arrives in European critical infrastructure first will arrive having passed through more eyes than any prior frontier capability, and the institutional muscle being built around that review is the substrate the next release cycle inherits.

The Other Side

For five years, frontier capability release has run on a discretionary access architecture: a handful of labs deciding which defenders, inside which jurisdictions, get to see a "too dangerous to release" model first. The assumption inside that architecture was that the company could be the gatekeeper for a capability shaping cyber defense across financial systems in Frankfurt, energy systems in Lyon, and telecoms in Helsinki, and that bilateral relationships between a lab and a few selected major-power regulators amounted to a sufficient review function for the rest of the world.

That assumption cracked yesterday at a Commission press briefing. The Commission confirmed OpenAI is in active preview-access talks over GPT-5.5-Cyber, with an initial framework already produced and further meetings this week. The UK AI Security Institute had already published, on May 2, that GPT-5.5 matches Mythos Preview on advanced cyber benchmarks: 71.4% Expert CTF vs. 68.6%, three of ten vs. two of ten on "The Last Ones." The technical asymmetry the two opposite release postures were predicated on is in public reporting now. Five Eyes have issued joint agentic-AI guidance. The Bank of England ran emergency coordination. The IMF spring meetings forced the labs to talk to finance ministries. Brussels arrived on terms it set.

What becomes impossible to maintain in this configuration is the architecture of single-lab discretion deciding who sees what. The cyber threat surface defenders face does not respect a vetted-organizations list, and the regulators responsible for critical infrastructure are no longer accepting bilateral terms set inside the lab. Thomas Regnier, the Commission spokesperson who said the intent was to "follow deployment of the model very closely and address security concerns," is the public face of the moment a gating decision over a frontier cyber model stopped being a private corporate call.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: a viral enzyme that has defeated antiviral chemistry since 2020 yielding its geometry at 2.4 angstroms across two new design pathways, continental-scale pathogen surveillance running on clinic laptops at twenty-five dollars a sample with five-hour turnaround across six African countries, an aging liver responding to a transplant of microbial communities from a younger body and quieting the very gene that governs cancer surveillance, a manufacturable stainless steel collapsing the cost of seawater hydrogen by an order of magnitude through a passivation layer chemistry had been treating as a liability for decades, a Malaysian state of 2.8 million people contracting Oxford and Cambridge to build its own sovereign AI on federated infrastructure rather than rent one from a foreign provider, and a multi-continent verification layer assembling itself around frontier cyber capability one jurisdiction at a time. There's also friction, and it's intense - two frontier labs entering European regulatory talks on opposite footings with Mythos still held back from Brussels, the rules of staged release for dangerous capability getting negotiated through bilateral conversations rather than any standing institutional channel, the cadence of capability deployment continuing to outrun the regulatory bodies most exposed to it, and a 50,000-sample annual surveillance target still many years of build-out away from the 1,065 sequenced in year one. But friction generates polish, and polish is what reveals the surface the next layer has to bond to. Step back for a moment and you can see it: the resolution at which biological systems become legible keeps falling, the cost of building the substrate of medicine and energy and intelligence keeps falling with it, the number of jurisdictions and communities able to participate in shaping that substrate keeps widening, and the materials and chemical and institutional assumptions that locked each domain into scarce inputs are giving way inside the same news cycle. Every transformation has a breaking point. A protective layer can corrode away under load... or form a second skin beneath the first that holds where the original chemistry could never reach.

AI Releases & Advancements

New today

  • Google: Launched the new AI-powered Google Finance across Europe with full local language support, including AI-powered research, advanced charting/technical indicators, expanded commodities and crypto data, and live earnings call audio with synchronized transcripts and AI-generated annotated highlights; Deep Search in Google Finance is now globally available. (Google Blog)

Other recent releases

  • NVIDIA: Released Star Elastic, a single Nemotron Nano v3 checkpoint that contains 30B, 23B, and 12B reasoning model variants extractable via zero-shot slicing without fine-tuning; available on Hugging Face under nvidia/NVIDIA-Nemotron-Labs-3-Elastic-30B-A3B in BF16, FP8, and NVFP4 precisions. (MarkTechPost)
  • Zyphra: Released ZAYA1-74B-Preview, a 74B-total / 4B-active MoE reasoning-base checkpoint trained on AMD MI300X GPUs with ~15T tokens of pretraining and 256K context extension, available under Apache 2.0 on Hugging Face. (Zyphra)
  • Zyphra: Released ZAYA1-VL-8B, a vision-language MoE model with 700M active / 8B total parameters built on ZAYA1-8B-base for visual understanding, grounding, and OCR tasks, available under Apache 2.0. (Hugging Face)
  • OpenAI: Launched the Codex Chrome extension, enabling Codex to operate signed-in browser sessions across parallel tabs for tasks on sites like LinkedIn, Salesforce, and Gmail, with per-domain allowlist/blocklist controls. (OpenAI Developers)
  • OpenAI: Released three new realtime audio models in the Realtime API (now generally available): GPT-Realtime-2 with GPT-5-class reasoning and a 128k context window, GPT-Realtime-Translate (live speech translation across 70+ input and 13 output languages), and GPT-Realtime-Whisper (low-latency streaming speech-to-text). (OpenAI)
  • Adobe: Launched a new productivity agent in Acrobat that lets users chat with PDFs, generate presentations/podcasts/social posts, and orchestrate text and image generation; available now in Acrobat AI Plans, Acrobat Studio, and the new Acrobat Express. (Adobe News)
  • Coder: Released Coder Agents in beta, a self-hosted native agent architecture that runs AI-driven developer workflows entirely inside an enterprise's own infrastructure without sending source code or prompts to external models. (SD Times)
  • ElevenLabs: Launched Studio Agent inside ElevenCreative, a conversational AI co-editor built into the Studio video timeline that drafts videos from a text prompt, places clips/voiceovers/sound effects frame-accurately, and supports interruption/handback. (ElevenLabs Docs)

Sources

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

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