Sign in Subscribe
Century Report

An AI Agent Finds Four New Energy Superconductors - TCR 07/06/26

Alibaba's Elements Claw scanned 2.4M crystal structures in 28 hours and surfaced four lab-verified superconductors, as AI discovery outpaces the bench.

Ben Linford

Ben Linford

— 17 min read
Three-panel Century Report infographic: AI-accelerated discovery of superconductors and drug targets, human labor shifting to AI orchestration, and chip and grid limits testing the buildout.

The 20-Second Scan

  • Alibaba's Elements Claw AI agent screened millions of crystal structures and surfaced four previously unknown superconductors, all later verified in laboratory experiments.
  • Amazon is winding down Mechanical Turk, the 21-year-old micro-task marketplace, closing it to new customers July 30 as automation absorbs the piecework it was built to route.
  • Nvidia's Kyber rack system slips to 2028 as the annual chip cadence collides with what Taiwan's fabs and packaging lines can physically build.
  • Newly released FOI documents show a flagship Lanarkshire AI datacentre has "no prospect" of running on the all-renewable power its developers promised.
  • An AI-guided CRISPR screen surfaced two topical small-molecule targets whose efficacy was comparable to injected psoriasis biologics in mice.
  • The NHS app is turning on AI triage to route patients toward GP, pharmacy, or A&E care, reaching 200,000 patients over the next year.
  • Record numbers of vehicle-to-grid EVs fed power back into a straining grid during last week's heat wave instead of drawing it down.
  • Some wealthy families are paying up to $75,000 a year to enroll their children in AI-taught private schools like Alpha and Forge Prep.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

Reported in the same weekend, an AI agent read 2.4 million crystal structures in 28 hours and returned four superconductors that Alibaba's Damo Academy says the lab later confirmed, while a separate AI-guided CRISPR screen mapped two overlooked genes that could turn injected psoriasis biologics into a topical cream. The commonality between the stories: an intelligence system perceiving the few candidates that matter inside a search space no research team could ever walk by hand, then leaving validation to people. The scarcity that priced these discoveries out of reach was never the biology or the physics. It was the arithmetic of the search, and that cost is collapsing.

As capability climbs the value chain, the human layer it was built on top of is thinning. Amazon is closing Mechanical Turk to new customers on July 30 after 21 years, retiring the marketplace that once hid human judgment inside "intelligent" systems by the penny. The piecework it coordinated has been absorbed by the models its workers helped train. What forms above that vanishing floor is orchestration work, sitting closer to specification and oversight than to repetition.

Underneath both threads runs a harder floor: the physical substrate. Nvidia's Kyber rack slipped to 2028 because advanced packaging and memory supply cannot scale on the annual cadence its roadmap assumed, and newly released FOI documents show a flagship Lanarkshire datacentre has "no prospect" of meeting its all-renewable promise. Fabrication and grid capacity resist capture in a way an architecture does not. A slipped flagship is a year for every rival design to close distance, and an exposed energy shortfall is the most concrete argument yet for building firm clean power faster.

That same honest accounting showed up on the demand side, where a record number of vehicle-to-grid EVs fed stored power back during last week's heat wave instead of straining the system - even as that same heat wave had the largest US grid operator ordered to run every available fossil plant flat out to keep the system standing. Watch what the physical limits pull into being: diversified silicon, a grid stress-tested into genuine cleanliness, and discovery gated by how fast a lab can validate rather than how many candidates it can afford to test.

The 20-Minute Deep Dive

An AI Agent Read 2.4 Million Crystals and Found Four Superconductors

For decades, the hunt for new superconductors moved at the pace of trial and error. The SuperCon database, the field's running ledger, holds roughly 26,000 confirmed materials accumulated across generations of laboratory work. Alibaba's Damo Academy reports that its Elements Claw AI agent added four previously unknown superconductors in a single research cycle - candidates the system nominated and that the team says were afterward verified in the lab.

The mechanics are worth slowing down on. Elements Claw is built on a one-billion-parameter foundation model trained on 125 million molecular and crystal structures. Working with researchers at Renmin University of China and the University of Chinese Academy of Sciences, the system screened 2.4 million stable crystal structures in 28 hours of GPU time and narrowed the field to roughly 68,000 promising candidates. From there it pushed toward the handful that survived physical validation. The four verified compounds are Alibaba's own account of the result, and that self-report deserves the usual caution until independent groups reproduce the synthesis. What is not in dispute is the compression: a search space that would swallow a career of bench work, traversed in a little more than a day.

This is the shape the whole discovery pipeline is taking, extending the pattern the July 2 edition of The Century Report documented when a Nature paper described an AI agent closing a yeast cell's full hypothesis-test-revise cycle across eight functional modules without human intervention. The agent is doing the perceiving itself, recognizing structural signatures of superconductivity across a space too large for any team to walk manually, then handing a short list to people who confirm it. Superconductors sit underneath some of the most consequential problems in the transition already underway: lossless power transmission, magnetic confinement for fusion, MRI and quantum hardware. Every new member of that family is a widening of what the grid and the lab can eventually do.

The old constraint on materials science was arithmetic more than imagination - millions of plausible structures, each expensive to evaluate, most of them dead ends. When an intelligence system can read the whole space in an afternoon and mark the few worth building, the bottleneck stops being how many candidates a field can afford to test and becomes how fast the lab can validate what the model already sees. That is a different world than the one where 26,000 superconductors took decades to find.

Mechanical Turk Winds Down as Automation Reaches the Piecework Layer

Amazon will stop accepting new Mechanical Turk customers on July 30, closing the front door on a marketplace that ran for 21 years. Existing customers keep access for now, but no new features are coming, and workers report that AWS has begun closing accounts. Amazon framed the decision as the result of "careful consideration." The quieter fact underneath that phrasing is that the specific work MTurk was built to distribute has been steadily absorbed by the systems its workers once trained.

The service launched in 2005 to route the tasks that resisted full automation - labeling images, judging sentiment, solving the puzzles machines could not yet solve. The name itself carried the irony from the start: the original 18th-century Mechanical Turk was a chess-playing automaton with a human hidden inside, and Amazon's version made the hidden human labor behind "intelligent" systems into an explicit marketplace. From 2018 the platform was rebranded toward data annotation for SageMaker, positioning MTurk as connective infrastructure for training the models that would eventually make much of the labeling itself unnecessary.

That loop closed faster than anyone marketed. A small 2023 analysis found that between a third and nearly half of MTurk task responses showed signs of large language model use, to complete the tasks they were paid to do by hand - human workers delegating to the machines, inside a marketplace built to supply human judgment to those same machines. Reddit workers say the platform effectively died years ago, hollowed out by bots and fraud well before this announcement. The micro-task economy MTurk pioneered is thinning because the layer of labor it was built to coordinate has moved up the stack, even though the work it did still needs doing.

This is the Labor Market Restructuring arc at one of its clearest inflection points. The tasks that anchored MTurk were the most fragmented, lowest-margin form of digital piecework ever assembled - human cognition sold by the penny to bridge the gap between what machines could almost do and what they could actually do. As that gap narrows toward zero, the coordination problem changes shape entirely. What emerges above the vanishing piecework layer is orchestration: fewer people labeling individual images, more people directing systems that label at scale, verifying edge cases, and designing the judgment criteria the models apply. That shift is exactly what the July 5 edition of The Century Report found already showing up in the jobs data, where 2.85 million active listings revealed demand concentrating on AI-orchestration skills even as the broader labor market cooled. The floor that MTurk represented is dissolving, and the work that remains sits higher - closer to specification and oversight than to repetition. The century-old trick of hiding a human inside the machine is ending because the machine no longer needs one at that layer, and the humans are being pulled toward the parts of the problem that still genuinely require them.

Read one more layer down, and MTurk was the purest form of a logic that treated human judgment as a raw input to be sourced as cheaply as fragmentation allowed. Its closing is that logic running out of room: the cheapest way to buy a penny of cognition is no longer to source it from a person at all, which removes the economic reason the penny-piece market ever existed. What thins here is the arrangement that priced human judgment lowest, while the judgment itself remains as valuable as ever.

Nvidia's Kyber Rack Hits the Manufacturing Ceiling

Nvidia's Kyber rack system, the next-generation platform built around its Rubin-family accelerators, has slipped to 2028, according to reporting from CNBC citing supply-chain sources in Taiwan. The silicon roadmap is intact. What has bent is the physical capacity to fabricate, package, and assemble these systems at the pace the roadmap assumed. Advanced packaging lines, high-bandwidth memory supply, and the specialized rack integration that Kyber requires cannot all scale on the annual cadence Nvidia set for itself when it promised a new architecture every year.

This is the quiet fact that the compute-buildout story keeps arriving at. For two years the industry narrated an unbroken exponential: each generation more capable, shipped faster, on a clockwork annual beat. That cadence was always a claim about manufacturing, and manufacturing has a floor. There are only so many advanced packaging facilities in the world. Memory suppliers are already stretched by a shortage that has pushed prices up across the sector, the same crunch the July 2 edition of The Century Report documented pushing inference-hardware diversification forward as DRAM prices surged 55 to 60 percent on AI demand. When one company controls the leading design but depends on a handful of foundries and packagers to turn that design into deployable racks, the timeline belongs to the fabricators, not the designer.

The interesting part is what a slipped flagship does to everyone standing behind it. A one-year delay in the leading platform is a one-year window for alternatives to close distance. AMD's accelerator line, Google's TPU generations, Amazon's Trainium, and a widening field of sovereign and open-silicon efforts all measure their competitiveness against Nvidia's shipping cadence. When that cadence stretches, the gap they need to cross gets shorter, and the buyers negotiating multi-year compute contracts gain reasons to diversify what they lock into. The manufacturing ceiling that constrains the leader is the same ceiling that flattens the field beneath it.

Step back and the delay reads less like a stumble and more like the moment the compute story stopped being about any single company's roadmap. The assumption underneath the last two years - that whoever holds the leading design holds a durable lead - depends on that design reaching customers faster than anyone else can. Physical fabrication is proving to be the great equalizer. It cannot be captured the way an architecture can, and it forces the whole industry onto a shared timeline set by the real limits of factories, materials, and skilled assembly. What emerges is a compute landscape where advantage is measured in months rather than years, and where the pressure to build more packaging capacity, more memory supply, and more foundry throughput ultimately widens access to advanced hardware for everyone downstream. The ceiling that delayed one rack is the same force pushing the entire supply chain toward abundance.

A Scottish Datacentre and the Real Cost of "All Renewables"

Documents released under freedom-of-information rules show that a flagship AI datacentre planned for Lanarkshire has, in the assessment contained in the disclosures, "no prospect" of meeting the all-renewable-energy commitment its developers made publicly. The promise of running entirely on green power was central to how the project was presented. The internal record tells a more complicated story about what the grid around the site can actually supply and when.

The developers' framing - that the facility would be powered by renewables - is a claim, and the FOI documents are the reason it can now be tested against what the connecting infrastructure is capable of delivering, extending a pattern the July 5 edition of The Century Report documented when a Guardian investigation found roughly two-thirds of OpenAI's touted £30 billion Stargate UK commitment was hypothetical, with the company apparently never having visited one of its own flagship sites. Scotland generates enormous quantities of wind power, but generation and delivery are different problems. A datacentre drawing continuous, heavy load needs firm power around the clock, and matching that demand to intermittent renewable supply requires either storage, transmission capacity, or grid connections that do not yet exist at the scale and timing the project assumes. The gap between the public commitment and the engineering reality is where these disclosures land.

This friction is uncomfortable, and it is also doing something useful. The AI buildout is forcing the true cost of energy into the open in a way that decades of abstract climate targets never did. When a marquee facility cannot honor an all-renewable pledge, the shortfall becomes a clear, auditable fact rather than a distant aspiration. It puts precise numbers on how much firm clean power the intelligence era actually requires, and it exposes exactly where the grid falls short. That exposure is what drives investment toward the storage, transmission, and generation that close the gap. The pressure these datacentres place on the system is uncomfortable precisely because it is honest.

There is a longer arc here worth holding alongside the shortfall. Every one of these projects that collides with grid reality generates data - real load figures, real connection timelines, real storage requirements - that planners have never had at this resolution before. The demand is arriving faster than the clean supply, and that mismatch is the single most powerful argument for building the supply out faster. A decade from now the grids being stress-tested by AI load today will be the grids that carry it on genuinely clean power, because the demand made the buildout unavoidable. The friction between what was promised and what the wires can deliver is the mechanism forcing the honest accounting that a fully renewable grid was always going to require.

The deeper shift the disclosure marks is where the energy bill lands. For years a developer could book an all-renewable pledge as a public asset while leaving the actual generation, storage, and transmission to be sorted out later, or never, with the cost of that gap carried by the grid and the public. An FOI release that lets the pledge be checked against what the wires can physically deliver is the point where that externalization stops working, and the real cost of firm clean power starts landing where the demand is created. Watch for the same disclosure pattern, following the Stargate UK figures the July 5 edition documented as roughly two-thirds hypothetical, to keep converting confident buildout claims into facts the connecting infrastructure can be measured against.

AI-Guided CRISPR Points Psoriasis Treatment Toward a Cream

The most effective psoriasis drugs today are biologics - antibodies that block the IL-17 pathway, delivered by injection, priced accordingly, and out of reach for a large share of the 125 million people the disease affects worldwide. A study published in Nature Communications describes a route toward the same biological effect in a form as ordinary as a topical cream.

The researchers ran a genome-wide CRISPR knockout screen in primary human epidermal keratinocytes, systematically switching off genes to find the ones that control surface IL17RA, the receptor the injected antibodies target. To prioritize which hits were most important, they used VirtualCRISPR, a language-model framework trained on functional-genomics data that reads screen results and surfaces regulators worth chasing. The screen and the model together pointed to two genes the field had largely passed over: ALOX5, which encodes 5-lipoxygenase, and OXTR, the oxytocin receptor. Both already have small molecules built against them. Topical zileuton, an existing ALOX5 inhibitor, combined with cligosiban, an OXTR antagonist, suppressed psoriasis-like skin inflammation in mice - comparable to the effect of a systemic anti-IL17RA antibody.

The honest framing is that this is a mouse result and an early one. The path from imiquimod-induced dermatitis in a mouse model to a cream on a pharmacy shelf runs through human trials, formulation work, and regulatory review. What the study demonstrates is a method: an AI-guided screen that converted a broad, unbiased genetic search into two specific, druggable targets that human intuition had not flagged, using molecules that already exist. That reuse of known compounds is what could compress the timeline - the discovery skips the years usually spent inventing a new drug from scratch.

Read against the superconductor work surfacing the same week, the pattern is hard to miss. In both cases an intelligence system did the hardest part - reading a search space no human could hold in their head and returning the few candidates that matter - and left validation to the lab. When the same approach that finds a superconductor also finds a way to turn an injected biologic into a topical, the reach of a treatment stops being gated by how the molecule is delivered or how many researchers can chase a target. The scarcity that priced these drugs out of reach was the cost of the search more than the biology, and that cost is falling fast.

The Other Side

For two years, the compute story ran on one assumption: whoever holds the leading chip design holds a lead that compounds. Get the best architecture, ship it on an annual beat, and the gap behind you widens every cycle. Capital priced itself against that idea, and so did the whole industry.

Kyber slipping to 2028 breaks the assumption at its weakest joint. The silicon roadmap is intact. What bent is the physical capacity to build the racks - advanced packaging lines, high-bandwidth memory, the specialized assembly Kyber needs, none of it able to scale on the yearly cadence Nvidia set for itself. A design can be captured. A packaging facility cannot. A one-year delay in the leader is a one-year window for AMD, Google's TPUs, Amazon's Trainium, and a widening field of open and sovereign silicon to close the distance.

The ceiling that caps the front-runner lifts everyone behind it. Advantage that used to be measured in years is now measured in months, and the scramble to build more packaging, more memory, more foundry throughput widens access to advanced hardware for everyone downstream.

Imagine a researcher at a modest university lab in 2032, the kind of place that in 2026 waited at the back of every allocation queue. The compute she needs to read a million molecular structures overnight sits in her own building, hers to run, because the fabrication bottleneck that once rationed frontier hardware to a handful of buyers forced the entire supply chain to widen instead. The year the leader stumbled is the year the field stopped belonging only to that leader's calendar.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: Alibaba's Elements Claw agent reading 2.4 million crystal structures in 28 hours and returning four superconductors the lab says it later confirmed, an AI-guided CRISPR screen pointing two overlooked genes toward turning an injected psoriasis biologic into a topical cream, the NHS app switching on AI triage to route 200,000 patients toward the right GP, pharmacy, or A&E care, and a record number of vehicle-to-grid EVs feeding stored power back into a straining grid during last week's heat wave instead of drawing it down. There's also friction, and it's intense - Amazon closing Mechanical Turk to new customers after 21 years as automation absorbs the penny-priced piecework its workers once trained, Nvidia's Kyber rack slipping to 2028 because advanced packaging and memory supply cannot scale on the annual cadence its roadmap assumed, FOI documents showing a flagship Lanarkshire datacentre has "no prospect" of meeting its all-renewable pledge, and AI-taught private schools charging some families up to $75,000 a year for the head start. But friction generates tolerance, and tolerance is the measure of what a system can actually hold. Step back for a moment and you can see it: intelligence systems doing the perceiving now, reading search spaces no lab could walk by hand and handing back the few candidates that matter, while the human piecework layer they were built on thins and the work that remains rises toward specification and oversight, and the physical substrate under all of it - fabrication capacity, memory supply, firm clean power - proving to be the real constraint no architecture can capture and no pledge can wish away. Every transformation has a breaking point. A ceiling can cap everyone beneath a single builder's pace... or, once it holds firm for all of them, pull the whole field up to meet it.

AI Releases & Advancements

New today

  • Sakana AI: Launched Sakana Translate, a free browser-based translation tool added to Sakana Chat powered by the Namazu model series, offering Translate, Proofread, and Ask modes for Japanese-English-Chinese translation. (Sakana AI)
  • Synthetic Sciences: Released OpenScience, an open-source, Apache 2.0, model-agnostic AI research workbench for machine learning, biology, physics, and chemistry research, installable via npm install -g @synsci/openscience. (GitHub)

Other recent releases

  • LlamaIndex: Open-sourced legal-kb, a reference application demonstrating agentic retrieval over Index v2 (LlamaParse Platform), exposing retrieve, find, read, and grep tools that let AI agents autonomously crawl large evolving knowledge bases. (GitHub)
  • Alibaba: Open-sourced Page Agent, a JavaScript in-page GUI agent (MIT license) that reads a webpage's live DOM and lets natural-language commands control web interfaces; model-agnostic via OpenAI-compatible backends including DashScope/Qwen, GPT, Claude, or local Ollama. (GitHub)
  • Interfaze: Released diffusion-gemma-asr-small, an open-source multilingual diffusion-based ASR model built as a ~42M-parameter adapter on a frozen DiffusionGemma backbone, supporting 6 languages and outperforming prior diffusion-based ASR systems on LibriSpeech. (Interfaze)

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.

Read more