Law was supposed to be easy for AI.

The profession runs on documents. Contracts, briefs, motions, discovery requests, regulatory filings. Every billable hour leaves a paper trail. And unlike medicine, where AI must contend with the complexity of biological systems, or finance, where microsecond arbitrage advantages disappear instantly, legal work operates on human timescales with human language. A contract dispute from 1982 reads much like one from 2024.

The pitch writes itself: AI systems that draft documents in seconds, review discovery in minutes, and catch errors that bleary-eyed associates miss at 2 AM. The technology exists. Westlaw’s Deep Research promises comprehensive legal research in under ten minutes. Clio’s Vincent AI will hand-craft you a personalized article from a treatise. Harvey.AI, trained on elite law firm work product, offers an agentic attorney assistant to BigLaw.

And yet.

Recent surveys report impressive AI adoption numbers in law, with up to 79% of attorneys claiming to use artificial intelligence at their firms. But these figures measure exposure, not integration. Having Copilot enabled or using the AI features baked into existing tools like Relativity counts as “adoption” in survey responses, even when actual workflows remain unchanged. The attorneys I speak with at conferences and Continuing Legal Education events across the country tell a different story: most firms have experimented with AI, few have transformed how they practice. The modal American lawyer in 2026 still works on a desktop computer, still pays for Westlaw or Lexis, and still approaches AI with the same wariness they brought to the cloud a decade ago.

Structural barriers make legal practice resistant to technological diffusion in ways that other industries don’t face. Understanding these barriers matters because law is where AI meets civic infrastructure. Courts, contracts, regulations, and rights flow through lawyers. If AI can’t diffuse through law, its broader social impact will remain constrained.

The Data Moat

Legal AI has a data problem that most industries don’t face, and it has two layers.

The first layer is raw legal data. To build useful AI for legal research, you need comprehensive databases of case law, statutes, regulations, and secondary materials. Only three entities in the United States have anything approaching complete coverage: Westlaw (Thomson Reuters), Lexis (RELX), and vLex/Fastcase, which Clio acquired in a $1 billion deal in November 2025. That deal pulled the third meaningful legal research database under a company focused on small and mid-size firm practice management, and Clio’s $5 billion valuation and $500 million Series G round suggest investors see the strategic value of owning one of only three complete legal datasets in the country. Everyone else either licenses from one of these three or works with incomplete data.

The second layer is what makes those databases worth paying for. Westlaw and Lexis don’t sell raw judicial opinions (much of that is publicly available). They sell the editorial infrastructure built on top: headnote taxonomies that organize millions of opinions into searchable categories, practice guides written by specialists over decades, and treatises that synthesize primary law into usable guidance. A California real estate attorney without access to Miller and Starr would be at a serious disadvantage, not because the underlying case law is hidden, but because navigating it without an expert-curated roadmap takes exponentially longer. Imagine being handed an encyclopedia to learn something vs. having a beautifully curated twenty-page guide from a panel of practitioners who have been through the procedure a thousand times. That’s the difference: substantive knowledge plus procedural shorthand, built up over years of practice in a single area of law.

The litigation around database access shows how fiercely incumbents defend this moat. Thomson Reuters sued Ross Intelligence not over case law itself, but over Westlaw’s headnote taxonomy, the editorial layer that organizes and summarizes judicial opinions. In February 2025, the court sided with Thomson Reuters, rejecting Ross’s fair use defense. The message: even if the underlying legal materials are free, the value-added structure built on top of them is proprietary and protected. Open-source alternatives like SALI have emerged in response, offering a vendor-neutral taxonomy that AI developers can use without licensing risk.

Cracks in the Moat

The data moat is real, but it may be more porous than it appears.

The Free Law Project’s CourtListener provides free access to millions of federal and state court opinions, oral arguments, and PACER documents. State-level open data initiatives, like Oklahoma’s, have made primary legal materials freely accessible. Harvard’s Caselaw Access Project digitized every official state and federal case through 2020. All U.S. state bar associations now provide members with free access to either vLex Fastcase or Decisis, which, for a solo practitioner handling state court matters, might be enough.

The editorial layer that was essential for human researchers may matter less for AI systems. vLex’s Vincent AI demonstrates a different approach: using AI to generate the synthesis layer rather than paying human experts to write it. Damien Riehl (Clio’s Tech Evangelist, perhaps best known for his viral TED Talk on music and copyright) calls this a “Me-Tise,” a personalized knowledge base rather than the traditional legal treatise. If AI can create practice guide-quality analysis from primary sources, the competitive advantage of having the best human-written treatises diminishes. The moat doesn’t disappear, but it gets shallower.

And there’s a whole category of legal technology that has no data moat at all. Legal research incumbents sit behind proprietary datasets. But a huge swath of legal tech (eDiscovery platforms, case management, billing, client intake, compliance, marketing, document automation) consists of traditional SaaS offerings where the value proposition is software engineering and workflow, not proprietary data.

When Anthropic launched legal skills as open-source plugins for its Claude Cowork platform on February 2, 2026, the market reaction was immediate and brutal: Thomson Reuters dropped nearly 16% in a single day (its worst on record), LegalZoom fell almost 20%, RELX lost 14%, and Wolters Kluwer shed 13%, roughly $285 billion overnight. The damage was concentrated in the SaaS-heavy segments. That’s why the vendor halls at Legalweek and Techshow are packed with AI startups attempting new, innovative ways to integrate AI into traditional workflows.

These segments face a threat that most legal tech vendors didn’t anticipate: frontier AI labs are no longer content to serve as infrastructure underneath vertical software. They’re increasingly building application-specific capabilities to directly serve users. Anthropic’s Claude Cowork, OpenAI’s Codex, and Perplexity’s computer-use agents automate entire desktop workflows, not just individual legal tasks. They draft documents, manage calendars, send emails, organize files, and handle billing without any legal-specific software in the stack. When the AI operates at the OS layer, the SaaS application sitting on top of it starts to look redundant.

Could frontier AI labs eventually purchase one of the Big Three legal datasets? They have the capital. The reason it hasn’t happened yet is simpler than the moat theory suggests: Thomson Reuters has a market cap of around $75 billion; RELX sits around $85 billion, which sounds enormous until you compare it to the markets these companies are already chasing in healthcare, finance, enterprise software, and consumer products. Legal data is a rounding error on their strategic roadmaps. For now.

The Messy Middle: Why Firms Can’t Execute

The barriers inside most law firms are organizational, not technical. Even firms that want AI can’t deploy it because their data is a mess and their governance structures punish change.

I learned this firsthand while consulting for a mid-sized firm eager to modernize. When I asked where their data was stored, the answer came in pieces: some was in iManage, some on SharePoint/OneDrive, some on an old local server (“the S:\ Drive”), and some was still paper. Before any AI system could leverage the firm’s accumulated wisdom, someone would need to locate, digitize, organize, and normalize years of fragmented work product scattered across incompatible systems and storage media. Even better if you could get it into a data lake and attach meaningful metadata.

This is a common story. Firms have spent years accumulating data in whatever system was convenient at the time, with no eye toward future interoperability. Even firms that have invested in document management systems find that adoption has been inconsistent: partners maintain personal filing systems, assistants save documents in non-standard locations, naming conventions are useless, and metadata is an afterthought. In niche practice areas where institutional knowledge is everything, a firm that has handled hundreds of similar transactions holds a substantial advantage, but only if that knowledge can be retrieved, synthesized, and deployed. Most firms can’t do that yet.

The governance problem compounds the data problem. A mid-sized, 30-person law firm has 10 to 15 partners, each with an equity stake and a vote on firm decisions. Unlike a corporation, where a CIO can mandate new tools across an organization, law firms operate as partnerships where every senior lawyer has veto power over changes that affect their practice. A partner who doesn’t want to learn new software can refuse, and the firm’s management has limited ability to force compliance. Technology decisions devolve to the lowest common denominator. The partner who complains loudest about change gets to block it. Enterprise legal technology vendors reinforce this pattern by focusing sales on large firms, leaving smaller firms with self-service products and no implementation support.

Small and mid-size firms face this most acutely because they lack dedicated technology leadership. A 200-lawyer firm might have a CIO with genuine authority. A 20-lawyer firm has a “technology partner” whose actual job is practicing law, with IT responsibilities layered on top. That partner’s time for evaluating AI tools competes with billable work, client development, and everything else. My hobby-horses, AI-literacy and AI-competency, take a backseat to hitting the requisite 1,900 annual billable hours.

After presenting at a Midwest state bar annual conference, I was approached by a young, AI-forward attorney who had been serving on his (fairly large) firm’s internal AI committee for a year. The committee’s senior partners had repeatedly deferred any decision on AI adoption, citing risk. Meanwhile, he’d been teaching himself Python and building tools in Cursor on his own time because he wanted to develop custom solutions for the firm. They had no appetite for any of it. He asked me what he should do. I told him he needed to leave. I gave him my slides from the presentation, told him to make them his own, and advised him to find local firms where he had connections and pitch himself as their in-house AI attorney. I told him to mark up his salary and spend his extra time evaluating the AI legal tools on the market while continuing to develop his own. He pitched the position to three different firms and got three offers. He’s working at a rising firm that is going all-in on AI as their in-house AI expert.

This pattern points toward increased mid-tier competition. The firms that adopt early gain a structural advantage in both efficiency and talent acquisition. Meanwhile, those that defer will watch their best associates walk out the door. And new firm structures are accelerating the shift. Arizona has pioneered Alternative Business Structure (“ABS”) programs that allow nonlawyer ownership of legal practices, opening the door for technology companies to co-own and operate law firms.

The results are already visible: Eudia Counsel, a Palo Alto-based AI startup with $105 million in funding, launched the first AI-augmented law firm under Arizona’s ABS program, embedding AI directly into M&A and contracting workflows for Fortune 500 clients. Virgil, co-founded by Answer.AI’s Jeremy Howard, Eric Ries (The Lean Start-up), and start-up attorney Luke Versweyveld, has created a law firm where developers and attorneys work hand-in-glove on a daily basis to automate the practice of law. Plymouth Street has carved out tech immigration as a near-fully automated practice area. And in the UK, Garfield.Law became the first firm authorized by the Solicitors Regulation Authority to deliver regulated legal services entirely through AI, handling small claims debt recovery starting at £2 per letter.

There is serious innovation at the margins, and 2026 does not look like 2024. In 2024, we saw a lot of firms talking about being “AI Native” and “leveraging AI,” but in 2026, we see firms truly innovating, whether it’s taking advantage of novel regulatory structures or rethinking entire workflows from the ground up with AI.

The Efficiency Paradox

When AltFee won the ABA TECHSHOW 2024 Startup Alley competition, company representative Scott Leigh offered a simple proposition: our product will help you divorce from the billable hour.

I’ve attended this competition for years, and this moment stuck out. The Startup Alley pitch competition is a live vote: attorneys in the room choose the winner in real time. No panel of judges, no curated selection committee. Practicing lawyers from all walks of life, voting with their phones. And they chose a tool designed to help them escape hourly billing.

The billable hour creates a misalignment between AI efficiency and law firm economics, but the misalignment is subtler than the usual narrative suggests. Every hour of associate time that AI eliminates is an hour that can’t be billed. The most routine, automatable legal work is also the most lucrative on a per-hour basis, as clients pay associate rates for tasks that don’t require partner judgment.

I was presenting at a State Bar conference, showing examples of AI services that can review tens of thousands of documents in minutes with a lower error rate than human attorneys. A senior attorney in the audience interrupted: “Why the hell would I want to do that?!” The question highlights the strain AI puts on traditional billing models. A majority of the industry still operates in a model where armies of junior associates grind through tasks that are ripe for automation.

But the efficiency narrative gets complicated by risk. When an AI misses a privileged document that gets produced in discovery, someone is responsible for that failure. When AI-generated analysis contains a subtle error that shapes litigation strategy, someone bears the consequences. An attorney can’t tell the General Counsel at a Fortune 500 company that a bot reviewed their M&A documents and expect that explanation to suffice if something goes wrong. The “human in the loop” is how lawyers sleep at night and how they retain the trust of their clients.

The billable hour is already under significant pressure from sources that have nothing to do with AI. Clients demand alternative fee arrangements, caps, fixed fees, and efficiency reporting. Realization rates have declined as clients push back on invoices. According to Clio’s 2025 Legal Trends Report, 59% of firms now offer flat fees exclusively or alongside hourly rates. The question is who captures the efficiency gains from AI: the firm, the client, or both? That uncertainty creates investment hesitation when combined with genuine concerns about quality and liability.

Some firms have stopped waiting for the answer. Whitney Harper and Gwen Griggs founded ADVOS Legal on a simple premise: stop measuring hours and start measuring value. Their subscription model rewards efficiency instead of penalizing it, and through ADVOS Pro they train other firms to do the same. Hello Divorce, founded by family law attorney Erin Levine, runs the same play at the consumer end. Levine sells DIY divorce starting at $99 and full attorney assistance averaging around $2,000, against a national average of $26,000 for a contested matter. Both firms make AI efficiency the product.

Individual attorneys run a smaller version of the same model. A bankruptcy lawyer offers a flat $1,500 package through the 341 meeting of creditors. AI shifts the attorney’s role from drafting and research to supervising micro-workflows and reviewing outputs, so an attorney who once handled 15 cases a month can manage 40 at the same fee. The bottleneck was never legal judgment. It was the document preparation around it. Productize the workflow and the efficiency gains land on the firm’s side of the ledger, not the client’s.

This model works for productizable offerings where discrete workflows map to flat fees. Large firms handling complex, bespoke matters (cross-border M&A, multi-district litigation, regulatory investigations) face a harder path to capturing AI efficiency gains. The work doesn’t break into neat packages. But for the vast middle of the legal market, value-based billing turns AI from a threat into an asset. Every process you automate improves your margins instead of creating awkward conversations about billing.

Risk, Trust, and the Supervision Gap

Lawyers have also learned from experience to be risk-averse to new technologies.

From 2013 to 2018, Google faced litigation over scanning Gmail content for advertising profiles, including emails sent to law firms’ clients who used consumer Gmail accounts. The 2018 settlement in Matera v. Google Inc. included injunctive relief requiring Google to stop scanning email contents for advertising purposes. For lawyers, these incidents confirmed suspicions about cloud services and confidential client communications. Many firms still refuse to use cloud technologies entirely, keeping all data on local servers. The risk aversion has a rational basis: a data breach at a high-profile firm could destroy the value of multiple client companies simultaneously.

This history helps explain why law remains one of the most Microsoft-entrenched industries. Firms migrated to Microsoft’s enterprise cloud partly because Microsoft offered clearer contractual protections around data handling, partly because Microsoft’s enterprise sales force understood compliance concerns, and partly because switching costs made staying with familiar tools the path of least resistance.

Microsoft’s Copilot was supposed to be the AI bridge that met lawyers where they worked. After a year of testing across multiple law firms, consultants at Clear Guidance Partners summarized Copilot as “minimally usable for legal work”. Most firms renewed pilot licenses but declined to expand beyond initial test groups. I was on the pilot for implementing MS Teams for a flagship state university, and the experience was nearly identical; they turned on Copilot and told us to kick its tires. I immediately tried to red-team it with prompts for data that I knew I shouldn’t have access to and found a ton of stuff. When I reported this back to the pilot group, the IT department screwed down the permissions to the point where it was lobotomized and useless, because it relied on individual users setting their SharePoint permissions carefully across the entire MS tenant, something that is very hard to police across a giant organization with decades of data and employee churn. They elected not to move forward with purchasing. Microsoft had only 8 million active Copilot users across 440 million M365 subscribers as of August 2025, a 1.8% conversion rate.

Bigger firms purchase enterprise or white-listed versions of frontier models from OpenAI, Anthropic, or Google, hosted on familiar AWS or Azure infrastructure with security policies their IT departments can stomach. But in doing so, they strip away the application layer that makes these models useful to non-technical attorneys: the agentic tooling, workflow integrations, and features like Claude’s Cowork or OpenAI’s Agent Mode that change the 2024 chatbot interface into a true digital assistant. Firms pay premium prices for hobbled versions of tools that are far more capable in their consumer-facing form.

When the world’s most entrenched enterprise software vendor struggles to sell AI tools to its existing customer base, lawyers’ skepticism looks less like technophobia and more like rational caution.

That caution hardened after June 2023, when lawyers Peter LoDuca and Steven Schwartz of Levidow Levidow & Oberman faced sanctions in the Southern District of New York for submitting a brief citing cases fabricated by ChatGPT. Mata v. Avianca became the legal profession’s cautionary tale about AI hallucination. The $5,000 fine was modest. The reputational damage was severe. Databases now collect these incidents (this one by Damien Charlotin now has over 1,200+ incidents), and these are paraded in front of attorneys at their ethics CLE’s as a warning. The broader impact on AI adoption was chilling.

But the nuance got lost. Legal media covered Mata v. Avianca as a story about AI gone wrong, not as a story about lawyers who didn’t do their jobs. The profession absorbed the lesson that AI could get you sanctioned, creating an incentive structure where risk-averse lawyers (which is most lawyers) decided the safer path was avoidance. The stakes in law are different from those in other professions. If you’re in marketing and AI makes a mistake on a social media post, you ask AI to generate ten more and pick the best. If a lawyer makes a mistake, people can lose their house, end up in jail, or lose custody of their children.

The hallucination problem has become a self-fulfilling prophecy. Any reported inaccuracy, regardless of context, reinforces the perception that AI can’t be trusted for legal work. Most of the attorneys getting sanctioned for fabricated citations are using free, consumer-grade models from Frontier Labs for legal research, which is the wrong tool for the job. The Big Three legal research platforms (Westlaw, Lexis, and Clio’s vLex) have addressed the citation hallucination problem by hyperlinking AI-generated references to primary sources, allowing lawyers to verify that a case exists before relying on it. But solving the “fake case” problem doesn’t solve the deeper concern: that AI might misrepresent a holding, overstate the strength of a legal theory, or miss a distinguishing fact that changes everything. For a profession where nuance is the product, “probably right” isn’t good enough. Frontier labs operate on a release-fast-fix-later cycle, shipping models at breakneck speed and resolving issues as they surface. Lawyers waiting for a hallucination-free model are going to be waiting a very long time by Silicon Valley standards.

And now a new problem is emerging. At Legalweek 2026, Danielle Benecke (Founder and Head of Applied AI at Baker McKenzie) presented a framework she called the “supervision gap.” As AI moves from an assistive tool to a primary work producer, the traditional model of attorney supervision (required by every state bar in the US) breaks. Full human review of AI output becomes economically irrational at scale. Our ethical rules say to treat AI as a junior associate, review everything it produces, and take responsibility for the output. But when an agentic system handles an entire workflow (all of your eDiscovery, drafting, or research – think OpenClaw for law), the attorney faces an impossible choice. You can rely on the vendor, or review all the work yourself, which is redundant and expensive. Or you can trust the system without a full review, which puts your license on the line in new and riskier ways beyond simply trusting an eDiscovery platform to tag documents.

Benecke called this the “outcome economy”: a world where liability pressure begins to shift from lawyers to AI vendors. If the lawyer is not meaningfully in the loop, the system must decide who owns the risk. That question gets more urgent if malpractice insurers or clients start demanding AI use. If your malpractice insurer won’t cover you unless you use AI-assisted review, or your client won’t hire you without it, the risk calculus flips. The profession needs new risk-sharing models between attorneys, vendors, and clients. We don’t have them yet.

The Access Question

Everything discussed so far explains why AI diffusion in law has been slow. None of it explains why slow AI diffusion in law is everyone’s problem.

According to the Legal Services Corporation’s 2022 report, 86% of civil legal problems faced by low-income Americans received inadequate or no legal help. Stanford research found that 75% of civil cases have at least one party without legal representation, roughly 15 million cases per year, where at least one side has no lawyer.

The average retainer for private legal representation runs $2,000 to $10,000. For a household living paycheck to paycheck, that might as well be $2 million. My favorite thing to do during our unit on access to justice in class is to ask students how many of them have $10,000 in their checking account for a retainer, to which none of them raise their hand, and I get to remind them that they (future lawyers) do not currently have access to an attorney. These cases aren’t marginal disputes about trivial matters. They’re evictions, custody battles, debt collection, disability claims, and domestic violence protective orders. The outcomes shape people’s lives.

AI could address this gap. The technology to automate intake, draft basic pleadings, and guide self-represented litigants through procedural requirements exists today. Hello Divorce demonstrates that productized legal services can work. Deployed at scale, AI-assisted legal services could reach millions of people navigating courts alone.

But AI isn’t being deployed at scale in legal aid. When I volunteered at the Legal Aid Society in San Diego, 5-7 interns were sharing two computers. The same barriers that slow BigLaw adoption (data moats, trust concerns, hallucination risks) apply with even more force to organizations serving vulnerable populations. A hallucinated case citation in a corporate lawsuit is embarrassing. A hallucinated case citation in a pro se eviction defense could cost someone their home. The legal profession’s AI investment flows toward the 14% of the population who can afford lawyers, not the 86% who can’t. Legal aid can’t afford Westlaw and Lexis enterprise licenses with all of the bells and whistles that an attorney has at Kirkland, and they’re not paying the extra $200-$500 per month, per attorney for advanced AI features.

Despite this underserved market, the legal profession has always been protectionist about who can practice law. LegalZoom has been sued in California for the Unauthorized Practice of Law for making forms that non-lawyers could use to represent themselves in court. Only in recent years have we seen a growing Access to Justice movement in places like Arizona and Utah, which have pioneered reforms allowing nonlawyer ownership and paraprofessional practice. California’s 2024 Justice Gap Study found that the situation has worsened since 2019. The gap between legal need and legal help isn’t closing.

• • •

Two open questions define the next phase of AI in law.

The first is the supervision question. As AI becomes capable of producing entire work products, the profession that has spent decades treating “I reviewed it myself” as the standard of care has no framework for what happens when that review becomes economically irrational. The ethical rules assume a human at the center. The technology is moving humans to the periphery. Someone has to reconcile those two facts, and the answer will reshape how legal services are delivered, priced, and insured.

The second is the access question. 86% of low-income Americans with civil legal problems don’t get meaningful help. AI-powered legal services could reach millions of them, and there are TONS of low-hanging fruit to help regular people. But the same structural barriers described in this article (data moats, trust deficits, governance paralysis, liability uncertainty) sit between the technology and the people who need it most. And the legal profession’s protectionist instincts, while sometimes well-intentioned, keep the drawbridge up.

The profession that runs on documents still can’t agree on who’s responsible when the documents write themselves. These are not questions of technology, but of liability, organization, and capital. Answering these questions can help ensure AI democratizes access to legal help for the people who need it the most.

Sean A. Harrington is Director of the AI & Legal Tech Studio at Arizona State University College of Law, where he teaches AI and the Practice of Law and researches the diffusion of artificial intelligence in legal services.

Research in materials science and chemistry presents a unique opportunity with specific bottlenecks that align with what LLMs are unusually well-suited to address. The friction points are addressable, and the vision of what is possible on the other side: better batteries, new classes of drugs, lighter structural materials, more efficient catalysts, is deeply important to the future of material abundance, economic prosperity, and good health that we hope for.

The Bottlenecks are Specific and Addressable

Work in materials science and chemistry spans extraordinary ranges in length scales (from a few atoms to manufactured components), times scales (from femtosecond reaction dynamics to years for material degradation), and methodologies (from quantum chemistry simulations to bench-scale synthesis and production scale up). The result is exceptional heterogeneity in data, tools, and workflows. A team optimizing a catalyst might need to combine campaigns of quantum simulations and manage data from dozens of instruments with differing file formats — all while synthesizing their own knowledge with figures, tables, and text from thousands of papers of varying quality.

Further, while materials science and chemistry have generated large datasets and repositories (e.g., Materials Project, OQMD, NOMAD, OMol25, the Materials Data Facility and more - see this list of 100s of other resources) there isn’t an equivalent to GenBank for genomics or Protein Databank for protein science because of the exceptional heterogeneity of materials and chemistry data. As such, there is no universal structured repository of synthesis procedures or process-structure-property relationships or molecular equivalent, and instead that information is often contained in personal expertise, millions of papers, figures, tables, hundreds of data resources, notebooks, and various repositories.

Researchers still spend significant time on mind-numbing tasks like data entry, extracting information from papers by hand, converting files between formats, and writing documentation and reports. They unfortunately must spend time on the small, grinding tasks that have nothing to do with the actual creative work of science but consume enormous chunks of mental capacity.

Dealing with such data and software heterogeneity and varying workflows are tasks that LLMs are particularly good at. LLMs have the potential to become connective tissue – a universal interface layer – because they are particularly adept at translating between heterogeneities in e.g., human intent/machine actions, multiple schemas, inputs/outputs across tools, and even narrative scientific context and structured data procedures. With such a broad new class of capabilities, there is a transformational opportunity for this connective tissue to be built between fragmented software, infrastructure, databases, papers, and more that currently don’t talk to each other. In such a vast application space of rapidly emerging LLM capabilities, it would take years to fully understand, specify, prioritize, and build such infrastructure if we waited for traditional research projects to be funded for this kind of work.

Instead, we decided to see what tools scientists could build for themselves through a hackathon.

The Hackathons: 2000 researchers across 4 continents

The first LLM Hackathon for Applications in Materials Science and Chemistry was held in 2023, and for this event, we didn’t know if anyone would show up or if the tools were mature enough to build anything real in 24 hours. It also wasn’t assured that materials scientists and chemists would have the requisite software development skills. But people showed up, and the latent capability to build was evident almost immediately. Many had never built an agentic system or worked with an LLM before.

Over three events, in 2023 and subsequently 2024, and 2025 (see figure), more than 2,000 participants, primarily graduate students, created over 170 publicly documented, open-source projects. The 2025 event alone had 16 in-person sites, plus an online hub for worldwide access. We provide in detail, a description of all of the projects and links here, but in short, they built complex software including: natural language interfaces that let non-experts control advanced instruments; agentic workflows that compress the time from idea to experiment; automated data management systems that make it possible to share data across labs; new ways to train researchers; new generative and predictive models; tools to extract structured information from a corpus of research papers; and much more.

These events produced moments of connection, learning, and building. This included teams spanning the globe and formed in interesting ways, e.g., researchers in India partnering with researchers in the United States and midnight “aha” moments on Slack that put a team on a prize-winning path. We purposefully designed this not as only a single day event, but as an ongoing community where these moments have a space to continue, enabling researchers to catalyze startups and worldwide collaborations, land fellowships, write papers together, and make discoveries of their own.

In these events, we primarily focused on building and learning. To be clear, these projects are prototypes, not yet production systems. But many of them are really good prototypes, and the speed at which they were built tells us something important about where the technology is right now and the latent capability of researchers to adapt these tools.

We next describe the outcomes and what we learned in more detail. We group the projects into four buckets for convenience, though many projects span multiple categories including:

1. Prediction and design: using LLMs alongside traditional ML to e.g., predict material properties or generate candidate structures or other distributions, especially when data are scarce.

2. Interfaces and automation: conversational control of instruments, simulations, databases, and development of closed-loop systems that compress the hypothesis-to-validation cycle.

3. Data management and knowledge extraction: compiling structured, computable data out of the messy world of PDFs, lab notebooks, patents, lightly documented datasets, and inconsistent databases.

4. Education and scientific communication: tutoring platforms, virtual lab simulators, tools for creating explanatory content to connect with audiences ranging from aspiring students to domain experts.

Prediction and design: LLMs as complements, not replacements

In materials science and chemistry, researchers are often working in low data regimes and with heterogeneous data as discussed previously. Importantly, the projects showed the strength of LLMs to work under such conditions.

The LLM4ConProp team explored whether LLMs can directly predict a material property of concrete. They assembled two curated datasets totaling nearly 4,000 records of concrete compressive strength and material composition and processing, and ran a head-to-head comparison of GPT-4.1 in zero-shot and few-shot settings versus tree-based models (random forests, XGBoost, LightGBM). The results showed that with enough in-context examples, the LLM approached ML baselines for low data regimes.

Unsurprisingly, LLMs also work well with multimodal and heterogeneous inputs. Researchers may wish to provide natural language description of desired properties, leverage research papers, or describe material or molecular characteristics and generate structures, synthesis pathways, or other distributions. The MIDAS team built a prototype showing structure-based drug design by conditioning a diffusion model on both protein pocket geometry and natural language instructions. With this system, users can say things like “generate a molecule with a hydroxyl group targeting this binding site” and get candidates that reflect both the structural constraints and chemical intuition. To train this, the team generated approximately one million molecule-text description pairs using GPT-3.5, covering functional groups, molecular properties, and pharmacophore descriptions. The whole system was wrapped in a chat interface with additional tools for docking analysis, similarity search, retrosynthesis, and iterative refinement through conversational feedback. Another project, SKY, tackled the inverse problem of how to make a material given specific property targets. The system takes a natural language description of a target structure, runs recursive similarity searches across Materials Project data, and uses LLM reasoning to propose grounded synthesis routes.

These projects illustrate how LLMs are effective in prediction and design especially when paired with traditional ML models as complements and to provide interfaces that handle the messy, heterogeneous inputs, outputs from these models, and to handle cross-service data connections. As foundation models improve at reasoning over mixed modalities like text, structures, spectra, images, this complementary role is likely to expand, particularly in the low-data regimes that define much of materials science and chemistry.

Interfaces and automation: making billion-dollar infrastructure more accessible

The US spends billions annually on national user facilities - synchrotrons, neutron sources, nanocenters, high-end electron microscopes, exascale supercomputers - that are too large and specialized for any single university to host. These are some of the most powerful scientific instruments on the planet, including publicly funded supercomputing facilities that provide compute access to researchers nationwide. Far fewer researchers use them than should, largely because accessing them requires months of specialized training. Making these facilities accessible without months of training could have outsized impact.

Natural language interfaces offer a clear path forward. Some of the most common applications run on scientific supercomputing are density functional theory (DFT) calculations and molecular dynamics (MD), but running simulation campaigns at high-performance computing facilities requires understanding of the specific system you are running on, optimizing for scale across many nodes, carefully monitoring the jobs, and more. LARA-HPC built an assistant that translates scientific goals, e.g., “compute the atomization energy of HCN”, into complete HPC DFT workflows. Similarly MINT LLM created natural language interfaces for MD analysis across codes.

Automated labs are gaining increased prominence within national user facilities and industry. The ACME team connected molecular design, quantum chemistry, and simulated robotic experimentation in a single feedback loop for discovering molecules for critical materials extraction. A reasoning model retrieves domain knowledge, e.g., ligand design rules, coordination geometry constraints from curated publications, then constructs candidate molecules. Those candidates pass through automated computational screening where the system builds 3D structures, runs semi-empirical quantum chemistry (GFN-xTB), and ranks by metal-ion binding affinity. In this hackathon, the top candidates didn’t enter the physical world, but rather a virtual stand-in that is being developed to synthesize and characterize coatings of these future molecules. Simulation results feed back to the reasoning model, which updates design rules and proposes improved candidates.

These prototypes are directly aligned with major national investments. The US Genesis Mission, launched by executive order in late 2025, aims to connect DOE’s 17 national laboratories, user facilities and scientific datasets into an integrated discovery engine. Canada’s Acceleration Consortium, various efforts in Europe and Asia, and industry efforts from Lila, Cusp, DeepMind, Periodic Labs, Radical AI, and others are pursuing parallel ambitions spanning automated labs and scientific intelligence. The hackathon projects show that the research community is ready to build on this infrastructure, and that many of these ambitious goals may be closer than expected. These working prototypes also point to a future where scientific user facilities have new interfaces for improved accessibility and efficiency.

Data management and knowledge extraction: liberating a century of science from PDFs and notebooks

Federal mandates increasingly require that scientific data be Findable, Accessible, Interoperable, and Reusable. In practice, that compliance falls disproportionately on graduate students and postdocs, the benefit to any single researcher rarely feels worth the effort, and the result is repositories that are technically FAIR but still unusable. LLMs offer a way to fundamentally shift this burden.

LLMs offer a way to fundamentally shift this burden and capture structured data more efficiently. For example, ExpAlign created a pipeline that parses hundreds of research PDFs, extracts key properties and hidden experimental details, flags inconsistencies, and combines results into clean, ML-ready data tables. MuMMIE tackled multilingual patent extraction across five languages, creating benchmarks for cross-lingual scientific information extraction. SuperconLLM built a four-agent pipeline that screens arXiv papers, performs named entity recognition and relation extraction, and generates structured database entries for superconductor properties. PolyNexus created a domain-specific knowledge base for electroactive polymers where you can ask natural language queries (”What is the conductivity of PEDOT:PSS?”) and get citation-backed results.

This extraction and information access at scale, when coupled with private sector efforts (e.g., Edison Scientific/Future House, Chemical Abstracts Service, ChemDataExtractor, Citrine Informatics), promises to collate and liberate hundreds of years of humanity’s collective scientific knowledge from PDF files into something structured, computable, and accessible data sources.

Education and scientific communication: the parts nobody talks about

Research impact depends not just on discovery but on how effectively knowledge transfers to the next generation of researchers and diffuses to adjacent fields, and to the public. These are areas where researchers typically receive little training and have few tools, and where LLMs may have an outsized impact.

For example, ChemTutor AI built a domain-specific tutoring platform that generates personalized problems with interactive 3D molecular visualizations and adaptive difficulty. It provides pedagogical scaffolding including e.g., step-by-step reasoning, worked examples, and Socratic questioning. MatSci LapLab, built a tool to simulate characterization techniques like TGA, SEM, and tensile testing, enabling students to train virtually.

On the communication side, AtomicShorts built a three-agent pipeline that helps scientists create short explanatory videos at a fraction of commercial costs and at different audience knowledge levels. By keeping scientists in the loop, the system ensures accuracy while reducing the enormous gap between researchers and the public.

These tools address an asymmetry, i.e., researchers are trained and incentivized to do science but less so to teach or communicate efficiently. If LLMs can lower the cost of creating high-quality educational and explanatory content, the pool of people who engage with, and eventually contribute to, scientific research could increase significantly.

The hackathon model: why this worked and how to replicate it

The hackathon series was, itself, an experimental test of what we call the “Social-First Hackathon Model”. This model is one that many research communities could replicate, and in fact, has already been used for two hackathons in microscopy and one in Bayesian optimization.

We designed the model around five principles:

1. Public and social by default. Teams submit entries via social media, LinkedIn, X, YouTube, with accompanying code repositories and a Google Forms entry. The social posts create immediate public visibility, provide verifiable credit for CVs, and eliminate the need for custom infrastructure.

2. Minimal central coordination. The entire event runs on free tools like Slack, Zoom, GitHub, Luma, and Google Workspace. Participants form teams, decide topics, and plan projects autonomously. Two to three central organizers coordinated an event for over a thousand participants.

3. Hybrid from the start. A global virtual cohort participates alongside physical hubs at institutions that volunteer to host. This decentralizes operations while maintaining a unified program and allows the best talent from anywhere to participate.

4. Time-boxed intensity. The 24h constraint forces rapid prioritization, prevents scope creep, allows researchers to participate without compromising their ongoing research, and creates urgency.

5. Leaning into academic incentives. Participants are primarily graduate students and postdocs. Co-authorship on papers, presentation opportunities, awards for CVs matter. After each event, we assembled teams to write articles including as many active participants as possible.

Three lessons emerged across events. First, team formation before the event is critical. The most common failure mode is participants showing up without teammates and being unable to crystallize both a team and a concept. The 2025 event addressed this by emphasizing early skill-matching for over a month before the event, using the shared Slack, virtual meetings, and a custom Miro board.

Second, domain breadth within teams matters. The events have no limits on team formation. Teams can be as big or small as needed, and combine in-person and virtual. Many of the strongest teams brought together broad teams including e.g., computational chemists, experimentalists, and ML engineers.

Third, the community persists. Over 1,400 researchers remained active in shared Slack channels after the events concluded, continuing collaborations, posting job opportunities, and building on each other’s work. The hackathon is a nucleation point that brings together high-agency researchers in a long-lived community that continues to percolate afterward.

Importantly, this model handles aspects that traditional funding mechanisms handle poorly: e.g., 1) rapid landscape mapping (170 projects surveying possibilities), 2) real-time workforce development via participants learning by building and forming lasting collaborations, 3) public-by-default outputs where every project is immediately available for others to learn from and build on, 4) and talent identification where two days of effort can produce a significant addition to the researcher’s CV.

The projects and community described here focused on materials science and chemistry, but illustrate a broader pattern about how AI capabilities spread through research communities. Properly structured hackathons may function as adoption accelerators, with the ability to compress the learning curve, give researchers permission to experiment (and fail), and produce reusable examples and visible proof that the tools work.

These events showed not just the diffusion, but the building of capability, showing how a capable workforce, able to effectively leverage AI can be built rapidly. The hackathons were views into the diffusion and capacity building happening in real time, suggesting that research institutions, national labs, and professional societies could drive meaningful AI adoption and diffusion of capabilities and knowledge by running similar events in their own domains. The model is lightweight, replicable, and documented. However, there are still areas for improvement, e.g., sustaining momentum after the event ends, securing compute credits and inference access for deeper development, and bridging the gap between a promising hackathon prototype and a tool that researchers use daily. These are tractable problems, but they require intentional investment in the community infrastructure, scientific middleware, and shared compute to further build researcher capacity and enable diffusion of tools into production use.

One important point sticks with me and fills me with hope. Many of the teams that participated, including those that built reasoning agents connected to simulation tools and external databases, autonomous NMR analysis pipelines, and natural language interfaces for software and hardware had never worked with LLMs or constructed an agentic system before. Yet, with a clear goal, and aligned incentives, they accomplished it together in just over a day. That fact alone should change our priors on the near-term trajectory of scientific research, even if these participants are perhaps exceptional due to selection biases.

Scientific tools often diffuse slowly through research communities, often taking years between invention and widespread use. The 170 projects built at the events are prototypes that help provide a concrete starting point to speed the diffusion. Yet, there is real work between a hackathon demo and a tool used daily by thousands of researchers that requires additional work and investment. From these events, researchers have catalyzed new collaborations, presented results at top international conferences, secured new funding, created teaching modules, and created software used by many other research groups. All of this, paired with the scale and breadth of working demonstrations show how much is now possible promising to draw more people in to build the next set.

This diffusion also requires building researcher expertise and familiarity. The hackathon events showed that with even a modest incentive structure and a focused community-building effort, thousands of researchers with minimal initial LLM training can be trained via “learn-by-shipping”. In this model each team produces (1) a concept, (2) a software repository, (3) a demo artifact (e.g., video), and (4) a short explainer. Those outputs and the community itself act as a diffusion substrate creating code others can build upon, workflows others can copy, and explorations of spaces relevant to many different research groups. As such, the training does not stay local to participants. This creates a virtuous cycle where increased visibility recruits the next cohort, reuse of the examples turns prototypes into shared resources or production software, and where normalization lowers perceived risk and increases adoption inside labs.

As these tools propagate, and e.g., the ability to run and interpret common techniques like XRD analysis, DFT, MD, molecular dynamics or leverage unique user facilities becomes conversational and ubiquitous, the barriers between disciplines become more porous with AI tools reducing transaction costs, e.g., meaning less time negotiating data formats, learning one-off software stacks, or translating jargon. Further, as data management overhead disappears, researchers can reclaim time for actual science and other research teams get access to better open data to use in their analyses. Towards training the next generation using AI, students can train on simulated characterization techniques expanding the researcher capacity dramatically. Breakthrough science has always thrived at the intersections between fields, institutions, capabilities, and expertise. In many ways, these projects are helping to create new intersections. AI is set to change dramatically as user expertise in AI tools diffuses, models improve, creating more of those intersections.

None of these tools or approaches are guaranteed to become widely adopted. The new technology creates an opening, but realizing the benefits depends on whether the research community and the institutions and agencies that support the research ecosystem lean into the opportunity. The experiences documented here suggest that relatively modest investments in inference infrastructure (or cloud credits), scientific middleware, documentation standards, and community coordination could yield outsized returns.

If you are an interested scientist, you can join the 1400+ researchers continuing to build in this space on the hackathon slack. Reach out to me if you’d like to sponsor this work, visit our community, and stay tuned for our next events!

Although the primary commercial potential of these new Chinese models is not yet in major large screen production, as Hollywood most fears, it is being widely adopted by a new media channel that China dominates: microdramas.

The MIT Tech Review last year reported on the rise of micro-drama exports from China last year. Microdramas are smartphone-native, have a fast turnaround, and are directly competitive for viewers’ attention and time. They are not “shorter TV shows,” they are a new vertical of their own.

These micro-dramas exploded in part due to timing: the phone became the living room, distribution became adept at buying attention, and production pipelines learned to optimize for retention at scale.

And because people don’t choose what to watch with their thumbs at 11:30 pm — they often turn on whatever is frictionless and emotionally legible in a phone-sized frame.

According to DataEye, a Chinese media analytics firm, over 300 million downloads of Chinese micro-drama apps globally pushed YouTube and Netflix off the top of app store rankings in the first half of 2025. Some of the most notable micro-drama Chinese platform names include ReelShort, DramaBox, GoodShort, and ShortMax. Most of the parent companies behind these platforms produce content in Chinese and English as they distribute across different social media platforms aimed at different audience demographics. But ReelsShort, ShortMax, and so focuses on the English-native market.

Similar to American short-form native platform Quibi, the leading Chinese micro-drama company Reelshort also targets mobile-first, short-form, vertical-video entertainment, but the main difference in their strategies lies in their approach. While Quibi failed with high-budget, A-list content, ReelShort thrived on ultra-low-budget, addictive, soapy, “guilty pleasure” dramas.

For the sake of research, I watched a few episodes and realized its addictive nature. The themes read like a speedrun of narrative dopamine: car crash amnesia, comeback revenge arcs, double lives, fake marriages, cheating spouses. Honestly, the more dramatic and emotionally triggering, the more engaging. They’re like the less polished, more spiteful cousins of K-drama tropes, and they remind me of the telenovelas that used to play in the background at my friend’s house. Each episode was a wild ride; it gave you dopamine, cortisol, oxytocin, adrenaline, all in 120 seconds of your life. Each episode feels short, but follow along for a whole storyline, and you’ve burned through an hour without noticing. Reelshort’s slogan “every second is drama” is quite literal in that time is attention and attention is the business.

And just when you’re super amped up and hooked, it displays a black screen to unlock the next episode, pay a few coins, or watch an ad. Micro-dramas are not just shorter TV series. The content business now looks more like the economics of gaming. As you’re already hooked, you are presented with two options to continue your show: pay upfront, or watch an ad to unlock the next episode.

The Business of Microdramas

Microdramas are both an advertising business and a subscription business. And it isn’t a small business. DataEye estimates that China’s microdramas + “manju” (AI/animated short dramas) generated over RMB 100 billion in annual output value—well above the earlier market expectation of ~RMB 60 billion. At that scale, the microdrama industry is 2× China’s domestic film box office annual revenue of RMB 51.832 billion.

Chart: The Hollywood Reporter

According to a Chinese business publication, The Paper’s report, Chinese micro-dramas have reached 200+ countries and regions. In 2025, Guangda Securities estimates that the overseas market generated ~US$2.38B in revenue on 1.21B downloads, with revenue growth accelerating to +263% year-on-year and download growth to +135% year-on-year. As of now, North America remains the main profit pool, contributing over 60% of overseas revenue for these companies, while Southeast Asia leads in share at ~35%, with the Middle East, Japan, Korea, and Europe also cited as expanding rapidly.

Amongst the top platforms by revenue are ReelShort, DramaBox, and GoodShort. The three platforms together account for over 53% of the market.

Embracing AI

Today, companies like ReelShort still mostly rely on Western actors when making content for Western audiences. But the next step, which many have told investors they plan to take, is to use AI to create characters to perform these manuscripts, thereby significantly reducing production and export costs and reshaping the economics of microdramas.

The most recent generation of Chinese video models certainly seems capable of making this leap. The internet has already been thrown into a frenzy as snippets of short period dramas leaked ahead of Seedance’s latest update announcements, as well as new action films created completely on Kling. Tutorials on how to turn written scripts into videos using these tools have also flooded X. Hugging Face’s head of the APAC ecosystem, Tiezhen Wang, is also sharing videos that resemble scenes from a Hollywood blockbuster created by Seedance 2.0. There was even a dedicated AI Film Festival hosted in India by InVideo as part of the broader India AI Impact Summit, which included PM Modi, President Macron, Jensen Huang, Sam Altman, and Sundar Pichai in attendance. Of course, AI sometimes still produces horrendous glitches. However, some argue that whether the protagonist was wearing the same t-shirt in two consecutive scenes doesn’t affect the experience of a one-minute short meant to be watched on a 1080p phone screen.

And this is no longer just hypothetical. In mid-March, ByteDance announced the release of 小云雀 ‘little sparrow, an AI-native drama platform. It accepts scripts up to 100,000 characters and is described as the first industry agent powered by Seedance 2.0. Industry leaders have reflected on how Seedance 2.0 has truly solved three previously perceived pain points for AI short-form drama: character/scene consistency, realism in complex physical motion, and continuity/rationality of camera movement.

In an interview with Chinese tech media Jiemian, Chinese filmmaker Gong Changhu told Jiemian that before Seedance 2.0 his 10-person team could produce a 120-minute short drama in 20 days; after Seedance 2.0, that time dropped to half, showcasing how the launch of such AI models could materially raise production efficiency and could trigger explosive industry growth.

A recent Kuiashou Kling press release stated that, since its launch in June 2024, their video generation models have served 60 million creators worldwide and produced over 600 million pieces of content. The press release frames Kling 3.0 as a shift from a “generation tool” to an “intelligent creative partner” that can grasp artistic intent and turn ideas into reality, effectively positioning it as an era in which anyone can turn ideas into films. AI is being sold less as a tool and more as something that takes intent and executes, a true partner in production.

What will AI-empowered micro dramas look like?

Microdrama exporters continue to rapidly capture traffic, scale, and distribution, both in China and abroad. The “second half” of the battle is the efficiency of AI adoption and the ecosystem flywheel effect. For companies like ReelShort, with a huge existing MAU base, the thinking is that once they can cut production costs, they can push out cheaply made content through existing distribution and monetize.

The likely new business model for AI-empowered micro-dramas will be built on a legacy supply chain. It’s about content ownership, character and language localization, lowering expectations, and using AI to cut costs and increase efficiency across the whole stack.

Let’s break down the layers of the ecosystem:

Layer 1: scale of writers. The massive popularity of web novels in China since the early 2000s has created a generation of writers trained to maximize retention at all costs: cliffhangers, rapid reversals, and emotional triggers engineered to drive binge behavior. Web novels, like micro dramas, typically lure readers in with a few free chapters before paywalling subsequent releases. Also, like micro dramas, the market is brutally competitive, with hundreds of thousands of active novels competing for readers at any given time. And the universally loved but loathed themes are all the same: Fifty Shades of Grey style CEO-turned-boyfriend, rags to riches, and Cinderella tales.

The Chinese digital literature ecosystem is a large commercial market; the latest official reports estimated that the number of online literature readers was ~575 million as of 2024 and that industry revenues were tens of billions of RMB. This digital literature space encompasses hundreds of thousands of writers. And that huge group of online novel writers has been pivoting to write micro-drama screenplays in swarms since 2023 as the vertical offers more monetary reward.

Layer 2: IP owners and product creators. The second layer of the supply chain comprises IP owners and product creators. Companies like ReelShort aren’t just microdrama producers. They are part of a larger ecosystem. The studio behind Reelshort is Crazy Maple Studio, which is 49% owned by COL Group (中文在线), a Chinese content company with an enormous inventory of stories ready to be adapted. And they work with writers in both forms: employees for the companies, as well as licensing/buying out webnovel IPs.

For context, COL was founded in 2000 and is listed on the Shenzhen Stock Exchange (stock code 300364). It is said to have more than 5.6 million pieces of digital content and more than 4.5 million digital-native authors. Its products include online books, movies, audiobooks, music, and more. The sheer scale of its IP ownership is an advantage of its own.

Layer 3: tools. The third layer comprises the tools. Chinese short-video platforms have positioned themselves to capture this market, leveraging their in-house data, such as ByteDance’s Seedance model and Kuaishou’s Kling. While both platforms have massive monthly user bases, the vast amount of video and image data they have for training has given them an advantage. Thus, their open-source/open-weight text-to-image models have become competitive with Western industry leaders, such as Veo, Sora, and Midjourney. Not only that, but tools such as Kling and Seedance are largely free for short-form video creation.

Kuaishou’s Kling is targeting independent filmmakers and indie studios to empower them with AI. They’ve also been partnering up with film festivals across Asia, from Hong Kong to Tokyo. ByteDance’s video generator is currently widely available across its AI apps, such as Douban, Jimeng, and Jianying/CapCut, and all at no cost.

Layer 4: distribution. And the final layer is the distribution. Although companies like Reelshort are still largely loss-making, their goal is to blitzscale distribution and then move to profitability. On the first goal, they are doing tremendously well: Reelshort is one of the top apps in the App Store entertainment category, directly competing with top-tier streaming apps for downloads. The logic of Reelshorts is to own distribution, reduce production costs through AI video generation apps, and then reach profitability.

Quibi demonstrated that cramming prestige TV Hollywood economics into short-form doesn’t work — but ReelShort’s bet is the inverse: own distribution, drive production costs toward zero with AI, and let one breakout hit cover the rest.

An article in the Hong Kong newspaper Ta Kung Pao offered a case study that reads like a DTC marketer’s dream: a micro-drama titled The Divorced Billionaire Heiress, reportedly costing under $200k and generating $35 million at the North American box office — 170× returns. Even if you treat that as an extreme example rather than the median, it explains why capital keeps wandering into this “unsexy” corner of entertainment. With AI video generation lowering per-unit production costs, Chinese microdrama platforms have the potential to be able to leap abroad into Western markets.

AI Micro-dramas, a New Soft Power

Television reshaped the movie industry: stories had to move from big screens to the little boxes in our homes, and with this new medium came the rise of the TV series, which challenged the two-hour movie formula. The internet and the rise of streaming changed how we consume those series. Now, micro-dramas are a new business model, natively built for the mobile era. While still niche, they represent a growing source of competition with mainstream streaming services for eyeballs.

While Western studios are still worried about AI taking actors’ or writers’ jobs, Chinese studios are building new micro-drama platforms that are uniquely positioned to leverage AI tools to more effectively capture user attention in the global attention economy. AI-powered microdramas represent a uniquely Chinese vertical, from AI models to distribution platforms, and potentially present a new form of soft power export, following the footsteps of Labubus.

We are excited to share that Issue 1 of Diffuse AI is now coming to an inbox near you. For our inaugural issue, we’re excited to have the following contributors:

• – Diffusion into the Feed: Mini-Dramas, AI-Native Entertainment, and the Attention Economy

• – Field notes from LLM hackathons for chemistry and materials

• – Why AI hasn’t reached the manufacturing floor

• – The Structural Barriers to AI Lawyers

• and – On the National Imperative for AI Diffusion, a joint interview with

• Anonymous – interviews a macro analyst on the impacts of AI diffusion for India’s economic development

We are aiming to ship published pieces every other week, alongside monthly round-ups of blogs, pieces, and tweets on AI diffusion. Send us your best takes!

And as always — we are always searching for in-depth, on-the-ground pieces about how AI is diffusing and shaping different parts of the economy. Pitch us through this form. We pay $1k USD for essays and reportage.

Everyone wants to know what the next few years of AI will look like. We have a different question: what is AI capable of right now?

Whether you’re worried about unemployment, deskilling, or human extinction – and even if you think AI is a flash in the pan – we all have a shared interest in understanding how it is already changing our world. But static benchmarks — the multiple choice questions or verifiable math and coding challenges typically used to evaluate models — are getting saturated as quickly as we can build them. The real test is how well AIs function in complex, context-heavy, dynamic environments — in other words, the real world.

What do we know so far? We’re starting to see serious interest in how AI is playing out across the economy, but the picture is messy. We’ve all seen the buzzy paper from Stanford arguing that AI is currently causing jobs to plummet among junior workers in “highly exposed industries” and the equally buzzy paper from Yale claiming there’s no discernable effect. Labs are investing in more sophisticated measures of how their models perform on the kinds of tasks that matter for real jobs, but there’s a big difference between a self-contained evaluation and the messy reality of day-to-day employment. And for every report on the wonders of vibe-coding, there’s a thread on hacker news insisting that AI productivity gains are a mirage. For every $30 million AI for science startup, there’s a grizzled computational biologist who’s ready to deflate the hype. What’s really going on? What are these systems actually capable of? What are the bottlenecks to realizing AI-assisted productivity gains? What does this all look like on the ground?

Diffuse AI is going to help answer all these questions, in as much nitty gritty qualitative detail as possible. We want:

• In-depth case studies of how AI is playing out in your industry — the more specific, the better.

• Interviews with experts about what current models are and aren’t useful for in their work.

• Stories of similar instances of historical tech diffusion.

• Thoughtful discussions of the strengths, weaknesses, and methodologies of economic impact benchmarks like GDPeval or the Anthropic Economic Index.

We don’t want:

• Predicting the future.

• Theorizing from first principles.

• Coming with an axe to grind.

Here are some examples of pieces we’d love to have published:

• Why AI isn’t replacing radiologists

• An interview with the AI Initiatives team at The New York Times

• 25 interviews with software engineers on how they use AI

• How Deepseek is actually diffusing in China: Shallow, Narrow, and Slow

• A case study on the first Compute Arms Race: weather forecasting supercomputer

Pitch us through this form, we pay $1k USD for essays and reportage. Help us figure out what’s really going on.