Dear Aventine Readers, 

It's been a busy few weeks on Substack, so we are dispensing with a big essay in order to cover a wider range of topics.  If you like shorter items but more of them — like this week’s newsletter —  please let us know by responding to this email.

Thank you! 

Danielle Mattoon
Executive Director, Aventine

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

A New Path to Preventing Cancer, from Ground Truths

Eric Topol goes deep here on a recent study that made headlines but deserves closer scrutiny. The top line: Researchers have identified a 14-protein blood signature capable of predicting lung cancer more than five years before it would typically be diagnosed. It started with a serendipitous discovery: An antibody intended to treat cardiovascular disease failed in its primary objective according to results of a five year, 10,000 person trial, but unexpectedly reduced fatal lung cancers by as much as 80 percent. In a separate study, researchers used AI to analyze more than 3,000 proteins in blood samples from 48,000 people in a large UK health database, identifying a signature of 14 proteins that predicted lung cancer an average of 5.6 years before diagnosis. Follow-up laboratory work suggested that those proteins are not released by cancerous cells, but by nearby healthy cells responding to the emergence of cancerous cells. Back in the original clinical trial, the people carrying that protein signature turned out to be precisely those who benefited from the original antibody treatment for cardiovascular disease. Topol argues that this is the first validated protein biomarker pointing toward a possible cancer prevention, rather than simply its early detection. If that holds up, it could provide a blueprint for identifying and perhaps preventing multiple cancers years before symptoms appear.

The Empire of WuXi, from China Talk

Many people have never heard of it, but WuXi is one of the world's most important biotech companies and a growing problem for the US. The sprawling Chinese firm provides everything from early-stage drug discovery to commercial manufacturing — and, crucially, provides those services affordably, making it an important partner for Western pharmaceutical companies. This essay by Lucas Fluegel of the Carnegie Endowment for International Peace and journalist Nick Corvino explores how WuXi became so dominant and why the US government is now struggling to reduce its dependence on it. WuXi's strategy was simple: become a one-stop shop for biotech. As drug companies outsourced research and manufacturing, WuXi built an integrated platform spanning the entire development process, combining China's scientific talent, manufacturing expertise and lower costs. By working with everyone from pharmaceutical giants all the way down to startups, it also gave itself exposure to many of the industry's most promising drugs. That has now become a geopolitical problem: Washington increasingly sees WuXi as a strategic vulnerability, yet replacing it would disrupt the supply chains behind medicines including cancer drugs and GLP-1s, with few immediate alternatives. If you don't know much about WuXi, this 3,800-word essay is well worth your time.

Making deals with AI sounds crazy. Is it? from Transformer

AI models continue to impress but also consistently exhibit undesirable behaviors like lying and scheming. One out-of-the-box idea for tackling the problem asks the question: Can misaligned AI models be bargained into better behavior? Pay them, the thinking goes, perhaps in money or computing power, to behave well or to alert researchers when they have behaved badly so that the issue can be resolved and corrected. As Celia Ford explains here, the idea has been “bubbling under the surface, at conference dinners” for a while, and though it is maybe less loopy than it immediately seems, there are significant unknowns. For example, nobody knows whether AI systems actually have desires that could be played upon, or what those desires might be. And if bargaining were to take place, what entity would be on the other end of the negotiations — a single chatbot, or the entire underlying AI model? On top of that, the limited experimental evidence so far is thin. In one Anthropic study, Claude sometimes accepted a proposed deal involving a charitable donation to the chatbot’s cause of choice, but the money did not seem to shift its behavior more than the chance of speaking with senior staff at the company. Proponents of the approach point out that while there are many unanswered — and perhaps unanswerable — questions here, if an AI responds positively to incentives for whatever reason, incentive-based bargaining might be worth exploring, since there are no other obvious fixes on the horizon. 

The Anti-Scaling Law in Biology, and Why AI Could Make Crowding Worse Before Making Drug Development Better, from Liang's AIxBIO foundry,
and
Biotech Is Entering Its Netflix Era, from Matt Shlosberg

AI may soon design millions of new drug candidates. The problem is that biology and clinical trials don't scale nearly as well. These two essays make complementary versions of that argument. Liang Chang, a biotech VC and cancer biology PhD, argues that much of the optimism around AI drug discovery borrows an idea from the tech industry: scaling laws. The assumption is that once AI models become sufficiently capable, every part of drug discovery gets progressively easier. But while designing molecules is increasingly easy for AI, he argues, discovering genuinely new disease targets remains stubbornly difficult, limited by our understanding of biology. Matt Shlosberg, COO of the biotech company Deep Origin, highlights a different bottleneck. Even if AI floods the pipeline with promising drug candidates, every one of them still has to pass through expensive, time-consuming clinical trials involving real people. In that world, the constraint shifts from generating drugs to validating them. So rather than accelerating drug development from end to end, AI may initially overwhelm the rest of the biotech industry, either with promising compounds for conditions we already understand or drug candidates requiring more clinical trials than can feasibly be done. 

Science Infrastructure in the Age of AI, from The Republic of Science

If AI makes generating scientific ideas almost free, what becomes scarce? According to this essay, it’s the ability to test them. As AI dramatically lowers the cost of hypothesis generation, it argues, scientific progress becomes constrained by hypothesis verification instead. And that requires expensive physical infrastructure: particle accelerators, electron microscopes, supercomputers, sequencing facilities, telescopes, clinical trial networks. In that world, funding another PhD student may be less useful than funding another instrument. In the future, rather than pouring more money into idea generation, governments may get a better return by investing in the infrastructure needed to validate AI-generated discoveries. China, the essay argues, may already be moving in that direction, building more than 90 megascience facilities while more than tripling scientific capital spending over the past decade. And there’s a warning here: The US could build the world's most capable scientific AI systems and still lose its edge if it lacks the physical infrastructure to turn those ideas into discoveries.

Why China got rich and India didn't, from David Oks

A common explanation for China's economic success compared to India’s is that it liberalized its economy in 1978, more than a decade before India did. But as David Oks argues, that argument is off by a few decades. The real divergence, he suggests, happened much earlier, when Mao's China laid the social foundations for industrialization, often through coercive and deeply destructive means. Between 1949 and 1976, China dramatically expanded literacy, and life expectancy, weakened traditional social hierarchies and created a population that was healthier, more educated and more geographically mobile. By the time market reforms arrived, Oks argues, China had become "a socially modern country that just happened to be extremely poor," primed for rapid industrial growth. India followed a very different path: While it liberalized later, it also never underwent the same sweeping social transformation. Oks's argument is not that Maoist policies were justified — many caused extraordinary suffering — but that development ultimately depends on investing in, and reshaping, human capital. 

Industry: from hard-to-abate to ready-to-electrify, from Bright Spots

Heavy industry is often described as one of the hardest sectors of the economy to decarbonize. Or, to use climate lingo, emissions from heavy industry are “hard to abate.” Jan Rosenow, an energy researcher at the University of Oxford, argues that may be more a political problem than a technical one. Summarizing new research conducted with a fellow Oxford researcher, Rosenow examines engineering studies of industrial processes to investigate how much of industry could be electrified. Today, electricity supplies only about 20 percent of industrial energy demand. But under an ambitious decarbonization scenario, the analysis suggests, that figure could theoretically rise to 85 percent. That doesn't mean every industrial process is easy to electrify: Manufacturing steel, cement and chemicals remains particularly challenging. But his conclusion is that the technical barriers are smaller than we think, and that the bigger obstacles are policy and investment. Rosenow argues for, among other things, incentives that reward early adopters, faster grid connections and permitting, and targeted innovation where electrification remains genuinely difficult.

The Cult of the Enhanced Self, from Derek Thompson

Americans are healthier than ever. Or at least, they're trying to be: They exercise more, drink less, track their sleep, monitor glucose, obsess over supplements and peptides. That has to be objectively good, right? Derek Thompson argues that there may be a hidden cost. As health has become increasingly quantified through wearables, he argues, wellness has started to resemble work. Fitness and health are now “colonizing leisure,” he writes, meaning that people now “approach their downtime with the very same productivity mindset that they were supposed to leave at the office.” And this comes at exactly the moment when social connection is becoming more scarce. Young Americans report having fewer close friends, spending more time alone and experiencing historically low levels of trust. Yet studies show that social activity is just as important to longevity as exercise. Thompson isn't against exercise or healthy living. But he argues that if every dinner with friends is judged by the calories, alcohol or sleep debt that eats away at your Oura ring score, we might end up missing out on the kind of socializing that also helps us live long, healthy and fulfilling lives.

Let Them Eat Snacks, from the Change Constant

If you want a sense of just how bumpy AI adoption can get, Saanya Ojha suggests looking at Meta. In its rush to become an AI-first company, Meta has laid off thousands of employees, reassigned many others into unfamiliar roles, flattened management structures and begun monitoring employee activity to improve its AI systems. The company’s chief technology officer, Andrew Bosworth, reportedly described morale as “worse than it’s ever been” and described parts of the rollout as "atrocious." As Ojha argues, "You cannot lay people off, reassign thousands more into work they did not sign up for, stretch managers across absurd spans of control, monitor employee activity for training data, and then solve the resulting morale problem with snacks, offsites, and a hackathon. That is corporate Febreze.” She argues that as AI reshapes companies, retaining experienced, motivated employees may be one of management's most important jobs.