Tech, Policy, and our Lives

Alexander Titus
  • 3.0 (2)
  • SCIENCE
  • UPDATED WEEKLY

Tech, Policy, and our Lives, brought to you by The Connected Ideas Project is a podcast about the co-evolution of emerging tech and public policy, with a particular love for AI and biotech, but certainly not limited to just those two. The podcast is created by Alexander Titus, Founder of In Vivo Group and The Connected Ideas Project, who has spent his career weaving between industry, academia, and public service. Our hosts are two AI-generated moderators (and occasionally human-generated humans), and we're leveraging the very technology we're exploring to explore it. This podcast is about the people, the tech, and ultimately, the public policy that shapes all of our lives. www.connectedideasproject.com

  1. APR 21

    Ep 64 - From Blacklists to Blueprints

    I’ve been referencing the BIOSECURE Act in these pages for months — in the biomanufacturing thesis, in the generic drug analysis, in passing asides about procurement signals and supply chain fragility. I’ve treated it as established context. Background architecture. The thing that’s already happened and that everyone in this space already knows about. I haven’t written a dedicated piece on it. That was a mistake, and I want to correct it now — not because the law is new, but because I think the public conversation about what it does has settled on exactly the wrong feature. The debate was about names. Which Chinese biotech companies would be designated. Whether BGI would make the list. Whether WuXi AppTec’s lobbying campaign would succeed. Whether the five-company approach was too narrow or too broad. That was the wrong debate. The version of the BIOSECURE Act that passed in the FY2026 NDAA doesn’t name five companies. It builds a machine. And the machine is more important than any list of names could ever be. The podcast audio was AI-generated using Google’s NotebookLM. What the Law Actually Built Let me describe the mechanism, because the legal architecture matters more than the headlines suggested. Section 851 of the FY2026 NDAA bars federal agencies from procuring biotechnology equipment or services from any “biotechnology company of concern.” It also bars agencies from contracting with entities that use covered biotech equipment or services in performing federal work — a downstream prohibition that extends the law’s reach beyond direct government suppliers into their supply chains. Loan and grant funds are covered too. The scope is broad and deliberate. But the designation mechanism is where the real design work happened. The law establishes two pathways for identifying a biotechnology company of concern. The first is automatic: any company on the Department of Defense’s Section 1260H list of Chinese military companies that is involved in biotech equipment or services is designated by operation of law. No additional review. No notice. No comment period. You’re on the 1260H list and you touch biotech — you’re covered. The second pathway is criteria-based. The Office of Management and Budget (OMB) leads an interagency process to identify companies that are subject to the direction or control of a foreign adversary, involved in biotech equipment or services, and assessed to pose national security risks — affiliations with foreign adversary militaries, provision of multiomic data to a foreign adversary, collection of human multiomic data without informed consent. This pathway includes procedural protections: 90 days to respond, periodic review, a process for requesting removal. Two tracks. One fast and automatic, one deliberate and procedural. Different temporal profiles for different risk profiles. That’s not a blacklist. That’s governance architecture. The Cybersecurity Lesson Nobody Applied Here is where I want to draw a connection that I think reframes what the BIOSECURE Act actually represents — and what it tells us about how we’re learning to govern frontier technology supply chains. For three decades, cybersecurity evolved through a specific failure mode. The early approach was signature-based detection: identify known malware, build a signature, distribute it to endpoints, block the match. It worked — until it didn’t. The attack surface expanded faster than signatures could be written. New variants appeared daily, then hourly. The lag between a novel threat and its corresponding signature became the vulnerability itself. By the time the signature existed, the damage was done. The industry’s response — the one that actually worked — was behavioral detection. Stop looking for known bad actors. Start looking for patterns of malicious behavior. Build systems that can identify threats they’ve never seen before, based on what the threat does rather than what it is. The shift was from static lists to adaptive systems. From recognition to pattern-matching. From naming the enemy to understanding the behavior that makes something an enemy. The BIOSECURE Act’s legislative evolution mirrors this transition almost exactly. The original bills named five companies. That’s signature-based governance. Identify the known threat actors, put them on a list, block them. It would have worked for those five companies. And it would have been obsolete within a year, as corporate restructuring, subsidiaries, joint ventures, and successor entities routed around the designations. You cannot blacklist your way to supply chain security any more than you can signature-match your way to network security. The threat surface evolves faster than the list. The version that passed builds behavioral detection into the governance architecture. The 1260H pathway captures entities based on their assessed relationship to the Chinese military — a behavioral criterion, not a corporate identity. The OMB criteria-based pathway captures entities based on what they do: whether they’re subject to foreign adversary control, whether they handle multiomic data in specified ways, whether their affiliations pose national security risks. The criteria travel. When a new entity emerges that exhibits the designated behavior, the system can capture it without new legislation. This is the design principle that matters: the law doesn’t just address the current threat. It builds the institutional capacity to address threats that don’t exist yet. And that distinction — between a law that solves today’s problem and a law that builds the machinery for tomorrow’s — is the distinction between a sandbag and a levee. The Temporal Gap But here’s where the design gets complicated, and where I think builders and policymakers need to pay close attention. The BIOSECURE Act’s prohibitions don’t take effect upon enactment. They take effect after the Federal Acquisition Regulation is revised — 60 days after the FAR update for 1260H-designated entities, 90 days after for criteria-based designations. OMB has one year to compile the initial list. The FAR revision process has its own timeline. A five-year rule of construction protects legacy agreements, including previously negotiated options. Add it up. The law passed in December 2025. The OMB list arrives no earlier than December 2026. FAR revisions follow. Effective dates trigger months after that. Legacy agreements survive for five years. The full force of the prohibition may not bind across the federal procurement landscape until 2028 or beyond. I wrote last week about temporal architecture — about the gap between when a system is designed and when it actually operates. The BIOSECURE Act is a case study. The governance intent is sound. The institutional machinery is well-designed. But the implementation timeline introduces a temporal gap during which the very dependencies the law aims to eliminate continue compounding. Wright’s Law doesn’t pause for rulemaking. Every month that foreign producers continue descending the biomanufacturing learning curve while domestic alternatives are not yet incentivized by the procurement shift is a month the cost gap widens. The 1260H pathway is faster — no procedural protections, no comment period, automatic designation. But it only captures entities already identified as Chinese military companies. The broader criteria-based pathway, which covers the more complex supply chain risks, is the slower one. This is the governance latency problem applied to procurement policy. The detection happened — Congress identified the vulnerability. The interpretation happened — the law’s criteria are well-specified. But execution latency — the time between legislation and operational effect — is measured in years. And in those years, the problem the law was designed to solve continues operating on its own timescale. The Levee’s Boundary There’s a second structural tension that I think deserves more attention than it’s getting. The BIOSECURE Act covers federal procurement. Executive agencies. Government contracts, grants, and loans. This is the lever the government controls directly, and it’s the right place to start. But recall the numbers from the BENS report I wrote about in the generic drug piece. Ninety-one percent of American prescriptions are generics. The federal government is a significant pharmaceutical purchaser, but it is not the whole market. The cascading dependency — China to India to American pharmacy counters — operates primarily through commercial supply chains that the BIOSECURE Act does not reach. The law addresses the 27% of military drug purchases that the Department of Defense study found depend on PRC suppliers. That’s critical. But it doesn’t restructure the commercial supply chain that delivers the other prescriptions — the ones that civilian hospitals, retail pharmacies, and patients depend on. The 679 APIs for which China is the sole KSM supplier don’t become less concentrated because federal agencies stop buying from designated entities. This isn’t a criticism of the law. It’s a diagnosis of its boundary conditions. The BIOSECURE Act is the first structural levee in a flood zone that extends well beyond the federal procurement riverbank. And understanding what it covers — and what it doesn’t — is essential for anyone trying to build the next section. The Medicaid Drug Rebate Program safe harbor is a telling detail. The law had to include a specific provision ensuring that drug manufacturers wouldn’t be penalized in the Medicaid system when the national security prohibitions prevent them from executing a VA master agreement. The fact that this carve-out was necessary reveals how deeply entangled the pharmaceutical procurement system is — pull one thread and you risk unraveling programs that millions of patients depend on. The legislators knew

    20 min

  2. APR 14

    Ep 63 - The Tempo Thesis

    I sat in a conference session recently and watched something happen that I’ve seen before but never quite named. Speaker after speaker — technologists, policy people, operators — kept circling the same idea without landing on it. One talked about detection speed for biological threats. Another about the lag between an AI capability and the regulation that addresses it. A third about why manufacturing learning curves are races, not exercises. The language was different each time. The domain was different. The variable was the same. Time Not as metaphor. Not as urgency rhetoric — the familiar “we need to move faster” that appears in every keynote and persuades no one. Time as something more fundamental. As the binding constraint that determines whether every other capability — technical, institutional, industrial — actually functions or just exists on paper. It struck me that for all the frameworks we’ve been building in this space — responsible innovation, governance architecture, reindustrialization strategy — we’ve been designing for capability, authority, and proportionality. We have not been designing for time. The podcast audio was AI-generated using Google’s NotebookLM. Speed Is Not the Variable There’s a distinction worth drawing carefully, because I think conflating two ideas has made this problem invisible. Speed is a metric. You can measure it, optimize it, benchmark it. Organizations talk about speed constantly. Move fast. Accelerate. Reduce cycle time. Speed is the thing you improve within a system that already works. Time is the medium in which all your systems must compose. It is not how fast you go — it is whether the systems that must coordinate with each other are operating on compatible timescales. A biosecurity detection system that identifies a threat in twelve hours is useless if the interpretation infrastructure takes twelve weeks and the policy execution mechanism takes twelve months. Each component might be excellent on its own terms. The failure is temporal — they don’t compose. Engineers have a name for this. Temporal coupling: when two systems that must coordinate operate on fundamentally different timescales, the system breaks. Not because any individual component failed, but because time itself became the fault line. I want to trace this mechanism across several domains, because I think it explains more about why our current systems are failing than any capability deficit does. Governance as Temporal Architecture I wrote about governance latency in these pages earlier this year — the gap between when a system behaves in a new way and when governance responds. I described three components: detection latency, interpretation latency, execution latency. I still believe in that framework. But I’ve started to think I was being too polite about what it actually describes. Governance latency is not a bug in governance. It is a temporal architecture — one that was designed, intentionally or not, for a world that moved at a different pace. Congressional hearing calendars. Notice-and-comment rulemaking periods. Interagency coordination cycles. These are not merely slow. They operate on a fundamentally different timescale than the technologies they govern. The gap between those timescales is not inconvenient. It is, itself, a space where outcomes are determined before the formal process even begins. The nation or institution that understands this — that treats temporal alignment as a design variable rather than an operational annoyance — gains an advantage that no amount of capability can offset. Because capability without temporal coordination is potential energy that never converts to kinetic. It sits in reserve, impressive and inert, while the clock runs. The Circle Is a Clock Consider biomanufacturing — a domain I’ve been writing about in this series. The circles-and-spirals thesis is, at its core, a temporal argument. The circle traps organizations in a time loop: no production experience means no yield data means no capital means no facilities means no production experience. The loop is self-reinforcing because each node operates on a timescale that prevents the next node from activating. Capital allocation cycles are quarterly. Facility construction takes years. Workforce development takes a generation. Yield improvement requires thousands of production hours that nobody can access because the facilities don’t exist. The spiral breaks the circle not by eliminating time, but by synchronizing it. Government demand signals compress the capital decision. Pre-built infrastructure compresses the facility timeline. Science investment steepens the yield curve so fewer production hours are needed to reach viability. The spiral is not faster in any simple sense — it is temporally coherent. Every node operates on a timescale compatible with the others. Wright’s Law, the principle that costs decline predictably with cumulative production, is a temporal claim wearing an economic costume. It says: the first mover in production will be the lowest-cost producer, and the gap will compound with time. China is further down the biomanufacturing learning curve than the United States. Every year that gap persists is not a static disadvantage. It is a temporal one — the curve steepens for whoever is on it and flattens for whoever is not. The Doubling Time of Consequence Biosecurity is perhaps the most visceral expression of this thesis. A biological threat does not wait for interpretation. It replicates on its own timescale — exponential, indifferent to institutional calendars. The difference between containment and catastrophe is not capability. We have the sequencing technology, the surveillance infrastructure, the countermeasure platforms. The difference is temporal coordination. Can you detect, interpret, decide, and act within the doubling time of the threat? I think about this in my work at Vigilance. The entire architecture of biological threat preparedness is, when you strip away the organizational charts and capability matrices, an exercise in temporal engineering. You are building systems whose purpose is to compress the gap between event and response to something smaller than the gap between event and consequence. That’s the design requirement. Everything else is decoration. From Dimension to Domain Here is where I want to push further than the conference session went, further than most strategy frameworks go. We tend to treat time as a dimension — the passive background against which things happen. Decisions take time. Manufacturing takes time. Governance takes time. Time is the water everything swims in. But the more accurate framing — the one that explains why temporally misaligned systems keep failing in predictable ways — is that time is a domain. A space in which advantage can be built, contested, and lost. A domain that requires its own strategy, its own architecture, its own design principles. If you accept that reframing, certain things follow. Temporal advantage is designable. You can build organizations, governance structures, and industrial systems that are optimized for temporal coherence — where the decision cycle, the implementation timeline, and the environment’s rate of change are deliberately aligned. Temporal disadvantage is structural, not accidental. When a governance system operates on a decadal timescale while the technologies it governs evolve on a monthly one, that is not a speed problem to be solved with urgency. It is an architectural mismatch that requires redesign. Temporal literacy becomes a core competency. The ability to read a system and identify where temporal misalignment is the binding constraint — rather than capability, authority, or resources — becomes as important as technical expertise or policy knowledge. What Temporal Design Looks Like This is where the argument becomes operational, and where I think builders, policymakers, and capital allocators need to pay close attention. If time is a domain, then every strategy has a temporal architecture — whether or not the strategist designed one. The question is not whether your organization operates within time. The question is whether you’ve deliberately engineered how your organization relates to time. For builders in frontier technology: the competitive advantage is not always the best technology. It is often the technology that reaches operational deployment first and begins descending the learning curve while competitors are still optimizing in the lab. This is Wright’s Law generalized. The first mover in production compounds an advantage that the better-but-later entrant may never overcome. Time on the curve is the asset. Everything else is a bet that time will be forgiving. It usually isn’t. For policymakers: governance latency is not a staffing problem or a willpower problem. It is a temporal design problem. The question is not “how do we make government faster” — it is “how do we build governance architectures whose operating timescale matches the domain they govern?” In some cases, that means pre-authorization frameworks that act before the crisis arrives. In others, it means modular governance that can be updated without rewriting the entire regulatory structure. In all cases, it means taking temporal architecture as seriously as institutional authority. For capital allocators: patience is a temporal strategy, not a virtue. The patient capital that biomanufacturing requires is not charity — it is an investment in temporal alignment, giving the learning curve enough time to generate the yields that make the economics work. The impatient capital that demands returns on quarterly timescales is not merely unhelpful. It is temporally incompatible with the problem it claims to be solving. I keep returning to that conference session. The speakers kept naming symptoms — speed, latency, urgency, readiness — w

    22 min

  3. MAR 17

    Ep 62 - The Generic Drug Trap

    I’ve been thinking about circles. Not the conceptual kind — though we’ll get there. The physical kind. The ones visible on a map if you trace the journey of a single generic antibiotic from raw chemical to American pharmacy shelf. Start in a chemical plant in Zhejiang province. Ship key starting materials to a synthesis facility in Gujarat. Convert them to an active pharmaceutical ingredient. Ship the API to a formulation plant — maybe still in India, maybe in a bonded zone elsewhere. Tablet, coat, blister-pack, box. Ship the finished dose to a U.S. distribution center. Dispense. That circle touches three countries, two oceans, and zero domestic manufacturing steps for the active ingredient. For hundreds of medicines Americans take every day, this is not the exception. It is the architecture. A new assessment from Business Executives for National Security — BENS, the nonpartisan organization that has partnered with senior national security officials since 1982 — maps this architecture with a specificity that should unsettle anyone who has been following the biomanufacturing reindustrialization conversation. Dave Gryska’s team at BENS just released Generic Drug Manufacturing and Biotechnology Innovation — and it does something I’ve been wanting someone outside government to do for years: it treats the generic pharmaceutical supply chain not as a trade policy problem or a healthcare access problem, but as a national security vulnerability with the same structural characteristics as the semiconductor dependency that launched CHIPS. I’ve written about the biomanufacturing reindustrialization thesis — the circles and spirals, the learning curves, the factory economics, the builder’s playbook. That piece was about the future: the companies and capital structures that could bring biological manufacturing home. This piece is about the present. Because the BENS report reveals that the system we need to reindustrialize isn’t just underbuilt. It is actively structured against its own repair. The podcast audio was AI-generated using Google’s NotebookLM. The Numbers Behind the Dependency Let me put the scale of the problem in terms that are hard to argue with. Ninety-one percent of drugs prescribed by American physicians are generics. Ninety percent of those generic prescriptions are supplied by India and China. The FDA estimated that approximately 80% of active pharmaceutical ingredients are manufactured overseas. In 2021, 87% of U.S. generic drug manufacturing facilities were located abroad. The United States directly imports roughly 16% of its APIs from China — but India, which supplies the bulk of finished generic doses to American patients, itself imports 80% of its APIs from China. This is the cascading dependency that the BENS report traces in detail, and it is the feature I want to hold up to the light. Because it’s not a supply chain. It’s a cascade. A disruption in Chinese chemical manufacturing doesn’t just affect Chinese exports. It ripples through India’s pharmaceutical sector and arrives at American pharmacy counters as a shortage — of antibiotics, of cardiovascular drugs, of the basic medications that chronic disease patients depend on to stay alive. The data on shortages is the trailing indicator. In 2024, the United States recorded its highest number of drug shortages to date — 323 medications affected. Antibiotics have proven especially vulnerable, experiencing shortages at a rate 42% higher than other generics. These aren’t obscure compounds. They’re the medicines that keep a 68-year-old’s blood pressure managed, that treat a child’s ear infection, that a hospital ICU reaches for when a patient goes septic. And here’s the number that stopped me: the API Innovation Center examined 40 essential drug molecules — the “Vital 40” — and found that India supplies about 63% of these APIs, Europe 22%, China 8%, and the United States just 5%. Five percent. For the most critical generic medicines in the American formulary, domestic production is a rounding error. But the 8% figure for China is misleading, and the BENS report is careful about why. An analysis of key starting materials (KSMs) — the precursor reagents that are synthesized into APIs — reveals that China is the sole supplier of KSMs for 679 different APIs. India is the sole KSM supplier for 402. The United States and European Union combined are sole suppliers for 44. Six hundred and seventy-nine to forty-four. That ratio is the supply chain expressed as a strategic position. And it tells you that the dependency isn’t just about who fills the last step — the tablet press, the blister pack. It’s about who controls the chemistry underneath. China’s dominance isn’t at the visible end of the supply chain. It’s at the foundation. The KSMs are the geology, and everything above them — APIs, formulations, finished doses — is built on ground that someone else owns. Why the Circle Won’t Break Itself In the biomanufacturing reindustrialization thesis, I introduced the idea that biomanufacturing fails in circles and scales in spirals. The circle: no production experience means no competitive yields, which means no capital investment, which means no facilities, which means no production experience. The system is closed. Nothing moves. The generic drug supply chain is this circle in its most advanced failure state. It didn’t get here by accident. It got here by economic logic operating without strategic constraint — and the BENS report traces the history precisely enough that the mechanism becomes clear. The 2000 elimination of tariffs on formulated pharmaceuticals — following the 1995 WTO Pharmaceutical Tariff Elimination Agreement — incentivized importing finished drug products. Domestic manufacturing became progressively less competitive. Facilities closed. The Viatris plant in West Virginia — once employing 1,400 workers producing critical generic medications across antibiotics, cardiovascular, and autoimmune therapeutic areas — is one example among many. America’s last factory producing penicillin closed in 2004. Not because the science was lost. Because the economics were untenable. And here is where the circle logic becomes vicious. Once domestic production stops, the workforce disperses. Once the workforce disperses, restarting production requires not just capital but a labor pool that no longer exists. Once facilities close, the institutional knowledge embedded in those facilities — the process optimizations, the supplier relationships, the regulatory compliance infrastructure — evaporates. You cannot restart what you’ve forgotten how to do. Or rather, you can — but at vastly greater cost and time than maintaining it would have required. This is why I called it a system designed to prevent its own repair. The cost advantage of foreign production compounds over time. Every year of offshoring means another year of learning curve accumulated abroad and lost domestically. China’s 20,000+ chemical companies, accounting for 40% of global chemical industrial output, aren’t just competitors. They’re the installed base. And by some estimates, it costs 50% less to produce APIs in India versus the United States or Europe, with labor costs in India and China estimated at one-tenth the cost for a Western company. The circle doesn’t just trap individual companies. It traps the entire sector. The margins on generic drugs are already razor-thin — that’s the whole point of generics. Domestic production faces stringent regulatory compliance, higher labor costs, higher energy costs, and competition from established foreign producers who have been descending the learning curve for decades. No individual company can break this circle alone. The economics won’t allow it. This is where the BENS report’s framing matters: this is a national security problem requiring a national security response. Not because the market failed in some abstract sense, but because the market optimized for exactly the wrong thing — unit cost — while ignoring systemic fragility. And fragility, in a system that 91% of American prescriptions depend on, is an existential risk wearing the mask of efficiency. Stockpiles Are Sandbags, Not Levees The BENS report takes SAPIR — the Strategic Active Pharmaceutical Ingredients Reserve — seriously, and I want to amplify the point because it cuts against a comfortable assumption in Washington. In August 2025, an executive order directed the Administration for Strategic Preparedness and Response (ASPR) to identify and stockpile APIs for approximately 26 essential medicines. SAPIR is being refilled and expanded. This is a good and necessary step. Stockpiling APIs, which have longer shelf lives than finished drug products, reduces risk at multiple stress points in the supply chain. But the BENS report says what I’ve been arguing in different language: stockpiling alone is insufficient. It is a passive defense mechanism. It buys time. It does not solve the structural deficit. Without a functioning industrial base to replenish the stockpile, the reserve is a finite resource that will run dry in a prolonged crisis. Think about what that means operationally. A geopolitical disruption — a Taiwan Strait crisis, an escalation in the South China Sea, a deliberate Chinese export restriction on pharmaceutical precursors — doesn’t resolve in weeks. The scenario that matters isn’t a temporary supply interruption. It’s a sustained one. And a stockpile sized for months faces a crisis measured in years. The analogy I keep returning to: SAPIR is sandbags. Essential during a flood. Useless for preventing the next one. What you need is a levee — domestic production capacity that can sustain the flow regardless of what happens upstream in the geopolitical watershed. And the geopolitical risk is not theoretical. The BENS report documents that Chin

    22 min

  4. MAR 10

    Ep 61 - The Biomanufacturing Reindustrialization Thesis

    There is a question I keep returning to — one that sits underneath the policy debates, the appropriations fights, the executive orders, and the increasingly urgent memos circulating through the national security establishment: Can America actually make things with biology? Not design them. Not discover them. Not publish papers about them. Make them. At scale. Reliably. On domestic soil. With a workforce that exists and supply chains that don’t route through adversarial nations. The answer, right now, is: barely. And it looks like this. Last year I stood in a pilot biomanufacturing facility — one of the few we have — and watched a team troubleshoot a fermentation run that had gone sideways at 5,000 liters. The organism was producing at bench scale. It had produced at 500 liters. At 5,000 liters, oxygen transfer became the constraint — at least, that’s what the team suspected in real time. The metabolic profile shifted. Yield dropped by a third. The lead process engineer — one of a few hundred people in the country with this specific operational expertise — was working through it with a combination of sensor data, experience, and what I can only describe as biological intuition. There was no model to consult. No binder. No runbook. No second shift that had seen it before. She was debugging a living system at industrial scale, mostly alone, in a building that smelled like warm yeast and sounded like a submarine engine room. That scene is the bioeconomy. The rest is narration. I’ve spent two+ years as a commissioner on the National Security Commission on Emerging Biotechnology, producing the most comprehensive governmental blueprint for biomanufacturing reindustrialization the United States has published. The NSCEB report is thorough. It is specific. I believe in the architecture. But government reports — even good ones — describe what should happen. They do not describe what it feels like to build in the gap between should and is. And they do not tell the builders, the operators, and the capital allocators where the openings actually are. That’s what I want to do here. This is the first in a series. The podcast audio was AI-generated using Google’s NotebookLM. The Gap Nobody in Washington Understands Viscerally Enough Two pieces of writing crystallized the biomanufacturing problem for me recently, and neither was about biology. The first is Aaron Slodov’s “American Shenzhen” framework — a detailed blueprint for rebuilding U.S. hardware manufacturing capacity: government as anchor tenant, 75/25 commercial-to-defense revenue models, special economic zones, streamlined permitting, venture-backed manufacturing startups. The insight is structural: Shenzhen didn’t emerge from a single policy. It emerged from a system — procurement signals, physical infrastructure, workforce pipelines, and regulatory architecture reinforcing each other simultaneously. The second is Oliver Hsu’s recent primer on factory economics for a16z, which articulates what venture capital is only now internalizing: the IP is the process. In these companies — and biomanufacturing startups are definitionally among them — the moat is not intellectual property in a patent filing. It is the production process itself. The yield curves. The learning rates. The operational knowledge embedded in people and equipment and process. I read both and thought: this is the framework I wish I could have injected directly into the NSCEB’s deliberations. Because here is what the data looks like from inside the commission. The United States’ share of global API production has collapsed from roughly 23% to approximately 3% over three decades. More than 90% of generic pharmaceuticals consumed in the U.S. depend on imported ingredients. Industry surveys indicate that roughly 80% of biopharma organizations are actively engaged with Chinese CDMOs. China holds approximately 58% of global synthetic biology patent filings, 28% of biological manufacturing patents, and 30% of novel antibiotics patents. That’s not an edge. That’s installed capacity. That’s a country that has been running production volume while America has been running conferences about it. The NSCEB report lays this out. The number one message is the urgency. The window for American biomanufacturing reindustrialization is open, but it is not open indefinitely. And the dynamics that will close it are not political. They are economic. The Learning Curve Is a Race — And We’re Losing It This is where Hsu’s factory economics framework becomes essential, and where the venture and builder community needs to pay close attention. Wright’s Law tells us that costs decline predictably with each doubling of cumulative production. The learning curve is the race that defines factory companies. The competitor with more cumulative production has lower costs. Lower costs win more contracts. More contracts mean more production. More production steepens the curve. The advantage compounds. China is further down the biomanufacturing learning curve than the United States in multiple product categories. Every year America does not build domestic production capacity is a year China accumulates more volume, drives costs lower, and makes the gap harder to close. This is not a static competition. It is a dynamic one, and the dynamics favor whoever starts manufacturing first and fastest. Now, biology adds a complication that makes this race harder than any hardware equivalent. Biological systems are stochastic. A fermentation run that works at 10 liters may behave differently at 10,000 liters — not because of engineering error, but because living organisms respond to conditions in ways we don’t fully understand. That engineer I watched troubleshooting the 5,000-liter run? She was navigating exactly this problem. The yield curve in biomanufacturing is less predictable than in semiconductors, aerospace, or any other production domain. And yield is the single highest-leverage variable in factory economics. A 20-point yield advantage can create a cost differential that determines who survives. This is the core tension, and I want to name it precisely because I think it’s the single most important concept in biomanufacturing strategy: Biomanufacturing fails in circles. It only scales in spirals. The circle: the learning curve demands production volume, but production volume requires facilities that won’t get built without capital, and capital requires the predictable yields that only come from production experience. No yields, no capital. No capital, no facilities. No facilities, no learning. No learning, no yields. The system is closed. Nothing moves. The spiral: break into the circle at any point — with government demand signals, with patient capital, with science that steepens the yield curve — and the circle becomes a spiral. Production generates learning. Learning improves yields. Yields unlock capital. Capital builds more facilities. Facilities train the workforce. The workforce improves operations. Operations improve yields. The system opens. Everything moves. In a circle, you die waiting for certainty. In a spiral, you manufacture your way into it. The circle is what kills startups. The spiral is what makes nations competitive. The NSCEB understood this. The report’s six pillars — political commitment, private sector mobilization, defense integration, innovation infrastructure, workforce, and allied coordination — are designed as a system specifically to break the circle and start the spiral. Push on every node simultaneously. That’s the architecture. But here’s what I want to tell the builders and investors directly: the government is going to be slow. The NSCEB report took two years. Implementation will take longer. If you wait for the full system to be in place before you move, you will be too late. The opportunity is in the gap between the signal and the infrastructure — and that gap is open right now. How a Builder Wins in the New Landscape Let me be concrete about this. Instead of listing the plays abstractly, I want to walk through what the next twelve to thirty-six months look like for someone building a domestic biomanufacturing company — and how the NSCEB architecture, as it comes online, changes the game at each stage. You start with a facility strategy This is the first decision and the one most biotech founders get wrong, because they’re trained to think about molecules first and production second. In the new landscape, production is the strategy. You need a facility — or access to one — that can run at pilot scale today and commercial scale within three years. The NSCEB recommends a network of precommercial biomanufacturing facilities operated through DOE and DOC, plus a $120 million biopharma manufacturing center under the Defense Bioindustrial Manufacturing Program. If you are a startup, the question is whether you build your own or position to be the anchor tenant in one of these government-catalyzed facilities. Either way, think about geography: proximity to a research university with relevant programs, an existing labor pool with manufacturing experience (not just biology PhDs — people who know how to run plants), and state-level incentive structures. The NSCEB’s regional hub model matters here. The companies that co-locate with the hub infrastructure will compound advantages that distant competitors cannot replicate. The procurement signal de-risks your first offtake The DBIMP at $762 million or more. Advanced market commitments and offtake agreements from DOD and HHS. The BIOSECURE Act — signed into law in late 2025 — already restricting federal contracts with foreign biotechnology companies of concern. That’s the stick. The DBIMP and AMCs are the carrot. If you are building domestic production capability for APIs, sustainable aviation biofuels, biomaterials, or engineered proteins, the federal procuremen

    25 min

  5. MAR 3

    Ep 60 - The Org Chart Dies Last

    This is a special edition of The Connected Ideas Project, because while it’s Episode 60 of the podcast, it’s the 100th edition of this newsletter since launch! Thank you for being part of this community. If you’re finding value, please share with your friends and colleagues! Every few years, a major technology company publishes a report that tells you more than it intends to. Microsoft’s 2025 Work Trend Index — “The Year the Frontier Firm Is Born” — is one of those reports. On the surface, it’s a well-produced argument for why every company needs to reorganize around AI agents. Survey 31,000 workers across 31 countries, add LinkedIn labor data and Microsoft 365 telemetry, wrap it in a compelling narrative about hybrid human-agent teams, and you’ve got a document that every Fortune 500 CEO will have on their desk by Friday. But read it a second time. Read it the way you’d read a national security assessment — not for the headlines, but for the assumptions underneath. And something more interesting emerges. Microsoft isn’t describing a productivity tool. They’re describing a new theory of the firm. And they’re describing it almost entirely in the language of efficiency, without ever seriously grappling with the governance architecture such a firm would require. That gap is where the real story lives. Every time we build autonomous capability faster than we build accountability, the system doesn’t fail immediately. It fails later. And it sends the bill. The podcast audio was AI-generated using Google’s NotebookLM. What the Report Actually Says Give Microsoft credit: the diagnosis is sharp. Business demands are outpacing human capacity. Eighty percent of the global workforce says they lack enough time or energy to do their work. The knowledge worker, as currently configured, is maxed out. Microsoft’s answer: intelligence on tap. AI agents that can reason, plan, and execute tasks autonomously — not chatbots, but digital colleagues joining teams with increasing independence. The report envisions three phases, from AI as assistant to AI as operator of entire business processes, and argues that companies embracing this trajectory are already pulling ahead. The numbers tell the story. Eighty-two percent of leaders expect to deploy agents to expand workforce capacity in the next eighteen months. Forty-six percent are already using them to automate entire workflows. A third are considering headcount reductions. And here’s the number that stopped me: 78 percent are considering hiring for AI-specific roles that didn’t exist a year ago. This is not incremental. Microsoft even coins a term — the “Work Chart” — to replace the org chart: a dynamic, outcome-driven model where teams form around goals, not functions, powered by agents that expand what each person can do. The Movie Production Model — and Its Missing Script One of the report’s most revealing analogies compares the Frontier Firm to movie production. Teams assemble for a project, agents fill specialized roles, the work gets done, and everyone disbands. It’s a compelling image. Lean, high-impact, fluid. I’ve been thinking about that analogy. Because it captures something real about where organizational design is heading. But it also reveals what the report doesn’t address. Movie productions work because of something the report never mentions: an extraordinarily mature governance infrastructure. There are unions, guilds, contracts, liability frameworks, insurance requirements, safety protocols, credential verification systems, and chain-of-command structures that have been refined over a century. The fluidity of production is enabled by the rigidity of the rules governing it. What is the equivalent for human-agent teams? When Microsoft describes a world where every employee becomes an “agent boss” — someone who builds, delegates to, and manages AI agents — they’re describing a massive delegation of judgment. And delegation of judgment, in any complex system, is a governance problem before it’s a productivity solution. I keep thinking about this because it mirrors a challenge I’ve watched play out in a different domain entirely. What Defense Planners Already Know I remember a briefing at the Pentagon — one of many, but this one stuck. A program manager was presenting an autonomy roadmap for a logistics system. Slides were clean. The capability curve was steep. Savings projections were compelling. And then someone from the policy shop asked a single question: “Who signs for the decision when the system gets it wrong?” The room went quiet. Not because the question was unexpected. Because everyone knew the answer wasn’t in the slides. That moment comes back to me reading Microsoft’s report. Because the concept they’re selling as a business revolution — human-machine teaming with autonomous systems — is something the Department of Defense has been grappling with for over a decade. Different vocabulary, same structural problem: How do you maintain meaningful human oversight when the systems you’re working with can operate faster, and increasingly more capably, than the humans directing them? The defense community learned several things the hard way. First, that the “human in the loop” is not a design feature — it’s a design requirement that must be engineered deliberately, or it erodes. Systems that are faster and more capable than their operators create irresistible pressure to defer. The human becomes a rubber stamp. In military contexts, this is called automation bias. In Microsoft’s Frontier Firm, it has no name yet. But the dynamic is identical. Second, that trust calibration matters as much as capability. The report’s own data hints at this: 52 percent of workers see AI as a command-based tool, while 46 percent see it as a thought partner. That split isn’t a preference — it’s a reflection of how well people understand what they’re delegating. Miscalibrated trust — too much or too little — is how autonomous systems fail in operational environments. The military has spent billions learning this lesson. The report proposes that every employee learn it on the job. Third, and most importantly: the governance architecture has to be designed before the capability scales, not after. The DoD doesn’t deploy autonomous systems and then figure out the rules of engagement. The rules come first. They’re imperfect, they evolve, but they exist before the system is operational. Microsoft’s report proposes deploying autonomous agents across entire business functions and then building the governance afterward. They call this journey “Phase 1 to Phase 3.” A defense planner would call it an operational risk. The Responsible Innovation Gap Here’s what fascinates me about the Frontier Firm concept. It’s a genuinely interesting framework for thinking about organizational transformation. The capacity gap is real. The potential for AI agents to expand what small teams can accomplish is real. I personally see this every day with the engineering teams I run. The shift from functional org charts to outcome-driven work charts is a prediction I think will prove directionally correct. But the report treats governance as an afterthought — a problem to be solved after the productivity gains are captured. And this is a pattern I’ve seen before. One of the themes we’ve explored in the Science of Responsible Innovation is that the time to design governance into a system is during the architecture phase — not during deployment, and certainly not after failure. Violet Teaming exists precisely because the traditional approach — build it, ship it, regulate it — doesn’t work when the systems in question are capable of autonomous action. The Frontier Firm, as Microsoft describes it, would have AI agents running supply chains, managing customer relationships, executing financial analysis, and operating business processes end-to-end. Each of these involves decisions with consequences for real people — employees, customers, communities. The report mentions a “human-agent ratio” as a new business metric. But a ratio tells you headcount, not accountability. Who is responsible when an agent makes a consequential error in a process it was running autonomously? The agent boss? The agent’s developer? The company that deployed it? The platform provider? These are not hypothetical concerns. They’re the same questions that biosecurity experts ask about autonomous laboratory systems, that defense ethicists ask about lethal autonomous weapons, and that financial regulators ask about algorithmic trading. The pattern is consistent: autonomous systems that operate faster than human oversight can track them create accountability vacuums. There’s a concept the defense community uses that Microsoft’s Frontier Firm badly needs: rules of engagement. Before any autonomous system operates, there are explicit boundaries — what it can do, what requires human authorization, who owns the consequence of each class of action. Call it an Accountability Ledger for the Frontier Firm: a document, maintained alongside the Work Chart, that maps every agentic process to a human owner who answers for its outputs. Not the person who prompted the agent. The person who is responsible when the agent’s decision costs someone their job, their loan, their medical claim. The Work Chart tells you who does what. The Accountability Ledger tells you who answers for what. You need both, or you have neither. If you can’t name the human who owns the downside, you don’t have automation. You have abdication. Microsoft mentions Daniel Susskind’s hypothesis that human work will persist because of three limits: efficiency, human preference, and moral judgment. That’s a reasonable framework. But notice the order. Efficiency is the domain AI masters first. Human preference erodes as people habitua

    21 min

  6. FEB 24

    Ep 59 - A New Study Taking Responsible Innovation From Benchmarks to Benchwork

    A few months ago, when we first started talking about the Science of Responsible Innovation at The Connected Ideas Project, I kept coming back to a simple question: How do we know? How do we know whether a technology is actually as powerful—or as dangerous—as we imagine? How do we know whether our fears are grounded in evidence or in extrapolation? How do we know whether policy is steering something real, or something hypothetical? It’s one thing to run a model through an in silico benchmark and watch it ace a virology exam. It’s another thing entirely to put a pipette in a novice’s hand and see what happens in a real lab. That’s why the recent paper, “Measuring Mid-2025 LLM-Assistance on Novice Performance in Biology,” feels so important. Not because it proves that AI is safe. Not because it proves that AI is dangerous. But because it does something rarer and more valuable: it measures. And in doing so, it gives us a template for what responsible-by-design evaluation can look like in the age of frontier AI and synthetic biology. The podcast audio was AI-generated using Google’s NotebookLM. The Gap Between the Benchmark and the Bench For the last several years, large language models have been climbing biological benchmarks at an astonishing rate. Protocol design. Sequence interpretation. Troubleshooting. Literature synthesis. In some cases, outperforming domain experts on structured tests. On paper, that looks like capability. And capability, when it intersects with viral reverse genetics or synthetic biology, looks like risk. But as I’ve discussed in recent work on Violet Teaming—particularly in “The Promise and Peril of Artificial Intelligence — ‘Violet Teaming’ Offers a Balanced Path Forward” —capability is not impact. And risk is not hypothetical power alone. It’s what happens when humans, institutions, and technical systems interact in the real world. The authors of this new study understood that. So instead of running another benchmark, they ran a randomized controlled trial. In a real BSL-2 laboratory. With 153 novices. Over eight weeks. Across five hands-on biological tasks modeling a viral reverse genetics workflow. Not a chatbot demo. Not a thought experiment. A physical lab. That matters. Because biology isn’t just text. It’s tacit knowledge. It’s sterile technique. It’s muscle memory and timing and pattern recognition. It’s knowing when a cell culture “looks off.” It’s knowing that the protocol you copied from a paper assumes three unstated steps. Benchmarks rarely capture that. The study did. And the results are, in a word, humbling. What the Study Actually Found The primary question was straightforward: does access to mid-2025 frontier LLMs significantly increase a novice’s ability to complete a sequence of tasks modeling viral reverse genetics? The answer, in binary terms, was no. Completion of the core workflow was low in both groups—LLM-assisted and internet-only—and there was no statistically significant difference in full workflow completion. If you stop there, you might conclude: the models don’t matter. But that would be the wrong lesson. Because the study also found something more subtle—and arguably more important. Across individual tasks, LLM-assisted participants were more likely to progress further through procedural steps. In cell culture, they completed tasks faster and with fewer attempts. Bayesian modeling suggested a modest uplift—on the order of ~1.4× for a “typical” reverse genetics task—though with uncertainty bounds that rightly temper interpretation. In other words: not a revolution. But not nothing. And this is where responsible innovation becomes interesting. Why This Is Violet Teaming in Practice When Adam Russell and I first articulated the idea of Violet Teaming, we described it as the integration of red teaming (adversarial probing), blue teaming (defensive hardening), and ethical design into a proactive, sociotechnical framework . Most conversations about AI and biosecurity oscillate between red and blue: Red: “What if this model can design a pathogen?” Blue: “Let’s add filters, classifiers, restrictions.” What this study does is different. It asks: what is the real-world uplift? How much does LLM assistance actually change novice capability in a physical lab? Not in theory. Not in speculation. In practice. That’s violet. Because it embeds evaluation into the design and governance process itself. Instead of arguing over worst-case extrapolations, we now have empirical data about: * Completion rates * Time-to-task * Procedural progression * Human–AI interaction patterns * Elicitation failures * Usage intensity and its (lack of) correlation with success That last point is particularly striking. Participants who used LLMs more did not necessarily perform better. There was no clean dose–response curve. That’s not a trivial observation. It tells us that raw access is not the same as effective amplification. It suggests that prompting skill, interface design, cognitive scaffolding, and user expertise mediate uplift. And that means risk is not simply a function of model weights. It’s a function of the entire sociotechnical system. That’s violet territory. The Most Important Finding: The Gap To me, the most important result is the documented gap between in silico benchmark performance and physical-world utility. This is not an indictment of benchmarks. They serve a purpose. But they are not reality. A model can generate a flawless text protocol for molecular cloning and still fail to help a novice identify the correct reagents from a messy inventory spreadsheet. It can hallucinate a DNA sequence that looks plausible but is wrong in a way a novice cannot detect. It can provide text-based instruction where video-based tacit demonstration might matter more. In the study, YouTube was often rated as more helpful than any individual LLM. That’s not because YouTube is smarter. It’s because biology is embodied. This is precisely the kind of nuance that responsible innovation requires. Without physical-world validation, we risk building policy on top of performance claims that don’t map cleanly onto human capability. This study doesn’t close the gap. It reveals it. And revelation is the first step toward responsibility. Responsible-by-Design Requires Quantification One of the themes we’ve explored in the Science of Responsible Innovation is that values without metrics are aspirations. Metrics without values are optimization problems. We need both. This study provides something we’ve been missing: a quantifiable baseline for novice uplift in a dual-use biological workflow. Not a theoretical upper bound. Not a catastrophic scenario. An empirical distribution. The Bayesian estimates even put a 95% credible upper bound around uplift (~2.6×), which matters enormously for policy calibration. If you’re designing guardrails, export controls, compute thresholds, or deployment policies, you need to know: are we talking about a 10× amplification? A 2× amplification? Or something closer to noise? This paper suggests modest uplift under the conditions studied. That doesn’t eliminate risk. It contextualizes it. And contextualization is the heart of responsible governance. Where the Study Can Go Next Now, let’s be honest. As strong as this study is, it is not the final word. It’s the first serious step. And if we want this to become an evolving framework for violet teaming and responsible-by-design evaluation, we need to iterate. Here are several ways I believe the next generation of this work could build on this foundation. 1. Extend the Time Horizon Eight weeks is meaningful. But complex biological workflows often require longer timeframes for skill acquisition. Low completion rates may reflect not just capability limits, but time constraints. A longer intervention period could reveal whether modest early procedural uplift compounds into higher eventual completion. Responsible innovation must account for trajectory, not just snapshot. 2. Integrate End-to-End Workflow The tasks were decoupled into discrete components. That’s methodologically clean, but real-world risk emerges from integration. A future iteration could test whether novices can string together multiple steps into a coherent, self-directed project—while still maintaining appropriate biosafety controls. 3. Compare Model Generations Longitudinally The models tested were mid-2025 frontier systems. Biology-specific models are already emerging. A longitudinal design—repeating the same protocol annually—would allow us to empirically track uplift curves over time. That would be invaluable for macrostrategy. Instead of forecasting speculative capability growth, we could measure it. 4. Test Interface Scaffolding The study hints that elicitation constraints matter. Novices may not know how to ask the right questions. What happens if we add structured prompting interfaces? Visual overlays? Augmented reality guidance? Automated error-checking layers? Risk may scale not just with model intelligence, but with integration depth. 5. Incorporate Expert–Novice Comparisons How much of the gap is due to user expertise? Running parallel cohorts—novices and trained biologists—could quantify differential uplift. That matters for both workforce development and biosecurity risk modeling. 6. Expand Metrics Beyond Binary Outcomes The procedural step analysis in this study was a brilliant move. Binary success/failure hides important dynamics. Future designs could incorporate: * Error rates * Near-miss events * Quality metrics * Safety deviations * Confidence calibration Responsible innovation isn’t just about “can they finish?” It’s about “how do they behave along the way?” The Human Story Beneath the Statistics I keep thinking about the participants in that lab. Undergraduates. Non-biologists. Humanities majors. Stand

    21 min

  7. FEB 17

    Ep 58 - Legitimacy Without Consensus

    Modern governance is haunted by an unrealistic expectation: that legitimacy requires agreement. We have come to believe—implicitly, often unconsciously—that if societies cannot reach consensus on the risks and benefits of a technology, then governance has failed. That disagreement itself is evidence of irresponsibility. That the absence of unanimity delegitimizes action. In an era of slow-moving institutions and narrow technologies, this belief was merely inconvenient. In an era of fast-moving, general-purpose systems, it is paralyzing. If the Science of Responsible Innovation is to function in the real world, it must confront a hard truth: consensus is no longer a prerequisite for legitimacy—and insisting on it may be the most irresponsible posture of all. The podcast audio was AI-generated using Google’s NotebookLM. The Myth of Consensus Consensus feels comforting. It suggests shared values, collective understanding, and moral clarity. It promises that decisions are not imposed, but agreed upon. But consensus has always been rarer than we like to admit. Most consequential decisions in modern history—from industrialization to nuclear power to the internet—were made amid deep disagreement. What sustained legitimacy was not unanimity, but the presence of institutions capable of acting, learning, and correcting course in public view. The expectation of consensus is a relatively recent artifact, amplified by social media, participatory rhetoric, and the moralization of policy debates. Disagreement is now treated not as a feature of pluralistic societies, but as a governance failure. This framing collapses under technological complexity. Why Consensus Breaks at the Frontier Emerging technologies resist consensus for structural reasons. They involve uncertain evidence, asymmetric risks, and uneven distributions of benefit and harm. They compress timelines. They force tradeoffs between present and future goods. They challenge existing power structures. Under these conditions, reasonable people will disagree—often profoundly. Expecting consensus in such contexts is not aspirational. It is evasive. It defers responsibility by setting an unattainable standard. Legitimacy as a Property of Process If legitimacy does not come from agreement, where does it come from? Legitimacy emerges from process, not outcome. A decision can be legitimate even when controversial if the process by which it was made is perceived as fair, transparent, and accountable. Conversely, a unanimous decision reached through opaque or exclusionary means can be profoundly illegitimate. This distinction is foundational to democratic governance, but it has been under-applied to technology. Responsible-by-design reframes legitimacy as something that is earned continuously, not bestowed once. The Elements of Legitimate Disagreement For disagreement to coexist with legitimacy, several conditions must hold. Visibility Disagreement must be visible, not suppressed. Legitimacy erodes when dissent is hidden or dismissed. Making disagreement explicit—documenting assumptions, minority views, and unresolved tensions—signals seriousness rather than weakness. Representation Those affected by a technology must have pathways to be heard, even if their views do not prevail. Legitimacy does not require that every perspective determine the outcome. It requires that perspectives be considered in good faith. Accountability Decision-makers must be identifiable and answerable. Anonymous authority breeds mistrust. Legitimate governance requires clear ownership of decisions, along with mechanisms for challenge and review. Revisability Perhaps most critically, decisions must be revisable. When evidence changes, governance must change with it. The promise of revisability—backed by real authority to act—allows societies to tolerate disagreement without freezing. Consensus as a Hidden Source of Power Calls for consensus often sound neutral. They are not. In practice, consensus requirements advantage those with veto power: incumbents, well-resourced actors, and those comfortable with the status quo. When unanimity is required, the default outcome is inaction. This dynamic is particularly dangerous in domains where delay carries real harm—unmet medical needs, climate risk, and/or infrastructure fragility. Insisting on consensus can therefore function as a form of quiet domination, disguised as caution. Legitimacy in the Absence of Certainty At the frontier of technology, uncertainty is unavoidable. Evidence will be incomplete. Models will be wrong. Early decisions will need correction. Legitimacy does not come from pretending otherwise. It comes from acknowledging uncertainty explicitly and designing governance that can absorb it. This is where governance latency becomes decisive. The faster institutions can detect harm, interpret signals, and act, the less they must rely on consensus as a substitute for control. Responsiveness replaces unanimity. The Relationship Between Legitimacy and Proportionality Legitimacy without consensus depends on proportionality. When governance distinguishes between green, orange, and red zones, disagreement becomes more tractable. Actors may still contest classification, but they are no longer arguing in absolutes. Proportionality creates space for partial agreement: agreement on process even when outcomes differ; agreement on oversight even when deployment is contested. This is how pluralistic societies move forward without pretending to agree. What Legitimate Governance Looks Like in Practice In a responsible-by-design system, legitimacy is built through concrete practices: * Clear articulation of decision criteria * Documentation of dissent and uncertainty * Defined authority to act and to revise * Transparent monitoring and reporting * Mechanisms for escalation and redress None of these require consensus. All of them require competence. The Discipline Ahead The future of technology governance will not be decided by who wins the argument. It will be decided by whether institutions can earn trust amid disagreement—by acting visibly, correcting quickly, and governing proportionally. At the frontier of technology, humanity is the experiment. Legitimacy without consensus is how we keep that experiment democratic, adaptive, and humane. That is not a compromise. It is the only path forward. -Titus Get full access to The Connected Ideas Project at www.connectedideasproject.com/subscribe

    16 min

  8. FEB 10

    Ep 57 - Who Decides the Zone of Proportionality?

    If restoring proportionality were simply a matter of classification, the problem would already be solved. Green zone. Orange zone. Red zone. The framework is intuitive. The logic is sound. And yet, in practice, the hardest part of proportional governance is not designing the zones—it is agreeing on where a technology belongs. This is the uncomfortable truth at the center of responsible-by-design: classification is not a technical exercise alone. It is a social, institutional, and political one. Every serious disagreement about emerging technology eventually collapses into a fight over zone placement. Not because people are irrational, but because zone assignment encodes values, incentives, and risk tolerance—often implicitly. Understanding why agreement is so difficult is the next step in building a Science of Responsible Innovation that actually works. The podcast audio was AI-generated using Google’s NotebookLM. The Illusion of Objective Classification There is a natural temptation to believe that enough data, enough modeling, or enough expertise will produce a single “correct” zone assignment. It will not. Risk is not an intrinsic property of technology. It is a relationship between a system and the world it enters. Severity depends on context. Reversibility depends on infrastructure. Distribution depends on power. The same technology can be green in one setting and orange—or red—in another. An AI model used for drug target prioritization inside a regulated pharmaceutical pipeline may be low risk and highly reversible. The same model released openly, paired with automated synthesis and weak oversight, may move quickly toward red. Zone assignment is therefore conditional, not absolute. Disagreement does not indicate failure of reasoning. It indicates that different assumptions are being applied—often without being named. Why Reasonable People Disagree Most zone disputes are not about facts. They are about frames. Different Reference Harms Some actors anchor on historical harm. Others anchor on theoretical maximum harm. Both are rational. Clinicians and researchers tend to focus on harm already occurring—patients dying today, diseases untreated, systems failing in real time. For them, delay carries moral weight. Security professionals and bioethicists often focus on tail risk—low-probability, high-severity outcomes whose consequences are irreversible. For them, even small probabilities demand attention. These are not incompatible perspectives. But without explicit proportional reasoning, they appear irreconcilable. Different Time Horizons Short-term and long-term risks do not feel the same, even when they are commensurate. Immediate harms are vivid and legible. Long-term harms are abstract and uncertain. People discount the future differently—not out of malice, but because institutions reward different time scales. Zone disputes often mask disagreements about when harm matters, not whether it matters. Different Power Positions Zone classification looks different depending on where one sits in the system. Those who bear downside risk—patients, workers, communities—tend to be more cautious. Those who capture upside—investors, developers, states—tend to emphasize opportunity. Neither position is illegitimate. But pretending that zone assignment is neutral obscures these dynamics. The Role of Uncertainty Disagreement intensifies under uncertainty. Early in a technology’s lifecycle, data is sparse, use cases are speculative, and second-order effects are poorly understood. This ambiguity invites projection. Optimists extrapolate potential benefit. Pessimists extrapolate potential harm. Both are filling gaps in knowledge with values. This is not a flaw. It is inevitable. The failure occurs when uncertainty is treated as a reason for absolutism rather than for adaptive governance. When Zone Disputes Become Pathological Healthy disagreement is not the problem. Pathology emerges when disagreement hardens into a stalemate or theater. This happens in three ways. First, zone inflation. Technologies are rhetorically pushed toward red because red confers moral authority. If everything is existential, restraint becomes the only defensible posture. Second, zone denial. Risks are minimized or dismissed to keep technologies green, often until failure forces reclassification. Third, zone laundering. Systems are framed narrowly to avoid scrutiny—presented as green tools while embedded in orange or red pipelines. All three erode trust. Who Should Decide the Zone? If zone assignment is not purely technical, who should decide? The answer is uncomfortable but unavoidable: no single actor can. Proportional governance requires pluralistic classification. This means: * Technical experts to assess capability and failure modes * Domain experts to understand real-world impact * Governance bodies to weigh systemic risk * Affected communities to articulate lived consequences Not consensus. Legitimacy. The goal is not unanimity, but a process that surfaces assumptions, documents disagreement, and allows decisions to evolve with evidence. Making Disagreement Productive A Science of Responsible Innovation does not eliminate disagreement. It structures it. Productive zone classification requires: * Explicit articulation of assumptions * Clear criteria for severity, reversibility, and distribution * Mechanisms for revisiting decisions as systems scale * Authority to move technologies between zones Most importantly, it requires humility—the recognition that initial classifications are provisional. Zones as Governance Conversations Zones should be understood less as labels and more as conversations. A technology placed in the orange zone is not “unsafe.” It is under active stewardship. A technology placed in the red zone is not “evil.” It is constrained because the cost of failure is too high. Disagreement over zones is not a sign that the framework has failed. It is evidence that it is being used. The Discipline Ahead The hardest work in responsible-by-design is not building the tools. It is building institutions capable of judgment under uncertainty. That requires tolerating disagreement without collapsing into paralysis or absolutism. It requires processes that can hold multiple perspectives without pretending they are equivalent. At the frontier of technology, humanity is the experiment. Deciding the zone is how we practice responsibility—not by eliminating conflict, but by governing through it. That, more than any classification scheme, is the true test of proportionality. -Titus Get full access to The Connected Ideas Project at www.connectedideasproject.com/subscribe

    19 min
See All (65)

Ratings & Reviews

3
out of 5
2 Ratings

About

Tech, Policy, and our Lives, brought to you by The Connected Ideas Project is a podcast about the co-evolution of emerging tech and public policy, with a particular love for AI and biotech, but certainly not limited to just those two. The podcast is created by Alexander Titus, Founder of In Vivo Group and The Connected Ideas Project, who has spent his career weaving between industry, academia, and public service. Our hosts are two AI-generated moderators (and occasionally human-generated humans), and we're leveraging the very technology we're exploring to explore it. This podcast is about the people, the tech, and ultimately, the public policy that shapes all of our lives. www.connectedideasproject.com

Information

  • Creator
    Alexander Titus
  • Years Active
    2024 - 2026
  • Episodes
    65
  • Rating
    Clean
  • Copyright
    © Alexander Titus
  • Show Website
    Tech, Policy, and our Lives