S2. The Nature Nanotechnology “Thesis” columns, 2014–2016: what they add to the map#

Supplementary reading for 05-maynard-risk-and-ai-map.md. Share S2: the eleven sole-authored “Thesis” columns in Resources/maynard-papers/Maynard supplied/, plus the arXiv submitted text of one of them. Every item was read in full. The map was read but not edited. There is no comparison here with Jensen Huang or with the EEA reports.

Scope, provenance and conventions#

Items read (in date order): nnano.2014.43, nnano.2014.116, nnano.2014.196, nnano.2014.294, nnano.2015.35, nnano.2015.120, nnano.2015.196, nnano.2015.286, nnano.2016.28, nnano.2016.99 and nnano.2016.167, all in Maynard supplied/. Also papers/2016_Spray-on-Carbon-Nanotubes_arXiv-1606.01439.pdf, the submitted text of nnano.2016.99.

Provenance, common to all items: - Evidence weight: highest. Each column is his own prose, and he is the only author. - Format. They are opinion columns in the journal’s “Thesis” slot, written between March 2014 and September 2016. - Affiliation. He wrote from the University of Michigan Risk Science Center until June 2015, and from the ASU Risk Innovation Lab from September 2015. - No AI involvement. All of them predate any AI assistance in his writing. - Editorial material is excluded. The journal will have edited headlines, standfirsts (for example “argues Andrew D. Maynard”) and pull quotes. These are probably editorial, so none of them is used as his words. For example, the standfirst of nnano.2015.196 says “radically change our approach to risk”, while his body text says “a radical new approach to risk”.

Citations: - Items are cited by filename and journal page, for example “nnano.2016.28 p.211”. In each two-page PDF, PDF p.1 is the first journal page and PDF p.2 is the second. - Every quotation was checked against text extracted from the PDFs. PDF-extraction artefacts (“Y et”, “co -evolved”, stray spaces before punctuation) have been silently corrected. - Curly quotation marks in the source are given here as straight ones. - “My reading” marks interpretation. † marks a label of mine, as in the map.

Overlap with the Substack corpus (checked against working/maynard/corpus/ and publications.md): - nnano.2014.294 was reposted in full, with a new introduction, as 2023-12-03 3d-artificial-brains-and-ai. The timeline (Phase 0) covers it; the map does not cite it. - nnano.2015.196 has a popular version, 2016-01-11 thinking-innovatively-about-the-risks-of-tech-innovation. - nnano.2015.286 is known in the corpus only through 2016-01-11 the-fourth-industrial-revolution-what-does-wefs-klaus-schwab-leave-out. That post links to the column for “actionable empathy” but does not reproduce it. - nnano.2016.99 has a popular companion, 2016-02-01 we-dont-talk-much-about-nanotechnology-risks-anymore (also on Vantablack S-VIS). publications.md records its coverage as NONE; PARTIAL is more accurate. - Reposts in web/. 2020science versions of nnano.2014.43 and nnano.2014.116, and a summary of nnano.2016.28, sit in Resources/maynard-papers/web/. They are outside this share and were not read. That he reposted them suggests he stood by them. - No corpus version exists for the other five columns: nnano.2014.43, 2014.116, 2014.196, 2015.120 and 2016.167. The web reposts noted above are not in the corpus.

1. nnano.2014.43, “A decade of uncertainty” (March 2014)#

Provenance. Sole author; University of Michigan. Reposted on 2020science in November 2014 (per the README).

Argument. - The report as a catalyst. The 2004 Royal Society and Royal Academy of Engineering (RS-RAE) report Nanoscience and Nanotechnologies was reputedly prompted by talk of “grey goo”. It set the global agenda for nanomaterial risk research. Peer-reviewed papers on the environmental, health and safety effects of nanomaterials rose from 139 in 2003 to 1,326 in 2013, with a jump in 2006 (p.160, Fig. 1). He calls the report “a voice of scientific reason and social concern” (p.160). - It neglected benefits. “Of the 21 recommendations arising from the report, 19 of them addressed potential implications” (p.159). - Two dangers, one ranked worse. Without the report, speculation “could have scuppered the nanotechnology enterprise or, worse, led to materials and products that showed a blatant disregard for health and environmental risks” (p.160). - Its darker legacy: a research rut. Materials such as silver, titanium dioxide and cerium oxide “don’t always seem to hold the dramatic risk surprises that some thought they would”. Carbon nanotubes are the exception. Researchers “dig ever-deeper to discover the assumed-to-exist evidence of novel risk”. The field has “worn a rut that is proving hard to get out of”, and “the relevance of this research has become less clear” (p.160). - In one line: “The speculation of possible risk has developed into an assumption of as-yet-to-be-discovered risk” (p.160). - The cost of the rut. By fixing on the materials that grabbed early headlines, “researchers run the danger of missing emerging materials and products that may present greater challenges than those initial nanoscale materials” (p.160). - Close. He asks whether researchers have created “a new, metaphorical grey goo”. If so, it may be time to “once again use speculation as a lever” to move research towards evidence-based approaches for the next generation of materials (p.160).

Key concepts. A research programme that hardens into a rut†; speculation as a lever; the assumption of undiscovered risk; omitted benefits as a flaw in risk advice; the next generation of materials.

What this adds to the map. - New idea: evidence-based programmes can mislead too. Well-meant, well-funded, evidence-based risk research can become path-dependent and self-confirming. It keeps generating data while drifting from the right question. The map has nothing on this. It bears directly on Andrew’s second note: a method and its investment can give comfort without addressing the real uncertainty. - Qualification to C7 (plausibility). The map treats speculation mainly as something to discipline. Here speculation is also a lever: a legitimate way to redirect research. Its failure mode is institutional (the rut), not only imaginative. - Qualification to “symmetric risk†” (§2, C6). He faults the report for neglecting benefits, so he counts both sides. But he ranks “blatant disregard for health and environmental risks” as worse than scuppering the enterprise. Counting both sides does not mean weighting them equally. The map’s flat “It is symmetric” (§2) is too absolute. - Deeper evidence for T2. The map’s template, from grey-goo fantasy to evidence-based risk (FFTF p.281; 2024-06-23), is here in 2014 with the RS-RAE report as its model. He adds a warning that the template can itself harden. - Andrew’s note 1. The argument is partly bibliometric (paper counts by hazard, exposure and fate). He treats the accumulated evidence as a real achievement: understanding “extends further than anyone might have imagined in 2004” (p.160).

2. nnano.2014.116, “Is novelty overrated?” (June 2014)#

Provenance. Sole author; University of Michigan.

Argument. - Why novelty drives the field. Novelty drives nano funding and commercialisation. Assuming that novel behaviour means novel risk is “In many ways … a reasonable assumption” (p.409), and it made sense when products first appeared. - Why novelty misleads. - “Novelty” is ambiguous. It can mean intrinsic nanoscale physics (quantum confinement) or extrinsic, size-enabled behaviour (tumour penetration; nanotube alignment). - Manufacturers care about utility, not novelty. - Products integrate materials that lose their identity, so “materials are mutable, and as a result their ‘novelty’ is frequently ephemeral” (p.410). - His verdict. “Novelty as a result is a subjective, transient, and consequently a rather unreliable indicator of potential risk.” It “favours the interesting (and possibly the headline-grabbing) over the important”, and “mundane risks are still risks” (p.410). Earlier: “even mundane risks are important” (p.409). - No size threshold. “there remains no indication of a bright line at 100 nm”. Harm follows “the hard reality of material–biology interactions, and not on an arbitrary determination of novelty” (p.410). - An alternative method. Study context-specific behaviour in plausible, commercially realistic products. Explore exposure potentials across the product life cycle to “map out plausible domains of risk”, “filtering out plausible modes of harm from the merely speculative and ‘novel’”. The goal is materials that are “safe — or safer — by design”, defined “by what they do rather than what they are called” (p.410). - Candour about the state of the field. “we are still struggling to develop the broader conceptual frameworks needed to ask useful, and timely, questions about material risks” (p.410). - Balance. “Some of these risks will be new or unusual; others may be conventional but not obvious” (p.410). - A pointer for the synthesis phase. Reference 1, for “over a decade of research and dialogue”, is his co-authored chapter in the EEA’s 2013 Late Lessons from Early Warnings volume. This is recorded as a fact about his citations and is not analysed here.

Key concepts. Novelty as an unreliable sentinel; mundane risks are still risks; behaviour, not labels; plausible products; life-cycle exposure mapping; safe(r) by design; no bright line.

What this adds to the map. - Earlier origin: behaviour, not labels (C13; §5.5). The map dates this to 2022-02-10 (“nature doesn’t care what we call a material”). It is here in June 2014, and again in March 2016 (item 9). He held it for at least eight years before the map’s date. - Earlier origin: “mundane but serious” (§5.7). The map’s calibration comes from AI in 2020-11-12. The same calibration governs his materials work in 2014. - Where the plausibility discipline came from (C7). “Plausible products”, “plausible domains of risk” and “plausible modes of harm” show the plausible-versus-speculative test at work inside technical risk assessment in 2014. The discipline grew out of risk science before FFTF gave it a film-based form. - Bearing on tension 1 (past lessons against “defies analogy”). My reading: his 2014 argument is that novelty misguides in both directions. It overplays some risks and hides mundane ones. Read that way, it cautions against building AI risk thinking on how unprecedented AI is, and in favour of asking what it does in plausible contexts of use. He does not dismiss novelty, though: it becomes one factor among several. - Andrew’s note 1. The method he proposes is exposure science extended to products and life cycles. It builds on quantitative assessment rather than leaving it. - Andrew’s note 2. He admits that conceptual frameworks are lacking, and he criticises letting an attractive framing (novelty) stand in for understanding.

3. nnano.2014.196, “Old materials, new challenges?” (September 2014)#

Provenance. Sole author; University of Michigan.

Argument. - A confession. Food-grade synthetic amorphous silica (fumed silica, sold since 1943) is something “I’ve often used … as an example of a safe nanomaterial”. New data at a Royal Society meeting brought evidence that “cast doubt on what I thought I knew to be true. I found myself, I must confess, facing a conundrum” (p.658). - Weighing industry evidence. The industry-prepared reviews may be “overly optimistic”, yet “the data, methodologies and analysis are sound” (p.658). He does the dose arithmetic himself: 5,000 mg per kg of body weight equals “a 70 kg adult eating 350 g”. - Plausible mechanism, weak inference. New studies (Zhang et al. 2012 in vitro; van der Zande et al. 2014 in rats) point to a plausible mechanism in strained siloxane rings. But “jumping to conclusions from in vitro studies would be at best naive” (p.659). - The central question. “how can we encourage exploratory risk research without it prematurely impacting consumer and regulatory decisions?” (p.659). - Precaution and its limits. Precaution “is a powerful argument given a long history of potentially avoidable health and environmental impacts from commercial products”. Yet “how are appropriate trigger points for action defined?” His judgement: “the balance of evidence on ingested fumed silica isn’t even close to a trigger point as yet, and that premature action could cause a lot more harm than good” (p.659). - Scientists’ duty. “the privilege of scientific insight should come with the responsibility to use this insight with care and consideration” (p.659). - Three irresponsibilities. - to question the material’s use on the strength of a cell study and a label; - “equally irresponsible to curb research” on the assumption that nothing is left to find; - “most definitely” irresponsible for researchers not to help others tell exploratory research from research meant to inform decisions. - Verdict. “Based on the available evidence, fumed silica in food is acceptably safe”, open to re-evaluation (p.659).

Key concepts. Exploratory versus actionable research; trigger points for action; balance of evidence; the privilege and responsibility of insight; “acceptably safe”; evidence judged on its merits, with its source noted.

What this adds to the map. - New idea: trigger points, and research versus action. This is his working model of when precaution should act. The map’s precaution entry (proportionate, scaled to irreversibility, never a default ban; §5.1) lacks this operational part. He asks the question openly and answers it case by case, not by rule. - Deeper evidence for C4 and §5.5. It is the fullest early statement of weight of evidence and of scientists’ responsibility for hype (map: 2014-12-14 researchers-should-take-more-responsibility). - Method. An early, clear case of self-implication and of revising his view in public (map’s method note): he is publicly unsettled by evidence against his own standard example. - Weighing costs. He counts “lives and livelihoods” harmed by premature regulation, in 2014. - A technology-neutral lens for later use. Industry-generated evidence is judged on its merits, with its source and possible optimism noted. - Andrew’s note 1. This is classic quantitative risk reasoning: dose extrapolation, animal-to-human differences, in vitro against in vivo. - Andrew’s note 2. It shows humility about his own earlier certainty. It also shows that his humility cuts both ways: he refuses to overclaim safety or harm.

4. nnano.2014.294, “Could we 3D print an artificial mind?” (December 2014)#

Provenance. Sole author; University of Michigan. Reposted in full with a 2023 introduction (2023-12-03 3d-artificial-brains-and-ai). According to the batch notes (B13), that introduction says his editor had doubts about running it and calls it “increasingly prescient”.

Argument. - Brains depend on their structure. “Our brains are analogue devices that are intimately dependent on their physical structure” (p.955). Two-dimensional fabrication limits brain-like computing. - A plausible convergence. 3D printing converging with nanotechnology could plausibly, within 10–20 years, produce dense 3D neuromorphic substrates, with printed power supply and microfluidic cooling. Such substrates could “blur the boundaries between programmed functionality and intelligence” (p.955). - Testing plausibility first. “a key first step is exploring whether there is a plausible likelihood of such a convergence occurring in the first place” (p.955). He labels his sketch “a naive thought experiment” (p.956). - Mind and awareness. “The idea of ‘mind’ implies a degree of awareness of self and environment”, an emergent property. Our only experience of massively interconnected 3D analogue processors is in living organisms. So he doubts the easy assumption that a printed “brain” is no more likely to show mind-like properties than a powerful computer (p.956). - Close. “if our technological capabilities are beginning to shift from the fanciful to the plausible in constructing an artificial mind that has some degree of awareness, how do we begin to think about responsibility in the face of such audacity?” (p.956).

Key concepts. Convergence of 3D printing, nanotechnology and neuromorphic computing; the substrate as a driver of AI capability; mind as emergent; from fanciful to plausible; responsibility in the face of audacity.

What this adds to the map. - Earlier AI engagement. The map says AI “arrived late”. In 2014 he wrote a column on AI capability and machine awareness. It came in through hardware convergence, which fits the map’s “one converging strand” but starts earlier than the map’s AI story. - Earlier origin: hubris. The map dates “technologies of hubris” to FFTF (2018). “responsibility in the face of such audacity” is here in December 2014, attached to artificial minds. - Bearing on tension 14 (substrate and consciousness). My reading: in 2014 his instinct was substrate-sensitive: mind as emergent from an analogue, three-dimensional physical architecture. That fits his 2024 turn to Seth’s biological naturalism (2024-06-30). It makes his 2023 endorsement of computational functionalism (2023-08-23) the outlier. He reposted this substrate piece in December 2023, months after that endorsement. - Qualification to superintelligence agnosticism. Here plausibility is used to open a speculative capability question, not to close it. His 2018 agnosticism about superintelligence did not extend to treating artificial minds “with some degree of awareness” as implausible. - Andrew’s note 2. He labels his speculation openly (“naive thought experiment”) and hedges it (“not beyond the bounds of plausibility”, p.955).

5. nnano.2015.35, “The (nano) entrepreneur’s dilemma” (March 2015)#

Provenance. Sole author; University of Michigan. He writes as a teacher on Michigan’s Master of Entrepreneurship programme, and mentions his son’s hackathon.

Argument. - Responsible innovation is gaining ground. Examples are the 2004 Alexandria dialogue, the EU’s Responsible Research and Innovation programme, the Journal of Responsible Innovation and the Virtual Institute for Responsible Innovation. Big business already has corporate social responsibility. - But entrepreneurs struggle with it. “Occasionally blinkered by their own optimism perhaps, entrepreneurs often have a deep-seated belief in the safety and efficacy of their creations; they need to in order to convince others to invest in their vision” (p.199). The threat of near-term failure makes attending to speculative future impacts “an ill-affordable luxury”. - The cost of neglect. Early disregard can leave technologies “locked into development trajectories that are highly susceptible to failure”. Early “‘course corrections’ early on in the innovation development process can help reduce or avoid liabilities later on” (p.199). Indiscriminate use of nano in consumer products in the mid-2000s led some businesses to shy away from the technology. - Entrepreneurs mean well. “it’s rare in my experience to find entrepreneurs who don’t care about the long-term consequences of their creations”; they are “more often than not driven by a desire to benefit society” (p.199). - The dilemma. “how can responsibility be built into the innovation process without it stymieing the very innovations it sets out to enable?” (p.199). - Why academic responsible innovation fails them. - Academic definitions (von Schomberg; Stilgoe, Owen and Macnaghten) are “intellectually elegant, but rather removed from the cut and thrust of most entrepreneurial environments”. - To an entrepreneur, responsible innovation “looks suspiciously like an attempt to regulate the very inventiveness that drives them”. - It must be “built into the fabric of entrepreneurship” rather than foisted from the top down (p.199). - Practical tools. - the Responsible Nano-Code, whose seven principles include board accountability, stakeholder involvement, and transparency and disclosure; - MATTER’s Principles for Sustainable Innovation; - mapping future technology trajectories; - informal networks with academics, regulators and publics; - “basic listening skills with a dose of humility can pay surprising dividends when public opinion is often as influential as technical viability”; - investors agreeing “metrics of worth-creation” (p.200). - The real barrier. The greatest barriers “are not necessarily time and cost, but imagination” and access to information (p.200).

Key concepts. The entrepreneur’s dilemma; optimism as a functional requirement†; lock-in and course correction; responsibility as added value; responsibility built from the bottom up; worth-creation.

What this adds to the map. - Earlier origin: the sincere-but-myopic innovator (C10; §5.2), with structural pressure included. The map dates the psychological account to FFTF (2018) and the structural account to 2022 (“super-consumers”). Here, in 2015, the innovator’s optimism is explained as required by the investment process, and neglect as a product of survival pressure. So structure and psychology run together from the start, which supports the map’s §7 point that structure is not a late turn. - Qualification to the story of responsible innovation (C12; §7 “Confidence in remedies”). The map and timeline have responsible innovation as “an endorsed framework (2015–2019)” before doubt sets in with “fiendishly hard to operationalize” (2023). In fact he already argued in March 2015 that its academic form was “removed from” entrepreneurial reality, while endorsing its aims. - The inferred shift in the timeline. The move from citing responsible innovation to proposing his own risk-innovation framing (“2015→2016”) happens within 2015. He reworks responsible innovation for entrepreneurs in March, and names risk innovation in September (item 7). - Deeper evidence for tension 4 (whose value?). The case for responsibility is made in enterprise terms: added value, liabilities avoided, investor confidence, “points of shared value”. This confirms the map’s reading of the frame’s early, enterprise-facing uses. It sits alongside his conviction that entrepreneurs want “to benefit society”. - New idea: lock-in and early course correction. Path dependence as a reason to intervene early. The map has “naive adoption and lock-in” only for education (2024-05-05). - Self-governance. He endorses voluntary industry codes “Used intelligently”, while saying “more is needed” (p.200). - Andrew’s note 1. Formal risk assessment is taken for granted (“can and do get expensive”). The question is access and timing: “Collaborating with academics studying risk may lead to low-cost risk analysis” (p.200).

6. nnano.2015.120, “Learning from the past” (June 2015)#

Provenance. Sole author; University of Michigan.

Argument. - A request for help. He receives “a small but steady stream of similar enquiries from people struggling to make sense of seemingly mysterious illnesses” (p.482). One correspondent had been told by an “expert” that only two or three electron microscopes worldwide could characterise particles from a fire extinguisher. That is false. “My e-mailer’s expert was a victim of the assumption that nanoparticles are so unique, and so small, that routine characterization methodologies have yet to catch up with them” (p.482). - A century of measurement. John Aitken’s 1888 condensation particle counter and his 1889 count from the top of the Eiffel Tower; Ruska’s electron microscope; his own early-1990s methods for collecting and analysing nanometre particles by scanning transmission electron microscopy and electron energy-loss spectroscopy. - The real gap. “there has been considerable uncertainty over which measurements are important from a health perspective” (p.482). “an inability to routinely monitor exposures with respect to specific particle attributes is far from being synonymous with not having the tools to characterize engineered nanoparticles” (p.483). - An anecdote. A student asks, “can we trust papers more than a few years old?” (p.483). - The principle. “seemingly novel challenges don’t always demand novel solutions, and sometimes, the key to moving forward safely, is to look back at what’s already known” (p.483). - What is still hard. “The harder challenge is working out what we should be measuring”, which attributes and at what levels. “Research is still needed here” (p.483).

Key concepts. Learning from the past; being able to measure versus knowing what matters†; the myth that novelty outruns method; the value of old literature; answering people’s worries.

What this adds to the map. - Andrew’s note 1, directly. He identifies as a measurement scientist and places his own 1990s work in a lineage going back to 1888. His foundations are quantitative, and they are explicitly built on. - Andrew’s note 2, directly. Being able to measure is not the same as knowing what matters, and “what we should be measuring” comes first. My reading: this is the materials-science root of his reluctance to put numbers on AI risk. Precise numbers on an attribute that does not matter give comfort, not understanding. - Bearing on tension 1. “seemingly novel challenges don’t always demand novel solutions” is his own 2015 counterweight to “defies analogy”. It warns against overclaiming novelty as firmly as the 2016 column (item 9) warns against forcing the new into old categories. - Method. He deals with a frightened member of the public with sympathy and plain evidence (“I would be surprised if…”). This is an early glimpse of the engagement practice the map says is under-represented.

7. nnano.2015.196, “Why we need risk innovation” (September 2015)#

Provenance. Sole author; his first column from the ASU Risk Innovation Lab. This is the founding printed statement of risk innovation. The 2016-01-11 Conversation piece is its popular version and drops most of the content below.

Argument. - The opening case. Google’s nanoparticle-sensor pill faces health and environmental questions. It also faces “outmoded or overly restrictive regulations”, “investor ambivalence, consumer suspicion, or social media backlash”. “Yet the probability of causing harm is not the only risk that could prevent these nanosensors from becoming a reality” (p.730). These “hint at a much larger and murkier risk landscape”. - Other examples. The Future of Life Institute’s grants of “close to US$7 million” for “robust and beneficial development of artificial intelligence”; CRISPR in human embryos; self-driving cars (p.730). - The underlying problem. “a growing disconnect between the rate at which we are innovating, and our ability to assess and manage the adverse consequences of this innovation” (p.730). - What conventional assessment misses. “Important as evidence-based health and environmental risk assessment and management are, they fail to capture the full panoply of personal, social, environmental, technological, economic, political and corporate risks” (p.730). - Innovation at the fringes. “Aba-made” repurposing in Nigeria, the Maker Movement and DIY biology. This innovation is “likely to flourish with little formal risk oversight”, can meet local needs “precisely because it exists at the fringes of formal regulatory frameworks”, and is “rarely risk-agnostic”. “at the heart of this disruption are shifts in what is considered to be of value”. What counts as value widens to “social justice, community resilience, fiscal independence, and personal discovery and pleasure” (p.730). - “The democratization of influence”. Influence is driven by “perceptions, beliefs and values that are not always grounded in scientific evidence, yet nevertheless have societal legitimacy”. And “even evidence-based approaches to assessing, managing and regulating risk are often grounded in values”. The EU’s regulatory definition of a nanomaterial “reflects a belief in what is important and implementable, not necessarily what has the potential to cause harm” (p.731). - The proposal: risk innovation. - “we need, I would argue, a new domain of research and practice: risk innovation”; - an organising framework for tools and practices that “protect social and environmental value, as well as enabling its creation and growth”; - “risk entrepreneurship, where the ultimate measure of an idea’s worth is whether it has an impact, not whether it adheres to convention”; - a “‘fail fast fail forward’ mentality”; - it “frames risk as a threat to existing or future ‘value’”, with value “broadly and multiply defined within personal, societal and organizational contexts” (p.731). - The range of risk innovation. - At one end, a book of seventeen haiku from a workshop with V2_ that “has more in common with arts and literature than it does risk analysis”. - At the other, the EPA and NIH’s Toxicology in the 21st Century programme, which uses high-throughput screening and computational biology to assess “tens of thousands of untested chemicals” (p.731). - Its place among other approaches. “It complements approaches to ensuring the safe and responsible use of emerging technologies”, naming responsible innovation and anticipatory governance. Risk is “not only endemic, but integral to progress”. “Without risk innovation, all we are left with is business as usual” (p.731).

Key concepts. Risk innovation; risk as a threat to existing or future value; the risk landscape; the gap between innovation and assessment; democratised innovation and influence; risk entrepreneurship; “fail fast fail forward”; value-laden risk definitions.

What this adds to the map. - Dates. Risk as a threat to value, “existing or future” value and the “risk landscape” were all in print in September 2015. The map’s earliest dates for them are 2016-01-11 and 2016-03-02. The map already notes that risk innovation comes “from a 2015 column”; the concepts should carry the 2015 date too. - Andrew’s note 1: the key text. The founding statement opens with “Important as evidence-based … risk assessment … are”. It sets the framework’s range from haiku to computational toxicology, so innovation in quantitative methods sits inside risk innovation. The map’s C3 (“necessary but no longer sufficient”) is right. The map should add that risk innovation was framed from the start as adding to quantitative risk science, not as an alternative to it. His rhetoric (“a radical new approach”) may be one source of the map’s absoluteness; his substance is additive. - A development in the relationship with responsible innovation. In 2015 risk innovation “complements” responsible innovation and anticipatory governance: they sit side by side. By 2020 he calls them “different approaches to applying the concepts that underlie risk innovation”, which is nested (the map, §3). The map presents the nesting as a fixed feature. It is a later development. - Earlier origin, secure in his own prose: definitions select risks (lens 6; orphan risks; the 2026 [mixed] selection apparatus). The EU-definition example says a risk definition encodes “what is important and implementable” rather than harm potential. This is his own idea a decade before the Fable-assisted paper, and it strengthens his claim to the idea that definitions select risks. - Earlier origins. The pacing problem (the gap between innovation and assessment), in September 2015; the social legitimacy of perceptions not grounded in evidence (map: 2016-03-12). - New idea: democratised innovation. Fringe and DIY innovation is treated both as a way to meet local needs and as a source of unmanaged risk. It is a small addition to the map’s thin coverage of the Global South (§8 gaps). - “Fail fast”, both ways. He endorses “fail fast fail forward” for risk methods. Three months later (item 8) he warns of systems “failing fast and failing spectacularly”. My reading: this is the seed of the reversibility test (2025-03-02): fast failure is acceptable in ideas and methods, not in tightly coupled systems. - AI. The founding statement of risk innovation uses AI safety funding as an example of the new risk landscape. So AI sits inside the frame from its first day, as one converging technology among several.

8. nnano.2015.286, “Navigating the fourth industrial revolution” (December 2015)#

Provenance. Sole author; ASU. publications.md records it as his most-cited governance piece.

Argument. - What converges. Industrie 4.0 and the fusion of digital, physical and biological technologies: the Internet of Things, 3D printing, and “cloud-based artificial intelligence and open-source hardware and software” (p.1005). - An ungoverned landscape. The risk landscape “lies dangerously far beyond the ken of current regulations and governance frameworks”. We risk “a global ‘wild west’ of technology innovation, where our good intentions may be among the first casualties” (p.1005). - Specific pressures. - “The risks of cyber ‘insecurity’ increase by orders of magnitude” as manufacturing is distributed. - Distributed manufacturing slips the regulatory net. - Hardware changed by software updates (Tesla’s autopilot) outpaces rules: “manufacturing regulations remain based on product development cycles that span years, not hours”. - “well-intentioned technologies are at some point going to fall through the holes in an increasingly inadequate regulatory net” (p.1005). - Convergence widens the gap. Robotics, nanotechnology and cognitive augmentation, and “artificial intelligence, gene editing and maker communities”, are “increasing the gap between what we can do and our understanding of how to do it responsibly”. - No frameworks, and the danger of fast failure. “we lack even the beginnings of national or international conceptual frameworks to think about responsible decision-making and responsive governance”. And “without rapidly emerging abilities to identify early warnings and take corrective action, the chances of systems based around converging technologies failing fast and failing spectacularly will only increase” (p.1005). - A knife-edge. In an “admittedly speculative assessment”, the revolution “stands on a knife-edge between great social good and widespread social harm”. It is “a revolution that we cannot turn the clock back on”: “Even without a top-down steer, entrepreneurs will continue to experiment with converging technologies”. “The result may be a slower revolution, but it will be a revolution nevertheless”. Still, “we have an opportunity to help steer” it (p.1006). - Six lessons, framed as personal (“These lessons are diverse, and vary from person to person and group to group”): 1. Multi-stakeholder dialogue. “we still lack the forums, the methodologies and the leadership necessary to ensure actionable outputs from multi-stakeholder dialogues”. 2. “Encouraging actionable empathy.” Empathy “is rarely taught and infrequently exercised”. It lets people “reflect and respect — without necessarily fully incorporating — all stakeholder perspectives”. 3. Education for everyone from consumers to CEOs, through science museums, citizen science, Twitter and YouTube. 4. Next-generation foresight, including “mechanisms for detecting early warnings of systemic instabilities around converging technologies that could signal local, or more widespread, catastrophic failure”. 5. Transforming risk. “Many established risk analysis methodologies are grounded in, and were developed as a consequence of, the impacts of previous industrial revolutions”. We must be “jolted out of our existing mental and procedural risk-ruts”. 6. Public–private investment, because the revolution could “increase the gap between the rich and the poor, amplify geopolitical tension, undermine environmental sustainability initiatives”. “These are all lessons learned from previous industrial revolutions, and there is no reason to suppose that the next one will be any different.” “the calculus is simple: the less that’s invested in avoiding failure, the more likely failure is” (p.1006).

Key concepts. Convergence; the “wild west”; years, not hours; early warnings of systemic instability; catastrophic failure; actionable empathy; next-generation foresight; risk-ruts; a knife-edge; “cannot turn the clock back”; steering.

What this adds to the map. - Earlier origins, all in December 2015. - Early warnings and tipping-point thinking. The map dates this to the 2018 FFTF draft and Future Rising (2020). - The timescale mismatch. The map dates it to 2021 and 2025. - The could/should gap. The map dates it to FFTF (2018). - Good intentions as casualties. A seed of myopic benevolence. - Inevitability joined to steering. The map has the obligation to innovate (2018) and hardening from 2025 to 2026. “Cannot turn the clock back … help steer” is his 2015 baseline, and its inevitability rests on decentralised entrepreneurs, not only on large firms. What hardens later is the application to AI in particular, and the narrowing of pauses (§7 “What changes”, 5). - A recurring diagnosis. “we lack even the beginnings of national or international conceptual frameworks” (2015) prefigures “the lack of even the beginnings of a framework” for AI hazard and exposure (2023-11-26). - New concept: actionable empathy. It is absent from the map, and in the corpus it survives only as a link. - Bearing on tension 9 (engagement as principle, thin as mechanism). He named the missing mechanisms himself in 2015 (“the forums, the methodologies and the leadership”). The tension is more “[he says so]” than the map records. - New: systemic and catastrophic failure. Systemic failure of converging technologies is framed in 2015. The systemic-risk thread (§5.7) is older than his 2023 AI writing. - Deeper roots of C17. Education appears in 2015 as a governance lever, including YouTube. - Gaps (§8). Cybersecurity gets short but real analysis here (“orders of magnitude”; distributed manufacturing). Justice appears as the rich–poor gap and geopolitical tension. - Andrew’s note 1. Transforming risk is one lesson of six. It rests on a critique of fit: methods built for earlier industrial revolutions. It is not a rejection of quantitative science. - Andrew’s note 2. “admittedly speculative”; how vulnerable we will be “is unclear”; there are no conceptual frameworks yet. This is the stance of humility towards a converging cluster that includes AI, ten years before the AI-specific version.

9. nnano.2016.28, “Navigating the risk landscape” (March 2016)#

Provenance. Sole author; ASU. There is no corpus version. It is addressed to early-career researchers.

Argument: seven guideposts. 1. “Risk starts with something that is worth protecting.” He starts from the conventional definition: “We usually think of nanotechnology ‘risk’ as the probability of disease or death occurring … This is a useful starting point.” Then: “When stripped down to fundamentals, risk concerns threats to something you or others value, or consider to be of ‘worth’.” This includes health and environment, and also “security, friendships, social acceptance, and our sense of personal and cultural identity”. “These broader dimensions of worth often depend on who is defining them”. A product perceived as threatening what someone values is a problem however safe you believe it is. Hence “engaging early and often” (p.211). 2. “‘Nanotechnology’ is an unreliable indicator of risk.” The attributes of a material predict harm; its labels do not (p.211). 3. “Nanomaterials are not just chemicals.” Chemical methods assume precise definition, but nanomaterials are described by statistical parameters. “These parameters are always a compromise, and never capture the full complexity of the material. And this opens up the possibility that a statistical parameter of choice may not adequately reflect a risk parameter of relevance.” “Treating nanomaterials as if they are well-defined chemicals can lead to substantial errors of judgment where risk is concerned”, “burying potentially important attributes in crude, and potentially irrelevant, metrics” (p.211). 4. “Benchmarking is important.” Benchmarking “helps tether speculative ideas to plausible realities”. “Without appropriate benchmarks, speculation can easily lead to conclusions and actions that aren’t supported by subsequent research” (p.211). Examples: ionic cadmium for quantum dots, and bleach, alcohol and soap for nanosilver. 5. “We have co-evolved with nanoscale materials.” Bodies are “surprisingly adept at managing materials that we consider to be highly novel”. Familiar-seeming materials such as fumed silica or carbon black may harm “precisely because those pathways are attuned to characteristics that these nanomaterials share with more biocompatible materials”. So the task is to understand how systems respond differently, since systems that have never met a nanomaterial do not exist (p.212). 6. “How you think about nanotechnology risk is probably incomplete.” “it’s hard for any one individual or disciplinary group to develop a complete picture of the causal chain between material design and the risk it is likely to result in”. Engineers have an “exquisite understanding” of their materials but may lack “even the language” to understand biological interference. Researchers need “the humility to recognize and respect expertise outside of their domain”. They also hold “well-meaning but ill-informed assumptions on citizen perceptions”, so that “the risk landscape researchers hold in their heads bears little resemblance to reality” (p.212). 7. “We need to be quick to question, and slow to respond.” Researchers need freedom for “what if?” research. They should avoid “hard-to-rescind decisions on potentially misleading immature science”, and “premature calls-to-action also run the risk of shutting down speculative research”. “Yet we also need the ability to respond proactively where early warnings of potential harm do begin to emerge — even before the science is mature” (p.212).

He closes by calling the guideposts “my own limited and probably occasionally blinkered experiences from working in the field for over 15 years” (p.212).

Key concepts. Worth; attributes, not labels; metrics as compromise; benchmarking; co-evolution; incomplete mental models and humility across disciplines; “quick to question, and slow to respond”; early warnings.

What this adds to the map. - Andrew’s note 1: the strongest single text. The threat-to-worth definition is introduced on top of the probability-of-harm definition (“a useful starting point”), in a column otherwise about toxicology practice. It is his own statement of “built on, not abandoned”. - Andrew’s note 2: the strongest single text. Metrics can miss what matters (“a statistical parameter of choice may not adequately reflect a risk parameter of relevance”). Forcing a new kind of thing into existing quantitative categories produces “substantial errors of judgment”. “Probably incomplete” is a guidepost in its own right. And humility runs across disciplines, including towards publics. My reading: this is almost exactly the structure of his later caution about quantifying AI risk, which bears out Andrew’s note that the lack of AI metrics comes from humility about method, not from neglect of it. - Earlier origin: analogy with its breakpoints named (C13; §5.5). “Nanomaterials are not just chemicals” (2016) is the template for “an algorithm is not a chemical” (2019-03-05). - Earlier origin, and an upgrade: comparison against a baseline (§5.1). The map has this as “Occasional” from 2021. In 2016 benchmarking is a named core practice, tied explicitly to plausibility. - Structural precursor of the cognitive Trojan horse (C14; §5.8). My reading. The co-evolution guidepost has the same shape as the 2026 argument: defences that evolved around familiar cues can be passed by something new that shares those cues. The 2026 essay (2026-01-10) invokes “evolutionary mismatch” with synthetic chemicals and vaccines; this is his own materials-science version, ten years earlier. - Earlier origin: technical experts’ blind spots (T5). The map says governing emerging technologies is a field of expertise that AI’s insiders mostly lack. In 2016 he makes the same point about materials engineers, and applies it to toxicologists and to researchers’ assumptions about publics. - Qualification to tensions 2 and 3. The 2016 guidepost holds both halves: be slow to act on immature science, and be ready to act on early warnings before the science matures. Together with the 2014 “trigger points” (item 3), this shows a deliberate balance held since 2014, not a drift between 2019 and 2026. The distinction between speculative research (free) and action (gated by evidence) also helps read his 2025–26 treatment of tails. “just on the off chance” (2025-04-06) concerns what to think about and research, not what triggers action. My reading. - Tension 4. “who is defining them” shows that he raised the question of whose value counts in 2016.

10. nnano.2016.99, “Are we ready for spray-on carbon nanotubes?” (June 2016), with arXiv 1606.01439#

Provenance. Sole author; ASU. The arXiv file is his submitted text. It matches the published column in substance. It adds only a sentence on Jun Kanno’s parallel Japanese research (Takagi et al. 2008) and labels the IARC outcome “group 3”, so it adds nothing further to the map.

Argument. - A new use. Anish Kapoor has exclusive rights to Vantablack S-VIS, a carbon-nanotube spray paint. Having worked on nanotube safety, he is “intrigued” to see nanotubes sprayed on objects people may touch. - An early warning, ignored. Within a year of Iijima’s 1991 description, the occupational hygienist Gerald Coles warned of asbestos-like effects. “At the time, his concerns did not gain much attention” (p.490). - His own part in the evidence. His own studies followed (2004 on exposure; 2005 on lung responses). In 2006 he was one of 14 authors calling for high-aspect-ratio nanomaterials to be investigated “within the next 5 years”. Poland, Donaldson and colleagues (2008) showed asbestos-like pathogenicity for long, straight multi-walled nanotubes. - Variety defeats generalisation. Nanotubes vary in “tens of thousands of such combinations”. Understanding them “has proved to be a herculean task”, and “it is still hard to predict how any given sample of carbon nanotubes will impact someone if they are exposed” (p.490). - MWNT-7. It took its name from a FedEx invoice: the “7” was chosen because the employee “loved Lucky-7 for gambles” (a personal communication, not his words). In 2014 IARC found that carbon nanotubes “cannot be classified due to a lack of data”, except MWNT-7, which had been studied so extensively that it could be placed in group 2B (p.491). - What the evidence says. Oberdörster and colleagues (2015): harm is likely “for some forms of the material (not all), at some exposure levels (as yet largely undetermined)”. The NIOSH recommended exposure limit is 1 µg/m³ (p.491). - Exposure decides. The S-VIS nanotubes are short and bound, but spraying produces respirable droplets, and whether coated surfaces shed particles is unknown. “everything hinges on the nature, form and concentrations of nanotube material that workers, users and others might actually be exposed to” (p.491). - Responsibility rests on firms and users. Surrey NanoSystems’ responsible approach is praised. “Yet so much of this responsibility currently lies with manufacturers and users. Considering this, I wonder what might happen when a less responsible company comes along with the next nanotube spray-paint.” “Without doubt, carbon nanotubes can be used safely, but without greater awareness of the potential risks they present, it is by no means a foregone conclusion that they will be” (p.491).

Key concepts. An early warning ignored, then taken up; exposure decides; variety that defeats generalisation; evidence concentrated where study concentrates†; responsibility that rests with firms and users; the less responsible entrant.

What this adds to the map. - Andrew’s note 1. This is his most technical exposure-science column. It was written between the launch of risk innovation (September 2015) and its second statement (March 2016). The value frame and quantitative exposure science ran side by side; one did not replace the other. - Andrew’s note 2. After roughly 24 years (1992–2016) of research, the effect of a given sample is “still hard to predict”. Formal classification was possible only for the one intensively studied form, named after a shipping invoice. My reading: this is the lived basis of the humility he brings to AI. Even a mature, well-funded quantitative programme lags behind the variety of its subject, and how evidence accumulates depends on accident and on where effort goes. - An early-warning story from his own record. Coles’s 1992 warning “did not gain much attention”, and he helped revive it (2006). The map mentions “nanotubes and asbestos” only as a reference case. This is recorded for the synthesis phase without comparison. - New technology-neutral lens (C10, C12). A regime that depends on the responsible firm’s responsibility is exposed to the next, less responsible entrant, and to users who do not know how to use the product safely. It sharpens his case against self-certified responsibility and for rules on specific harms. - Deeper evidence for lens 6 (how risk systems select risks). The IARC outcome shows evidence accruing where study effort is concentrated. The same point appears in item 1 (headline materials) and item 7 (the EU definition).

11. nnano.2016.167, “Is nanotech failing casual learners?” (September 2016)#

Provenance. Sole author; ASU.

Argument. - A paradox. Nanotechnology is heavily funded (US federal spending of US$1.4 billion in 2015), yet casual, self-motivated learners struggle to find good material. - An informal survey. Prompted by work with a sixth-grade teacher, he asked eight family members to search. He admits: “Not that any of this has much significance given the sample size or ‘methodology’” (p.734). - What they found. Good resources exist but are not easy to find. The big online learning platforms (Khan Academy, TED, Crash Course) carry little on nanotechnology. - Why it matters. “Nanotechnology is already affecting the lives of billions of people”. The point is “not so much about building trust with publics around the technology or just about public education”. It is about “ensuring that curious and motivated individuals have the ability to find information that is accurate, understandable and relevant to them”. Without this, “it becomes easier for nanotechnology development that is not accountable to citizens to occur”, and “easier for opportunists to fill the information-vacuum” (p.735). - Remedies. - “placing casual learning on a par with formal education”; - training early-career scientists; - letting faculty substitute online content for teaching; - partnerships with established content creators; - “counting excellence in developing online casual learning resources toward academic tenure and promotion”. - Candour. “some researchers are probably not cut out for engaging with casual learners”. Risk Bites is “essentially me, a whiteboard, and a camera” (p.735).

Key concepts. Casual learners; discoverability; the information vacuum; accountability through access to information; incentives for public scholarship; Risk Bites.

What this adds to the map. - Earlier origin. The “information-vacuum” filled by opportunists (2016) comes before the map’s “insights vacuum” (2023-04-10) and “understanding-vacuum” (2023-11-26). - Deeper roots of C17 and of public scholarship (method). Self-directed learning is treated as a matter of democratic accountability in 2016, with academic reward structures as the barrier. This links to universities’ public duty (2016-01-31). - Against the deficit model. He says explicitly that the aim is not persuasion (“building trust”) or public education. It is access to information for the curious citizen. This supports C5 and §5.3. - Method. He is candid about the weakness of his own evidence (a sample of eight). It bears little on AI.

Cross-cutting findings#

A. Earlier origins (the map’s date against the earliest date in this share)#

B. Andrew’s two notes#

Note 1: quantitative risk assessment remains a foundation, built on. - For. The evidence here is plentiful and explicit: - The founding statement of risk innovation begins “Important as evidence-based … risk assessment … are” and puts computational toxicology inside the framework (nnano.2015.196 pp.730–731). - The value definition is introduced as extending a definition that is “a useful starting point” (nnano.2016.28 p.211). - The same months produce his most technical exposure-science writing: dose arithmetic, life-cycle exposure mapping, benchmarking, exposure limits (nnano.2014.196, 2014.116, 2016.28, 2016.99). - He places his own 1990s measurement work in a century-long lineage (nnano.2015.120). - Complication. His rhetoric sometimes signals a break: “a radical new approach to risk”, “not whether it adheres to convention”, “risk-ruts”. A reader who takes the rhetoric over the substance will make the map’s mistake. The substance is extension: “It complements approaches”.

Note 2: humility about method, and the hubris of numbers. - For. The technical roots are clear: - metrics that “may not adequately reflect a risk parameter of relevance” (nnano.2016.28); - being able to measure versus knowing “what we should be measuring” (nnano.2015.120); - a funded programme hardening into “an assumption of as-yet-to-be-discovered risk” (nnano.2014.43); - after 24 years, it is “still hard to predict” what a given sample will do, with evidence concentrated on one arbitrarily named sample (nnano.2016.99); - “probably incomplete” as a guidepost; “we lack even the beginnings of … conceptual frameworks” (nnano.2015.286). - Complication. In nanotechnology his answer to deep uncertainty was more and better-targeted quantitative research (the 2006 research agenda; “Research is still needed here”), not stepping back from quantification. His humility also cuts against acting on thin evidence (“quick to question, and slow to respond”), balanced by readiness to act on early warnings. My reading: for AI, the faithful reading is not “numbers are hubris”. It is “know what matters before you measure; do not force a new kind of thing into metrics built for an old one; label speculation; and still grapple”. This is consistent with Andrew’s own gloss.

C. Effects on the map’s tensions (§8)#

• Tension 1 (past lessons against “defies analogy”). He has held both halves since 2014–16. Don’t overclaim novelty: “a rather unreliable indicator of potential risk” (2014) and “seemingly novel challenges don’t always demand novel solutions” (2015). Don’t force the new into old categories: “Nanomaterials are not just chemicals” (2016). The 2026 “defies analogy” leans towards the second half. His 2014 self supplies a test for the first half: judge by what a technology does in plausible contexts, not by how new it is.
• Tensions 2 and 3 (plausibility against tails; weight of evidence against acting early). These are a balance he has held on purpose since 2014: trigger points (nnano.2014.196); “slow to respond” alongside readiness to act on early warnings “even before the science is mature” (nnano.2016.28). He also distinguishes free speculative research from evidence-gated action. He asks the trigger-point question openly, so tension 3 could be tagged “[partly his]”.
• Tension 4 (whose value?). Enterprise-facing justifications dominate in 2015 (nnano.2015.35). “who is defining them” appears in 2016 (nnano.2016.28).
• Tension 5 (symmetry and steering against inevitability). “Cannot turn the clock back” plus “help steer” is his 2015 position, grounded in decentralised entrepreneurship.
• Tension 9 (engagement thin as mechanism). He named the gap himself in 2015 (“the forums, the methodologies and the leadership”) and proposed “actionable empathy”.
• Tension 14 (substrate). His 2014 intuition was substrate-sensitive, which fits 2024 and makes 2023 the outlier (my reading).
• §2 and C6, “It is symmetric”. He counts both sides but does not weigh them equally (“or, worse”, nnano.2014.43 p.160). The map should soften this.

D. Suggested corrections to supporting files (not made here)#

• publications.md, item 9: it says the two risk-innovation columns introduce “orphan risks”. Neither nnano.2015.196 nor nnano.2016.28 uses the term. The map’s 2018 date for orphan risks stands.
• publications.md, B4: nnano.2016.99 coverage should read PARTIAL (the 2016-02-01 companion post); nnano.2014.294 is covered in full by 2023-12-03.
• Map §1 “Limits”: the Nature Nanotechnology columns, including “Why we need risk innovation”, have now been read directly.
• Timeline, Phase 1: the move from citing responsible innovation to risk innovation happens within 2015 (March → September), and his doubts about the practicality of responsible innovation are present from March 2015.

Digest: the most important additions from S2#

The eleven Nature Nanotechnology columns (March 2014 to September 2016) are sole-authored, predate AI and carry the highest provenance. They are the formative layer the map is missing. Many of his signature ideas are already at work in his technical prose two to eight years before the dates the map gives them, and they grew out of materials risk science, not out of films or AI.

1. Quantitative risk science is the foundation, built on rather than left behind (Andrew’s note 1). The founding statement of risk innovation (nnano.2015.196, September 2015) opens “Important as evidence-based health and environmental risk assessment and management are”. It places high-throughput computational toxicology inside risk innovation, at the opposite end of a spectrum from a book of haiku. “Navigating the risk landscape” (nnano.2016.28) introduces risk as a threat to “worth” only after calling the probability-of-harm definition “a useful starting point”. The same years produce his most technical columns: dose arithmetic for fumed silica, life-cycle exposure mapping, benchmarking, carbon-nanotube exposure limits. The value frame and the numbers ran side by side. His rhetoric (“a radical new approach”, “risk-ruts”) sometimes sounds like a break; his substance is extension (“It complements”).

2. His humility about numbers has a technical origin (Andrew’s note 2). - In 2016 he wrote that nanomaterials are described by statistical parameters that “are always a compromise”, so “a statistical parameter of choice may not adequately reflect a risk parameter of relevance”. Treating a new kind of thing as a well-defined chemical “can lead to substantial errors of judgment”. - In 2015 he pointed out that particles have been measurable for a century, and that “The harder challenge is working out what we should be measuring”. - In 2014 he described a well-funded evidence programme hardening into “an assumption of as-yet-to-be-discovered risk”. - In 2016, after 24 years of nanotube research, it was “still hard to predict” what any given sample would do, and IARC could classify only the one form studied most.

The complication is that his response in nanotechnology was more, and better-targeted, science. The faithful reading for AI is “know what matters before measuring, and do not force the new into old metrics”. It is not “numbers are hubris”.

3. Many of the map’s dates are too late. - Behaviour over labels: 2014, not 2022. - Benchmarking against a baseline: 2016, as a core practice. - “Nanomaterials are not just chemicals” (2016): the template for “an algorithm is not a chemical” (2019). - The pacing gap (“years, not hours”): 2015. - Early warnings of systemic “catastrophic failure”: 2015, not 2018–20. - Inevitability paired with steering: 2015. - Sincere entrepreneurs whose optimism the investment process requires: 2015, not 2018–22. - “Responsibility in the face of such audacity”: 2014.

For provenance the most important is this. The idea that risk definitions select which risks count is in his own 2015 prose, a decade before the Fable-assisted orphan-risks paper: the EU nanomaterial definition reflects “what is important and implementable, not necessarily what has the potential to cause harm”.

4. A precursor of the cognitive Trojan horse (my reading). “We have co-evolved with nanoscale materials” (2016) argues that familiar-seeming materials can harm “precisely because those pathways are attuned to characteristics” they share with benign ones. That is the same shape as his 2026 argument about fluent AI getting past evolved epistemic vigilance.

5. Corrections to the map’s absolutes. - Symmetric risk. He counts both sides but does not weigh them equally: blatant disregard for health risks is “worse” than scuppering an industry (2014). - Evidence against early action. This is a balance held since 2014, not a drift from 2019 to 2026: “quick to question, and slow to respond”, yet ready to act on early warnings “even before the science is mature”. - Responsible innovation. He doubted its practicality in March 2015 even while endorsing it. - Engagement. He named the missing mechanisms himself (“we still lack the forums, the methodologies and the leadership”) and proposed “actionable empathy”, a concept the map lacks. - Novelty. In 2014 he called novelty “a rather unreliable indicator of potential risk”, and in 2015 wrote that “seemingly novel challenges don’t always demand novel solutions”. These are his own early counterweights to “defies analogy”.

6. AI. A 2014 column on 3D-printed artificial minds treats machine awareness as moving “from the fanciful to the plausible”. It treats mind as dependent on its physical substrate, which fits his 2024 turn to biological naturalism. The founding risk-innovation column already uses AI safety funding as an example.

Lenses for the next stage (technology-neutral): founders’ optimism as something investment requires; responsibility resting on a responsible firm is exposed to “a less responsible company”; industry evidence judged on its merits; trigger points for action; ask what a technology does in plausible use, not how new it is.