Sarah Constantin
Traction Was Never Evidence

One strange thing: a multibillion-dollar biotech whose best product was a peach smell

In January 2025, Sarah Constantin sat down to conduct a triple autopsy.

The deceased were three synthetic-biology unicorns: Amyris, Zymergen, and Ginkgo Bioworks. Between 2021 and 2023 all three saw their share prices collapse, and two entered bankruptcy. Zymergen had been valued at nearly $5B in 2021.

Reaching Ginkgo, she supplies a pair of numbers: in fifteen years of operation the company reported 219 "programs," of which six were commercialized. Its heavily automated labs claimed the ability to "create 50,000 different genetically modified cells in a single day." And its most successful product line was aroma chemicals sold to the fragrance firm Robertet.

Then comes a parenthesis that contains the whole personality of the piece:

I happen to know that the peach scent gamma-decalactone has been cheaply chemically synthesizable since the turn of the 20th century. (Lactones are responsible for the rich, long-lasting peach scent of Guerlain's Belle Epoque masterpiece Mitsouko.)What's Behind the SynBio Bust?, 30 January 2025

She "happens to know" because, besides this blog, she writes a perfume blog.

Sit with the shape of that. A company built on the premise that biology will manufacture everything burned fifteen years and an astronomical amount of capital, and its proudest commercial result was making a smell — one a chemist could already make more cheaply in 1900. She immediately twists again: Ginkgo's nitrogen-fixing bacterial product for agriculture may also struggle to find customers, because chemical fertilizer is so cheap. "Again, the Haber-Bosch process is a hundred years old!"

The point is not that the company was stupid. The point is that the arithmetic was sitting there from day one — public, elementary, available to anyone — and evidently nobody ran it. Billions of dollars of capital, an entire industry's enthusiasm, and an unknown quantity of due-diligence reports all went around it.

Who she is

Sarah Constantin studied math at Princeton (2010) and took a PhD at Yale in applied harmonic analysis (2015). She then worked a series of jobs that all amount to pointing mathematics at a real industry: medical-literature research at MetaMed, data science at Palantir (2016–2017), drug discovery at Recursion, self-driving trucks at Starsky Robotics, AI-controlled manufacturing at Nanotronics. She founded a longevity research nonprofit (the Longevity Research Institute) and a startup using machine learning to screen for aging-modulating drugs (Daphnia Labs). She lives in Brooklyn with two children, the second born in 2025.

One detail tells you more than the résumé: twice a year she counts the colors on fashion runways. Each spring/summer and fall/winter ready-to-wear season she works through the runway images Vogue posts free — over thirteen thousand of them for SS26 — and produces two counts: one entirely by hand, one by asking GPT-4o to identify the colors in each image and aggregating. Then she spends several hundred words laying out exactly how the manual and automated counts systematically diverge: LLMs lump shades of green under "green" where she splits more finely; the model counts a red belt and red shoes as two instances of red where she counts one; "gray" and "grey" get tallied separately.

It's a quixotic little hobby, but it's mine (and, strangely, I have never seen anybody else do it.)SS26 Color Stats, 14 October 2025

A person who does scientific due diligence on gene therapies, counting runway colors with the same instrument and attaching a methodology section. That is not an eccentricity; it is the same machine pointed somewhere else.

The blog is called Rough Diamonds, subtitled "Underrated opportunities in science, technology, and society." It opened in August 2021 and got serious in March 2022; the 2019–2021 material in the archive was carried over from her earlier blog, Otium. There is one conspicuous silence: in April 2024 she left Nanotronics to consult independently, announced she would write more, and then stopped after May 2024 for roughly thirteen months, returning in June 2025. The blog is still active; the most recent post is dated 19 August 2026.

What she is wrestling with: can someone without inside information find out the truth?

In December 2024 she wrote a post titled "What Goes Without Saying," subtitled "a little rant." It lists premises that strike her as beyond dispute but which she has discovered many people do not share. First on the list:

Concepts like Goodhart's Law, cargo-culting, greenwashing, hype cycles, Sturgeon's Law, even bullshit jobs are all pointing at the basic understanding that it's easier to seem good than to be good, that the world is full of things that merely appear good but aren't really, and that it's important to vigilantly sift out the real from the fake.What Goes Without Saying, 20 December 2024

That is the root problem. But if she stopped there she would be one more cynic with a Substack. What has made these eight years worth something is that her position makes the problem genuinely hard, and she refuses to pretend otherwise.

She has written a first-hand account of how she ran into it. At Palantir she was embedded with a customer. What she found: a data science department of maybe ten people whose entire work product was one number — a risk score — sent to another department, whose head did not trust statistics and threw it away.

The entire work output of those ten data scientists was irrelevant to the company.The Great Data Integration Schlep, 13 September 2024

The footnote gives his reasoning, which reads like a joke and isn't: he did not believe in probabilities between zero and one. "He would say things like 'either it is, or it isn't' and didn't buy it when we tried to explain that a 90% chance and a 10% chance are both uncertain but you should treat them differently."

What this taught her was not "big companies are dumb." Her generalization is both more precise and more fair:

Every company has something fucked-up and dumb going on somewhere, no matter how admirable they are in other respectsThe Great Data Integration Schlep

And the thing is invisible from outside. Most of the people you meet present as "reasonable and highly competent professionals," and everybody has a different story about the Badness and why it is unavoidable or someone else's fault. She also freely admits she is not good at this corporate detective work: she asks questions bluntly, which sometimes gets results and sometimes ruffles feathers, and then she backs off when she gets negative reactions, which is itself often the wrong move in retrospect. Her summary: she has just enough experience to know the thing exists, and to be humbled by how hard it is to see the whole picture of what is going on in an organization.

Which sets the problem in its final form. If people inside can't see it, what business does an outsider have claiming to? She states her disadvantage without hedging — a rando layman, no private data, no sidechannel academic gossip about which studies are fatally flawed, sitting in front of a screen, where it is "not obvious that I can learn a real thing at all."

But she also names her one structural advantage, and this sentence is the reason the blog exists:

Sure, I'll know less about any given field than the experts; but I'm also not pitching my own technology or my own scientific research. I have less conflict of interest.In Search of Opportunity, 25 April 2022

That is the trade: give up domain depth, buy disinterestedness, then claw back some of the depth through exhaustive, unglamorous labor. Everything she has done since is a test of whether the trade pays.

Her ideas, told as she found them

1. Funded, sold, staffed — all three together still are not evidence

In November 2022 she published a post whose subtitle is the entire conclusion: "Stop giving things credence for getting traction."

She opens with a concession that matters, because it keeps the piece from being a gripe: social proof and market validation are not wrong. "If people I trust believe this, it's more likely to be true" and "if people paid money for this, it's more likely to be useful" aren't totally wrong.

Then the real question: how hard is it, actually, for an absolutely terrible idea to accumulate enough market validation and social proof to look plausible on first inspection?

Her method is that of a mathematician. She builds a placeholder and runs the chain forward. She calls it BUZZWORD for APPLICATION. Stipulate that it adds no technical value. Now watch.

The first link is magnitude. The "industry" is a few years old with, say, $500M in revenue this year. Her reaction: "That's nothing." Big companies in the application industry can easily spend that much on a probably-useless technology just for the value of information in case it turns out to be awesome.

The second link is correlated information sources, and it is the crucial one. A handful of top management consulting firms and enterprise software firms are used by basically all Fortune 500 companies. If one of them recommends or partners with a buzzword startup, that startup can honestly claim endorsement or use by many Fortune 500 companies — "without those companies having independently and individually vetted its technology." Meanwhile all the major players send executives to the same conferences, where the same thought leaders explain the Future of the Industry and the importance of integrating the buzzword into your operations.

Here she declines the lazy explanation. Executives are not stupid: they know it's marketing and includes puffery. But they are short of time, unfamiliar with the technical fields, and aware that they will have to adopt new technologies faster than they can personally understand them. Worse, the buzz itself functions as a coordination mechanism — because other stakeholders read the same white papers and attend the same conferences, a technology being buzzword-y makes it more likely they will support adoption too.

The third link is budget mechanics, the hardest-hitting passage in the piece. Large companies allocate budgets — sometimes billion-dollar budgets over several years — toward buzzword technologies. The executives in charge of spending them have strong individual incentives to spend down the whole pot, or they look like failures to their own bosses. Which implies the marginal product purchased need not be useful or cost-effective at all; it may be bought only because it fits the budget category and helps pad out the spending.

Fourth, she supplies the strongest counterargument herself: large companies often do have rigorous vetting. In manufacturing it is the Factory Acceptance Test, taken very seriously, entirely about meeting objective technical specifications. Then she dismantles it: there are also "special isolated magisteria for pilot projects or other experimental things that aren't expected to be really integrated into the main production process." So a startup claiming ACME MEGACORP as a customer does not mean the product is in daily use in ACME's core activities — and therefore does not mean it has been rigorously evaluated at all.

The derivation closes: even if the thing is objectively terrible on the merits, expect some of these startups to raise several VC rounds, land multiple Fortune 500 clients, make the news, and IPO. She adds the timescale: in the long run the market usually does penalize stupid ideas, but "that can take more than a decade to shake out; in many cases, only after a BUZZWORD-for-APPLICATION company has gone public and short sellers notice that customers are no longer buying." Meanwhile the founders are decamillionaires.

Only now does the uncomfortable corollary arrive: you will meet many individuals who are successful and admirable in many ways, and whose entire careers rest on a premise that is 100% false. She is careful to point out that none of this requires anyone to be a villain — the dynamics hold even if every participant is a sincere believer. And you can find knowledgeable, hardworking, honest technical experts inside such an industry; ask them "so, is this even useful?" and sometimes they will have a plausible-sounding rationale, and sometimes they will not even have asked themselves the question, especially where it touches issues outside their field.

common human social proof heuristics like "I talked to a guy and he seemed smart and good and he does this for a living" totally fail to function relative to the vastness of the economy.Bad Ideas Can Get Surprisingly Far, 5 November 2022

What raises this above its cynical cousins is that she ends by turning herself in. For a long time, she says, she failed to integrate this into how she saw the world, because she was holding two bad heuristics.

The first she calls the "laziness heuristic." It runs: people are lazy, nobody likes working very hard; nobody would go to the trouble of starting a company or working long hours if the whole premise were faulty. Wouldn't they get demoralized and quit? Her verdict: "This is, empirically, just not how everyone works." There are a lot of high-energy, career-motivated people willing to do tons of work if they expect to be paid or credentialed well for it, and not very concerned with whether their entire field rests on faulty assumptions or whether the technology makes economic or technical sense. Then the twist: "In fact, maybe that kind of 'tunnel vision' actually makes you better at accumulating legible career accomplishments."

The second she calls an implicit "duty to be humble": I shouldn't judge someone to be wrong unless I can match all their accomplishments, skills, and virtues. It sounds silly once said explicitly, she grants, but we have all heard "it's easy to be a critic" and "don't confuse your Google search with my degree." Disagreement is often received as setting oneself above someone. Her resolution is not counter-arrogance but a separation of two things: the world is vast and being right is really hard, so most people — including her "superiors" — should be expected to be wrong about lots of things; and "disagreement need not imply personal contempt or a claim of personal superiority."

She ends by handing the instrument to the reader, with one question: how easy would it be for a terrible idea to produce the signals of traction that you are seeing? If the answer is "pretty darn easy," then the traction should not reassure you.

2. So run the numbers yourself: the arithmetic of synthetic biology

If traction isn't evidence, what is? The previous idea has to answer that or it is merely a solvent. Her answer: arithmetic — and usually arithmetic embarrassingly close to elementary.

Back to the three corpses. She dissects them not from governance or market timing but from one physical fact that determines everything downstream.

Biomanufacturing runs on living things. Living things cost money to keep alive.

An artificial synthesis only needs enough materials and energy to run one chemical reaction; a biological route involves all the "overhead" of keeping the organism alive!What's Behind the SynBio Bust?

And then you still have to extract the compound from the organism, which she notes "can itself be a messy business." So although microbial syntheses can be locally very efficient — billions of years of evolution designing the enzymes — traditionally it is almost impossible for a biomanufacturing process to compete on price with an artificial synthesis of the same compound. She uses this in passing to explain something else: why biofuel companies have such a shaky track record. It is simply hard for any biofuel to beat fossil fuels or solar on cost per joule.

From which a hard boundary follows: in practice, biomanufacturing is only used for compounds complex (and expensive) enough that they can't be made thermochemically at all. Proteins and peptides — insulin, growth hormone, antibodies, the viral vectors in vaccines and gene therapies — qualify, because complex molecules break down at high heat and need finesse and mild temperatures, whereas simple molecules can be synthesized faster by cranking up heat and pressure, which fossil fuels are excellent at providing.

So a biomanufacturing startup's sweet spot ought to be novel, high-value, low-volume complex biologics. "Unfortunately, as we'll see…they kinda didn't do this."

Three case files:

Amyris — "(Energy) Math is Hard, Let's Go Shopping!" (her heading). Founded in 2002 around a genuine scientific advance: synthesizing the antimalarial artemisinin from genetically engineered yeast. But malaria is a disease of poor countries and there was never serious money in it. In 2006, on investors' advice, it pivoted to biofuel — fermenting Brazilian sugarcane into "green" diesel. The diesel tested well. Then reality: yields were poor, the yeast kept dying, bacterial contamination recurred, and the fuel industry's standard production quotas run to 40–50 million liters. How bad were the costs? In its first large contract, supplying the Rio de Janeiro and São Paulo bus systems, Amyris was selling the fuel at a loss even with Brazil's hefty public subsidy. In 2012 it pivoted again, to personal care and beauty ingredients — smaller volumes, higher price per liter. But those were still relatively simple molecules with alternative production routes, so it had no real moat. It generated substantial sales anyway, launched its own skincare line in 2016, went on an acquisition spree of celebrity beauty companies chasing revenue, and reached $153M in revenue in 2019 — without ever reaching profitability. Bankruptcy filing in 2023. Her verdict is one line: "Biomanufacturing math is unforgiving."

Zymergen — yeast ain't everything. Valued near $5B in 2021, share price collapsed the same year, bankrupt in 2023. On the surface it made real things too: high-throughput lab automation, machine-learning methods for genetically optimizing microbes. But she points out all of that was concentrated on the first part of the process — generating a yeast strain that produces the chemical of interest and performs well in the lab. It was a strain engineering company, not a manufacturing company: little in-house capacity, production outsourced to contract manufacturers who struggled to meet goals. Its first product, Hyaline, a film for foldable phone screens, was a commercial failure, and there was no other revenue.

The line she extracts from the 2021 prospectus is lethal: Hyaline was not even yet being produced by fermentation, but was "a non-fermentation produced biomolecule sourced from a third party" which they planned to "convert to a fermentation-produced molecule" in 2022. The same document warned of limited access to contract manufacturing capacity: "we may not have the manufacturing capacity required to meet our commercial needs after the end of this year."

Her response is short enough to sound like a hand hitting a table: "This is not how you develop a manufacturing process. Anywhere." And the explanation that follows is the lesson an outsider most needs. Process development and manufacturing have to work closely together with tight feedback loops. Manufacturing departments are archetypally conservative; their job is to execute a given procedure carefully and not screw it up. Tinkering and innovation are process dev's job. If your whole company is process dev and your "manufacturing department" is an external contractor, the situation is exactly analogous: you need to be onsite doing continual testing and adjustment. Your vendors are your coworkers. And, she stresses, none of this is arcane — "It's repeated ad infinitum in Deming's Out of the Crisis, which has defined how manufacturing works since the 1980s."

Ginkgo Bioworks — "You May Have To, Like, Actually Sell Things." Technically still in business, but its share price fell almost to nothing in 2021 and has not recovered, following a short seller's report revealing that virtually all apparent revenue came from the accounting shenanigans of "selling" to internally funded and operated "customers." Fifteen years, 219 programs, six commercialized. Surely all that automation bought it a speed or price advantage? No: customers complained of high costs and delayed deliveries, and internal documents revealed policies of padding invoices. The result was that even some of the "customers" that were Ginkgo-owned sockpuppets were unhappy and switched to competitors.

And the most successful line was the peach smell from the opening.

What shows her character is the ending. She does not let this become a hit piece:

None of these criticisms needs to reflect badly on the vast majority of current or past employees of these three companies. By all accounts, they hired serious scientists and engineers who did a great deal of quality work.What's Behind the SynBio Bust?

"The problems were at the strategic and management level" — which is to say, what failed was not the science. What failed is that nobody ran the arithmetic. Her two conclusions are almost boringly plain, and that is the point: only biomanufacture products that can't be made more cheaply some other way; and don't specialize exclusively in strain engineering when the products you're developing will necessarily require manufacturing process development. She closes with a lovely control case: of the three, Amyris did best on manufacturing performance and revenue — and Amyris reports having had the most difficulty achieving satisfactory production results precisely when it was working with a contract manufacturer that didn't allow the participation of Amyris staff. "It really pays to be hands-on."

3. Exponential progress is not a law of technology; it is one industry's shadow

A solvent, however sharp, can only tell you what doesn't count as evidence. In September 2023 she offered her own positive criterion, and its shape is beautiful.

She starts from an observation everyone has seen: some things are getting exponentially cheaper. Compute is the cliché; the cost of sequencing a human genome has also fallen exponentially; so has the cost of solar power. So: is this some kind of universal law of technology development? Is exponential improvement just how technology evolves, in the Kurzweilian sense? "Or are these just three unrelated, hand-picked examples, which likely have different explanations?"

Her answer is a third option, and it is startlingly blunt:

Every tech that is improving exponentially in the 21st century is explained by improvements in semiconductor fabrication.The Enchippening, 13 September 2023

To make you believe it she does something characteristic: she refuses to explain how chips work and insists on how they are made. Look up "how computers are made," she notes, and you get semiconductor physics or AND/OR/NOT gates — "but that only tells you the principles of how a computer works. It doesn't tell you how a computer is made."

And how they are made compresses to one sentence: flat layers of stuff on top of other flat layers of stuff. A transistor is microscopically thin flat layers — semiconductors, insulators, metals — deposited or grown in a pattern on a flat wafer; wiring transistors into circuits is just more of the same, flat rectangles on flat rectangles. And all those layers come from two mechanisms: deposition (vaporize the material and let it condense or chemically bond onto the surface; if you need a crystal, grow a perfect molecular lattice layer by layer, which is epitaxy) and etching (coat with photoresist, project UV in precise patterns so the resist hardens where exposed, then etch the exposed areas with corrosive chemicals or plasma ions).

fundamentally, Moore's Law is about semiconductor manufacturers getting better at making smaller rectangles of flat stuff.The Enchippening

Once you see it as a manufacturing capability rather than a computational miracle, the reach opens up. An LED is one layer of p-type and one of n-type semiconductor. A laser diode is a p-n junction plus a waveguide. A power inverter is two transistors and two capacitors. Capacitors and resistors themselves can be made from flat layers of metals and insulators using the same processes. That, she notes, is the entire point of an integrated circuit: everything that can be reduced to layers of flat rectangles and manufactured on the same equipment, should be.

Then the prettiest inference chain — the life sciences. A gene sequencer chops DNA into a bajillion microscopic wells containing materials that fluoresce in the presence of A, T, C, or G; a photodetector measures when a well lights up. Photodetectors are semiconductor devices, fabricatable on a chip. And the wells themselves — the flow cell — are a microfluidic device: patterned silicon, glass, or plastic. And microfluidics needs feature sizes measured in microns, not nanometers, so like solar cells it can piggyback off older and cheaper semiconductor technology.

Most of the experimental methods that are transforming today's life sciences are downstream of the manufacturing techniques of the semiconductor industry.The Enchippening

mRNA vaccines are possible because genetic sequencing is ubiquitous. She doesn't wave away the biology, either — for each new microfluidic sensing device, someone has to find a way to sense the quantity of interest in a microscopic sample with a microscopic sensor, and ensure the materials and the biological sample survive contact with each other. But the field has an enormous tailwind: "Once you know what you want to make, the tools for making it at scale generally already exist."

The payoff is an operational, positive rule — note how exactly it complements "traction isn't evidence": "you should be bullish on any useful tech that is newly fabricatable on a chip." She even supplies the mechanism of the cost decline rather than asking you to trust a trend line: very new, experimental semiconductor processes usually start on small wafers; as demand grows and the process becomes reliable, the industry moves to larger wafers; larger wafers mean cheaper devices, because you get more devices for the same number of deposition and etch steps, and thus roughly the same machine and operator time.

Read in reverse, the rule is the verdict on those three synbio companies: if your thing is not downstream of that manufacturing base, do not expect it to get exponentially cheaper merely because it is new and exciting.

4. The bottleneck is never where the excitement is

In September 2024 she wrote the most-shared post on the site. It opens like this:

This is a little rant I like to give, because it's something I learned on the job that I've never seen written up explicitly.The Great Data Integration Schlep, 13 September 2024

She begins with a question nobody bothers to ask: "digital transformation," "the Fourth Industrial Revolution," "Industrial Internet of Things" — what do those actually mean? Do they mean anything at all?

Her answer: yes, and only one thing — the process of putting all of a company's data on computers so it can be analyzed. This is the prerequisite to any "AI" or even basic statistical analysis. Before you apply your fancy algorithms, you need the data in one place, in a tabular format.

Then the subheading that stops you: "Wait, They Haven't Done That Yet?"

No. Her inventory is concrete. Some data is not digitized at all — literally on paper: lab notebooks, QA reports, work orders. Some is "barely digitized," meaning scanned PDFs of those papers: fine for records, impossible to search or analyze. Her example is unforgettable: a major aerospace manufacturer kept all the results of airplane quality tests as scanned handwritten PDFs of filled-out forms. Her comment is an exclamation: "Imagine trying to compile trends in quality performance!" Still other data is siloed inside machines on the factory floor — modern automated machinery generates plenty of sensor measurements, actuator logs, and process-setting changes, and that data is stored in that machine, and only that machine. She notes that process engineers have used data to run factories for nearly a century under statistical process control, but in practice far more data is generated and collected than is used; only a few process variables get tracked and optimized, and the rest is "data exhaust," to be ignored and maybe deleted. In principle the excess may be relevant to the facility's performance, but nobody knows how, and they're not equipped to find out.

So when somebody proposes "using AI" to optimize operations, manufacturing companies are often skeptical. To use AI you need to build a model around a big dataset. And they don't have that dataset.

She breaks the job into four steps — obtain, digitize, clean, build infrastructure — and shows that each is hard for human reasons.

Obtaining first. "Obtaining the data is a hard human problem. That is, people don't want to give it to you." Hence the line that circulates in the industry: it is not at all unusual for it to be easier to make a multi-million dollar sale than to get the data access necessary to deliver the finished tool. Part of this is security: the semiconductor industry's paranoia about industrial espionage is entirely justified, so they are not putting factory data on the cloud, may have fully airgapped facilities, and do not want images of in-progress chips or process details reaching "the wrong hands." And the bigger the company, the likelier you are to meet multiple IT policies from different departments, and the likelier that some of them contradict each other or are fundamentally incompatible with any large-scale data integration. Her extreme case: a company that required even prospective vendors to spend over $1M of their own money building a special ultra-secure room to store data in — before the sale was even closed.

The rest of the reasons are more political and more human: people worried a big-data tool will replace their jobs or make their performance look bad, and obstructing; inter-departmental rivalries; and people who are simply busy, for whom carving time out of the workday to do setup for a vendor is an inconvenience.

This is where she gives her most interesting account of Palantir — and the key is that she thinks its distinction is not the software. The databases, ETL pipelines and analytics dashboards "work fine, but they are not, as far as I know, unique or exceptional in the tech industry." What is remarkable is this:

they have invested in a large number of people — at least a third of the company by headcount — to descend en masse to the customer's facilities, set up and customize their database and associated tools, teach the users to work the UI, and, crucially, negotiate for data access.The Great Data Integration Schlep

She reconstructs the playbook in four moves: only sell to companies facing an existential threat, where something has gone so wrong they might go out of business; get buy-in from the C-suite, who care about company-wide objectives and whose reputations gain from leading a transformation; win the hearts of front-line workers by being genuinely helpful, making their jobs easier, and building personal connections on site; then use your allies at the bottom and the top to squeeze out your opponents in the middle — managers, often in IT or data science, who would rather not have an external vendor on their turf. The higher-ups can demand you be given speedy assistance and data access; the front-line can pressure their bosses to get the cool new tools running.

Then the most perceptive sentence in the piece: "The Palantir Way is labor-intensive and virtually impossible to systematize, let alone automate away. This is why there aren't a hundred Palantirs." You have to throw humans at the persuasion problem — well-paid, cognitively flexible, emotionally intelligent humans who can cope with corporate dysfunction. Which makes them a lot like management consultants, and in fact, she says, data integration at large scale is inherently a little like management consulting.

The second obstacle is cleaning, and here she gives what I think is the deepest example in the post. Every working data scientist will tell you they spend more time on data cleaning than on running models. What does cleaning mean? Removing commas, formatting, normalization, imputing missing data consistently — and understanding what your data sources actually refer to, so that your judgment calls make practical sense.

Her example: two columns are identical. Are they literal duplicates of the same sensor output, so one can safely be deleted — or are they two different sensors that happened to give the same readings this run because the setting that would let them differ was switched off?

The answer can't be derived from the dataset, because the question pertains to the physical machine the data refers to; the ambiguity is inherently impossible to automate away using software alone.The Great Data Integration Schlep

That is an epistemological point, not an engineering complaint: the information you need is not in the data. It is on the machine, or in the head of someone who knows about that setting.

On whether large models could take this over she gives an honest 2024 observation: she expected LLMs to make substantial if not total improvements in automating data cleaning, but "my preliminary experiments with commercial LLMs (like ChatGPT & Claude) have generally been disappointing; it takes me longer to ask the LLM repeatedly to edit my file to the appropriate format than to just use regular expressions or other scripting methods myself." And she immediately leaves herself open: she may be missing something simple about prompting, or LLMs may need more surrounding software scaffolding or specialized fine-tuning before they dent the task.

Her general observation is worth more than any single example: some kinds of work are harder than they look — tedious and seemingly routine, but often requiring judgment calls that are surprisingly hard to outsource to a junior staff member, let alone automate. Data cleaning is one. So are data labeling, data entry, and summarizing key facts and numbers out of large piles of documents. Apart from the basic programming, none of it requires special education; it is "just" reading, writing, arithmetic and common sense — "and yet, somehow, not just anyone can do it right."

Which delivers the final inference, under the heading "AI is Gated On Data Integration." Her argument: for AI to make really serious economic impact, once the low-hanging fruit of public Internet data is exploited, it has to start learning from business data and substantially improving the productivity of large companies. If you are imagining an "AI R&D researcher" inventing technologies, that means integrating into corporate R&D — mainly big manufacturing firms with heavy science and engineering investment: semiconductors, pharmaceuticals, medical devices and instruments, petrochemicals, automotive, aerospace. You would need enough access to private R&D data to train on, and enough credibility built through pilots to gradually win free rein, and you'd need to start virtually from scratch with each new client. "This takes time, trial-and-error, gradual demonstration of capabilities, and lots and lots of high-paid labor, and it is barely being done yet at all."

She refuses very explicitly to be read as an AI bear: "I'm not saying 'AI is overrated', at all." All of this work can be done and can ultimately be extremely high ROI. Her analogy is the arrival of computers in business decades ago: there was nothing fake about that transformation; it happened, it mattered, many fortunes were made — but it happened at the speed of human negotiation and learning. Executives had to decide to buy computers, and many individuals inside each company had to work out what to do with the new machines, and that had to happen independently for each use case. It wasn't instantaneous, and it didn't hit everywhere at the same speed; some applications were still doing things manually long after others were fully digitized.

"So too with AI."

How she thinks

The best entrance to her method is the 2019 post, because there she points the instrument at herself.

GPT-2 had just been released and she read the samples. Her first reaction was that they were "disturbingly good": if you are skimming, reading for writing style and evocative words, they flow. The unicorn sample reads like a real science press release; the Miley Cyrus sample like a real gossip post; the JFK sample like a real politician's speech.

"But if I read with focus, I notice that they don't make a lot of logical sense." Then she picks them apart, and it is funny. The unicorns are named "Ovid's Unicorn" after their distinctive horn — but "Ovid" doesn't refer to a horn, so why would that be naming them after one? And you just said they had one horn, so why four in the next sentence? From the Tolkien sample: "As they marched out through the thicket the morning mist cleared, and the day turned to dusk." Her annotation is one line: "Yeah, day doesn't turn to dusk in the morning." The resurrected-JFK sample is worse: brain tissue harvested from his chest? A human brain about one and a half cubic centimeters?

The point, crucially, is not about GPT-2 but about readers:

The point is, if you skim text, you miss obvious absurdities. The point is OpenAI HAS achieved the ability to pass the Turing test against humans on autopilot.Humans Who Are Not Concentrating Are Not General Intelligences, 25 February 2019

What makes this the foundation stone of her method is the self-observation that follows. She noticed how often her own brain slides from focused attention into skimming: she read the sample on Spanish history with interest, and the generated text was obviously absurd; her eyes glazed over during the sample about video games, since she doesn't care about video games, and the machine text looked totally unobjectionable to her. Her comment: the brain is constantly evaluating what's worth focusing on and what's fine to tune out — so GPT-2 is actually useful as an instrument for measuring your own level of attention.

From which comes the rule I think is the single most portable thing she has written. The mental motion of "I didn't really parse that paragraph, but sure, whatever, I'll take the author's word for it" is, in her introspective experience, absolutely identical to "I didn't really parse that paragraph because it was bot-generated and didn't make any sense so I couldn't possibly have parsed it" — except that in the first case she assumes the error lies with her rather than the text. And "this is not a safe assumption in a post-GPT2 world."

So: instead of defaulting to humility (assume the passage is true and you're missing something), default to null — if you don't understand a passage, assume you're in the same epistemic state as if you'd never read it at all.

That rule is the seed of her whole working method. The rest can be read as versions of it at different scales.

Completionism combined with baseline skepticism. Her phrase, and note the humility of the framing — she calls it "one of my few available tactics," something grown out of a disadvantage rather than a clever trick. In practice it is a heuristic: "try to find a literature that's reasonably small and then go through all of it." Her example carries a rare note of pride: it is satisfying that she knows of literally every experiment where anyone has done something to extend lifespan in a mammal.

Alongside it sits a decision that says a great deal about her. She once tried to be more "professional" and use PubMed searches, the way systematic reviews in journals do. Her verdict: "This was a terrible idea." PubMed does not find nearly all the papers on a topic and includes tons of extraneous, utterly unrelated ones. Then the line: "I don't know why PubMed is the standard for academics' lit reviews, but not being an academic I am free to use a search engine that works." She uses Google Scholar, and pages through it often enough to get accused of being a bot.

Her heuristics for finding opportunities appear under a heading that is itself a statement: "Hulk Prefer Big Simple Thing." The entries beneath are as terse as passwords. Measurement and data collection is good. Actual experiments almost always beat simulations or theories. Look for big effect sizes on outcomes that matter. Complicated statistical models are easy to game. Proxy outcomes are suspect. Simple interventions are best; fewer moving parts = fewer things to screw up. Avoid anything that looks like make-work for bureaucrats. And a cluster she labels "indications of sincerity": accidental discoveries; forgotten discoveries from before 1990; the unglamorous and little-hyped; use in competitive industries with fast feedback loops; a researcher personally affected by the problem they're solving, like treating one's own disease. The last entry is the most her: "if you're bored, you have the wrong framing question."

The word that ties all of this together finally surfaces in August 2026, when she writes about evaluating grant applications: unfakeable. She has a bias toward projects that are "technical," which she defines as "demonstrating unfakeable, rare skill at structured processes." She likes donations to community infrastructure partly because it is concrete, and thus harder to fake — if you're operating a physical space or running an event, people can come and evaluate it directly.

It forces contact with reality in a way that doesn't happen by default with theoretical research and writing. I place a lot of value on "doing a real thing" compared to doing a potentially-bigger but potentially-unreal thing.Some Ways I Think About Evaluating Grant Applications, 13 August 2026

In the same piece she says she almost always likes projects that simply compile public information to answer a straightforward question ("How many people die from every disease, worldwide?"), for exactly the same reason: "It's easy to be confident these projects aren't 'fake'; they are what they say on the tin."

The last layer is prose itself, and here her view is unexpectedly un-engineerish. She says you have to write for logic and vibe at once. Logic: is what I'm saying coherent and true, are my arguments valid. Vibe: are the associations and emotions my words will evoke the ones I intend? A common failure among logical thinkers, she says, is writing that works logically and fails on vibe. Her rule is concrete enough to use immediately: never describe something as "not sugary" or "not fraud." Avoid "not X" where X is the most vivid, high-emotional-intensity word in the sentence, because attention lands on the vivid word and much less on the neutral "not" — readers come away with an implicit feeling that the thing maybe is sugary, or is fraud. She offers a personal case: years ago a friend sent her a piece by an academic historian on the Enlightenment origins of science that described science as "Luciferian." She was put out and asked why he thought she'd enjoy a piece condemning science as devilish — and it turned out she had misread; the author meant it neutrally or positively, as transgression of the established order in pursuit of human power. Her conclusion: don't play that kind of clever game with lay readers. "If you think something is good, you should make it sound good — and not compare it to Lucifer."

Where she might be wrong

The strongest testimony about this instrument's systematic bias comes from her, in public, in a post written for the purpose.

In April 2023 she published a long piece, "Why I am Not An AI Doomer," with a specific position: of the four links in Yudkowsky's argument she accepted the first three and rejected the fourth — that is, she did not think current machine learning was on track toward the dangerous kind of AGI. Her reasoning was genuinely original. The dangerous property is not intelligence but agency, and for agency she took William James's definition: "fixed aim, varying means." A frog trying to escape a jar will try various strategies; iron filings "attracted" to a magnet will not devise ways around an obstacle. Her claim was that a model pursues a function in its own mind while an organism pursues an outcome in the world; reality corrects organisms ("If a robot's model says it's doing fine, but then an anvil falls on it, then the robot is not fine, it is smashed") and does not correct a loss function. Further, she argued that the very same technical obstacle to aligning an AGI is an obstacle to creating one.

In May 2024 she published "Rethinking AI."

But I have to admit that I no longer stand by that post as it's written.Rethinking AI, 24 May 2024

She kept part of it: she still thinks mastery of the physical world must flow through interaction with the physical world, and is still willing to bet no drug will enter clinical trials without in vitro testing first. But she gave up the trunk: "I think the way to bet is that the AI/tech industry will, sooner-or-later, create bona fide 'minds', fully intelligent in every sense that we are."

More valuable is her diagnosis of why she was wrong, because it lands precisely on the blind spot of the toolkit. She had reasoned that nobody would build AI-trained industrial robots: simulation training has a track record of not working well, real-world training is expensive, and an AI robot carrying GPUs for inference is not price-competitive with a traditional hard-coded industrial robot — and even traditional industrial robots are rare, with electronics and automotive basically the only manufacturing fields using significant numbers of them. Every link in that chain is exactly the kind of arithmetic she is best at. Then Figure AI raised $675M from investors including Bezos, Nvidia, Microsoft and OpenAI; Skild reportedly raised close to $300M; 1X raised over $125M; and Fei-Fei Li started a spatial-reasoning company.

"I feel a little silly about this," she writes. "As it turns out, they absolutely will build the AI robots and they are willing to pay what it takes to train them."

And then the diagnosis: what her former worldview was completely missing is selection effects — taking markets and incentives seriously. She quotes her own old model in full so you can inspect it: "Just because something is possible in principle, and would be really valuable, doesn't mean it'll get done; if it did, we'd have widespread nuclear power…" The replacement: "'It's too hard' or 'it's too expensive' or 'it's too weird' are not permanent blockers, absent prohibition by force, to a sufficiently lucrative thing happening."

That is the systematic tilt, and it is not a one-off slip. A gauge built to ask "does the arithmetic close today, and is anybody faking it?" is uncannily accurate on questions like the synbio bust, and will read persistently and predictably low on questions where enough money will eventually be thrown to change the arithmetic. Her negative judgments are far more reliable than her negative predictions. The distinction is worth keeping in hand as you read her: when she says "this does not work today," listen closely; when she says "this will not happen," discount.

There is another blind spot, which she also found and wrote down, and with real unease. In "Big-World Intuitions" (July 2026) she inventories the heuristics she trusts — don't worry about competitors, just serve your customers; as a small trader, bid your true price; early in the game, develop your resources; in repeated interactions, be consistently trustworthy; working on a hard problem far from solved, you don't have to worry about "what if we succeed and it's bad?" What they share, she notes, is that the world is big enough relative to you that your best strategy is not to explicitly calculate how your action affects the environment. Then she admits she pretty much only ever uses this class of intuition: she doesn't know what to do with questions like "what if I, personally, have the power to shape society," and doesn't know where to begin with adversarial chains of "if I do this, she'll do that."

it is disquieting to notice that none of my grounding intuitions or usual ways of thinking about what "healthy" or "prudent" or "ethical" look like, are fit for such situations.Big-World Intuitions, 30 July 2026

Set that beside her subject matter and it acquires weight: she has spent eight years assessing technologies that might be endgame technologies, and she has told you in advance that if one of them ever is, she has no tools for it.

A third item is not an error in argument but the cost of the method to its owner, and it may be the most candid thing she has written. In April 2024 she left her job to consult and announced she would write more. What followed was nearly thirteen months of near-silence. On returning in February 2026:

It feels like "writer's block" or like I "can't think of anything to say", but obviously that's suspect, and the real thing is that I can't think of anything to say that's impeccable and beyond reproach and definitely won't get criticized.Getting Back To It, 26 February 2026

The sentence just before is shorter: "The fact is, I've been hiding." A method whose entire purpose is not being fooled can, at some point, swing around and become a fear of being caught out. She also identifies the loop's self-reinforcement: the less she participates in public life, the fewer discussions she's part of, and the fewer opportunities she has to riff off what other people are saying.

The last item is my own reading, and she has not said it. "Traction isn't evidence" is a solvent, not a ruler. It can tell you what doesn't count; it can't tell you what does. Her positive criteria — the enchippening, techno-economic analysis — operate only in domains where a clean number exists: cost per liter, cost per base pair, feature size, production quota. And she writes a great deal about domains with no such number: institutions, regulation, organizations, education, philosophy, ethics. There the solvent runs unchecked.

To her credit, she knows. In "How I Write" she says her non-research posts "stray more into the realm of opinion and I wind up feeling disgusted with my not-quite-right lens," especially on general-interest, non-STEM, emotionally fraught topics, "where I worry that I'm not saying objectively true things so much as airing my personal biases." She notes those posts are more likely to be duds that linger as unpublished drafts. Her own verdict on herself deserves more weight than any I could offer.

(One bibliographic note: her 2023 "Why I am Not An AI Doomer" drew a public discussion thread on LessWrong, where named commenters pushed back on the agency step. I have not read those arguments in the original and so do not summarize them here. Her own correction a year later is the weightiest criticism I did read.)

Sources

• Bad Ideas Can Get Surprisingly FarNovember 2022. A fictional "BUZZWORD for APPLICATION" industry, run forward step by step to show how a technically worthless idea collects VC rounds, Fortune 500 clients and an IPO. Plus the two bad heuristics she confesses to.
• What's Behind the SynBio Bust?January 2025. Autopsies of Amyris, Zymergen and Ginkgo. The cost arithmetic of biomanufacturing, the lethal sentence in Zymergen's prospectus, and the peach scent that has been cheap to synthesize for a century.
• The EnchippeningSeptember 2023. Her most original positive criterion: everything improving exponentially in this century is downstream of semiconductor fabrication. From deposition and etching to the microfluidics inside a gene sequencer.
• The Great Data Integration SchlepSeptember 2024, the most-shared post on the site. AI is gated on data integration; a third of Palantir negotiating for data access; the department head who did not believe in probabilities between zero and one.
• Humans Who Are Not Concentrating Are Not General IntelligencesFebruary 2019, among the earliest in the archive. A general rule extracted from reading GPT-2 samples: default to null — if you don't understand a passage, assume you never read it.
• Rethinking AIMay 2024. Her public retraction of the previous year's "Why I am Not An AI Doomer," and her diagnosis of why she was wrong: "The thing my former worldview was completely missing is selection effects."

Updates