In the Pipeline · started January 2002; by the date this page was checked it has run for twenty-four and a half years, "continuously" by the author's own account, and well past five thousand posts · the blog's year filter spans 2002–2025 and the page reports "Last Year 199" posts, i.e. still close to daily · still publishing; most recent post as checked was 22 September 2026 · entirely free, no paywall · two visible breaks: a near-silent stretch through most of 2003, and the move from Corante to Science/AAAS in August 2015 (followed by a platform migration to science.org in August 2021, with thirteen years of archives carried over intact) · this page is built on 20 posts spanning 1 February 2002 to 24 July 2026, including 7 from the earliest period (2002–2005) and at least 5 on each side of the 2015 move · the site has no author-curated "best of" page, so selection follows the author's own standing categories (Things I Won't Work With, How Not to Do It, Birth of an Idea, AI and Machine Learning) plus outside discussion volume
In August 2015 a man moved a blog he had been writing for thirteen years to a new address. In the introductory post, this is how he described his career:
I have never once worked on any compound that has made it all the way to the market, not in 26 years and counting. If you want to know how come your prescriptions cost so much, it's me.Welcome to "In the Pipeline", 2015-08-03
His name is Derek Lowe. He is a medicinal chemist — nearly thirty years across Schering-Plough, Bayer and Vertex, and since 2017 a director in chemical biology at Novartis. His blog In the Pipeline has run from January 2002 to today: twenty-four and a half years, five days a week, well past five thousand posts, one of the longest-lived science blogs on the internet.
And his central argument is hiding inside that self-deprecating line. It is not modesty. It is a balance sheet.
In July 2026 he read a paper pulling together every estimate of clinical failure rates since the 1960s, and opened with the sentence he has written more often than any other:
I'm fond of saying that the most important single statistic about the drug industry is the clinical failure rate, which is (by any reasonable standard) appallingly high.Clinical Failure Rates Over the Decades: Yikes, 2026-07-24
The numbers: 90% in the 1960s, 81% in the 70s, 80% in the 80s, 88% in the 90s, 91% in the 2000s, 91% in the 2010s. Sixty years, essentially flat. His comparison:
No cutting-edge restaurant could survive if 91% of its innovative dishes proved inedible or outright poisonous. What other industries operate under these bizarre conditions?ibid.
What Lowe has done for twenty-four years is insist on treating that number as an environmental constant of the industry, not as a management problem waiting to be fixed. Every few years someone arrives from outside announcing that they are here to fix it. His job is to stand in the doorway and explain, one at a time, that they have misunderstood what they are up against.
In 2007 the arrival was Intel's Andy Grove, who said publicly that pharma could use some semiconductor urgency — chips had transformed the world, while in Parkinson's "we have gone from levodopa to levodopa." Lowe's answer was fast and hard. He first conceded that everything Grove said was true, then pointed out the thing Grove had not considered: we built the chips.
Ever seen one of those huge blow-ups of a chip's architecture? … And guess what - it's nothing, it's the instructions on the back of a shampoo bottle compared to the complexity of a living system. … Mr. Grove, you can print out the technical specs for your chips. We don't have them for cells.Andy Grove: Rich, Famous, Smart and Wrong, 2007-11-06
Fourteen years later he turned the same argument on AI drug discovery. This time he had a simulation in hand: cut the cost of the preclinical stages by 20%, and the savings vanish next to what you would save by cutting the clinical failure rate by 20%.
All that other stuff is a roundoff error compared to the clinical failure rate.AI and Drug Discovery: Attacking the Right Problems, 2021-03-19
And there are only two roads to a lower clinical failure rate: pick better targets, and know in advance whether the thing is toxic in humans. Those happen to be where machine learning has made the least headway — because everything on the pipeline that generates data in bulk (binding constants, cell activity, animal models) is a proxy. All necessary. None of them a substitute for the one measurement that counts.
But in the end, they're all still just proxies. … if you don't hold your breath when the first human doses start, then you haven't been doing this stuff long enough.ibid.
In 2023 he compressed it into a claim time can settle: there are no existing AI/ML systems that reduce clinical failure risk from target choice or toxicology. By 2026, evaluating a system that claimed to generate biomedical hypotheses on its own, he was still using the same ruler — not benchmark scores, but the literature. The MEK-pathway idea the system "found" for acute myeloid leukemia had been worked on for years; cerivastatin's unusual ability to kill AML cells had been described twenty-five years earlier. None of that work was cited in the paper.
We definitely need all the help with the literature that we can get, but hype we have enough of already.Evaluating "Co-Scientist", a New AI Science System, 2026-05-27
That is his fixed move: never ask which side you are on, ask what the actual number is and what that citation actually says. When the Alzheimer's image-fraud scandal broke in 2022, his first sentence was not outrage but a correction — to people saying the fraud had wasted the field's time since 2006, he replied that the inaccuracy in that is that we have been wasting our time for longer than that. When he took apart a million-view listicle about "eight foods banned in other countries" in 2013, he did not argue the opposite case; he clicked through the article's own citations and found they did not support it. (The links were quietly removed afterwards.) Out of that wreckage he pulled two principles:
There is no essence, no vital spirit. A compound is what it is, no matter where it came from. … Paracelsus was right: the dose makes the poison.Eight Toxic Foods: A Little Chemical Education, 2013-06-21
Incidentally, the thing he is most famous for on the internet is a column called Things I Won't Work With. It covers chlorine trifluoride, one ton of which leaked in the early 1950s and burned through a foot of concrete before eating a metre of sand and gravel below it; and dioxygen difluoride, which sets solid ethanol on fire at 90 kelvin. He calls it Satan's kimchi, and tips his asbestos-lined titanium hat to the 1962 author who reacted it with everything he could find.
It reads like a joke, but it is the other face of the food-chemistry piece. The abstract noun "chemicals" manufactures both fear and contempt; actual substances have actual tempers, actual doses, actual ways of killing you. Twenty-four years in, that is what he has been teaching.
On 3 August 2015 a thirteen-year-old blog changed addresses. The author wrote an introductory post explaining what he did for a living. Those are usually written as a CV. He wrote this:
I have never once worked on any compound that has made it all the way to the market, not in 26 years and counting. If you want to know how come your prescriptions cost so much, it's me.Welcome to "In the Pipeline", 2015-08-03
In the same post, his case for why the blog should exist at all:
I've been a researcher in the drug industry since 1989, at several different companies. Over that time, I've worked on projects directed against schizophrenia, Alzheimer's, diabetes, osteoporosis, many types of cancer, resistant gram-negative bacteria, and several nasty viruses (among others). You may notice that we still have all these diseases with us, and that might be the best single reason for this blog's existence.ibid.
He also offered a definition of his own thirteen-year archive: "thirteen years worth of being stuck with me in an elevator."
None of this is modesty. It is a balance sheet, and it is the thesis of everything he writes. A man who has never produced a drug has spent twenty-four years writing about drug discovery and become the most-read person in the field — that is not an irony. It is the evidence.
Derek Lowe, born in Harrisburg, Arkansas. BA from Hendrix College, PhD in organic chemistry from Duke (total synthesis of natural products), then a postdoctoral year in Germany on a Humboldt Fellowship. Into industry in 1989: eight years at Schering-Plough on central nervous system drugs, nine at Bayer on metabolic disease, ten at Vertex on chemical biology and new technologies, and since 2017 a director in chemical biology at Novartis. He also writes a column for the Royal Society of Chemistry's Chemistry World and is the author of The Chemistry Book.
He started In the Pipeline in January 2002 — with prior approval from his supervisor and his company's legal department. Per Wikipedia, he was among the first people to blog publicly from inside the pharmaceutical industry.
One detail says more than any biography about how close he stands to all this. In May 2026, assessing an AI system's drug-repurposing picks, he noted that one of them was cerivastatin, and added a parenthesis: its withdrawal from the market "ultimately led to the closure of the entire research site I used to work at!" He did not elaborate. He went back to the pharmacology.
The blog itself has moved twice: in August 2015 from Corante to Science/AAAS, and in August 2021 along with the whole Science site to science.org, carrying thirteen years of posts and comments across without losing one. It is still being updated at close to a daily cadence (the site reports "Last Year 199"); the most recent post when this page was checked was 22 September 2026. Along the way he has been "threatened with lawsuits once or twice; things should work out fine."
One number: nine in ten.
In July 2026 he read a paper that gathered every estimate of clinical failure rates since the 1960s and graded each source for scope and reliability. The opening sentence is the one he has written more often than any other:
I'm fond of saying that the most important single statistic about the drug industry is the clinical failure rate, which is (by any reasonable standard) appallingly high. To my mind, everything that is infuriating or inexplicable to outside observers (or patients!) about the biopharma industry traces back to that number, which represents a rate of failure that no other major industry operates under.Clinical Failure Rates Over the Decades: Yikes, 2026-07-24
The numbers themselves: 90% in the 1960s, 81% in the 70s, 80% in the 80s, 88% in the 90s, 91% in the 2000s, 91% in the 2010s. He adds that error bars exist and are wider on the older figures, but the last three decades all fall within the same bounds, "so if you quote a failure rate of 91% you cannot be far wrong."
His comparisons:
…what the auto industry would look like of 91% of new car designs proved unable to roll out of the factory, or if 91% of new airliner models were unable to leave the ground - and if you only found that out after spending all the R&D money to build them at full size and trying to fly them. No cutting-edge restaurant could survive if 91% of its innovative dishes proved inedible or outright poisonous.ibid.
The crucial move comes next. Someone could read those numbers nostalgically: the 70s and 80s only ran at 80%, so the old ways were better, let's turn the clock back. Lowe closes that road on the spot. Those two decades were probably lower because they were the last pre-molecular-biology era, when compounds were more often evaluated in tissue slices than against isolated proteins — and more importantly, because the low-hanging fruit was still on the tree.
"Resetting the calendar is not going to bring back ACE and HMG-CoA reductase as new clinical programs," he writes — two targets that produced the blood-pressure drugs and the statins, two of the most lucrative harvests of the twentieth century. The diseases that are left are left precisely because they are harder. "It's not that we don't see them! Alzheimer's disease is a perfect example."
How did he arrive at the number? Not by reading it. By living it: three decades, four companies, dozens of projects, each one worked through by hand and each one dying — and then written up as a black joke in a 2015 introductory post. Twenty-four years of this blog are, at bottom, an answer to one question: when the default outcome in your industry is failure, what epistemology actually holds up?
In November 2007 Andy Grove of Intel spoke up. His line: the semiconductor industry had done unbelievable things over forty years, while in Parkinson's "we have gone from levodopa to levodopa"; ALS has no good treatment, Alzheimer's none at all. Why? Because chips run on a deeply felt conviction that time to market is time to money, whereas no pharma executive ever says "before the end of the year I'm going to have drug xyz" the way Steve Jobs said the iPhone would ship on schedule.
Lowe's reply was titled Andy Grove: Rich, Famous, Smart and Wrong.
His first move is not rebuttal but concession: everything Grove listed, "hey, it's all true." The second move dismantles the causation — and the way he does it is characteristic. He does not reason. He describes meetings he has sat in:
I've been in this industry for eighteen years, and I cannot count the number of year-end goals I've had to deal with. … But here's the rough part: although drug company people talk like this, they're full of manure when they do. … I'll bet some of my readers have sat in a few meetings - I sure have - and looked up at the screen thinking "Why on earth are we recommending this drug to go on?", only to have the answer be "Because it's early November".Andy Grove: Rich, Famous, Smart and Wrong, 2007-11-06
In other words: the medicine Grove was prescribing had already been swallowed for years, and the side effects were visible.
The third move is the centre of gravity, and the argument he would return to for the next twenty years:
Mr. Grove, here's the short form: medical research is different than semiconductor research. It's harder. Ever seen one of those huge blow-ups of a chip's architecture? It's awe-inspiring, the amount of detail that's crammed into such a small space. And guess what - it's nothing, it's the instructions on the back of a shampoo bottle compared to the complexity of a living system.ibid.
Why different? His reason is the plainest one available, which is exactly why it lands: we built the chips. "Making the things from the ground up is a real advantage when it comes to understanding them, but we started studying life after it had a few billion years head start." And chips are (presumably) designed to be comprehensible and efficient, whereas living systems "have been glued together by relentless random tinkering." Then the close:
Mr. Grove, you can print out the technical specs for your chips. We don't have them for cells.ibid.
The piece ends on a dare: you have the seed money, you can raise more by waving your hand — go start your own drug company, hire bright no-nonsense researchers, tell them you will have a Parkinson's drug by year end. "Prove me and the rest of the industry wrong."
In 2012 the same argument came back wearing statistics. A piece in Nature Reviews Drug Discovery coined Eroom's Law — Moore backwards — for the steady decline in new drugs per billion dollars of R&D. Lowe carried it to a wider readership, and particularly enjoyed one "scrupulously cruel" passage: every proposed fix in the original list — more management attention to cost and speed, reorganising into smaller focused units or larger ones with economies of scale, outsourcing to lower-cost countries, R&D performance scorecards, making scientists more "entrepreneurial" — is footnoted to the paper that proposed it.
Of the four causes the authors identify, Lowe takes the first, the "Better Than the Beatles" effect:
Nothing goes away unless it's well and truly superseded. Aspirin is still with is. Ibuprofen still sells like crazy. … But we're in the position of someone trying to come up with a better comb.Eroom's Law, 2012-03-08
This is the other face of the 91%: not only do nine in ten fail, the surviving one has to beat a market already stocked with cheap answers that work. In 2012 he wrote the blunt forecast: "We are heading, at speed, for a world in which fewer and fewer useful medicines are discovered, while more and more people want them."
In March 2021 he read a review by Andreas Bender and singled out its Figure 2: a simulation. Suppose you make the preclinical stages 20% cheaper, or 20% faster. How much do you save? Now suppose you cut the clinical failure rate by 20%. How much do you save?
The second swamps the first.
…the real problem is having drug candidates fail in the clinic. All that other stuff is a roundoff error compared to the clinical failure rate.AI and Drug Discovery: Attacking the Right Problems, 2021-03-19
So how do you cut it? The best ways, he says, are picking better targets and picking compounds that don't throw up unexpected toxicity in humans. "Those, sadly, are exactly the areas that AI/ML approaches are currently making the least traction in, because it's so hard to think up a useful way to attack them."
Speeding up screening or estimating physical properties is difficult too, but far more tractable — "which accounts for the press releases talking these up as if they're removing gigantic stumbling blocks to fast and easy drug discovery."
Why are the two hard things so hard? Because everything on the pipeline that generates numbers in bulk is a proxy: mechanistic biochemical rationales, cell assays, animal assays, scoring schemes for structures and properties. We have to have them, because the things that actually matter — does this work against a human disease, to what degree, without causing more problems than it solves — can only be measured in the clinic.
But in the end, they're all still just proxies. … The bottom line is that they're all useful, and they're still not enough. We all go into the clinic having done a lot of stuff that's Necessary But Not Sufficient, and if you don't hold your breath when the first human doses start, then you haven't been doing this stuff long enough.ibid.
He also notes a fact nobody bothers to state: no one even tries to predict the side effects a compound will show in a two-week rodent tox study. "You're about to find out what they are, and pretty much anything that's a real concern is going to come as a surprise to you anyway."
In May 2023 he compressed it into a claim time can settle. BenevolentAI had announced layoffs after its pan-Trk inhibitor missed its endpoints in atopic dermatitis. Lowe said he had followed the company for years, "mostly caught my eye about them was the level of hype in their press releases." Then: here is something I have said many times, and will now say louder for the folks in the back of the room:
There are no existing AI/ML systems that mitigate clinical failure risks due to target choice or toxicology.AI and the Hard Stuff, 2023-05-25
And then the knife. BenevolentAI used to say it had "created a bioscience machine brain… Proprietary algorithms perform sophisticated reasoning on over 50 billion ingested and contextualised facts":
But then, their stock also used to be worth 10 euros/share. It's now under 2, and that's still a higher per cent value remaining in it than there is in those claims about bioscience machine brains.ibid.
By 2026, evaluating a system called Co-Scientist — advertised as generating biomedical hypotheses and mechanisms on its own — he was using the same ruler. The method is worth copying: he ignores the benchmark scores and goes to the literature for precedent.
The paper ran drug repurposing against acute myeloid leukemia; a human panel picked five candidates out of thirty proposals, three of which showed cell activity. Lowe checked them one by one. Targeting the MEK pathway in AML with binimetinib "is not exactly a new idea, with a number of lines of research having been investigated"; pacritinib had already been into the clinic in some AML subtypes; and cerivastatin's "unusually strong ability to induce apoptosis in AML cell lines was noted in the literature 25 years ago." Of three unsupervised proposals, the only active one — KIRA6 — hit a mechanism "actually been proposed before as an AML therapy." Vorinostat, used for liver fibrosis, is an HDAC inhibitor, and the whole HDAC class had been studied in fibrosis models up to animal work.
Then the quiet blow: "I do not see citations of any of those references in the manuscript in its current form."
I think it's good that the system identified the mechanisms that it did, but I don't think that the paper does enough to point out that none of these ideas are without precedent - in some cases, a lot of precedent. … We definitely need all the help with the literature that we can get, but hype we have enough of already.Evaluating "Co-Scientist", a New AI Science System, 2026-05-27
Notably he does not dismiss it. The most promising use, he says, is precisely as "a focused results digester" — hand it a large corpus of experimental data and see what it makes of it, ideally testing it on cases where you already know the answer. He wants to know whether, walked along with the experiments, it would have converged faster. "Or not! It would be quite useful to know."
In July 2022 Science published an investigation: a series of highly influential Alzheimer's papers contained doctored images, with suspicion centred on Sylvain Lesné of the University of Minnesota and his 2006 paper — cited over two thousand times — on a soluble amyloid oligomer called AB*56.
The online reaction was predictable: because of this fraud, we have wasted our time since 2006.
Lowe's first move was to correct that — but not in the comfortable direction:
First off, I've noticed a lot of takes along the lines of "OMG, because of this fraud we've been wasting our time on Alzheimer's research since 2006". That's not really the case, as I'll explain. But don't get your hopes up: from one point of view, the main inaccuracy in that statement is that we've been actually been wasting our time in Alzheimer's research for even longer than that.Faked Beta-Amyloid Data. What Does It Mean?, 2022-07-25
What follows is an exercise in separating three tangled questions and giving each its own evidential weight.
First, the fraud is real and disgraceful. He relays the conclusions of several independent reviewers, including Elisabeth Bik, and notes that key experimental evidence looks like "nothing more than cut-and-paste jobs assembled to show the desired result." Then he turns the point on himself and his fellow reviewers:
When I review a paper, I freely admit that I am generally not thinking "What if all of this is based on lies and fakery?" It's not the way that we tend to approach scientific manuscripts.ibid.
Second, AB*56 probably does not exist. Dennis Selkoe's lab failed to find it early on; so did others.
Third — where he parts company with everyone — the amyloid oligomer hypothesis was not invented by that paper. He lays out the timeline: papers on pathogenic soluble oligomers were plentiful in the late 1990s and early 2000s. The trials were going to happen regardless. "The idea of amyloid oligomers as a key driver of AD is not a crazy one in any way, and people were going to put it to the test whether the *56 paper came out or not." He concedes the other half too: the excitement "surely did accelerate things. We probably got more clinical trials, sooner, than we would have otherwise."
And his verdict on the field is harsher than that of the people shouting about twenty wasted years:
Every single one of these interventions has failed in the clinic. Every last damn one. … We have to put money and effort down on other hypotheses and stop hammering, hammering, hammering on beta-amyloid so much. It isn't working.ibid.
The same move had appeared nine years earlier on an entirely different battlefield. In June 2013 a BuzzFeed listicle, "Eight Foods That We Eat in The US That Are Banned in Other Countries", had a million views. Lowe's response was not to argue that chemicals are safe — that is just the same false binary, run backwards. What he did was click through the article's own links: to the American Cancer Society, to NPR. They did not support the claims being made.
The links were subsequently removed, replaced with a note that "some studies linked in the original version of this article were concerning unrelated issues." Lowe added an update at the top pointing out that the studies were not unrelated at all — "all I can say is that they used to be there, until someone apparently realized how embarrassing they were."
From the rubble he salvages two general principles, and they are the point of the piece:
You cannot assume anything about the effects of a chemical compound based on whether it came from a lovely rainforest orchid or out of a crusty Erlenmeyer flask. … There is no essence, no vital spirit. A compound is what it is, no matter where it came from.Eight Toxic Foods: A Little Chemical Education, 2013-06-21
Paracelsus was right: the dose makes the poison. You can illustrate that in both directions: a beneficial substance, taken to excess, can kill you. A poisonous one, taken in very small amounts, can be harmless.ibid.
The numbers do more work than any amount of arguing: chicken meat came in at 2.3 parts per billion of inorganic arsenic; white rice at about 100 ppb and brown rice at 170 — and that is all-natural arsenic the plant took up from the soil. "If you're having chicken and rice and you want to worry about arsenic, worry about the rice."
The last line is his temper: "There are facts in this world, you know. Learn a few."
The thing Lowe is best known for online is a column called Things I Won't Work With. It looks like pure entertainment. It is the other face of the section above.
The 2008 instalment began domestically: he had written about putting out fires, a commenter mentioned the old sand bucket and wondered whether anything would set sand itself on fire. A reader wrote in: yes. Chlorine trifluoride.
It is apparently about the most vigorous fluorinating agent known, and is much more difficult to handle than fluorine gas. That's one of those statements you don't get to hear very often, and it should be enough to make any sensible chemist turn around smartly and head down the hall in the other direction.Sand Won't Save You This Time, 2008-02-26
Then the specifics: a one-ton spill in the early 1950s burned through a foot of concrete floor and chewed up another metre of sand and gravel beneath it. That process, he adds, "would necessarily have been accompanied by copious amounts of horribly toxic and corrosive by-products: it's bad enough when your reagent ignites wet sand, but the clouds of hot hydrofluoric acid are your special door prize if you're foolhardy enough to hang around and watch the fireworks." He gives the last word to rocket-fuel chemist John Clark's out-of-print classic Ignition!:
"It is hypergolic with every known fuel, and so rapidly hypergolic that no ignition delay has ever been measured. It is also hypergolic with such things as cloth, wood, and test engineers… For dealing with this situation, I have always recommended a good pair of running shoes."John Clark, Ignition!, quoted by Lowe in the above
Two years later came dioxygen difluoride, O2F2, known in the literature by its shape as FOOF. You make it by running a mixture of oxygen and fluorine through a heating block at 700°C — "the heater block glowed a dull red color, observable with room lights turned off." Lowe's gloss: "And yes, what happens next is just what you think happens."
The real protagonist is A. G. Streng of Temple University, who in 1962 not only prepared multiple batches and kept them around but was apparently charged with finding out what it did to things. All sorts of things:
"It reacted instantaneously with solid ethyl alcohol, producing a blue flame and an explosion. When a drop of liquid O2F2 was added to liquid methane, cooled at 90°K., a white flame was produced instantaneously…"A. G. Streng (1962), quoted by Lowe in Things I Won't Work With: Dioxygen Difluoride
Streng went on to react it with ammonia, water ice, chlorine, red phosphorus, bromine fluoride, chlorine trifluoride, perchloryl fluoride and tetrafluorohydrazine. Lowe's overall assessment:
If the paper weren't laid out in complete grammatical sentences and published in JACS, you'd swear it was the work of a violent lunatic. I ran out of vulgar expletives after the second page. A. G. Streng, folks, absolutely takes the corrosive exploding cake, and I have to tip my asbestos-lined titanium hat to him.Things I Won't Work With: Dioxygen Difluoride, 2010-02-23
Even Streng gave up somewhere: sulfur compounds defeated him, because hydrogen sulfide plus four molecules of FOOF sheds 433 kcal. "So if you feel like whipping up a batch of Satan's kimchi, go right ahead."
The sister column is How Not to Do It, on laboratory disasters. The 2008 entry covers diazomethane, a reagent nobody will ship you, so you make it fresh:
That diazo group is looking for an excuse to revert back to nitrogen gas, which process comes with an inevitable no-substitutions side order of kaboom. The chemist's job is to not give it that excuse.How Not To Do It: Diazomethane, 2008-04-30
He gives two scenes. One is an institutional consequence: somebody graduates and leaves a big flask of yellow solution behind — "Is this yellow stuff what it looks like it is? How long has it been in here? And who the hell made it, anyway?" — and the school adds a new inspection rule. "You couldn't leave until someone determined that you weren't passing on any explosive bequests." The other is his own: a summer thunderstorm knocks the power out, the head of the safety committee rolls a cooler of dry ice down the hall calling "Anybody need to store something in the cold?" — "Well," said Lowe, "we've got a couple of liters of diazomethane solution." "That's not very funny." "That's because it's not a joke" — and they went to the front of the line.
Why is this more than a gag reel? Because it is the same epistemology as the food-chemistry piece, seen from the other side. The abstract noun "chemicals" manufactures both fear and contempt — the frightened assume anything synthetic is poison, the dismissive assume the bottles are just liquids. Lowe's whole column dismantles that abstraction: real substances have specific tempers, specific temperature ranges, specific ways of killing you. After reading about FOOF you will never again think "chemicals" is a thing you can have a single opinion about. That may be the most useful thing a lay reader takes away from this blog, and it arrives wrapped in a joke.
Laid out across twenty-four years, the transferable part is larger than the conclusions.
Check the citations, not the positions. This is his first move and his most consistent one. With the BuzzFeed piece, he clicked through the article's own links. With Co-Scientist, he went looking for precedent behind each "discovery." Neither time was he arguing with an opinion; both times he was inspecting the floorboards the opinion stood on. This method has a useful property: it does not require you to be smarter than your opponent, only more willing to spend the ten minutes. "Ten minutes of unbiased reading would have served to poke holes all through most of the article's main points," he wrote. "I've spent more than ten minutes."
Always ask for the magnitude first. Not "is there arsenic" but "how many ppb, and how does that compare with rice"; not "is the failure rate high" but "90, 81, 80, 88, 91, 91." His judgements almost always land on a number first, and the opinion grows out of the number.
Keep proxies and endpoints apart. This is his most portable idea and it has nothing to do with pharmacology. Are the beautiful, abundantly collectable numbers you have actually the thing you care about? If not, optimising them harder may only be polishing the roundoff error. Wherever "what can be measured" and "what matters" fail to coincide, this applies.
Treat your own failure record as a credential, not a stain. "I have never once worked on any compound that has made it all the way to the market" would be suicide from someone trying to sound authoritative. He puts it in the first paragraph of his introduction, because in an industry with a 91% failure rate he is the representative sample. His credibility comes not from his successes but from the scale of his failures.
Not knowing, in public. This is his least-known and most worth copying stretch. Starting in the summer of 2002 he serialised, on the blog, an experiment whose content he refused to disclose — because the intellectual property belonged to his employer, and because of scientific priority. He found it absurd himself: "Now there's some great blog material, eh?" He pre-emptively compared himself to the South Sea Bubble operator in Charles Mackay who sold shares in "an undertaking of great advantage and no man to know what it is." "At least I'm not selling shares."
The serial ran, on and off, for more than two years, under a category called Birth of an Idea. Most of it is failure. A January 2004 post is titled If This Doesn't Work, There's Only Reality To Fall Back On; when the results came in days later, "rarely have I seen less encouraging data." By March the idea was "last seen crashing to earth in a rain of feathers and melting wax." In April 2005 he went back for more, and the reason is extremely honest: he could now think of confounding variables in the earlier runs, and "I still - in the face of a fair amount of evidence, I admit - believe that this whole thing should work."
In the waiting periods he wrote the best sentences on this blog:
It can be more psychologically comforting to keep them in the "untried but promising" category, rather than find out if they're real.Per Fits and Starts, Ad Astra, 2003-08-03
These experiments will carve great swaths through Explanation Space, blanking out whole regions of potential false-positives.Good News, Backing Slowly Through the Door?, 2003-12-07
As I've said, scientific progress depends, to an unappreciated extent, on the willingness to look like a fool. It's tough work, waiting to find out if you are one or not.Data, At Last, 2005-06-21
On 23 June 2005 the post was titled News Flash: I May Not Be a Fool. He finally had results that looked right. Two months later he set the whole thing up again with fresh material, repeating the best run "just to be sure that there wasn't something odd about the last batch" — and mentioned in passing that "none of this is officially what I'm supposed to be doing, so I have to work these things in as I can."
This is a kind of writing that has almost vanished: one person, not knowing how it ends, narrating the life and death of an idea in real time with the self-doubt left in. It is closer to research than any essay about "the scientific method."
Treat the comment section as a co-author. He says so in the 2015 introduction: the comments have been what he enjoys most, "I realize that this is not the usual comments-section experience, but the signal/noise at the old site was always high." This is not politeness. The chlorine trifluoride piece exists because of a reader's question and another reader's tip; the two crucial updates on the Thiola price-hike post came from a patent point raised in the comments and a company presentation someone sent him on Twitter.
Twenty-four years of near-daily posting is itself a method. In 2015 he admitted that when he started he worried about running out of material; "that has never been the problem. Things just keep happening." Someone forced to reach a verdict on the day's news every day develops a craft for locating magnitudes quickly — a craft you cannot acquire writing one long essay a year.
The structural objection is posted on his own door. The masthead says the blog is editorially independent, all content is Derek's own, and he does not in any way speak for his employer. But the industry he writes about is also the industry that pays him. He knows how that sounds — in a 2021 post on price controls he reached for a very English reference:
Longtime observers of the industry that pays my bills will know that when that fifth question comes up, pharma companies always talk about bad effects on drug development and innovation under a price-control system. One response to that is "Well, they would, wouldn't they?", but just as with the canonical Mandy Rice-Davies response, it doesn't necessarily mean that the objection is invalid prima facie.A Drug Pricing Proposal and its Effects, 2021-09-08
Declaring a conflict of interest is honest. Declaring it does not dissolve it.
From 91% to "companies should set their own prices" there is a missing step. A high failure rate is a factual claim; no particular price follows from it. And that is where he stands:
I think that drug companies should be able to set the price of their drugs, because they have a lot of failures to make up for…The Most Unconscionable Drug Price Hike I Have Yet Seen, 2014-09-11
To be fair, this is not blanket defence. That sentence comes from the angriest thing he has written about his own industry — a September 2014 dismantling of Martin Shkreli's Retrophin raising the price of an obscure old drug, Thiola, from $1.50 a pill to over $30, a full year before Shkreli became a national villain over Daraprim. "This is exactly the sort of thing that makes people go into fist-clenching rages about the drug industry, and with damn good reason. This one enrages me, and I do drug research for a living."
At the policy level, though, his direction is consistently one-way. The 2021 piece criticising the Congressional Budget Office's model draws on Steve Usdin's reporting, Harvard Business School's Amitabh Chandra, and an open letter — all pointing the same way, that price controls will mean fewer drugs. He does not stress-test the opposite case with equal force, and concludes: "It seems impossible to me that it can change it any other way but for the worse." For a writer known for demanding the specific number before forming a view, the evidence on this particular question is noticeably thinner.
A named opposing view. Donald W. Light and Rebecca N. Warburton's "Demythologizing the high costs of pharmaceutical research" (BioSocieties, 2011, DOI 10.1057/biosoc.2010.40) attacks the first link in that chain directly: they re-examine the widely cited $802 million per new drug figure, argue that roughly half of it consists of "profits foregone," and offer a far lower median estimate of $59.4 million. In the interests of honesty: I have read this paper's publication details and abstract-level conclusions, not its full text, and I could not find a post in which Lowe replies to these authors directly. So what is reported here is bibliographic fact only. I am not reproducing their derivation, and not adjudicating between them and him.
A bet that time will settle. "There are no existing AI/ML systems that mitigate clinical failure risks due to target choice or toxicology" is a good bet, because it is falsifiable and because he says he is "very willing to stand by that statement." It also means that the day it is overturned, his entire account of AI needs rewriting. Note the door he leaves open, and how frankly he leaves it: "Someday we may well. But that day is not today."
The edges of his field of view. Molecules, companies, clinical trials, laboratories: this is his ground, and on it he has no equal. Access to medicines, pricing in low- and middle-income countries, patent waivers, how health systems allocate — these are neither his focus nor his strength. His best craft (go check, go count, go compare magnitudes) is rarely applied on that side, which is precisely where the consequences of the 91% argument land.
One last item; readers can judge for themselves. Since 2025 his writing on US federal research funding has changed register. Discussing NIH restrictions on foreign co-authors in May 2026, he wrote: "I don't trust the intentions of my own government, because my own government is now a vile collection of incompetents, hacks, and grifters" — along with a swipe at those who believe none of it matters because artificial general intelligence is about to invent everything humans would have invented anyway. Some will say this is inconsistent with twenty-four years of looking at the data before forming a view. It can also be read as the same person responding to a situation where there is no longer any data to argue about. His own stated reason is this: every single person working in industrial science and engineering came up through academic scientific training.
Other posts quoted on this page, in order: Welcome to "In the Pipeline" (2015-08-03), Eroom's Law (2012-03-08), Sand Won't Save You This Time (2008-02-26), How Not To Do It: Diazomethane (2008-04-30), Eight Toxic Foods: A Little Chemical Education (2013-06-21), The Most Unconscionable Drug Price Hike I Have Yet Seen (2014-09-11), Coronavirus Vaccine Prospects (2020-04-15), Blog Switchover In Progress (2021-08-17), A Drug Pricing Proposal and its Effects (2021-09-08), AI and the Hard Stuff (2023-05-25), Evaluating "Co-Scientist", a New AI Science System (2026-05-27), The Current Crisis: What's Happening to Science in America (2026-05-28). The 2002–2005 Corante-era posts (Medicine Man; Per Fits and Starts, Ad Astra; Good News, Backing Slowly Through the Door?; Data, At Last; Experimental Update, For Those Who Care; Mismatched Socks; News Flash: I May Not Be a Fool) were read from that site's archive; the category pages did not survive the migration, so they are cited by title and date only. Blog home: science.org/blogs/pipeline.