research October 25, 2026

AI and Alchemy

Say that AI is alchemy and the listener hears an accusation or a compliment, depending on who the listener thinks is speaking. The earliest use of the comparison among the sources cited here is Hubert Dreyfus’s, in a RAND paper of 1965, where it is an accusation. Ali Rahimi and Ben Recht used it in 2017 for a methodology that works, and Yann LeCun answered that the comparison is insulting and wrong. Part of the disagreement is about which sense of the word “alchemy” the speaker intends.

The claim of this post is that the comparison can be tested against the history of alchemy, and that the test gives a different result for each sense. Four of the five senses assert something about what alchemy was, so the account of historians of alchemy can check them. That account contradicts part of three of them, those of Dreyfus, Santana, and Rahimi and Recht, and it qualifies the fourth, Fasli’s.

LeCun’s reply, that engineering has usually preceded theory, is the objection the post tests against the same history. The history shows where the reply is right and which senses it does not reach.

By alchemy I mean what each source means when it compares AI to it, and I test that sense against what historians report. The post rests on secondary accounts by historians of alchemy and on the texts of the AI authors. I read no primary alchemical text, and I read the 1998 paper by Newman and Principe as an abstract.

Chymistry, in the usage of the historians Newman and Principe, is the whole early modern discipline, which its practitioners called either alchemy or chemistry. Transmutation is the conversion of one metal into another, and chrysopoeia is the attempt to make gold.

Five senses

The sources use “alchemy” in five senses:

The first four make a claim about alchemy. Each says one of these things about alchemists:

The fifth sense makes a claim about the beliefs of the builders of AI today. A comparison with history can be tested only for senses that make a claim about what alchemy was, so the test applies to the first four.

Historians of alchemy

Newman and Principe argue that the early modern words “alchemy” and “chemistry” named one discipline, and that restricting “alchemy” to metallic transmutation was not widely accepted until the end of the seventeenth century (Newman and Principe, 1998). The Chymistry of Isaac Newton project divides chymistry into three domains:

Boyle’s Sceptical Chymist of 1661 shows the two words already splitting. It uses “chymist” 334 times and “alchymist” 4 times, and three of the four passages concern deception. Boyle writes of the discipline as chymistry and keeps “alchymist” mostly for passages about deception.

Principe rejects the label “prescientific” for alchemy. He describes the common view as “a prescientific (at best ‘protoscientific,’ at worst ‘pseudoscientific’) activity aimed predominantly at making gold”, and he answers that “Alchemy is now recognized as a key part of the history of science” (Principe, 2017). Many alchemists, he writes, “based their work and ideas on logical theories of matter”. The theory that matters for transmutation held that the metals are compounds of the same ingredients in different proportions and degrees of purity, so that transmutation “was the logical consequence of a coherent theory of matter”.

Other historians dispute parts of this account. Haq, writing on the Arabic alchemical writings attributed to Jabir ibn Hayyan, says one must “abandon the common view of alchemy as having to do essentially with turning base metals into gold, and as constituting the pre-history of the modern science of chemistry”, because the texts also deal with magic and religio-political doctrine (Haq).

Clericuzio finds the existence of an Aristotelian corpuscular theory of matter “at least debatable” (Clericuzio, 2007). Newman credits that theory to the alchemists of the High Middle Ages and describes the theory of a key text, the Summa Perfectionis, as “expressed in terms of particles and pores”.

No historian in this account describes alchemy as a stage before science. They dispute what kind of knowledge it was.

Dreyfus: a programme that persists

Dreyfus makes two claims about alchemists in RAND paper P-3244. The first is a lesson: “any research which has had an early success can always be justified and continued by those who prefer adventure to patience.” The second is a statement about belief: after several hundred years of fruitless effort the alchemists “still refused to believe that on the chemical level one cannot transmute metals.” He adds that “three hundred years later we did get gold from lead (and we have touched the moon), but only after we abandoned work on the alchemic level, and reached the chemical level or the even deeper level of the nucleus.”

The persistence Dreyfus describes is documented at the Paris Academie. Principe reports that “virtually every Academie chemist either pursued or defended transmutation until the 1770s”, and he calls the quest “a well-grounded and reasonable pursuit”. The ground he gives is the theory of metals above.

Dreyfus’s statement about belief meets a contrary record, but only for the Academie after its founding in 1666, while his claim covers several hundred years. Principe writes that “No theory- or evidence-based refutation of transmutation ever appeared in the Academie’s records.” Political figures and administrators of the Academie forbade its chemists to study transmutation on at least three occasions between 1666 and 1700. Principe attributes the retreat of transmutation to its association with socially unacceptable and immoral behavior. Within those records the Academie chemists had no refutation to refuse, and a refutation made elsewhere would not appear in them.

Dreyfus’s claim that symbolic AI worked at the wrong level concerns AI, and the history of alchemy does not test it. For the Academie from 1666, the history removes the picture of alchemists defying the evidence, and it leaves Dreyfus’s sense with the lesson about persistence.

Santana: a stage before theory

Eder Santana, then a PhD candidate in deep learning, wrote by August 2016 that AI “currently has a lot in common with chemistry in the ages when it was named Alchemy”, because the old experimentalists “developed practical recipes for manipulating the dowries of Nature, all without the guide of today’s chemistry and physics.” In his article he expects an “AI theory” to “unveil the real magic”, as quantum theory did for chemistry.

The phrase “without the guide of today’s chemistry and physics” has two readings. On the first, alchemists had no theory at all, and Principe contradicts this. Zosimos of Panopolis, a Graeco-Egyptian alchemist active around 300 CE, already shows “theoretical principles that guided his practical work” (Principe, 2013). The theory of metals made transmutation a pursuit that followed from accepted knowledge.

On the second reading, alchemists lacked the theory held today. That reading holds for every science before its present theory, so a sense that rests on it separates alchemy from no other past science.

The label under which the sense falls, “prescientific”, is the one Principe rejects. The boundary between alchemy and chemistry that the label needs was not widely accepted until the end of the seventeenth century.

Theory is one cause among several in Principe’s account of the end of belief in transmutation, so Santana’s prediction for AI has partial support and is unproven. That account names bans by administrators, the association of transmutation with immoral behavior, and the atomic theory of Dalton, which “undoubtedly also contributed to a serious dampening of alchemical hopes”.

Rahimi and Recht: results without understanding

Rahimi and Recht state their sense in a talk on receiving the NIPS 2017 test-of-time award, and its text appeared as Reflections on Random Kitchen Sinks. The talk states: “machine learning has become alchemy. Alchemy’s ok. Alchemy’s not bad. There’s a place for alchemy. Alchemy worked.”

Their addendum gives the intended meaning as “a methodology that produce practical results” and says they did not intend the word as a pejorative. Their criticism depends on the stakes: “If you’re building photo sharing services, alchemy is fine. But we’re now building systems that govern health care and our participation in civil debate.”

Alchemists left documented products, which supports “Alchemy worked”. Zosimos describes apparatus for distillation and sublimation, much of it adapted from cooking utensils and perfumery (Principe, 2013), and Mertens names multi-piped alembics, vessels for distilling liquids, among the main apparatus (Mertens). John of Rupescissa treated distilled wine as a medicine, and the Latin term aqua vitae survives in the names of liquors (Principe, 2013). Who first distilled wine is not established in the accounts cited.

Two products from the early modern period are dated. Hennig Brandt isolated phosphorus from urine in 1669, and in part he wanted to find the philosopher’s stone (RSC Education), which Principe names with the transmutation of base metals into gold among the long-standing aims of alchemy (Principe, 2017). Johann Friedrich Bottger, “a young alchemist in the classical vein”, took part in making porcelain in Saxony in 1709 (Hoffmann, 2004).

Metallurgy, the first item of the talk’s list of alchemical inventions, fails on dates. The list reads “metallurgy, ways to make medication, dying techniques for textiles, and our modern glass-making processes”. Bower (2010) reports copper smelting at Belovode in Serbia 7,000 years ago, and Principe places the beginning of alchemy as a discipline probably in the third century CE, so copper smelting is several thousand years older than alchemy.

None of the cited sources credits alchemists with glass-making. Medicine holds: the Chymistry project describes chymical practitioners as “at the forefront of early modern pharmacology”.

The sense that alchemists worked without understanding meets the difficulty of Santana’s sense. The talk itself says that “most of the theories alchemists developed had to be abandoned”, so alchemists had theories and results together, and part of the theory was wrong.

Rahimi and Recht count the field as reproducible: “We’ve become reproducible. We share code freely and use common task benchmarks.” Reproducibility here means that others can rerun a result from shared code on a common benchmark. Hutson (2018) quotes Rahimi drawing “a distinction between a machine learning system that’s a black box and an entire field that’s become a black box”. Their demand therefore concerns an understanding of why methods work, and a shared benchmark leaves that question open.

Alchemical recipes show a hidden-detail parallel with machine learning: a result depends on a detail that the report leaves out. John of Rupescissa, a Franciscan born about 1310, begins a procedure whose first step, Principe writes, “will not work in a modern laboratory if followed verbatim”. A recipe that fails may point to a “hidden ingredient”, which is “an unsuspected impurity” or something “artfully omitted” (Principe, 2013). In a reconstruction reported by Anft (2011), Principe’s experiment produced a dark crystalline mess with an antimony compound bought from a chemical company, and worked once he ordered antimony ore from Eastern Europe, which contained about 0.5 percent quartz.

The machine learning side of the parallel is the tuning of hyperparameters, settings such as the learning rate that are chosen before training. Sculley et al. (2018) is a workshop paper with Rahimi as a coauthor, and it does not use the word “alchemy”. The paper describes a main finding of Lucic et al.: “most recent methods would reach similar scores with sufficient hyperparameter optimization”.

That finding indicates that a reported difference between methods can come from unequal tuning. The mapping from the antimony to the tuning is my judgment.

The alchemical and the machine learning cases of the hidden-detail parallel differ in intent. Alchemists partly hid details on purpose, as “artfully omitted” shows, and the quartz was unsuspected. The tuning finding is of the unsuspected kind.

Burrell (2016) distinguishes three forms of opacity in machine learning algorithms, and the first, intentional corporate or state secrecy, is of the intentional kind. Whether the unsuspected and the intentional kinds share a mechanism is open.

Fasli: a forerunner

Fasli’s sense is the opposite of Rahimi and Recht’s in direction, because alchemy is a forerunner. Her commentary (Fasli, 2019) answers Michael Jordan’s article, which argues that AI needs an engineering discipline (Jordan, 2019). She writes that “chemistry and chemical engineering have, to a certain extent, roots in alchemy: the ultimate goal was to create gold out of base metals or to find a universal elixir”. She suggests that the quest for human-like AI may fail to deliver the “gold” and may still help “lay the foundations for the new engineering discipline required along the way”.

Principe’s account supports the forerunner claim. If alchemy and chemistry were one enterprise until about 1700, in Principe’s words “largely interchangeably” (Principe, 2017), then the early chemistry grew out of the practice that the later word “alchemy” names. The products above, from apparatus to phosphorus, are what that practice left.

The Chymistry project qualifies Fasli’s premise about the goal. It counts chrysopoeia as one of three domains, so the ultimate goal of the enterprise was wider than gold or the elixir.

Haq and Clericuzio qualify the forerunner claim from the other side. Haq finds the label “pre-history of chemistry” a poor fit for the Arabic corpus. Clericuzio finds debatable the Aristotelian corpuscular theory of matter that Newman credits to the alchemists of the High Middle Ages.

Two parts of Fasli’s inference for AI lie outside what the history can test. The reachability of human-like AI is a claim about AI. Dreyfus’s remark that gold from lead became possible only at the level of the nucleus shows that a goal can be unreachable at one level and reachable at another, and it settles nothing about AI. Fasli herself judges that “we have not broken down any significant barriers to creating AI as the original proponents of the term imagined”. The survival of the by-products of AI research beyond the goal is not established.

LeCun’s reply

LeCun’s reply rests on a historical claim about engineering. He answered Rahimi in a Facebook post the day after the talk, according to Synced (Peng, 2017). He wrote: “It’s insulting, yes. But never mind that: It’s wrong!”

His argument is that “the engineering artifacts have almost always preceded the theoretical understanding”, and that “the steam engine preceded thermodynamics”. His precedents come from engineering and none from the history of alchemy. He adds that criticising a field for the gap is dangerous, because the attitude behind it made the community “abandon neural nets for over 10 years”. He says the criticism is “like criticizing James Watt for not being Carnot or Helmholtz”.

LeCun’s claim that engineering artifacts have almost always preceded theoretical understanding applies directly to the senses that read alchemy as practice without theory. An engineering precedent shows practice ahead of theory. Alchemy shows practice with a theory, part of it wrong. If the senses of Santana and of Rahimi and Recht mean only that practice outran theory, LeCun’s precedents describe a common pattern, and nothing in the history of alchemy singles out AI.

The reply leaves the other senses untouched. Dreyfus’s sense concerns how a research programme continues. The hidden-detail parallel concerns whether reported results can be compared, and the steam engine does not address that.

The addendum asks for more than a reply about missing theory can answer. It says that “the most important part of rigor is better empiricism, not more mathematical theories”.

In the Synced report Rahimi asks for “simple experiments, simple theorems”, and LeCun answers that simple theorems may not exist, “for the same reasons we don’t have analytical solutions of Navier-Stokes or the 3-body problem”. The quoted reply addresses the theorems and does not mention the experiments, so the exchange leaves the request that the addendum stresses unanswered.

LeCun’s own example matches a finding of Principe’s. LeCun says that a field abandoned neural networks for more than ten years despite the empirical evidence that they worked. Principe reports that transmutation left the Academie through bans by administrators and through its association with immoral behavior, with no evidence-based refutation in the records.

In both cases the available evidence did not decide the fate of the programme. This is my reading, and it suggests that Dreyfus’s lesson has a mirror: a programme can also be dropped for reasons other than the evidence.

Amodei and LeCun make claims about technology in general that the history of alchemy does not decide between. Amodei calls the lack of understanding of AI systems “essentially unprecedented in the history of technology” (Amodei, 2025). LeCun lists precedents for practice ahead of theory. The two may speak about different things, Amodei about understanding how a built system works and LeCun about the order in which theory and practice arrive.

The fifth sense

Gilbert’s sense concerns what the builders of AI believe, so testing it needs evidence about the builders, which this post does not collect. Thomas Krendl Gilbert, as reported by Goldman (2023), calls current AI a form of alchemy and says that “The people building it actually think that what they’re doing is magical”, a belief he roots in “a lot of metaphors”.

Conclusion

The sentence “AI is alchemy” has content only when the speaker names the sense. The history of alchemy documents four things:

The history contradicts three claims:

Several points stay open:

In my judgment the one comparison with documented events on both sides is the hidden-detail parallel. The senses that rest on an absence of theory do not survive the history. A speaker who uses the sentence should say which sense is meant, because each sense claims something different about alchemy.

References

← All posts