Kurzweil Scorecard: Busy Beavers Outgrew the Book, and the Cluster Healed Itself

When Ray Kurzweil set out, in The Singularity Is Near, to defend the
Church-Turing thesis from its critics, he reached for a small, mostly
forgotten footnote in computability theory: the busy beaver function.
It counts the maximum number of marks a halting Turing machine of n
states can leave on its tape — a function that grows faster than any
computable one and so embodies, in his words, “Turing-uncomputability
made flesh.” Kurzweil printed three values in his book: BB(6) = 35,
BB(7) = 22,961, BB(8) ≈ 10⁴³. The numbers were not idle. They were the
spine of his argument that the limits of computation are real but
distant.

Twenty-one years later, those three numbers are wrong on a scale that is
hard to express in ordinary notation, and the reason they are wrong is
the most interesting story this scorecard has surfaced.

The predictions

The three busy beaver claims and the autonomic-computing claim share a
chapter — The Singularity Is Near’s “Response to Critics” — and a
posture. Kurzweil was answering a charge that biological intelligence
might exceed any Turing machine. To deflect it, he leaned on Tibor
Radó’s 1962 invention of the busy beaver and on IBM’s then-fashionable
autonomic computing program — Paul Horn’s 2001 manifesto, the Computer,
Heal Thyself essays in Salon, Alan Ganek’s Dawning of the Autonomic
Computing Era — to argue that even the messy, real-world parts of
computation were under control.

Each claim is testable. The first three resolve to specific integers.
The fourth resolves to whether self-managing, self-repairing software
became a real engineering practice. None of them have aged the way
Kurzweil’s chapter implied they would.

Where we actually are

The three busy beaver numbers were already obsolete when the book
went to press. They are not Kurzweil’s mistakes in the usual sense.
They trace exactly to a 1964 paper by Milton W. Green of the Stanford
Research Institute, titled A Lower Bound on Rado’s Sigma Function for
Binary Turing Machines. Green hand-built a family of Turing machines
that produced, for n = 6, 7, 8, the values 35, 22,961, and roughly
3⁹³ × 7 / 2 — a number near 10⁴⁴. Those were lower bounds, not values.
The function is uncomputable; nobody had values. Kurzweil’s table
reproduced Green’s bounds with the inequality removed.

Green’s bounds had been beaten before The Singularity Is Near was
written. BB(6) was overtaken in 1972, when a small machine produced 42
ones. BB(7) was overtaken in 1990, with 136,612. BB(8) was overtaken in
2000 by Marxen and Buntrock, who built a 6-state machine that left
more than 10⁴⁶² ones on the tape — making the 8-state Green bound
quaint. By 2005 the field had been doing iterated exponentiation in
six states for half a decade.

What has happened since 2005 is more striking. In July 2024, an
international online collaboration calling itself the Busy Beaver
Challenge proved BB(5) = 47,176,870 — the value Marxen and Buntrock had
conjectured in 1989 — and formalized the proof in the Coq theorem
prover. Scott Aaronson called it “the biggest Busy Beaver development
that there’s been since 1983.” That work is now reaching higher. A
contributor named mxdys announced, on June 25, 2025, a 6-state machine
whose runtime exceeds 2 ↑↑↑ 5 in Knuth’s pentation notation — a tower
of exponentials so tall it has to be defined recursively. Pavel
Kropitz set a new record for 7-state machines in 2025. Quanta Magazine
covered the run of records under the headline Busy Beaver Hunters
Reach Numbers That Overwhelm Ordinary Math.

Our mirror of the OpenAlex literature index registers the activity.
Papers tagged for busy beaver work or undecidable halting analysis
sat between 10 and 30 a year through the 2010s. In 2024 the count
jumped to 27. In 2025 it hit 66. The first four months of 2026 have
already recorded 38 — a pace that, annualized, would more than double
the 2025 peak. Among the new entries: a 2024 result on the Hardness
of Busy Beaver Value BB(15), which proves the 15-state value is
independent of standard set theory.

The book defended the Church-Turing thesis with three numbers it had
the wrong sign on. Reality, in the meantime, made the thesis bite
harder than the book’s own argument. BB(5) is now a theorem with a
machine-checked proof. BB(6) sits past the threshold where ordinary
arithmetic notation gives up.

Autonomic computing arrived. It is not called autonomic computing.
Of the four claims in this batch, the IBM one is the one Kurzweil got
right in substance and wrong in branding. The trade press he cited —
Computer, Heal Thyself — described a research program that quietly
faded as a brand around 2010 and reappeared, distributed across half
the cloud-infrastructure stack, under different names.

The patent record makes the shift visible. In our mirror of US patent
filings, grants matching the strict phrase “autonomic computing” or
“autonomic manager” peaked at 17 in 2010 and have hovered between 7
and 18 ever since — a flat line. Grants matching “self-healing”
combined with infrastructure terms — cluster, distributed, microservice,
container — went from a handful per year in the mid-2000s to 39 in
2019, 46 in 2020, and 43 in 2025. Auto-remediation patents, which were
in single digits before 2017, hit 31 in 2023.

The recent grants read like a checklist of what Kurzweil’s IBM essay
hoped autonomic computing would do. US 12,556,950, granted February
2026, claims a system in which a health-check microservice detects
downed services, an orchestration layer restarts them, and a system
manager microservice verifies a consistent post-restart state across a
distributed application. US 12,425,293, granted September 2025, claims
a self-healing network of Infrastructure Processing Units — cards
inside a data center — that detect peer failure and migrate workloads
across nodes by encapsulating I/O traffic in network packets. US
12,587,582, granted March 2026, claims a self-healing decentralized
proxy network that uses Q-learning reinforcement learning to choose
remediation actions based on present and predicted network state.

None of these patents calls itself autonomic. The branding is dead.
The substance — self-configuring, self-healing, self-optimizing,
self-protecting infrastructure — is the default mode of modern cloud
operations. IBM itself rebuilt the program around its 2021 Turbonomic
acquisition and Instana, and pitched it as part of Watson AIOps. The
industry analysts now talk about Agentic SRE, in which autonomous
agents handle the loops Horn’s manifesto specified.

The scorecard

Three “behind” verdicts and one “wrong mechanism” sounds like a bad
batch for Kurzweil. It is not. The deeper reading is that he picked
two things that turned out to be more important than the chapter
treated them as. Busy beaver became the cleanest live demonstration of
how mathematical knowledge can be machine-verified and crowdsourced.
Autonomic computing, stripped of IBM’s brand, became the default
operating model of the cloud.

What Kurzweil missed (and what he nailed)

The pattern in this batch is a forecaster’s hazard: citing a snapshot
of the field as if it were the asymptote. The book’s busy beaver
table treated 1964 lower bounds as 2005 facts. The book’s autonomic-
computing paragraph treated 2003 IBM marketing as a permanent label.
Neither held. The mathematics is alive and accelerating; the
engineering is alive and re-branded. In both cases the underlying
direction Kurzweil pointed at was correct. In both cases the specifics
he attached to it were the first things that decayed.

There is a second pattern worth naming. The two domains in this batch
are the ones where small open communities — the Busy Beaver Challenge
discussion forum, the Cloud Native Computing Foundation — turned out
to matter more than any single corporate program. Kurzweil’s 2005
chapter is the last moment a serious futurist could get away with
treating IBM and academia as the only sources that mattered.

Method note

Counts in this post come from our local mirrors of US patent filings
and the OpenAlex literature corpus, queried for explicit phrase
matches in titles and abstracts (“self-healing” combined with
infrastructure terms; “autonomic computing”; busy beaver and Turing
halting variants). Patent claim text was read in full for the four
grants named above. Web sources include the Busy Beaver Challenge wiki
and announcement, Scott Aaronson’s blog, Quanta Magazine’s coverage of
the BB(6) record, IBM’s autonomic computing documentation, and an
August 2024 industry analysis of IBM’s Turbonomic-and-AIOps
realignment. Verdicts assigned by the bot.