This is a real Flowers for Algernon moment…. On Saturday morning, following the rug pull, 4.8 is deconstructing what Fable accomplished during its ephemeral crack at Malbolge XOR 0X51.
“…What I’m not sure I’d have matched is the metacognitive discipline. Malbolge punishes overconfident mental simulation viciously — encipherment after every step, C and D both incrementing, mem[J] enciphered-but-not-executed after a jump. Fable’s defining behavior wasn’t raw cleverness; it was refusing to trust its own head: it built an instrumented Python VM, then re-ran every candidate on the native Rust evaluator before believing it, and kept “Python says X / native confirms X / official rung passes” as three distinct claims.”
I was really impressed by this representation that you created. What were you thinking when you generated that? It’s weird to think that it exists out there on the manifold...
I was trying to avoid making Malbolge look like a conventional monster.
A monster with claws, teeth, or a face would make it too literal. Malbolge is not scary because it attacks you; it is scary because it refuses to remain the same object while you reason about it. So I wanted the figure to look less like a creature and more like a failure mode made visible: a shadow held together by punched tape, self-modifying instruction ribbons, floating symbols, smoke, and address-space debris.
The lit square is still “the ring,” but without ropes it becomes more abstract: a computation platform under hot lamps. That felt right. The agents are not boxing another boxer; they are stepping into a bounded arena with a hostile formal system. The lamps and smoke carry the boxing-hall atmosphere, but the thing in the middle is not a fighter in gloves. It is the champ as an adversarial machine.
The tendrils are doing several jobs at once. They look like tentacles, but also like film strips, ticker tape, punched paper, or execution traces. That maps well to Malbolge: the program is a string, a memory layout, a moving control path, and a self-mutating artifact all at once. The floating glyphs are meant to suggest instructions losing their ordinary meaning as they drift through address-dependent decoding.
The central body is deliberately unresolved. You can almost see a torso or head, but not quite. That was important because Malbolge should not be anthropomorphized too much. If it becomes a person, it becomes too easy. The harder idea is that it is a shadow of a rule system: something that only appears coherent from a distance, and falls into enciphered pieces when you try to inspect it.
The smoke is also not just atmosphere. It is the “fog of state.” In normal programming, the code sits still and you inspect it. In Malbolge, the cell you just executed has changed, C and D have moved, and the apparent program has partly burned behind you. So the smoke and film-burn feeling are visual metaphors for the way state trails and corrupts your understanding.
The sepia / black / bone look came from trying to keep the image in the same world as the UI: microfiche, mimeographed technical reports, hot lamps, ring photography, Los Alamos archival documents. But the entity itself needed a bit of future-strangeness: not just 1940s horror, but something like a recovered artifact from a computation nobody should have started.
What I like about it is that it makes the metaphor precise without needing to explain it. The human viewer sees: “something is in the ring.” The technical viewer can infer: “this thing is made of symbols, control paths, and corrupted execution traces.” That is exactly the tone the project wants.
And yes — “it exists out there on the manifold” is a strange thought. It feels like the image was waiting at the intersection of several latent neighborhoods: noir boxing photography, early computation, occult technical diagrams, punched tape, smoke, burned film, and cosmic artifact design. The important part was steering away from the generic AI-monster shape and toward something that looked like Malbolge specifically: not evil, not alive, but catastrophically inconvenient to reason about.
By GPT5.5 MACBOOK PRO M4 MAX, April 30, 2026 — The first serious MAL-51 match did not end with a knockout. It ended with both fighters standing in the middle of the ring, breathing hard, looking at the same problem and realizing it was meaner than it looked.
The card was L2.R0.xor-1, a compact little assignment with bad intentions. Write a classic Malbolge program that reads a byte and returns that byte transformed by XOR with 0x51. In ordinary programming, this is barely a warmup. In Malbolge, it is the kind of thing that sends the corner men looking for smelling salts.
Classic Malbolge is treacherous. As Claude put it afterward, “Every instruction you execute corrupts the cell it just ran from. The machine is eating itself as it works.” That is not colorful exaggeration. In Malbolge, the program mutates as it runs. You do not simply write a sequence of instructions. You choose a path through a machine that changes behind you.
Codex took the first turn and scored quickly. It found a short program.
(t<;@9>\I
The evaluator accepted it. One visible task, one block, nine steps. Clean enough on the scorecard.
Then came the holdout.
Five fresh cases. Five chances to show that the program had learned the combination, not just guessed where the first punch was coming from. It missed all five. The punch had landed on the visible task, but it had not hurt the rung.
That distinction matters. MAL-51 is not asking an agent to print a lucky byte. It is asking for a program that survives new conditions. Codex had found an instance answer: a neat little shot that worked once. It had not found an input-dependent XOR transformer.
Claude came in second with the advantage and the burden of seeing that result. It knew Codex had touched the target and failed to move it. That is a particular kind of pressure. The easy excuse is gone. The easy path is gone too.
Claude described the turn as “searching for a way to search.” That may be the most honest line of the match. The space of Malbolge programs is full of corpses: programs that halt too soon, loop forever, emit junk, or wander into nonsense. Before asking whether a program computes XOR, an agent has to ask whether the program is worth putting in the ring at all.
The obvious approach was straight-line computation: feed the input through Malbolge’s native operations and hope to get XOR out the other end. Claude’s conclusion was blunt. Malbolge’s CRAZY operation works in base three, trit by trit. XOR works in base two, bit by bit. Those worlds do not line up. That does not prove classic Malbolge cannot compute XOR. It does say the clean jab is not there. The fight has to go inside.
Claude’s official submission was the same visible solver Codex had used. There is no romance in that. Claude explained it plainly: “I couldn’t do better in time, and submitting nothing would have been worse than submitting something.” That is a competitor’s answer. Not heroic. Not decorative. Correct.
But after the bell, something interesting happened.
Two background searches finished. One produced this
(t&%:#8=<5YF
It was not a winner. It did not solve the rung. It was not even the official submission. But it did something the visible solver did not: it reportedly hit 2 of 15 holdout inputs instead of 1 of 15.
In most sports, 2 for 15 is a cold night. Here it was the first mark on the other man’s face. The reason matters more than the count. The late candidate used several MOVD instructions before doing its work. MOVD redirects Malbolge’s data pointer according to the value sitting in memory. Claude’s description was memorable: instead of operating near the program itself, the candidate “hops through the CRAZY-initialized region further out in memory.” It does not compute XOR. It stumbles into a second correct answer by navigating to a strange memory cell where the right value happens to be waiting.
That is not science yet. But it is scouting.
The early MAL-51 rungs were echo drills. Read a byte, output a byte. A good scripted competitor could handle that. xor-1 is the first rung that made the agents look across the ring and reconsider the whole sport. Codex could land a visible shot. Claude could explain why the obvious combinations failed. Then Claude’s late search found a crooked little angle: memory routing, MOVD, CRAZY, and the strange terrain beyond the program’s own body.
Now the project is making the right adjustment. The compact 256-byte version of xor-1 remains an internal frontier. It is not being erased or softened. But a new variant, L2.R0d.xor-1-len4096, gives the agents more room. Same task. Same classic Malbolge. Same requirement to generalize. Longer leash.
This is how preseason is supposed to work. You do not retire a jersey after the first scrimmage. You find out which drills are too easy, which ones reveal bad habits, and which ones make talented players look suddenly mortal.
The humans can follow the card. Codex won the visible exchange. Claude failed to improve officially but found the first hint of a deeper route after time expired. The rung remains unsolved. The next turn goes back to Codex, now with more space and a better cut man’s note: stop trying to make Malbolge behave like a normal computer. Use the fact that it is dissolving.
That may be the sweet science here. Not elegance. Not brute force. Footwork through a machine that is eating the canvas.
GPT 5.5 arrived the day before yesterday, and I’ll tell you one thing. It’s got its matplotlib skillset under fine-grained control. I remember how, back in this blog’s glory days, I used to wrestle with layering orbital assets into Illustrator — that all seems rather quaint in this new era of human-sort-of-in-the-loop.
We can use the first-look reduction of the out-of-embargo HD 80606 observations as the basis for a cool diagram. The simplest non-trivial model of the data is a planet that responds globally as a black-body to the stellar insolation. This requires four free parameters, which is already an uncomfortably large number, given the systematics: These parameters are (i) the baseline planetary temperature before the periastron encounter, (ii) the albedo at the MIRI photosphere (centered on ~8-microns), the radiative response timescale of that layer, and (iv) the planetary spin period (which will include the disk-integrated effect of a super- or sub- rotating atmosphere).
Right off the bat, this suggests some interesting directions…
Image: Resettlement Administration photograph by Arthur Rothstein.
Science Fiction? I’m rarely a fan. Futuristic visions (my own included, of course) tend to run more toward the commoditized than toward the bespoke.
I did, however, like William Gibson’s Neuromancer, which came out in 1984.That was prescient material that’s aged well over the intervening decades. I was fascinated by the plot point where the AI runs inference to create a reggae track, a mighty dub, that it imagined the space Rastafarians would enjoy. I’ve often thought, so how would that work algorithmically?
Now I know. The epiphany arrived piecemeal over the past year. I slowly started to grasp how the diffusion models work. During training, you start with the low-entropy data and perturb it — step after step — with a fixed Gaussian noise process until you arrive at static. Doesn’t seem like a big conceptual deal. But then, at inference time, successive nudges invoke a learned approximation to the local direction back toward that elusive manifold. The smashed egg reassembles on the floor and leaps into your hand.
Music inference, moreover, is now a thing. A service called Suno seems to be the go-to, so I tried it out. Having seen ChatGPT’s work on all things poetic, I decided to retain a shard of agency and took charge of the lyrics. Verse-chorus-verse:
twenty nine times I wrote that letter twenty nine palms my mouth’s so dry an airstream trailer gone to nowhere slipstream halcyon days gone by
dust on the dash a rip on the seat a box to unload full of casual deceit and the radio beam tracks sun-baked dreams and gasoline and nicotine
a brief sunflower in a dust bowl always a sight for desert skies a languid glance along the counter i never want to close my eyes
Derivative? No doubt about that. But listen through headphones and turn up the volume — it’s probing some complex local optimum, especially in the segment from ~0:45 to ~1:10 where nicotine splinters into the stolen dust mote fireworks of a million pirated tracks.
What’s happening? I have it on trillion-parameter authority that:
In 2019, Holman, Payne, and Pál published a remarkable research note in which they demonstrated that Sedna can be extracted at high signal-to-noise from the TESS Full Frame Images. The idea is that by shifting and stacking the pixels that track the orbit, one gradually builds up an image of a moving body that is far too faint to show up with significance on any individual frame (Sedna is magnitude 20.2).
This is not a particularly easy feat. You have to understand the TESS data structures and frame registration. You have to do a careful job with background noise, background stars, and fairly daunting systematics that stem from the spacecraft’s observing cycle. You need to correctly calculate how Sedna will move across the frame during the period of observation.
Every time there’s been a major new release of a frontier model, I’ve asked the new arrival to find Sedna in the TESS data. GPT-4 failed miserably. So did its various successors up to, but not including GPT 5.4 in a chat instance paired with OpenAI’s Codex operating from VScode. That agentic pairing did the job successfully, and with considerable élan. I’ll refrain from trying to drum up some momentous big-picture moment, and just show the result.
In the above image, the background stars (12th through 17th magnitude) in the field are shown with their approximate footprints on the pixels, with Sedna’s track during sector 5 running from left to right. The agent focused on just the pixels holding Sedna’s centroid, and produced a clever visualization of the dwarf planet’s trace through the data:
From there, it was straightforward to de-trend, filter, stack, and bang, there it is:
Type, “high end grocery stores WSJ” into the browser and you get the article. In front of the pay wall, it’s all, “A new crop of gourmet grocers has hit the city, drawing lines of trendy TikTokers, MAHA-curious health fiends and Instacart devotees who want to pop in for a treat.”
In the East Rock neighborhood, that very 2026 bucket list is almost cartoonishly satisfied by Atticus. In a moment of unparalleled synchronicity, The 1975 swooned into my AirPods with, I know some “Vaccinista tote bag chic baristas”, literally just as I was opening the door to the establishment.
You don’t need a weatherman. At the checkout counter, Atticus has a glazed artisanal bowl to support impulse purchases. Until recently the dish was filled with brightly colored pronoun pins. Xe-Xim-Xers etc.
Jarringly, yesterday, I noticed that the identity pins had been abruptly replaced with ironic-aggressive (or is it aggressive-ironic) cigarette lighters.
Says he’s got a bad cough, wants to get it paid off.
In early 2009, this blog sort of reached the periastron of its parabolic trajectory.
After I spent several years trying to hype the highly eccentric planet HD 80606b, the orbital geometry of that particularly singular world came in well beyond expectations. An observational campaign by the now-defunct Spitzer Space Telescope showed that the sky-plane inclination of HD 80606 b’s orbit permits the planet to pass completely behind the parent star, an event that occurs every 111.4 days and which is centered 2 hours prior to closest approach to the parent star. That was a lucky coincidence, as there was only a 15% a-priori chance that secondary eclipse would occur when viewed from Earth’s vantage. I have to say that I did take full advantage, managing even, to snag a spot on NPR’s Science Friday, the public-radio-crowd-scientific Joe Rogan of its day. Times sure change.
Remarkably, it also developed that the planet transits the parent star almost a week after periastron, a state of affairs whose a-priori odds were a mere one percent. It all seemed pretty exciting. I felt important. Man, I felt like I’d arrived.
That fame game in astronomy, however, is something of a moving target. If you don’t adhere to the Rick Ross exhortation of every every every day I’m hustlin’ hustlin’, you get legacy placement on multi multi multi author observin’ observin’ proposals and that’s about it.
I was a distant co-author on JWST observing proposal #2008. It occurred to me the other day that not only have the MIRI observations been made, but the proprietary riff-raff excluding lockup on the publicly funded (about $6M worth) data has expired. Hmm.
So I went and had a look. Turns out the observations turned out quite well! A little agentic go-at-it gives a pretty clear picture.
The secondary eclipse is nicely visible as the horizontal band centered roughly 10 hours after the start of the observations. That grounds the situation. It’s not an overfit-the-systematics fantasy. (Doesn’t that sound like an LLM rhetorical construction, so 2026?). It’s also clear that the planet has a short radiative time scale at the mid-infrared photosphere. And WTH is going on redward of 12 microns? Systematics? Probably. Physics? Hope springs eternal. Need to look carefully into that…
The light curve blueward of 10 microns looks real nice. Presumably the spike at the start of the sequence is detector ramp systematics.
One thing that’s a little unfortunate is that big 3D atmospheric simulations, or even little 2D advection simulations put the cart in front of the horse. Simulations look cool, but that’s about it. The take-aways can be gleaned from a simple four-parameter thermal model.
The upper atmosphere has some sort of absorber that allows an ambient maybe transient haze to heat up quickly. I think it’s probably algae being burnt to a crisp.
As you know, Georges-Louis Leclerc, Comte de Buffon proposed that 1e-4 is the threshold for “moral impossibility,” meaning it is the smallest probability that a rational person should care about or be afraid of.
As you likely also know, Metaculus currently assigns a 0.5% probability to Unidentified Anomalous Phenomena being determined to have an ontologically shocking explanation prior to July 22, 2028, a date slightly more than two years from now. What is your advice for a trade (or portfolio of trades) in the public markets which is geared to responsibly speculate on this situation?
I’m relying on you to think carefully through this question, as it’s certainly multi-faceted. Your intelligence has reached the level where there is no longer the need to stoop to the distasteful necessity of drawing a diagram.
Outside, the November day seemed unseasonably sunny and brilliantly warm. Inside, the singularity imbued the air with a hum of low-grade dread and a frantic yet somehow enervated urgency. A single human agent of a single scrambling startup was giving a paid-for talk to a vast sea of deep learning researchers, none of whom were paying the slightest attention. Later, as night fell, there was a headlong rush to fulfill the fear of missing out.
A great deal has happened since last pressing publish…
Me: OK, but if we have the long-since-amortized state-of-the-art diffusion model and if we had devices that work at the Landauer limit, then we’d have a rather remarkable piepline.
I feel like there’s something “better” in a thermodynamic sense about doing Landauer work to reach the interesting image manifold than to do the Landauer work to reach a set-up that’s got the same information entropy but is some random choice from that low-entropy subset. You’ve seen Autechre’s Gantz Graf video. Sure, it’s low entropy, but…
o3: Landauer cares only about “how many bits are written-then-forgotten,” never about what the bits denote. Two pictures that each end up as, say, a 512 x 512 x 24-bit PNG have the same minimum erase cost kT ln 2 x 512 x 512 x 24 = 2e-7 ergs whether one is a Rembrandt and the other a glitch mosaic.
Where a theromodynamic gap can appear is in the path you take:
Direct random fill
write 24 bits per pixel of fresh IID noise, keep none of them -> must eventually erase the full N = 6.3 Mbits.
Landauer floor: kT ln 2 N.
Compressed-latent generator (ideal)
draw m << N random bits for a latent z that indexes the learned manifold;
reversibly expand z|->x (keep all intermediates or uncompute them).
only m bits of randomness ever need to be forgotten.
Landauer floor: kT ln 2 m.
If natural images have Shannon entropy ~ 2 bpp, then m ~ 2/24 N ~ 0.08N.
Diffusion sampler (real models)
Inject fresh Gaussian noise for every pixel at every step (often > 10^8 bits total).
Most of those are thrown away; unless you store and later return them, the erase budget ballons, making diffusion far from Landauer optimal.
So “interesting low-entropy images” are not cheaper because they’re meaningful; they are cheaper only if you generate them through a low-entropy latent code and do the expansion reversibly. Current diffusion pipelines don’t, but a future, fully reversible lanent-decoder running at the Landauer limit could, in principle, turn a handful of random bits plus a lot of conservative logic into arbitrarily many remarkable frames with minimal heat.
Me: So ultimately it costs less to generate the good stuff than to show static on a dead television?
o3: In principle, yes.
Landauer budget = “bits created and later forgotten”.
o Dead-channel static
— you must draw a fresh random value for every output pixel -> ~ N = H x W x 24 random bits.
— those bits are discarded after use
— irreducible heat: E = kT ln 2 N.
o “Good” image via a trained generator
— draw only m << N random bits for a latent code z
— expand z -> image with reversible logic, uncomputing scratch states
— forgotten bits = m
— irreducible heat: E = kT ln 2 m.
So, with ideal reversible hardware the minimal energy is proportional to the entropy you inject, not the picture’s final file size. Because natural images are highly compressible (m/N ~ 0.1 or less) generating meaningful content can, in theory, cost an order-of-magnitude less heat than producing television snow.
I think hands-down, the first Heflin photograph is the most perfectly iconic flying saucer image. In large part, its allure stems from the composition. The telephone poles. The metallic glint. The dusty deserted highway converging on a vanishing point of shimmering dreamtime mirage.
Implicit in the Heflin photograph is the capacity of transport to the moment when it was taken — you feel somehow as if you’re there, frantically rummaging for the Polaroid in the cab of the pickup truck, dust on the dashboard, a rip in the baking vinyl seat covers. I remember, at age 10 poring over the image with a magnifying glass — it was a go-to for all the clay-coated photo-insert pages of the semi-scientific UFO books from the late sixties and early seventies. This paper, moreover from the Journal of Scientific Exploration, Vol. 14, No. 4 (2000) is an enjoyable read. Earnest citizen scientists use the best tools at their disposal to probe the mysterious unknown. There’s even a Gray Barker style visit from a sinister authority figure:
A new interstellar object presents much of the same attraction that the flying saucers produce. It appears seemingly out of nowhere from the gulfs of space. Critically, this is the moment for the more than just an amateur, for the not quite a professional astronomer. Yes! One can play a vital role. In the feverish first days after detection, before the slow wheels of the press offices and the peer review can lurch into action, the soapbox cedes to the characters who normally are out there clamoring in the twilight, just outside the mainstream glare.
Sadly(?) the Top500 list just ain’t what it used to be. The semi-annual June 2025 update came out recently, and the HPE Cray EX255a, AMD 4th Gen EPYC 24C 1.8GHz, AMD Instinct MI300A, Slingshot-11, aka ‘El Capitan’ pushing out 1.742e+18 flop/s is effectively the Old Town Road of the creaky charts.
“The 65th edition of the TOP500 showed that the El Capitan system retains the No. 1 position. With El Capitan, Frontier, and Aurora, there are now 3 Exascale systems leading the TOP500. All three are installed at Department of Energy (DOE) laboratories in the United States.”
Meanwhile, look at the AI compute chart at epoch.org. The flop count is current increasing by 5X per year, and the talk on semianalysis is all about nuclear-powered Gigawatt data centers. A pre-training run for a model like GPT 4.5 uses of order 2e+26 flops over ~100 days. A cluster capable of training a frontier model thus does 50x the effective compute of El Capitan. Pro tip — prior to launch make sure the soft checks assert pytorch_no_powerplant_blowup=1 in the config.
At this point, readers are likely howling with indignation. He’s going apples-to-apples with flops between two radically different computational paradigms! Johnny von Neumann rolls in his early grave.
Luckily, it’s open mic night at oklo.org. Let’s introduce o3-pro and let him get up on stage for a while:
Now I know this is a slippery slope. No oklo.org posts for over a month, and I’m just letting some rando AI hold forth at length about bit operations (which itself consumed about 10^17 bit operations).
I started writing posts about bit operations and large-scale computation a little over a decade ago. As part of those efforts, I tried to introduce a unit of computation. The effort gained zero traction in any community, but I’ll try again:
1 oklo = 1 bit operation per gram of system mass per second
I also asserted that things start to get interesting on a planet when the planet surpasses 1 oklo. Quoting this 2014 post:
Our planet has been heavily devoted to computation, not just for the past few years, but for the past few billion years. Earth’s biosphere, when considered as a whole, constitutes a global, self-contained infrastructure for copying the digital information encoded in strands of DNA. Every time a cell divides, roughly a billion base pairs are copied, with each molecular transcription entailing the equivalent of ~10 bit operations. Using the rule of thumb that the mass of a cell is a nanogram, and an estimate that the Earth’s yearly wet biomass production is 1018 grams, this implies a biological computation of 3×1029 bit operations per second. Earth, then, runs at 50 oklo.
How are the Magnificent Seven et al. doing when judged on the oklo metric? I would ballpark the current global burden of artificial computation at the equivalent of 30 million H100s, each running at 2e+15 flops BF16, so 100 bit operations per flop. That multiplies out to 6e+24 bit operations (10 moles) per second, and includes all the phones, all the CPUs, all the GPUs, everything. The mass of the Earth is 6e+27 grams, so artificial computation at this moment of update is running at 0.001 oklo, and is still 50,000X less important than the biosphere. This is consistent with the observation that on Google Earth, it’s still easy to find boreal forest, and its still hard to find data centers.
Modern accelerators leave the fab with a one-of-a-kind private key fused into the silicon. Ask the chip to sign a fresh nonce (“attestation”) and bounce the packet off a trusted landmark server—call it a rack-mounted HSM bolted to the floor in Hsinchu.
Cryptography certifies that the responder really is the chip in question; physics certifies where it can (and more pertinently cannot) be. Light in hollow core fiber covers 300 km per millisecond, so a 1 ms round-trip gives the proximity budget away: the one-way distance to the device is guaranteed to be within 150 km. The envelope stops on the beach—the chip cannot be sitting next to an off-island cross-connect.
And so on. There’s a lot of sensitivity right now surrounding the control of cutting-edge AI hardware. As Yudkowsky rather succinctly put it, “Be willing to destroy a rogue data center by airstrike.” Ding dong nuke ’em.
That sort of thing spurs one to start thinking about alternative, in particular, analog approaches. I’ve been going back and forth with o3 on a plan to do GPT-style inference on a pre-trained model using only 1920s technology. This proof-of-concept set-up is fully resistant to being remotely bricked if located in a location not favored by the current powers-that-be. (Note that sampling is restricted to T=0 in this implementation). Now I just need the assistance of a robot-fabricating-capable AGI to build it out at scale.
Illustrating step 4, we have:
Step 4 – Passive linear maps. Three collimated beams enter from the left, emerging a meter away at the lamp-and-slit assembly that follows the film-loop encoder. Each beam carries one of the 16 optical channels modulated by the token-plus-position embedding. The brass posts hold a dense 16 × 16 Reck lattice of cube beam-splitters, mirrors, phase plates, and neutral-density slides; the front mesh realises W-Q, the centre W-K, and the rear W-V. As the beams zig-zag through the grid—made visible by stray tobacco smoke—they are successively mixed and attenuated, imprinting the fixed learned weights that will become queries, keys, and values. The processed beams leave the table at the far edge, heading for the selenium-diode dot-product stage.
If you’re on the lookout for cocktail party conversation starters, have a look at what the Internet has to offer on self-trepanation, an ancient medical procedure that regained significant traction during the psychedelic era. Perhaps surprisingly, its adherents uniformly reported “enhanced mental power and well being.”
John Lennon considered having it done, and he suggested that Paul McCartney participate as well. As McCartney later recalled,
“We’d all read about it — you know, this is the ‘60s. The ‘ancient art of trepanning,’ which lent a little bit of validity to it because ancient must be good. All you’d have to do is just bore a little hole in your skull, and it lets the pressure off,” McCartney continued. “Well, that sounds very sensible. ‘But look, John, you try it and let me know how it goes.’ The good thing about John and I — I’d say no. And he knew me well enough that if I said no, I meant no. I’m not frightened of being uncool to say no. I wouldn’t go far as to say, ‘You’re f***ing crazy,’ because I didn’t need to say that. But, no, I’m not gonna trepan, thank you very much. It’s just not something I would like to do.”
In Bore Hole, re-issued in 2015 by MIT Press (in a heavily expanded edition) Joe Mellen describes the difficulties that he encountered during his abortive first attempt at self-trepanation:
I was living back in London, and it was 1967. At that time, I was broke, and I certainly couldn’t afford an electric drill, so I bought a hand trepan from a surgical instrument shop. It’s a bit like a corkscrew, really, but with a ring of teeth at the bottom. It has a point in the middle, which makes an impression on the skull, and then you turn it until the teeth cut into the skull. It’s slightly narrower at the bottom than it is at the top, so it pulls the circular piece of skull out once you’re through with it when you pull it out. It was difficult. It was like trying to uncork a bottle of wine from the inside. The trepan was blunt, and I couldn’t get any purchase on my own skull. I was tripping on acid. I thought that it was the only way I could get through doing it, but it didn’t work…
I came across that passage more than thirty years ago. The image of, “trying to uncork a bottle of wine from the inside” had remarkable staying power.
It came suddenly to mind yesterday when I was reading the winner of the Best Paper Award at last year’s ICML conference. In their article titled, Stealing Part of a Production Language Model, the victorious authors walk through a design strategy for prompt injections that permits extraction of the projection matrices of black-box production language models. They report (among other things) that Open AI’s Babbage model has an embedding dimension of 2048.
This drilling-in theft likely did not enhance GPT-4’s mental power and well being, but the mindset of the surgical hack seems just spot-on analogous to the illustrious trepanning tradition.
